European Spatial Data Research

EuroSDR Projects

Evaluation of Building ExtractionReport by Harri Kaartinen and Juha Hyyppä

Change DetectionReport by Klaus Steinnocher and Florian Kressler

Sensor and Data Fusion Contest: Information for Mappingfrom Airborne SAR and Optical Imagery (Phase I)

Report by Anke Bellmann and Olaf Hellwich

Automated Extraction, Refinement, and Updateof Road Databases from Imagery and Other Data

Report by Helmut Mayer, Emmanuel Baltsavias, and Uwe Bacher

Official Publication No 50

November 2006

The present publication is the exclusive property of European Spatial Data Research

All rights of translation and reproduction are reserved on behalf of EuroSDR. Published by EuroSDR

Printed by Gopher, Utrecht, The Netherlands

EUROPEAN SPATIAL DATA RESEARCH

PRESIDENT 2006 – 2008: Stig Jönsson, Sweden

SECRETARY-GENERAL: Kevin Mooney, Ireland

DELEGATES BY MEMBER COUNTRY:

Austria: Michael Franzen Belgium: Ingrid Vanden Berghe Cyprus: Christos Zenonos; Michael Savvides Denmark: Joachim Höhle; Lars Jørgensen Finland: Risto Kuittinen; Juha Vilhomaa France: Marc Pierrot-Deseilligny; Franck Jung Germany: Dietmar Grünreich; Günter Nagel; Dieter Fritsch Hungary: Arpad Barsi Ireland: Colin Bray Italy: Carlo Cannafoglia; Riccardo Galetto Netherlands: Jantien Stoter; Aart-jan Klijnjan Norway: Jon Arne Trollvik; Ivar Maalen-Johansen Portugal: Berta Cipriano Spain: Antonio Arozarena, Francisco Papí Montanel Sweden: Stig Jönsson; Anders Östman Switzerland: Francois Golay; André Streilein-Hurni United Kingdom: Keith Murray; David Chapman

COMMISSION CHAIRPERSONS:

Sensors, Primary Data Acquisition and Georeferencing: Michael Cramer, Germany Image Analysis and Information Extraction: Juha Hyyppä, Finland Production Systems and Processes: Eberhard Gülch, Germany Core GeoInformation Databases: Keith Murray, United Kingdom Integration and Delivery of Data and Services: Mike Jackson

OFFICE OF PUBLICATIONS:

Bundesamt für Kartographie und Geodäsie (BKG) Publications Officer: Andreas Busch Richard-Strauss-Allee 11 60598 Frankfurt Germany Tel.: + 49 69 6333 312 Fax: + 49 69 6333 441

CONTACT DETAILS:

Web: www.eurosdr.net President: president@eurosdr.net Secretary-General: secretary@eurosdr.net Secretariat: admin@eurosdr.net

EuroSDR Secretariat Faculty of the Built Environment Dublin Institute of Technology Bolton Street Dublin 1 Ireland

Tel.: +353 1 4023933

The official publications of EuroSDR are peer-reviewed.

H. Kaartinen and J. Hyyppä: EVALUATION OF BUILDING EXTRACTION ............................ 9

ABSTRACT ........................................................................................................................................ 11

1 INTRODUCTION ....................................................................................................................... 11

2 MATERIALS............................................................................................................................... 13

2.1 Test Sites and Delivered Data............................................................................................. 132.1.1 General.......... ........................................................................................................... 13 2.1.2 FGI Test Sites Espoonlahti, Hermanni and Senaatti ................................................ 132.1.3 IGN Test Site Amiens .............................................................................................. 18

2.2 Reference Data.................................................................................................................... 192.2.1 Field Measurements ................................................................................................. 192.2.2 Aerial Images ........................................................................................................... 192.2.3 Ground Plans............................................................................................................ 19

2.3 Produced 3D-models .......................................................................................................... 20

3 METHODS................................................................................................................................... 20

3.1 Building Extraction............................................................................................................. 203.1.1 Summary of Applied Data ....................................................................................... 203.1.2 Building Extraction Methods ................................................................................... 21

3.1.2.1 About method descriptions ........................................................................ 213.1.2.2 CyberCity AG............................................................................................ 213.1.2.3 Hamburg University of Applied Sciences ................................................. 233.1.2.4 Nebel+Partner GmbH ................................................................................ 263.1.2.5 Stuttgart University of Applied Sciences................................................... 283.1.2.6 IGN ............................................................................................................ 303.1.2.7 ICC ............................................................................................................ 303.1.2.8 FOI............................................................................................................. 303.1.2.9 C+B Technik.............................................................................................. 333.1.2.10 Delft University of Technology ................................................................. 343.1.2.11 University of Aalborg ................................................................................ 363.1.2.12 Dresden University of Technology............................................................ 37

3.1.3 Time Used for Building Extraction.......................................................................... 393.1.4 Level of Automation ................................................................................................ 40

3.1.4.1 Summary.................................................................................................... 403.1.4.2 CyberCity AG............................................................................................ 403.1.4.3 Hamburg University of Applied Sciences ................................................. 403.1.4.4 Nebel + Partner and ICC............................................................................ 403.1.4.5 Stuttgart University of Applied Sciences................................................... 413.1.4.6 IGN ............................................................................................................ 413.1.4.7 FOI............................................................................................................. 413.1.4.8 C+B Technik.............................................................................................. 413.1.4.9 Delft University of Technology ................................................................. 423.1.4.10 University of Aalborg ................................................................................ 423.1.4.11 Dresden University of Technology............................................................ 42

3.2 Accuracy Evaluation........................................................................................................... 423.2.1 Analysis Using Reference Points ............................................................................. 423.2.2 Analysis Using Reference Raster Ground Plans ...................................................... 43

4 RESULTS..................................................................................................................................... 43

4.1 Level of Detail .................................................................................................................... 434.1.1 Number of Targets ................................................................................................... 434.1.2 Extracted Building Vectors ...................................................................................... 44

4.2 Geometric Accuracy ........................................................................................................... 484.2.1 Building Location .................................................................................................... 484.2.2 Building Height........................................................................................................ 654.2.3 Building Length ....................................................................................................... 684.2.4 Roof Inclination ....................................................................................................... 71

4.3 Total Relative Building Area and Shape Dissimilarity....................................................... 73

5 DISCUSSION AND CONCLUSIONS....................................................................................... 75

ACKNOWLEDGMENTS.................................................................................................................. 76

REFERENCES ................................................................................................................................... 77

K. Steinnocher and F. Kressler: CHANGE DETECTION ............................................................ 111

ABSTRACT ...................................................................................................................................... 113

1 INTRODUCTION ..................................................................................................................... 113

2 PROJECT HIGHLIGHTS ....................................................................................................... 114

3 METHODOLOGY.................................................................................................................... 114

3.1 Segmentation .................................................................................................................... 1153.2 Object-oriented classification ........................................................................................... 1153.3 Comparison classification – land use data ........................................................................ 1173.4 Contributions of project partners ...................................................................................... 118

4 CASE STUDIES ........................................................................................................................ 118

4.1 Austria .............................................................................................................................. 1194.1.1 Data and study area ................................................................................................ 1194.1.2 Segmentation.......................................................................................................... 1224.1.3 Classification.......................................................................................................... 1234.1.4 Change map ........................................................................................................... 1264.1.5 Evaluation of change map...................................................................................... 128

4.2 Switzerland ....................................................................................................................... 1314.2.1 Data and Study Area .............................................................................................. 1314.2.2 Segmentation, classification and comparison ........................................................ 133

4.3 Germany ........................................................................................................................... 137

4.3.1 Data and study area ................................................................................................ 1374.3.2 Segmentation, classification and comparison ........................................................ 1374.3.3 A comparison of change detection methods and analysis of results (by Andreas Busch from BKG)............................................................................. 140

4.4 Denmark ........................................................................................................................... 1454.4.1 Data and study area ................................................................................................ 1454.4.2 Segmentation, classification and comparison ........................................................ 146

4.5 Additional applications ..................................................................................................... 1484.5.1 Digital Scanner....................................................................................................... 148 4.5.1.1 Data and study area.................................................................................. 148 4.5.1.2 Segmentation, classification and comparison .......................................... 148

5 CONCLUSIONS AND OUTLOOK......................................................................................... 156

REFERENCES ................................................................................................................................. 157

ANNEX 1 EUROSDR CHANGE DETECTION WORKSHOP VIENNA,

JUNE, 16TH

-17TH

2005 MINUTES ............................................................................ 159

ANNEX 2 PAPER PRESENTED AT IGARSS05, JULY, 25TH

-29TH

2005, SEOUL

KRESSLER, F., FRANZEN, M. AND STEINNOCHER, K. ................................. 165

ANNEX 3 EUROSDR CD WORKSHOP NICOSIA 2005, OCTOBER, 25TH

2005

MINUTES ................................................................................................................... 173

ANNEX 4 COMMENTS ON CHANGE MAP BY BEV........................................................... 179

A. Bellmann and O. Hellwich: SENSOR AND DATA FUSION CONTEST: INFORMATION

FOR MAPPING FROM AIRBORNE SAR AND OPTICAL IMAGERY (PHASE 1) ............. 183

ABSTRACT ...................................................................................................................................... 185

1 INTRODUCTION ..................................................................................................................... 185

1.1 Aim of the Project............................................................................................................. 1851.2 Contest Phases .................................................................................................................. 1861.3 Participants ....................................................................................................................... 186

2 TEST SITES AND TEST DATA ............................................................................................. 187

2.1 Specification of the Test Data........................................................................................... 1872.2 Additional Data................................................................................................................. 190

3 REFERENCE DATA ................................................................................................................ 192

3.1 Generation of Reference Data........................................................................................... 1923.2 Accounting for Different Acquisition Times .................................................................... 1933.3 Reference Maps ................................................................................................................ 193

4 ANALYSIS................................................................................................................................. 195

4.1 Basic Information ............................................................................................................. 1954.2 Evaluation Procedure........................................................................................................ 195

4.2.1 Areal Areas ............................................................................................................ 1954.2.2 Linear Objects ........................................................................................................ 1974.2.3 Qualitative Comparison ......................................................................................... 198

5 RESULTS OF PHASE I ........................................................................................................... 199

5.1 Test Site Copenhagen ....................................................................................................... 1995.1.1 Areal and Linear Objects ....................................................................................... 1995.1.2 Qualitative Comparison ......................................................................................... 2015.1.3 Summary – Copenhagen ........................................................................................ 201

5.2 Test Site Fjärdhundra........................................................................................................ 2025.2.1 Areal and Linear Objects ....................................................................................... 2025.2.2 Qualitative Comparison ......................................................................................... 2045.2.3 Summary – Fjärdhundra......................................................................................... 204

5.3 Test Site Oberpfaffenhofen............................................................................................... 2055.3.1 Areal and Linear Objects ....................................................................................... 2055.3.2 Qualitative Comparison ......................................................................................... 2075.3.3 Summary – Oberpfaffenhofen................................................................................ 208

5.4 Test Site Trudering ........................................................................................................... 2085.4.1 Areal and Linear Objects ....................................................................................... 2085.4.2 Qualitative Comparison ......................................................................................... 2115.4.3 Summary – Trudering ............................................................................................ 211

6 CONCLUSION – PHASE I ...................................................................................................... 212

ACKNOWLEDGMENTS................................................................................................................ 213

REFERENCES ................................................................................................................................. 213

H. Mayer, E. Baltsavias and U. Bacher: AUTOMATED EXTRACTION, REFINEMENT, AND

UPDATE OF ROAD DATABASES FROM IMAGERY AND OTHER DATA ........................ 217

ABSTRACT ...................................................................................................................................... 219

1 MOTIVATION AND IMPLEMENTATION ......................................................................... 219

1.1 Motivation ........................................................................................................................ 2191.2 Implementation ................................................................................................................. 220

2 QUESTIONNAIRES................................................................................................................. 221

2.1 Summary of responses to the Questionnaire for Producers of Road Data ........................ 2212.2 Summary of responses to the Questionnaire for Researchers and Manufactures ............. 222

3 TEST DATA AND EVALUATION CRITERIA .................................................................... 225

4 APPROACHES AND EVALUATION OF RESULTS .......................................................... 226

5 ANALYSIS OF RESULTS ....................................................................................................... 230

6 CONCLUSIONS AND RECOMMENDATIONS................................................................... 237

7 ACKNOWLEDGMENTS......................................................................................................... 239

REFERENCES ................................................................................................................................. 239

APPENDICES................................................................................................................................... 243

APPENDIX 1: PROJECT PROPOSAL ......................................................................................... 243

APPENDIX 2: QUESTIONNAIRE FOR PRODUCERS OF ROAD DATA.............................. 249

APPENDIX 3: QUESTIONNAIRE FOR RESEARCHERS AND MANUFACTURERS ......... 257

APPENDIX 4: GENERAL CHARACTERISTICS OF THE APPROACHES........................... 267

APPENDIX 5: README FILE ...................................................................................................... 271

APPENDIX 6: DOCUMENTATION BY BEUMIER AND LACROIX...................................... 275

APPENDIX 7: DOCUMENTATION BY HEDMAN AND HINZ ............................................... 277

APPENDIX 8: DOCUMENTATION BY ZHANG AND COULOIGNER ................................. 279

EuroSDR-Project

Commission 3 “Evaluation of Building Extraction”

Final Report

Report by Harri Kaartinen and Juha Hyyppä

Finnish Geodetic Institute, Masala, Finland

11

Abstract

The objective of the EuroSDR Building Extraction project was to evaluate the quality, accuracy, feasibility and economical aspects of semi-automatic and automatic building extraction based on laser scanner data and/or aerial images. Data sets from four test sites were delivered to the eleven participants of the project. For each test site the following data were available and could be downloaded from the FGI ftp-site: aerial images, camera calibration and image orientation information, ground control point coordinates and jpg-images of point locations (with the exception of Amiens), laser scanner data and cadastral map vectors of selected buildings (vector ground plans). In this report the used test sites, data sets as well as used reference data are described. The building extraction methods used by the participants and methods used in accuracy evaluation are explained. Produced 3D city models are depicted and analysed especially in respect to accuracy, used data and the level of automation. Finally, discussion and conclusions are given summarizing the results of the project.

Presently, photogrammetric techniques and hybrid techniques provide the highest accuracy and level of detail in 3D city reconstruction. Despite the high amount of research during the last decade the level of automation is still relatively low. Improvement in automation can be achieved most significantly by utilising the synergy of laser and photogrammetry. The photogrammetric techniques are powerful for visual interpretation of the area, measurement of the building outlines and of small details (e.g. chimneys), whereas laser scanning gives height, roof planes and ridge information at its best.

1 Introduction

Three-dimensional geographical information systems are of increasing importance in urban areas in various applications such as urban planning, visualization, environmental studies and simulation (pollution, noise), tourism, facility management, telecommunication network planning, 3D cadastre and vehicle or pedestrian navigation. In the late 1990s, OEEPE conducted a survey on 3D-city models. The scope of that study was to find out the current state of generating and using 3D city data. The study was based on a questionnaire sent out to about 200 European institutions (Fuchs et al., 1998). Fifty-five responded to the questionnaire from seventeen countries (curiously, 30 % of the replies came from Germany). The most important objects of interest according to the users were buildings (95%), traffic network (76%) and vegetation (71 %). It was stated that there was a definite lack of economical techniques for producing 3D-city data. The lack of knowledge of information sources, as well as high costs for building and vegetation acquisition hindered broader use. Data sources used by the producers were aerial imagery (76%), map data (54%), classical survey data (46%) and aerial range data/laser scanning data (20%). At that time range data was regarded by many producers as too expensive.

Semi-automatic and automatic methods for 3D city models are aimed at reducing the costs of providing these data with reasonable level of detail. Today, aerial photogrammetry is still one of the main techniques for obtaining 3D building information. Digital aerial photogrammetry supports accurate measurement of points and structures, which are usually defined by a human operator (Brenner, 2005). Despite significant research efforts in the past, the low degree of automation achieved has remained the major problem. Thus, the majority of development work has recently focussed on semi-automatic systems, in which, for example, recognition and interpretation tasks are performed by the human operator, whereas modelling and precise measurement are supported by automation.

12

Due to the development of scanning systems and improvements in the accuracy of direct georeferencing, airborne laser scanning (ALS) became a feasible technology to provide range data in the early 1990s. At that time, ALS was already considered as a mature technology (Baltsavias, 1999). ALS provides dense point clouds (more than 10 returns per m2 are possible today) with 3D coordinates. This makes range data segmentation relatively easy (Brenner, 2005).

The integration of laser point clouds and photogrammetric processes with aerial photos also provides new technological solutions. By combining the good height assessment accuracy of laser scanner and good planimetric accuracy of aerial images, both high accuracy and higher automation can in theory be obtained. However, despite the progress that has been made with integrating laser scanning systems and digital images, automated processing of the resulting datasets is at a very early research stage (Brenner, 2005).

A short summary of the state of the art in building extraction methods can be obtained from Baltsavias (2004) and Brenner (2001, 2005).

A more detailed summary of building extraction methods with emphasis on achievements of the last years can be obtained from Alharthy and Bethel (2004), Ameri and Fritsch (2000), Baillard and Dissard (2000), Baillard and Maïtre (1999), Baltsavias (2004), Braun et al. (1995), Brenner (2001, 2003, 2005), Brunn et al. (1998), Brunn and Weidner (1997, 1998), Centeno and Miqueles (2004), Chen et al. (2004), Cho et al. (2004), Cord and Declerq (2001), Cord et al. (2001), Dash et al. (2004), Dold and Brenner (2004), Elaksher and Bethel (2002), Elaksher et al. (2003), Fischer et al. (1998), Forlani et al. (2003, 2006), Fraser et al. (2002), Fuchs and Le-Men (2000), Fujii and Arikawa (2002), Förstner (1999), Grün (1997, 1998), Grün and Wang (1998a,b, 1999a,b), Grün et al. (2003), Haala et al. (1998), Haala and Brenner (1999), Haithcoat et al. (2001), Hofmann et al. (2001, 2003), Hofmann (2004), Jaynes et al. (2003), Jutzi and Stilla (2004), Jülge and Brenner (2004), Khoshelham (2004), Kim and Muller (1998), Lee and Choi (2004), Li et al. (2004), Maas and Vosselman (1999), Maas (2001), Madhavan et al. (2004), Mayer (1999), Morgan and Habib (2001,2002), Neidhart and Sester (2003), Niederöst (2003), Noronha and Nevatia (2001), Oda et al. (2004), Oriot et al. (2004), Overby et al. (2003), Paparoditis et al. (1998), Peternell and Steiner (2004), Rottensteiner (2000, 2003), Rottenstener and Briese (2002), Rottensteiner and Jansa (2002), Rottensteiner et al. (2003, 2004, 2005), Sahar and Krupnik (1999), Schwalbe (2004), Sequeira et al. (1999), Shufelt (1999), Sinning-Meister et al. (1996), Sohn (2004), Stilla and Jurkiewicz (1999), Süveg and Vosselman (2004), Söderman et al. (2004),Taillandier and Deriche (2004), Takase et al. (2004), Tan and Shibasaki (2002), Tsay (2001), Tseng and Wang (2003), Vosselman (2002), Vosselman and Dijkman (2001), Vosselman and Süveg (2001), Wang and Grün (2003), Weidner and Förstner (1995), Zhan et al. (2002), Zhang and Wang (2004), and Zhao and Shibasaki (2003).

Due to the rapid development of sensors and methods during the last years, it was proposed and accepted that under the EuroSDR Commission III “Production Systems and Processes”, a joint test would be undertaken in order to compare various methods. The objective of the EuroSDR Building Extraction project was to evaluate the quality, accuracy, feasibility and economical aspects of

1. Semi-automatic building extraction based on photogrammetric techniques with the emphasis on commercial and/or operative systems

2. Semi-automatic and automatic building extraction techniques based on high density laser scanner data

3. Semi-automatic and automatic building extraction techniques based on integration of laser scanner data and aerial images

This paper reports the results obtained in the study. Initial results have previously been reported in Kaartinen et al. (2005a,b).

13

The analysis of the structural quality of the reconstructed buildings was not performed in the project at this stage. Even though, it is an important aspect for the end user how well do the reconstructed structures reflect real life structures.

2 Materials

2.1 Test Sites and Delivered Data

2.1.1 General

The project consisted of three sites provided by the Finnish Geodetic Institute (FGI), namely Senaatti, Hermanni and Espoonlahti and one site, Amiens, provided by the Institut Geographique National (IGN).

For each test site the following data were available and could be downloaded from the FGI ftp-site: - aerial images - camera calibration and image orientation information - ground control point coordinates and jpg-images of point locations (with the exception of

Amiens) - laser scanner data - cadastral map vectors of selected buildings (vector ground plans)

- Espoonlahti: 11 buildings (white vectors in Figure 2-6) - Hermanni: 9 buildings (white vectors in Figure 2-4) - Senaatti: 6 buildings/blocks (white vectors in Figure 2-2) - Amiens: 7 buildings

Participants were requested to create the vectors of the 3D city models from the given areas in all of the four test sites using the materials described above. Participants could use any methods and data combinations they wished. The 3D-model should consist of permanent structures of the test area: buildings modelled in as much detail as possible (this mainly concerns roof structures) and terrain so that it would be possible to measure building heights using the model. In addition the participants were requested to describe the process used and economical aspects, time spent for the process etc. Reporting instructions were given after the data was delivered.

2.1.2 FGI Test Sites Espoonlahti, Hermanni and Senaatti

Each of the test sites had their own characteristics: - Espoonlahti: with a significant variety of houses, undulating terrain and a large number of

trees, Espoonlahti is located in Espoo, about 15 km west of Helsinki with high-rise buildings and terraced houses.

- Hermanni: a large, simple block of flat houses with low vegetation, Hermanni is a residential area about 3 km from the main city centre with four to six storey houses built mainly in the 1950’s.

- Senaatti: a typical European city centre with some complex buildings and no vegetation, Senaatti includes the area around the Senate Square in Helsinki main city centre, three to six storey houses and Lutheran Cathedral built mainly in the 19th century.

In addition to the construction type, that differs from site to site, the test sites have also been flown with different laser scanners (TopEye, TopoSys-I, TopoSys-Falcon) and with different pulse densities

14

(from 1.6 to about 20 pulses per m2). Last pulse data was not available for all the test sites and therefore, first pulse data was considered as the major data.

Figure 2-1: Senaatti test site.

Figure 2-2: Senaatti test site (TopoSys-1). As can be seen, ground plans are shown in advance

for selected buildings.

15

Figure 2-3: Hermanni test site.

Figure 2-4: Hermanni test site (TopEye).

16

Figure 2-5: Espoonlahti test site.

Figure 2-6: Espoonlahti test site (TopoSys Falcon).

17

Espoonlahti Hermanni Senaatti

Photos Stereo pair Stereo pair Stereo pair Date 26th of June 2003 4th of May 2001 24th of April 2002 Camera RC-30 RC-30 RC-30 Lens 15/4 UAG-S, no

1335515/4 UAG-S, no

1326015/4 UAG-S, no

13260Calibration date 22nd of November

200218th of January

200014th of April

2002Flying height, scale

860 m, 1:5300 670 m, 1:4000 660 m, 1:4000

Pixel size 14 microns 15 microns 14 microns Pixel size on the ground

7.5 cm 6 cm 5.5 cm

Table 2-1: Aerial images.

Espoonlahti Hermanni Senaatti

Acquisition 14th of May 2003 End of June 2002 14th of June 2000 Instrument TopoSys Falcon TopEye TopoSys-1, pulse

modulated Flight altitude 400 m 200 m 800 m Pulse frequency 83 000 Hz 7 000 Hz 83 000 Hz Field of view ± 7.15 degrees ± 20 degrees ± 7.1 degrees Measurement density

10-20 per m2 7-9 per m2 on the average

1.6 per m2 on the average

Swath width 100 m Ab. 130 m Ab. 200m Mode First pulse 2 pulses First Pulse

Table 2-2: Laser scanner data.

Helsinki City Survey Division granted permission to use their material (aerial photos, laser scanner data and cadastral maps) for the Senaatti/Hermanni data. Terrasolid Oy also gave permission to use the Hermanni laser scanner data (partly owned by them). FM-Kartta Oy gave permission to use the Senaatti laser scanner data in the project. The Espoonlahti laser scanner data is owned by the FGI and permission to use the data in the project was granted. Espoonlahti aerial imagery was purchased from FM-Kartta Oy for this project using internal funding of FGI. Espoo City Survey Division granted permission to use cadastral map vectors.

Airborne images were in TIFF-format, with orientation parameters and ground point coordinates that could be used as control or check points. Laser scanner data was in ASCII format (XYZ point data). Cadastral map vectors were in DXF-format.

In the preprocessing of Senaatti and Hermanni data, laser scanner data was matched visually to map vectors in the XY-directions and height was shifted using known height points. These shifts were done after the original laser scanner observations were transformed from WGS84-coordinates to a local rectangular coordinate system. The coordinate system for all data was local rectangular, with origin at X(easting)=50000, Y(northing)=20000 and ellipsoid WGS84. In Espoonlahti, the coordinate system for all data was the Finnish National Grid zone 2 (Kartastokoordinaattijärjestelmä KKJ) projection Transverse Mercator, ellipsoid International (Hayford) (scale factor at central meridian 1.00000, longitude of central meridian 24 00 00 E, latitude of origin 00 00 00, false easting 2500000.000 meters, false northing 0.000 meters).

18

2.1.3 IGN Test Site Amiens

This study site, with a high density of small buildings is located in the city of Amiens in Northern France. Airborne images were part of a digital acquisition (June 23rd 2001) using IGN’s digital camera. The images were in TIFF-format. The ground pixel size is around 25 cm. The main characteristic of these acquisitions is the large overlap rate (around 80%) between the different images: each point of the terrain appears on a large number of images. Amiens consists of four strips with a total of eleven images. Laser acquisition with a density of 4 points per m2 came from the TopoSys system. A digitised cadastral map, with 2D description of buildings is in DXF-format.

Figure 2-7: Amiens test site.

Figure 2-8: Amiens test site (TopoSys).

19

2.2 Reference Data

2.2.1 Field Measurements

On FGI test sites Espoonlahti, Hermanni and Senaatti reference data was collected from November to December 2003 using a Trimble 5602 DR200+ tacheometer. Measured targets included corners of walls, roofs, chimneys and equivalent constructions as well as ground points next to building corners. Altogether, about 980 points were measured in Espoonlahti, 400 points in Hermanni and 200 points in Senaatti.

Known points were used to orientate the tacheometer to the test site’s coordinate system. In Espoonlahti, points provided by Espoo City Survey Division were used, but the coordinates were transformed from Espoo’s own coordinate system to Finnish National Grid using transformation parameters by the National Land Survey of Finland. After the transformation, one known point was used to determine a small systematic shift to correct the error in the transformation; same point was used as reference point in the laser data and the aerial image data. All known points used in Espoonlahti were also measured using RTK-GPS equipment to ensure that there were no gross errors.

In Hermanni, points provided by Helsinki City Survey Division were used. In Senaatti, the tacheometer was orientated to Helsinki’s rectangular coordinate system by measuring corners of buildings (coordinates were obtained from the cadastral map) and height was obtained by measuring a known height point in the area.

On all three FGI test sites, repeated observations to the same targets from different station setups were made to control the uniformity and accuracy of reference measurements. The differences in these repeated measurements were on average 4.7 cm in plan (max 8.3 cm) and 1.2 cm in height (max 3.5 cm) based on nineteen control observations in total. It is a fact that distance measurement directly to the surface of e.g. a building wall is affected by the angle between the wall and the measuring beam, especially on long distances (as the laser footprint expands). This effect can also be seen in those control measurements where measured distances were somewhat longer than those mainly used in this test, thus, giving a slightly more pessimistic value of the total accuracy expected.

2.2.2 Aerial Images

Reference data for the Amiens test site included thirty-two roof points measured manually from aerial images. The given accuracy (standard deviation) of these points was 25 cm for X, Y and Z. Reference data was measured and delivered by IGN.

2.2.3 Ground Plans

After some minor updating and refinements on the basis of reference measurements in the field, the cadastral maps of Helsinki and Espoo City Survey Division were used to produce raster images of building ground plans with a pixel size of 10 X 10 cm2 for the test sites Hermanni and Espoonlahti.

20

2.3 Produced 3D-models

3D-models were obtained from ten participants (eleven organisations) (Table 2-3). The first models were delivered in February and the last ones in October 2004. The Swedish Defence Research Agency (FOI) was also asked to extract buildings from Espoonlahti in October 2005.

TEST SITE

Espoonlahti Hermanni Senaatti Amiens PARTICIPANT

Building

vectorsDEM

Building

vectorsDEM

Building

vectorsDEM

Building

vectorsDEM

CyberCity AG, Switzerland

DXF ArcView DXF ArcView DXF ArcView

Delft University of Technology, Netherlands

DXFand

VRML VRML

DXFand

VRML VRML

Hamburg University of Applied Sciences and Nebel+Partner, Germany

DXF DXF DXF DXF

Institut Geographique National, France (IGN)

DXFand

VRML

DXFand

VRML

DXFand

VRML

DXFand

VRML

DXFand

VRML

DXFand

VRML

DXFand

VRML

DXFand

VRMLSwedish Defence Research Agency (FOI)

DXF DXF DXF DXF

University of Aalborg, Denmark

VRML

C+B Technik, Germany DXFand

ASCII

DXFand

ASCII

DXFand

ASCII

DXFand

ASCIIInstitut Cartografic de Catalunya, Spain (ICC)

DGN DGN DGN (DGN)

Dresden University of Technology, Germany

VRML VRML

University of Applied Sciences, Stuttgart, Germany

DXF DXF DXF

Table 2-3: Participant-generated 3D-models and data formats.

3 Methods

3.1 Building Extraction

3.1.1 Summary of Applied Data

Table 3-1 summarizes the data use, degree of automation and time use. A more detailed description of the building extraction methods employed is presented in Chapter 3.1.2, of time use in Chapter 3.1.3 and of the level of automation in Chapter 3.1.4.

21

Used data Time use

Participant Laser

data

Aerial

images

Ground

plan

Level of

automation Espoonlahti Hermanni Senaatti Amiens

Cybercity 100 % low medium medium medium Hamburg 100 % low medium high Stuttgart 100 % low-high low low ? IGN (Amiens) 100 % medium low IGN (Espl, Her,Sen)

50 % 50 % medium low low low

ICC laser+aerial 80 % 20 % low-medium high medium high Nebel+Partner 90 % 10 % medium high high medium high ICC laser 100 % low-medium high high high ? FOI 100 % high low low low low FOI outlines 100 % X high low low low low

C+B Technik 100 % X medium-

high ? ? ? ?

Delft 100 % X medium-

high ? medium

Aalborg 100 % X high ? Dresden 100 % X high low low ? = not available

Table 3-1: Summary of used data, level of automation and time use of building extraction.

3.1.2 Building Extraction Methods

3.1.2.1 About method descriptions

Participants of the building extraction project were requested to describe the methods used for building extraction. The authors have edited following process descriptions using these reports.

3.1.2.2 CyberCity AG

CyberCity used aerial images, camera calibration and exterior orientation information. Work was done by experienced personnel.

CyberCity methods are based on concepts reported by Grün and Wang (1998a,b; 1999a,b) and commercialization of the methods in CyberCity AG. The modelling is based on two steps. First, manual photogrammetric stereo measurements are carried out using an in-house developed digital photogrammetric workstation Visual Star. The point measurements are taken in a certain order and points are attached certain codes using CyberCity's Point Cloud Coding System which will help the more automatic process. Then, the resulting point clouds are imported into the main module CC-Modeler, which automatically triangulates the roof points and creates the 3D building models. The building footpoints and walls are generated by projecting and intersecting the boundary roof points with the Digital Terrain Model (DTM). Alternatively, cadastral building footprints can be projected back to the roof, which results in realistic roof overhangs and fulfills the consistency between 2D cadastral maps and 3D city models. With CC-Edit geometric corrections can be carried out to get accurate and correct 3D city models with planar faces and without overlaps or gaps between neighboring buildings etc.

22

Figure 3-1 shows the work- and dataflow in CC-Modeler™.

Used module from the CC-ModelerTM

X Visual Star

Figure 3-1: Work- and Dataflow CC-Modeler™

Step 1: Preparation of the data

The operator prepares the existing data so that it can be used in CC-ModelerTM which includes for example image transformation, converting the existing orientation files and generating the internal orientation file.

Step 2: Data capturing with Visual Star

The operator classifies the captured points as boundary points and interior points, based on their different functionality in one object (Figure 3-2). Different codes are assigned to these points, which help the CC-Modeler to generate the roof faces and an accurate topology. Each roof type can be generated and there are no constraints regarding a roof type library. The interpretation of the data depends on the operator which gives the opportunity to achieve different level of detail. The result (point cloud) is stored as an ASCII file.

X

23

Figure 3-2: Point definition in CC-Modeler.

Step 3: Import to CCModeler

As a topology generator, CC-Modeler generates roof topology from manually measured point clouds. The collected data from Step 2 is imported to CC-Modeler. This workflow can be followed automatically or in step mode for detecting face mistakes that can occur due to inadequate point measurement or very complicated structures.

The vector data of the 3D objects (roof polygons) are projected onto the original aerial image or orthophoto to check how the measured data fits to the ground.

To generate vertical walls, the roof border polygon is projected onto the DTM and intersected with it.

Step 4: Refinement in CC-Edit

Besides other functions, the measured objects, that are as close as possible to their existing size and shape, can be regularized with some functions with CC-Edit. CC-Edit uses two strategies: automatic adjustment based on least squares and an approach of CAD editing. The aim of the editing process is the generation of a 3D model that meets the following conditions: 1) Same height for groups of eaves and ridge points, 2) Roof faces with more than 3 points should be planar, 3) Collinearity of eaves and ridge lines, 4) Right angles in main structures.

Step 5: Final control of the data

To avoid errors, the data is checked as a last step another time by overlapping model and aerial data in the stereo mode of Visual Star. This helps to avoid errors in location and completeness of the dataset.

Step 6: Export to the desired format in CC-Edit

Through an interface in the CC-Edit, the files can be written directly to the DXF format.

3.1.2.3 Hamburg University of Applied Sciences

Hamburg University of Applied Sciences used a digital photogrammetric workstation DPW770 with SocetSet from BAE Systems / Leica Geosystems to extract buildings using aerial images, camera

24

calibration and exterior orientation information (Figure 3-3). Photogrammetrically measured data was corrected and improved by an AutoCAD 2000 system. All measurements and modelling was performed by an unexperienced person.

After import of the digital aerial images in the digital station using the given camera calibration and exterior orientation data, the automatic interior orientation of each image was performed in order to achieve the necessary transformation parameters from pixel to image coordinates. The quality of the exterior orientation of each stereo pair was checked for existing y-parallaxes and by comparing the measured and given control points. If the exterior orientation was accepted, each building was measured manually by the operator using the software module Feature Extraction of SOCET Set. The operator could select three different roof types (flat, peaked and gabled). The measurement sequence of the peaked roof points is illustrated in Figure 3-4. The operator first measured the points at each roof corner (points 1, 2, 3), then the fourth point at the roof top and finally one point at ground elevation. For gabled roofs the operator measured the point sequence 1, 2, 3 and 4 clockwise, the two top points (5 and 6) and finally one point at ground elevation (Figure 3-5). Feature Extraction (Auto Create Mode) automatically creates the remaining sides of the buildings based on the value specified for the ground. Complex features were broken down into simpler components, which were later combined in AutoCAD. Finally, the measured data was transferred to AutoCAD 2000 for the correction of some measurements and for modelling of the complex buildings.

Figure 3-3: Workfow of photogrammetric data processing using SOCET Set on the DPW770.

25

Figure 3-4: Sequence of manual photogrammetric point measurements for peaked roofs (BAE

Systems, 2000).

One of the best ways to extract complex features is to break them down into simpler components. Thus, complex buildings were separated into several parts and combined later in AutoCAD (Figure 3-6). Due to the limitations of manual measurements, the photogrammetrically measured points were not always homogenous, i.e. roof points were often not at the same height or the 3D models of the extracted buildings were often not rectangular, which required a significant amount of manual corrections of the 3D models in AutoCAD 2000 (Figure 3-7). Furthermore, separately measured buildings parts were modelled and combined to one building (Figure 3-7).

Figure 3-5: Sequence of manual photogrammetric point measurement for gabled roofs (BAE

Systems, 2000).

26

Figure 3-6: Breaking down a complex building into simpler components (BAE Systems, 2000).

Figure 3-7: Geometrical corrections and modelling of photogrammetrically measured buildings

using AutoCAD 2000.

3.1.2.4 Nebel+Partner GmbH

Nebel+Partner used TerraScan software by Terrasolid to extract buildings using laser scan data (Figure 3-8). All measurements and modelling was performed by an inexperienced person. Aerial images were not used for any measurements. Image crops were only used as superimposed images for better visual interpretation of the laser point clouds during the measurement of the roofs. Before manual measurements with the point clouds, each laser point set was automatically classified as buildings, ground elevation, and high, medium and low vegetation by TerraScan. In the TerraScan software tool - Construct Building - an algorithm automatically finds roofs based on laser hits on planar surfaces of the roofs, which results in vectorized planes of each roof. Roof boundaries can also be created or modified manually (Figure 3-9).

27

Figure 3-8: Workflow of laser scanning data processing using TerraScan.

Figure 3-9: Superimposed aerial images and generated roof boundaries. Manual change of the

position of the yellow roof.

28

In TerraScan the operator can select three different building boundary types (rectangle, rectangular and polygon) (Figure 3-10). A ‘Rectangle’ boundary consists of four edges, while a ‘Rectangular’ boundary has more than four edges.

Figure 3-10: Building boundary types of roofs.

3.1.2.5 Stuttgart University of Applied Sciences

Stuttgart used inJECT1.9 (pre-release) software to semi-automatically extract buildings using aerial images, camera calibration and exterior orientation information. For a description of the used automated tools see (Gülch and Müller, 2001). One student performed the building extraction as a part of a diploma thesis.

The general workflow was as follows: - Derivation of image pyramids with MATCH-B software (demo version sufficient) - Import of given exterior orientation data into OrthoMaster (demo version sufficient) using

BLUH format (required a small change of columns in the given text file) - Automatic interior orientation in OrthoMaster1.4 (demo version sufficient for manual IO) - Automatic derivation of MATCH-AT project file - Direct import to inJECT1.9 (pre-release) - Measurement of buildings without stereo-viewing using parametric (Figure 3-11) or

polyhedral (Figure 3-12) 3D building models with one common ground height for a building part or for a building composite. Partly the building models were measured as CSG structure forming composite buildings. These were not merged inside inJECT, but were merged externally. The results presented here give the pure measurements inside inJECT. Snapping function, rectangular enforcement for polygonal measurements, and automatic enforcement of planarity were frequently used for the polyhedral buildings. Basically all ground height measurements were done by image matching. The rooftop height and some shape features of saddleback and hip-roof buildings were measured by area based and feature based image matching.

- Storage in GML3 - Export to DXF and partly VRML (Figure 3-13)

29

Figure 3-11: Parametric building models (here saddleback-roof building).

a) b)

Figure 3-12: Polyhedral flat-roof building. a) First five corners are measured. b) The whole roof

outline is measured and with one matched ground point the walls are derived.

In addition to the derived GML3 and DXF files a visualization using VRML was automatically derived with automatically extracted texture (see example in Figure 3-13).

Figure 3-13: Building extraction with inJECT with parametric and polyhedral building models

and textured 3D view in a VRML Browser (Student exercise by A. Novacheva and S.H. Foo).

30

3.1.2.6 IGN

IGN used calibrated aerial images in multiview context in the Amiens test site and calibrated aerial images and laser DSM in the Espoonlahti, Hermanni and Senaatti test sites. All work was carried out by experienced personnel, one person working with Amiens, Espoonlahti and Hermanni and one with Senaatti. For each test site IGN created a pseudo-cadastral map manually using aerial images.

The workflow was divided into preparation (fully automatic procedures such as DSM and true ortho processing), cadastral map edition and pruning, 3D reconstruction of buildings and quality control. Materials and methods varied between test sites.

Amiens test site - First a DSM and the true ortho image were processed by correlation using the multi-view context. Then two methods of reconstruction were used. For prismatic models, the 2D shape was edited and the median height in the DSM was measured. For other models, the skeleton of the central ridge was edited manually in one image and the system automatically reconstructed the 3D shape. Finally, the quality was controlled with a difference image of 3D polygons and DSM.

Espoonlahti test site - Pseudo-cadastral maps were edited on a single image with height adjustment using the roll-button of the mouse. For each 2D polygon a median height was measured on the laser DSM. Finally the quality was controlled with a difference image of 3D polygons and DSM.

Hermanni test site - Pseudo-cadastral maps were edited on a single image with height adjustment using the roll-button of the mouse. Reconstruction was fully automatic using these 2D polygons and laser DSM.

Senaatti test site - On this test site, several modules were used. Sometimes pseudo-cadastral maps were edited on a single image with height adjustment using the roll-button of the mouse. After this reconstruction was carried out using a model driven approach (Flamanc et. al. 2003). Moreover manual tools were used for complex buildings. Finally special tools such as dome edition mode were used to extract specific structures.

3.1.2.7 ICC

ICC used TerraScan, TerraPhoto and TerraModeler software by Terrasolid to extract buildings using laser scan data with and without aerial images. All work was carried out by one person.

The major difference of the applied approach with the Nebel+Partner, who also used TerraScan for the test, was that in the ICC process, the orientation information of aerial images was applied. Thus, the building outlines could be derived using the image data.

3.1.2.8 FOI

FOI used their own in-house software and methods to extract buildings using laser data with (‘FOI outlines’ in results) and without (‘FOI’ in results) the given ground plan. All work was carried out by an experienced person.

Without a ground plan the three pre-processing steps of the building extraction algorithm were: 1. Rasterize the data to obtain digital surface models (DSMzmax and DSMzmin) 2. Estimate the terrain surface (DTM) 3. Classify the data above the ground surface (vegetation / buildings)

31

Each group of connected pixels classified as buildings were used as the ground plan for the building extraction algorithm.

When using the given ground plan, instead of classifying the data (step 3 above), the outlines were used to create a classification image with buildings. The classification image is used to specify the ground plans of the buildings. The outlines were not used when estimating the roof polygons along the contour of the building.

The raw laser scanner data was resampled into a regular grid (0.25 x 0.25 m2). Two grids were saved, one containing the lowest height value in each cell (DSMzmin), and one containing the highest z-value (DSMzmax). The building reconstruction methods are based on DSMzmin in order to reduce noise from antennas for example. The ground surface (DTM) is then estimated (Elmqvist, 2002). Having estimated the ground surface, the remaining pixels are further classified as buildings and vegetation. Each group of connected pixels classified as buildings are used as ground plan for the building extraction algorithm. If the given ground plan is used for the building extraction algorithm, the classification of buildings and vegetation is not needed.

For each ground plan (Figure 3-14) detected in the classification or obtained otherwise, elevation data is used to extract planar roof faces. Figure 3-14b shows the interpolated height data (DSMzmin) of the area classified as a building.

Figure 3-14: (a) Classified building ground plan. (b) DSMzmin.

The planar faces of the roof are detected using clustering of surface normals. The algorithm works on gridded but not interpolated data. For each pixel within the building ground plan a window is formed. The surface normal is estimated by fitting a plane to the elevation data within the window. The plane parameters are estimated using a Least Squares adjustment.

A clustering of the parameters of the surface normals is then performed. In the clustering space, the cell with the largest number of points is used as initial parameters of a plane and an initial estimation of what pixels belong to the plane is made using these parameter values. Finally, an enhanced estimation of the plane parameters using a reweighted Least Squares adjustment is performed and a final estimation of the pixels that belong to the plane is obtained. These pixels form a roof segment. This clustering process is repeated on the remaining non-segmented pixels within the ground plan. Only roof segments having a certain minimum area are kept. After the clustering process, region growing is performed on the roof segments until all pixels within the ground plan belong to a roof segment. Figure 3-15 shows the segmentation result.

32

Figure 3-15: The roof segments.

Having segmented the roof faces of a building, the relationship between the faces is defined. By following the outlines of roof faces, so called ‘topological points’ are first added. Topological points are defined as vertices where a roof face’s neighbour changes. In Figure 3-16, the topological points are marked with triangles. Next, each section of an outline between any two topological points is defined either as an intersection line, height jump section or both. For sections defined as both an intersection line and jump section, a new topological point is added to split the section. In Figure 3-16, the intersection lines (thicker) and the height jump sections are shown.

Figure 3-16: Topological points (triangles), intersection lines (thicker) and height jump sections

of the building.

As shown in Figure 3-16, the jump sections are often noisy and difficult to model. Therefore, for each jump section, lines are estimated using the 2D Hough transform (Figure 3-17 (left)). The slopes of the estimated lines are then adjusted according to the orientation of the plane (Figure 3-17 (right)).

Figure 3-17: Left: Estimated lines for a jump section. Right: Lines adjusted according the

orientation of the plane.

The intersection points of the estimated lines are used as vertices along the jump sections (Figure 3-18). These vertices along the jump sections are used together with the topological points to define the roof polygons.

33

Figure 3-18: Topological points and estimated vertices along height jump sections (triangles).

After having defined the roof polygons, a 3D model of the building can be created (Figure 3-19). The roof of the model is created using the defined roof polygons. The height value of the vertices is obtained from the z-value of the plane at the location of a vertex. Wall segments are inserted between any two vertices along jump sections.

Figure 3-19: The 3D model output from the building reconstruction algorithm.

3.1.2.9 C+B Technik

C+B Technik used their own software to extract buildings using laser scan data and given ground plans. The program CB-LaserCity was developed by Dr.-Ing. Emil Wild for the extraction of buildings from laser scanner data for 3D –city models. The program is written in C++ for Windows operating systems.

The primary goal of the program is to handle large areas and to derive the main shape of buildings automatically. In a first step a triangle net is created from the laser scanner data, which builds the basic data structure for the modelling process. The basic computation method selects the laser scanner points within a building polygon. Triangles are combined to create surfaces and triangle sides are classified as edge lines. The resulting surfaces, edges and corners are analyzed and edited to achieve the typical building objects, which are for example inner vertical walls, horizontal or tilted roof planes and horizontal ridges. The surfaces are also adapted to the building polygon, so that for example an edge line and a house corner fit together. For each situation a complete automatic solution is not possible, so that a complete check and possible editing of the result must be performed

The extraction of buildings from laser scanner data is split first into an automatic computation step and second an interactive check and editing step.

34

Figure 3-20: Work flow used by C+B Technik in building extraction.

The results for each building have the following logical structure: the ground polygon, the outer walls, the inner walls and the roof surfaces. The results can be output as ASCII- or DXF-data.

3.1.2.10 Delft University of Technology

Delft used their own software and methods to extract buildings using laser data and ground plans based on studies by Vosselman and Dijkman (2001) and Vosselman and Süveg (2001). All work was carried out by an experienced person. If building outlines were not available, they were manually drawn in a display of the laser points with colour-coded heights. In the Hermanni test site, only buildings with given ground plan were modelled.

If the point cloud within a building polygon can be represented by a simple roof shape (flat, shed, gable, hip, gambrel, spherical, or cylindrical roof) the model of this roof is fitted to the points with a robust Least Squares estimation. Often building models can be decomposed interactively such that all parts correspond to the above mentioned shape primitives. If the point cloud is such that all roof planes can be detected automatically, an automatic reconstruction is attempted based on the intersection of detected neighbouring roof faces and the detection of height jump edges between roof faces.

If the two situations above do not apply, the building polygon is split into two or more parts until each part fulfils one of the two above conditions. Optionally, point clouds can be edited to remove outlier points that would disable an automatic roof reconstruction.

Preparation

Building Polygons, Triangle Net, Assignment Points to Building

Automatic Computation Step

Loop over all building polygons

Geometric Building Model Building Information

Interactive Step (Check, Editing)

Computation-Database for Complete Re-Computation

Check, Visualization Computation Parameters Line- and Point Editing

35

Detailed Description (Letters refer to steps in the block diagram (Table 3-2))

Common steps for both methods:

A: Polygon splitting is done interactively. Based on the outline of the building, preferred directions of lines for splitting the polygon are generated. These lines can be moved interactively over the polygon. At the optimal position, the polygon can be split by clicking a mouse button.

Shape primitive fitting method (left side of diagram):

E: After selection of the optimal model by the operator, the approximate values of the parameters of this model are derived from the height histogram of the points within the polygon.

F: The height of a point is expressed as a function of the model parameters and linearised. In a robust estimation (ignoring points more than 0.5 m away from the roof surface) the roof shape parameters are obtained.

Automatic roof reconstruction (right side of diagram)

K: Splitting of polygon into concave segments by extending the building outlines at the concave corners of the outline. See (Vosselman and Süveg, 2001) for details.

L: Detection of planar faces in the segments using a 3D Hough transformation. After detection the planar faces are grown and merged over the borders of the ground plan segments. See (Vosselman and Dijkman, 2001) for details. Planes that are nearly parallel or nearly horizontal are constrained to be exactly parallel or horizontal.

M: Detection of intersection lines between roof faces by intersecting all pairs of detected planes and checking if both planes of a pair have points near (the same segment) the intersection line. Also detection of height jump edges between neighbouring roof faces.

N: The segments are further split at the positions of the detected intersection lines and height jump edges until the points in each segment can be represented by one plane.

O: All adjacent segments that can be represented by the same plane are grouped to one segment.

T: Sometimes, a small part of the building (e.g. a dormer) may be missed by the automatic reconstruction procedure, whereas the rest of the building is modelled correctly. In this case, the model is accepted. The missing feature is extracted from the point cloud and fit to the model of a shape primitive.

36

Normal: Manual task of the operator Italic: Task of the operator that is supported by task specific tools Bold: Automatically performed task

Table 3-2: Block diagram of Delft processing strategy

3.1.2.11 University of Aalborg

The modelling of buildings was carried out by three university students as part of their studies (Frederiksen et al., 2004). The applied method used laser scanner and 2D vector map data to extract and model the buildings. The applied method can be used to model rectangular buildings with gable roofs and rectangular buildings with gable roofs in different levels, see Figure 3-21.

Figure 3-21: Types of buildings to be modelled.

A 2D vector map is used to extract the data within the buildings and the 2D vector map is also used as building outline.

After the extraction of the data, the data is interpolated into a rectangular grid. For every point in the grid, a surface normal based on the point and its neighbouring points are calculated (see Burrough and McDonnell, 1998, p. 190-192). Now all points belonging to a roof plane have to be found. An

37

assumption is made: the two longest parallel sides of the building each have one roof plane orientated towards the side of the building, meaning a surface normal on the roof plane is perpendicular in the xy-plane to the side of the building. This means that every point is examined regarding its surface normal, and surface normals perpendicular to the side of the building and with approximately the same orientation are grouped. The result is two groups of surface normals – each group representing a roof plane.

The points of each roofplane are now adjusted by the principles of Least Squares adjustment and a method called "the Danish method" (Juhl 1980). The Danish method is a method used in adjustments and is used to automatically weight down points with large residuals compared to the determined roof plane from the Least Squares adjustment. This means that points not belonging to the roof plane, e.g. points on chimneys, are sorted out of the group automatically and will not have influence on the final determination of the parameters of the roof planes (Juhl 1980).

Gable roofs at different levels will also show only two groups of points with approximately similar surface normals, because the method cannot distinguish between roof planes at different levels. To find the border between the buildings a gradient filter is applied to the grid and large gradients are grouped by means of a connected components analysis. Groups are adjusted by Least Squares adjustment to a line and thus a building with gable roofs at different levels is divided into rectangular buildings with two roof planes.

3.1.2.12 Dresden University of Technology

The Dresden method (Hofmann, 2004) only uses point clouds obtained by a pre-segmentation of airborne laser scanner data. All work was carried out by an experienced person. It is a plane-based approach that presumes that buildings are characterized by planes. It utilizes a TIN-structure that is calculated into the point cloud. The method only uses point clouds of the laser scanner data that contain one building. In order to get such point clouds polygons coarsely framing the building can be used to extract the points (e.g.in ArcGIS). The polygons can be created manually or map or ground plan information can be used. See Figure 3-22 to Figure 3-25.

Figure 3-22: Building model algorithm by Dresden.

38

Figure 3-23: A building’s point cloud with TIN-structure.

The parameters of every TIN-mesh, which define its position in space uniquely, are mapped into a 3-D triangle-mesh parameter space. As all triangles of a roof face have similar parameters, they form clusters in the parameter space.

Figure 3-24: Clusters in the parameter space.

Those clusters that represent roof faces are detected with a cluster analysis technique. By analyzing the clusters, significant roof planes are derived from the 3-D triangle parameter space while taking common knowledge of roofs into account. However, no prior knowledge of the roof as, for example, the number of roof faces is required. The obtained roof planes are intersected in accordance to their position in space. Analysing the intersected roof faces, determines the roof outlines and the ground plan is derived.

39

Figure 3-25: A 3D building model with TIN-structure.

3.1.3 Time Used for Building Extraction

Table 3-3 summarises the time used for building extraction as reported by participants. They were also asked to estimate the time needed for extracting buildings of 10 km2 of urban area.

Espoon-

lahti Hermanni Senaatti Amiens 10 km

2 Comment

Cybercity 20.3 10.0 16.8 1521 Hamburg 19.0 23.0 2305-3090 Stuttgart ? 2.0 ? ?

IGN 4.0 0.3 7.0 0.4 110 10 km2 refers to

Amiens ICClaser+aerial

32.0 5.0 19.0

Nebel + Partner

27.0 11.0 15.0 12.0 1156-31250

ICC laser 31.0 8.0 14.0 8.0 Delft 12.5 1516 Aalborg ? C+B Technik ? ? ? ? ?

Dresden 1.8 1.8 ? Hermanni: 1 building, Senaatti: 5 buildings

FOI 5.4 1.6 4.6 0.3 282

Table 3-3: Time used per test site and estimated time needed for extraction of 10 km2 area

(hours).

40

3.1.4 Level of Automation

3.1.4.1 Summary

The degree of automation was estimated as a value from 0 (0 = no automation in the process) to 10 (10= fully automatic) based on the workflow charts and the procedure descriptions and comments provided by the participants (Table 3-4).

Test site

Participant Espoonlahti Hermanni Senaatti Amiens

CyberCity 3 3 3 Hamburg 3 3 Stuttgart 5 7 3 IGN 5 5 5 6 Nebel+Partner 6 6 6 6 ICC laser+aerial 4 6 4 ICC laser 4 6 4 FOI 10 10 10 10 FOI outlines 10 10 10 10 C+B Technik 6 6 6 7 Delft 7 6 Aalborg 8 Dresden 9 9 * It should noticed that the degree of automation has been subjectively estimated. The accuracy of the estimation is not high and more information about the automation level can be obtained from each individual methods depicted in Chapter 3.1.2.

Table 3-4: Estimated degree of automation in building extraction*.

3.1.4.2 CyberCity AG

CyberCity can use automatic procedures when creating roof topology from manually measured points and when refining the extracted building, but the level of automation is considered low. This is also a choice made by CyberCity: “Due to the complexity of natural scenes and the lack of performance of image understanding algorithms, the fully automated methods cannot guarantee results stable and reliable enough for practical use. Therefore, CyberCity has developed a semi-automated approach, which gives the human operator strong computational support in order to generate 3D City models from aerial images efficiently“ (Grün and Wang, 1998a,b).

3.1.4.3 Hamburg University of Applied Sciences

Although there is some automation included, most procedures required work by the operator (see Figure 3-3), and the level of automation is considered to be low.

3.1.4.4 Nebel + Partner and ICC

Nebel + Partner used automatic and semi-automatic procedures for building extraction after automatic classification of laser scanner data (see Figure 3-8). The level of automation can be considered medium as time-consuming manual corrections were carried out.

41

ICC used the same methods as Nebel + Partner, but according to them the automatic procedures in building extraction gave very bad results and most of the work had to be redone by hand (90%). The Hermanni test site was simpler and automatic results were better, so less editing was needed (30%). The use of aerial images did not modify the automatic procedures. They only helped the manual editing. The level of automation was considered medium for Hermanni and low for other test sites.

3.1.4.5 Stuttgart University of Applied Sciences

The level of automation differed in the test sites.

In Hermanni the level of automation was high. Roof-top matching, form adaptation and mostly parametric building models were used. Most often a saddleback-roof-building was used. In one case the walls were taken away to combine three composite primitives. Very little manual adaptation of parameters was required after the form adaptation and the roof-top matching. Ground height was generally determined by matching. Copying capabilities were also used.

In Espoonlahti the level of automation varied from low to medium. Mostly polyhedral buildings were measured. Ground height was generally determined by matching. Automatic snapping, rectangular and planarity enforcement were frequently used for the polygon measurements. For the very few buildings of type saddleback or hip-roof type roof-top matching, form adaptation and ground height matching were used.

In Senaatti the level of automation was low. Mostly polyhedral building models were applied. Matching along gutter lines did not always work. Automatic snapping, rectangular and planarity enforcement were frequently used for the polygon measurements. Ground height was generally determined by matching.

3.1.4.6 IGN

In the IGN method the preparation (such as DSM and true ortho processing) is fully automatic. Cadastral plan editing and pruning is done manually. The level of automation in 3D reconstruction varies: the model driven approach is fully automatic, straight skeleton edition and 3D shape reconstruction is semi-automatic as is the dome edition module. Quality control is done automatically, but also visual control is performed. The Senaatti test site is more complex so there was more human interaction.

3.1.4.7 FOI

FOI used fully automatic procedures.

3.1.4.8 C+B Technik

In the C+B Technik method the building ground plans are needed, and if they are not available, they have to be created manually. The degree of automation differs according to the area type and the shape of the buildings. Especially in rural and suburb areas, where simple buildings and roof shapes can be found and the building polygons fit to the laser scanner data a very high degree of automation can be reached. In other areas, especially in downtown areas, where complex areas can be found, the subdivision of complex buildings into several computation parts can support the automatic process.

42

3.1.4.9 Delft University of Technology

The level of automation was considered medium. There are many automatic steps in the processing strategy (see Table 3-2), but an operator is needed to guide the extraction process and for quality control.

3.1.4.10 University of Aalborg

In the applied method the building ground plans are needed, and if they are not available, they have to be created manually. The level of automation in building extraction is high, but some operator involvement was needed as the procedures were at the development stage. Some of them were tested for the first time.

3.1.4.11 Dresden University of Technology

In the Dresden method the building polygons are needed. If ground plans or such are not available, they have to be created manually. After this point cloud extraction can be performed automatically with a batch algorithm in e.g. ArcGIS. The user has to define the mean point spacing and the accuracy of the laser scanner data. Then, the whole building model reconstruction runs automatically. One by one, it reads the point clouds and processes them.

3.2 Accuracy Evaluation

3.2.1 Analysis Using Reference Points

Reference points were used to analyze the accuracy of the location (single point measurement), length (distance between two points) and roof inclination (slope between two points) of the modelled buildings. Single points were analyzed separately for planimetric and height errors. If a DEM was included, building heights were measured to the DEM using bi-linear interpolation. If not, wall vectors were used. On some models the same ground height was used for whole buildings resulting in large errors in building height determination. These cases were marked at the building height analysis stage.

Root mean squared error (abbreviated to RMSE, Equation 3-1), was calculated for building length, height and roof inclination.

( )

n

ee

RMSE

n

iii

=

−= 1

221

(Equation 3-1)

where e1i is the result obtained with the described retrieved model, e2i is the corresponding reference measured value, and n is the number of samples. Additionally, minimum, maximum, medium, mean, standard deviation and interquartile range (IQR, Equation 3-2) values were calculated. Interquartile range values represent the range between the 25th and 75th quartiles.

thth ppIQR 2575 −= (Equation 3-2)

43

where p75th is the value at the 75th quartile and p25th is the value at the 25th quartile. For example, if the IQR is 20 cm and the median value is 0, 50% of the errors are within ±10 cm. Significantly deviating measurements (outliers) were detected using threshold levels: the lower bound at the 25th quartile minus 1.5*IQR and the upper bound at the 75th quartile plus 1.5*IQR. IQR is not as sensitive to large deviations as standard deviation/error. Additionally, the coefficient of determination R2 was calculated to help to separate cases between low variability of the reference data and high estimation accuracy and high variability of the reference data and low estimation accuracy.

Descriptive statistics are presented before and after outliers have been removed.

3.2.2 Analysis Using Reference Raster Ground Plans

In the Espoonlahti and Hermanni test sites reference raster ground plans were used to compute the total relative building area and total relative shape dissimilarity (Henricsson & Baltsavias, 1997). Total relative building area gives the difference between modelled building area and reference building area. Total relative shape dissimilarity is the sum of the area difference and the remaining overlap error, i.e. the sum of missing area and extra area divided by the reference area. The vector models delivered by participants were rasterized (pixel size 10 cm) and compared to reference raster ground plans. For those participants that modelled only buildings with given ground plans, the reference raster ground plans consist only of these buildings. Total relative building area and shape dissimilarity were computed in two ways. Firstly all buildings in the test sites were included, giving the values for the whole test site. Secondly only those buildings that exist on both reference and delivered model were included, giving the values for modelled buildings.

The eaves of the roofs are not included in reference raster ground plans thus making the reference buildings somewhat smaller than extracted buildings in the Hermanni test site.

4 Results

4.1 Level of Detail

4.1.1 Number of Targets

One measure of the level of detail is the amount of points that could be measured for geometric model analysis using the produced 3D-models. It should be noticed that this measure only gives an indication of the level of the detail due to the low number of buildings extracted by some of the partners. The total amount of reference points found (in %) on roofs (eaves, ridges and other structures) is presented in Table 4-1.

44

Test site

Participant Espoonlahti Hermanni Senaatti Amiens

CyberCity 87 76 75 Hamburg 69 33 69 Stuttgart 60 25 31 IGN 66 26 28 75 ICC laser+aerial 80 33 36 Nebel+Partner 76 47 34 ICC laser 62 24 24 Delft 35 1) 33 Aalborg 20 FOI 41 23 28 63 FOI outlines 73 1) 29 1) 100 1)

C+B Technik 59 1) 32 1) 23 1) 50 1)

Dresden 47 2) 57 3)

1) Only buildings with given ground plan extracted. 2) Only one building extracted. 3) Five buildings extracted.

Table 4-1: Total amount [%] of reference points that could be found from produced 3D-models.

4.1.2 Extracted Building Vectors

Extracted building vectors (top view) are shown in Figure 4-1 (Espoonlahti), Figure 4-2 (Hermanni), Figure 4-3 (Senaatti) and Figure 4-4 (Amiens). CyberCity produced the most detailed models in the FGI test sites, especially in Hermanni and Senaatti, as can be seen from Table 4-1 and the figures below. They were the only participant that also modeled chimneys and equivalent smaller objects on roofs, which would be impossible to identify using only laser scanner data on test sites where the point density is low. Test site visualizations delivered by participants are shown in Appendix 1.

45

CyberCity Stuttgart Hamburg

IGN ICC laser+aerial ICC laser

Nebel+Partner C+B Technik FOI

FOI outlines

Figure 4-1: Extracted buildings of Espoonlahti.

46

CyberCity Stuttgart IGN

Delft ICC laser+aerial Nebel+Partner

ICC laser Aalborg C+B Technik

Dresden FOI FOI outlines

Figure 4-2: Extracted buildings of Hermanni.

47

CyberCity Hamburg Stuttgart

IGN ICC laser+aerial ICC laser

Nebel+Partner Delft C+B Technik

Dresden FOI FOI outlines

Figure 4-3: Extracted buildings of Senaatti.

48

IGN Nebel+Partner C+B Technik

FOI FOI outlines

Figure 4-4: Extracted buildings of Amiens.

4.2 Geometric Accuracy

4.2.1 Building Location

Building location accuracies for all modelled buildings in Espoonlahti are given in Table 4-2 and for selected buildings (same four buildings modelled by all participants) in Table 4-3. In Hermanni and Senaatti the accuracies are given for all targets (Hermanni Table 4-4, Senaatti Table 4-8), eaves (Hermanni Table 4-5, Senaatti Table 4-9), ridges (Hermanni Table 4-6, Senaatti Table 4-10) and other targets (Hermanni Table 4-7, Senaatti Table 4-11). In Amiens the accuracies are given for all targets, eaves and ridges (Table 4-12).

Planimetric and vertical single point deviations are depicted in Appendix 2.

49

HeightCyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI 2005

N 318 256 137 243 293 277 225 54 67 152 163 Min -270 -215 -56 -193 -255 -83 -239 -489 -202 -252 -351 Max 16 194 282 373 269 315 461 84 810 324 103 Median -14 -10 -3 -18 -13 -13 30 -18 -16 -16 9 Mean -16 -13 2 -18 -18 -15 29 -33 17 -18 7 STD 21 47 41 46 35 27 61 93 136 70 43 IQR 12 51 14 11 16 16 17 19 18 20 12 R2 0.9993 0.9985 0.9969 0.9983 0.9994 0.9979 0.9992 0.9985 0.9939 0.9941 0.9991 Outliers

removed

N 308 247 123 199 261 243 178 49 52 127 141 Min -36 -108 -30 -40 -44 -45 -6 -35 -44 -57 -13 Max 9 80 22 4 12 16 51 7 16 22 32 Median -14 -10 -4 -18 -12 -11 30 -19 -16 -16 9 Mean -13 -13 -5 -18 -13 -12 29 -20 -15 -18 9 STD 9 35 11 9 10 10 10 11 11 14 8 IQR 11 47 11 8 13 13 12 19 12 14 11

NorthCyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI 2005

N 329 261 137 250 301 285 233 57 71 156 163 Min -35 -89 -39 -115 -45 -98 -118 -75 -352 -820 -663 Max 70 125 40 154 82 122 80 34 98 622 779 Median 7 -1 -5 -10 8 -1 1 -4 -22 -32 23 Mean 8 0 -5 -11 9 2 1 -4 -28 -36 22 STD 13 22 13 35 16 40 31 19 68 137 122 IQR 15 22 15 39 19 58 43 16 53 117 81 Outliers

removed

N 322 252 133 241 294 284 232 54 65 147 147 Min -23 -45 -32 -88 -24 -98 -79 -31 -109 -246 -144 Max 37 44 20 59 45 109 80 30 86 205 162 Median 7 -1 -5 -10 8 -2 1 -4 -21 -29 22 Mean 7 -1 -6 -11 9 2 1 -2 -17 -24 17 STD 11 18 11 29 14 39 30 14 45 86 64 IQR 14 22 15 36 18 57 43 15 50 111 76

East Cyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI 2005

N 329 261 137 250 301 285 233 57 71 156 163 Min -128 -79 -25 -100 -58 -120 -98 -66 -214 -802 -246 Max 51 128 40 230 47 127 164 64 102 311 814 Median 12 6 3 -4 4 -4 -3 -13 -13 -10 -4 Mean 11 7 3 0 4 -3 -2 -11 -18 -15 4 STD 14 26 11 36 16 40 36 22 57 118 106 IQR 13 25 15 43 20 47 44 25 66 106 92 Outliers

removed

N 318 249 135 246 295 277 228 55 70 145 153 Min -16 -42 -25 -82 -34 -88 -83 -60 -143 -208 -174 Max 37 54 27 82 44 87 83 25 102 195 183 Median 12 6 3 -4 5 -4 -3 -13 -12 -10 -9 Mean 11 5 3 -2 5 -3 -3 -11 -15 -10 -8 STD 10 19 10 32 15 36 31 19 52 76 69 IQR 13 23 14 43 19 46 44 22 65 99 86

Table 4-2: Espoonlahti building location accuracy in cm, all buildings.

50

HeightCyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI2005

N 57 44 45 47 49 45 33 32 37 26 28 Min -123 -215 -14 -30 -42 -42 -46 -31 -76 -29 -90 Max -4 25 15 373 -4 2 39 53 149 130 22 Median -17 -12 -3 -20 -16 -16 23 -24 -17 -18 14 Mean -19 -16 -3 -3 -18 -15 17 -18 -14 -12 10 STD 15 38 5 80 10 9 19 20 41 31 20 IQR 6 16 3 6 14 14 14 15 10 6 4 Outliers

removed

N 52 31 39 43 44 31 30 31 24 26 Min -25 -27 -8 -30 -35 0 -31 -29 -29 7 Max -6 21 4 -9 2 39 -12 -2 -14 21 Median -17 -12 -2 -20 -16 23 -28 -17 -19 14 Mean -17 -10 -2 -20 -15 21 -23 -17 -20 13 STD 4 11 3 4 8 11 8 6 4 3 IQR 5 14 3 5 13 11 14 6 5 4

NorthCyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI2005

N 58 44 44 47 48 44 32 35 41 25 27 Min -11 -82 -26 -71 -18 -60 -55 -75 -181 -187 -133 Max 46 125 40 47 82 122 66 34 79 75 107 Median 7 -3 -5 -3 3 1 7 -4 -29 -1 18 Mean 9 1 -5 0 4 10 10 -5 -31 -22 6 STD 11 31 12 23 15 38 33 22 55 70 54 IQR 13 18 12 29 14 60 58 16 43 73 51 Outliers

removed

N 56 40 42 46 47 31 36 23 23 Min -11 -39 -26 -41 -18 -31 -96 -145 -44 Max 33 29 15 47 24 19 38 75 73 Median 7 -4 -6 -3 2 -4 -28 2 18 Mean 8 -4 -6 1 3 -4 -26 -8 8 STD 10 14 8 21 10 13 36 53 32 IQR 13 16 10 30 14 14 37 50 49

East Cyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI2005

N 58 44 44 47 48 44 32 35 41 25 27 Min -8 -79 -13 -65 -20 -106 -45 -66 -214 -372 -196 Max 37 46 37 56 47 123 79 64 68 243 87 Median 12 3 8 -16 4 -16 -8 -15 -17 -10 -4 Mean 13 1 6 -16 6 -20 -2 -13 -23 -18 -10 STD 9 24 10 26 13 40 31 25 53 104 64 IQR 11 25 12 23 13 43 45 30 69 108 85 Outliers

removed

N 56 42 43 42 45 43 34 40 23 26 Min -8 -42 -13 -50 -20 -106 -66 -132 -127 -138 Max 29 46 27 27 26 47 25 68 65 87 Median 12 4 8 -16 3 -16 -16 -17 -10 -3 Mean 12 5 6 -18 4 -23 -16 -18 -14 -3 STD 8 18 9 17 11 34 22 44 54 53 IQR 10 22 11 20 14 41 30 64 87 90

Table 4-3: Espoonlahti building location accuracy in cm, selected buildings.

51

HeightCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 165 54 57 71 103 54 44 44 41 8 37 49 Min -20 -42 -52 -20 -20 -11 -18 -17 -22 -2 -37 -91 Max 66 112 567 25 32 26 268 63 267 42 1142 21 Median 14 33 5 0 0 3 4 1 12 13 6 -5 Mean 15 34 38 1 0 4 20 9 25 14 86 -11 STD 13 39 128 8 8 8 57 19 59 15 261 24 IQR 16 50 21 9 9 7 19 19 38 22 45 19 R2 0.9995 0.9949 0.9728 0.9998 0.9998 0.9998 0.9940 0.9983 0.9914 0.9957 0.8959 0.9983Outliers

removed

N 162 52 68 97 47 39 40 39 31 44 Min -17 -12 -18 -19 -11 -18 -17 -22 -37 -38 Max 38 42 14 14 15 37 40 54 65 21 Median 14 3 0 0 2 3 -1 12 -3 -4 Mean 16 7 0 -1 2 5 4 13 7 -5 STD 12 14 7 7 6 12 13 22 24 13 IQR 16 20 9 9 6 11 17 38 26 16

NorthCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 165 54 57 71 103 54 44 44 41 8 37 49 Min -42 -47 -181 -50 -82 -284 -103 -103 -103 -68 -127 -204 Max 75 58 114 38 56 54 113 113 113 117 60 38 Median -2 5 -22 -6 4 1 -15 -11 -18 -2 -40 -27 Mean -2 4 -29 -6 3 -3 -14 -9 -19 6 -35 -43 STD 12 17 53 16 25 44 43 46 45 63 48 63 IQR 12 17 52 23 27 27 39 38 50 76 59 82 Outliers

removed

N 159 52 51 100 52 40 39 39 48 Min -24 -26 -107 -49 -36 -91 -80 -103 -191 Max 20 37 57 56 54 57 69 69 38 Median -2 5 -22 5 2 -15 -11 -19 -24 Mean -2 4 -32 5 4 -20 -8 -25 -40 STD 9 14 34 21 20 31 33 36 59 IQR 11 17 46 27 26 37 26 48 79

East Cyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 165 54 57 71 103 54 44 44 41 8 37 49 Min -35 -62 -66 -24 -85 -52 -63 -63 -64 -19 -98 -193 Max 61 49 171 31 54 115 100 100 101 101 132 174 Median 9 3 21 11 7 7 5 13 13 22 6 -5 Mean 10 0 24 10 2 7 7 15 9 28 9 1 STD 13 25 51 12 24 28 33 35 34 46 54 66 IQR 15 31 56 17 31 26 46 44 44 80 53 80 Outliers

removed

N 158 53 54 70 101 50 43 40 33 47 Min -11 -56 -66 -16 -55 -37 -63 -64 -67 -130 Max 37 49 126 31 54 43 57 57 0.99 145 Median 9 3 18 11 7 6 4 12 9 -5 Mean 10 1 17 11 4 3 5 6 15 2 STD 10 23 41 11 22 20 30 31 42 55 IQR 15 31 49 15 30 22 40 42 49 76

Table 4-4: Hermanni building location accuracy in cm, all targets.

52

HeightCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 54 38 43 50 53 40 23 28 24 5 22 34 Min -20 -42 -52 -20 -20 -11 -6 -17 11 13 -37 -91 Max 36 107 567 25 21 26 268 63 267 42 1142 21 Median 4 29 9 1 -3 4 15 13 31 21 20 -7 Mean 5 29 52 2 -1 6 39 15 49 23 130 -14 STD 11 39 144 9 9 9 74 22 68 12 331 28 IQR 14 58 19 14 10 12 28 30 30 12 55 32 R2 0.9997 0.9949 0.9714 0.9998 0.9998 0.9998 0.9912 0.9980 0.9925 0.9919 0.7976 0.9976Outliers

removed

N 53 38 52 21 22 19 32 Min -20 -9 -20 -6 11 -37 -71 Max 34 42 19 51 54 65 21 Median 4 8 -3 14 31 16 -6 Mean 5 11 -2 18 29 15 -9 STD 10 13 8 18 15 28 22 IQR 13 17 10 28 29 32 30

NorthCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 54 38 43 50 53 40 23 28 24 5 22 34 Min -30 -47 -181 -50 -55 -284 -103 -103 -103 -68 -127 -204 Max 20 58 114 38 52 54 113 113 113 117 60 38 Median -2 5 -29 -7 6 4 -25 -19 -24 22 -49 -45 Mean -3 4 -37 -7 5 -4 -26 -19 -28 8 -48 -54 STD 11 19 56 18 24 51 49 49 50 74 49 70 IQR 14 27 50 26 30 34 49 35 59 91 47 107 Outliers

removed

N 38 52 39 21 24 23 21 Min -107 -49 -59 -103 -91 -103 -127 Max 57 52 54 7 41 69 33 Median -31 8 5 -28 -19 -25 -52 Mean -39 6 3 -37 -21 -34 -53 STD 35 23 24 33 34 41 44 IQR 48 27 34 51 25 53 48

East Cyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 54 38 43 50 53 40 23 28 24 5 22 34 Min -31 -62 -66 -24 -85 -52 -63 -63 -64 -19 -98 -193 Max 61 49 171 31 43 115 100 100 101 101 132 174 Median 4 6 23 12 7 6 -2 13 1 26 3 2 Mean 5 1 26 11 1 6 4 10 5 39 6 9 STD 14 26 57 13 27 31 41 38 42 47 62 74 IQR 15 33 65 16 35 29 54 58 60 89 74 88 Outliers

removed

N 52 42 49 52 38 33 Min -24 -66 -16 -65 -52 -130 Max 29 148 31 43 59 174 Median 4 22 12 7 5 2 Mean 5 22 12 2 2 15 STD 11 53 12 24 24 66 IQR 15 65 23 34 16 88

Table 4-5: Hermanni building location accuracy in cm, eaves.

53

HeightCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 16 16 14 16 16 14 10 14 10 3 10 15 Min 7 -42 -12 -9 -10 -6 -4 -6 -20 -2 -7 -14 Max 34 112 3 6 3 4 3 1 7 1 -1 -2 Median 19 41 -7 -1 -1 0 0 -1 -5 -1 -4 -4 Mean 19 46 -7 -1 -2 -1 0 -1 -6 -1 -4 -5 STD 8 36 4 4 4 3 3 2 10 2 2 3 IQR 12 33 6 6 8 3 5 2 20 1 4 4 R2 0.9998 0.9955 0.9999 1.0000 0.9999 1.0000 1.0000 1.0000 0.9995 0.9998 1.0000 1.0000Outliers

removed

N 14 13 2 14 Min 13 -3 -2 -11 Max 95 1 -1 -2 Median 41 -1 -2 -4 Mean 47 -1 -2 -5 STD 24 1 1 3 IQR 23 2 1 4

NorthCyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 16 16 14 16 16 14 10 14 10 3 10 15 Min -15 -26 -68 -21 -46 -16 -21 -66 -58 -31 -65 -74 Max 8 22 84 10 56 24 87 82 92 63 58 21 Median -6 5 -13 -7 4 -3 -6 -2 -9 -27 -7 -18 Mean -6 3 -7 -5 3 0 5 9 -6 2 -8 -20 STD 6 11 34 9 23 11 32 36 42 53 35 34 IQR 8 10 13 11 28 14 23 32 31 4 22 64 Outliers

removed

N 15 11 9 12 9 2 7 15 Min -8 -34 -21 -11 -58 -31 -16 -74 Max 22 -2 22 69 23 -27 19 21 Median 5 -13 -9 -2 -15 -29 -7 -18 Mean 5 -13 -4 9 -17 -29 -4 -20 STD 8 9 13 23 26 3 12 34 IQR 10 6 18 30 50 4 9 64

East Cyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 16 16 14 16 16 14 10 14 10 3 10 15 Min -10 -52 -23 -16 -20 -23 -23 -26 -20 -18 -63 -70 Max 26 30 50 19 36 43 38 66 42 62 51 75 Median 10 1 16 8 12 7 8 24 14 -18 7 -18 Mean 9 -3 18 8 10 8 9 23 12 9 4 -16 STD 10 22 21 9 14 16 22 29 18 46 37 39 IQR 11 36 29 10 22 23 41 43 28 1 35 42 Outliers

removed

N 15 2 14 Min -3 -18 -70 Max 19 -18 45 Median 9 -18 -24 Mean 10 -18 -23 STD 7 1 31 IQR 9 1 42

Table 4-6: Hermanni building location accuracy in cm, ridges.

54

HeightCyber-

City Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 95 0 0 5 34 0 11 2 7 0 5 0 Min 1 -7 -20 -18 -7 -22 -10 Max 66 1 32 16 -5 0 141 Median 19 -5 2 -1 -6 -8 117 Mean 21 -4 2 -3 -6 -11 74 STD 11 3 9 10 1 8 75 IQR 19 3 5 12 2 15 138 R2 0.9996 0.9922 0.9996 0.9973 na 0.9997 0.9975 Outliers

removed

N 94 Min 1 Max 38 Median 19 Mean 20 STD 10 IQR 19

NorthCyber-

City

ICC

laser+

aerial

Nebel+

Partner Delft Aalborg

C+B

Technik

FOI

outlines

N 95 0 0 5 34 0 11 2 7 0 5 0 Min -42 -9 -82 -65 4 -24 -73 Max 75 19 48 57 4 17 49 Median 0 10 1 -10 4 -15 -46 Mean -1 7 -2 -7 4 -10 -28 STD 13 13 27 34 0 17 49 IQR 10 26 25 35 0 9 62 Outliers

removed

N 90 32 5 Min -20 -36 -24 Max 19 48 -14 Median 0 2 -22 Mean -1 3 -20 STD 8 18 5 IQR 10 23 8

East Cyber-

City

ICC

laser+

aerial

Nebel+

Partner Delft Aalborg

C+B

Technik

FOI

outlines

N 95 0 0 5 34 0 11 2 7 0 5 0 Min -35 -7 -49 -24 28 -8 -7 Max 52 31 54 35 35 37 83 Median 10 0 3 16 32 20 25 Mean 12 6 1 9 32 17 38 STD 13 15 24 21 5 18 38 IQR 16 12 26 41 7 34 79 Outliers

removed

N 91 4 33 Min -7 -7 -49 Max 37 5 50 Median 10 -1 3 Mean 12 -1 -1 STD 10 5 23 IQR 16 7 32

Table 4-7: Hermanni building location accuracy in cm, other targets.

55

HeightCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 110 46 26 41 52 49 35 48 16 8 17 41 Min 2 -78 -54 -81 -117 -127 -129 -134 -221 1 -146 -149 Max 65 94 54 65 50 50 134 1462 1369 30 115 237 Median 26 10 2 7 2 1 3 4 33 6 16 -15 Mean 29 10 4 1 -3 -6 4 57 209 10 14 -18 STD 14 37 32 28 30 31 46 297 467 10 57 64 IQR 20 35 47 31 17 17 35 34 152 4 34 47 R2 0.9995 0.9961 0.9965 0.9977 0.9967 0.9971 0.9929 0.8410 0.8632 0.9959 0.9929 0.9875Outliers

removed

N 42 38 44 43 32 44 14 6 14 37 Min -55 -38 -29 -31 -51 -50 -221 1 -34 -113 Max 60 39 31 28 75 44 271 7 44 86 Median 10 8 2 2 5 4 25 5 14 -15 Mean 9 3 1 1 8 2 43 5 12 -15 STD 29 20 13 13 28 23 126 2 23 37 IQR 31 28 11 12 30 31 150 4 29 41

NorthCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 110 48 26 41 52 49 35 48 16 8 17 41 Min -81 -98 -79 -150 -96 -164 -133 -231 -75 -91 -100 -318 Max 39 75 60 72 68 259 196 102 82 -3 78 219 Median 6 4 7 -23 3 -11 -43 -17 -10 -26 -18 -50 Mean 2 0 6 -19 0 -8 -39 -17 6 -32 -11 -64 STD 19 32 25 41 29 71 69 69 59 29 53 124 IQR 22 34 22 41 33 78 98 0.99 114 30 92 182 Outliers

removed

N 107 46 24 37 49 47 34 47 Min -33 -69 -23 -90 -58 -132 -133 -166 Max 39 65 50 58 41 107 70 102 Median 6 4 7 -23 3 -11 -47 -14 Mean 4 1 7 -23 2 -10 -46 -13 STD 15 26 16 28 22 56 56 62 IQR 20 26 18 36 32 64 98 0.99

East Cyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 110 48 26 41 52 49 35 48 16 8 17 41 Min -24 -120 -59 -100 -38 -125 -92 -108 -107 -54 -102 -178 Max 41 34 25 73 61 61 87 62 62 43 91 214 Median 4 -19 -8 -5 -15 -10 -11 -21 -39 -10 -28 1 Mean 5 -24 -9 -9 -13 -14 -10 -23 -29 -10 -24 13 STD 13 38 19 34 17 39 34 44 57 27 57 90 IQR 15 37 30 29 22 47 39 64 79 22 80 119 Outliers

removed

N 107 44 37 51 47 33 7 Min -24 -86 -62 -38 -77 -58 -54 Max 32 34 51 15 61 39 0 Median 3 -15 -5 -15 -10 -11 -12 Mean 4 -16 -4 -14 -9 -10 -17 STD 12 28 24 14 33 26 18 IQR 15 33 27 22 37 39 16

Table 4-8: Senaatti building location accuracy in cm, all targets.

56

HeightCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 56 34 21 29 42 38 28 39 13 5 15 32 Min 2 -78 -54 -81 -112 -112 -113 -122 -53 3 -146 -149 Max 59 94 50 39 50 50 134 1462 1369 30 115 237 Median 21 14 1 11 2 1 9 13 54 7 20 -18 Mean 24 13 1 3 0 -4 13 75 276 13 16 -16 STD 14 39 31 26 27 27 42 327 495 11 60 68 IQR 16 42 57 34 16 16 30 37 239 17 27 44 R2 0.9996 0.9959 0.9969 0.9982 0.9975 0.9979 0.9947 0.8355 0.8898 0.8580 0.9923 0.9871Outliers

removed

N 55 33 28 36 34 25 36 11 12 27 Min 2 -60 -38 -28 -31 -28 -47 -53 -34 -67 Max 52 94 39 31 28 61 44 271 44 59 Median 20 14 13 2 2 9 10 36 18 -15 Mean 23 16 6 2 1 11 4 77 14 -16 STD 13 37 20 13 14 23 24 109 25 32 IQR 16 42 33 10 13 30 35 110 27 30

NorthCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 56 35 21 29 42 38 28 39 13 5 15 32 Min -61 -98 -79 -150 -66 -164 -133 -231 -75 -91 -100 -318 Max 39 75 60 72 68 259 196 102 82 -3 78 163 Median 6 5 6 -18 3 -9 -44 -31 -20 -29 -18 -57 Mean 3 1 6 -19 2 -5 -36 -23 7 -40 -11 -73 STD 17 33 28 45 26 76 74 74 62 34 54 117 IQR 20 26 24 41 34 73 111 111 127 67 76 163 Outliers

removed

N 55 31 19 27 41 36 27 Min -24 -42 -23 -90 -58 -132 -133 Max 39 53 50 59 68 107 70 Median 6 5 6 -18 3 -9 -51 Mean 4 2 7 -18 4 -8 -44 STD 15 20 17 35 24 58 59 IQR 20 22 18 36 34 62 102

East Cyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 56 35 21 29 42 38 28 39 13 5 15 32 Min -24 -120 -59 -100 -38 -125 -92 -108 -107 -54 -102 -178 Max 32 21 25 73 15 61 87 62 62 43 91 214 Median 5 -20 -8 -5 -14 -10 -10 -20 -47 -2 -42 27 Mean 6 -26 -7 -11 -13 -13 -8 -20 -30 -8 -25 20 STD 12 38 20 38 14 43 35 45 62 35 61 99 IQR 12 51 31 33 22 55 38 68 129 54 80 129 Outliers

removed

N 52 25 26 Min -19 -62 -58 Max 27 51 39 Median 4 -5 -10 Mean 5 -4 -9 STD 10 24 26 IQR 11 28 36

Table 4-9: Senaatti building location accuracy in cm, eaves.

57

HeightCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 14 12 5 12 10 11 7 9 3 3 2 9 Min 7 -46 -33 -68 -117 -127 -129 -134 -221 1 -2 -142 Max 50 38 54 65 10 18 6 8 -4 6 3 12 Median 27 4 23 -2 -3 0 -31 -2 -22 5 1 -8 Mean 31 1 18 -4 -19 -14 -35 -20 -82 4 1 -25 STD 11 26 36 32 39 42 47 46 120 2 3 47 IQR 13 44 79 38 36 25 45 52 198 3 4 49 R2 0.9997 0.9977 0.9973 0.9959 0.9941 0.9931 0.9901 0.9910 1.0000 0.5714 1.0000 0.9896 Outliers

removed

N 13 11 9 10 6 8 8 Min 21 -68 -50 -49 -51 -50 -46 Max 50 15 10 18 6 8 12 Median 27 -2 1 2 -18 0 -5 Mean 32 -10 -8 -3 -19 -6 -10 STD 9 25 20 19 23 19 19 IQR 13 36 36 14 36 11 25

NorthCyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 14 13 5 12 10 11 7 9 3 3 2 9 Min -21 -48 -13 -53 -96 -104 -127 -54 -52 -28 -57 -287 Max 21 47 26 58 38 57 4 61 62 -3 27 219 Median -2 -4 8 -28 -1 -21 -43 19 -1 -24 -15 -10 Mean 1 -1 7 -20 -8 -17 -52 7 3 -18 -15 -33 STD 14 29 14 30 39 54 47 39 57 13 60 150 IQR 24 39 21 33 47 64 60 77 51 5 84 241 Outliers

removed

N 11 2 Min -53 -28 Max 8 -24 Median -31 -26 Mean -27 -26 STD 18 3 IQR 28 5

East Cyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 14 13 5 12 10 11 7 9 3 3 2 9 Min -11 -106 -31 -53 -38 -76 -43 -97 -58 -18 -15 -73 Max 25 34 0 39 61 7 25 5 -2 -9 -11 51 Median -2 -16 -20 -5 -18 -13 -26 -21 -9 -12 -13 -14 Mean 2 -20 -17 -4 -13 -16 -16 -37 -23 -13 -13 -12 STD 10 39 13 24 29 22 28 37 31 5 3 40 IQR 13 38 25 24 19 18 20 84 49 7 4 51 Outliers

removed

N 12 9 10 5 Min -85 -38 -27 -43 Max 34 2 7 -21 Median -14 -18 -10 -28 Mean -13 -22 -10 -32 STD 31 13 11 10 IQR 33 22 15 17

Table 4-10: Senaatti building location accuracy in cm, ridges.

58

Height CyberCity

N 40 Min 11 Max 65 Median 33 Mean 35 STD 13 IQR 20 R2 0.9995

North CyberCity

N 40 Min -81 Max 28 Median 7 Mean 0 STD 23 IQR 25 Outliers

removed

N 38 Min -33 Max 28 Median 9 Mean 4 STD 16 IQR 21

East CyberCity

N 40 Min -22 Max 41 Median 4 Mean 6 STD 16 IQR 20

Table 4-11: Senaatti building location accuracy in cm, other targets.

59

All targets Eaves Ridges

Height IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI

N 24 22 2 20 10 11 1 6 14 11 1 14 Min -153 -.99 -71 -213 -153 -.99 -71 -213 -151 -72 -37 -67 Max 115 5 -37 -7 115 -8 -71 -7 -12 5 -37 -9 Median -34 -38 -54 -39 -9 -40 -71 -52 -57 -18 -37 -32 Mean -37 -33 -54 -47 -6 -41 -71 -78 -59 -25 -37 -33 STD 59 26 24 45 69 25 73 40 25 15 IQR 54 34 34 24 58 10 59 54 39 13 R2 0.9436 0.9846 1.0000 0.9562 0.9194 0.9826 0.9180 0.9649 0.9903 0.9945 Outliers

removed

N 21 21 18 8 8 5 13 Min -108 -72 -67 -48 -65 -102 -49 Max 40 5 -7 40 -21 -7 -9 Median -31 -38 -37 -9 -41 -52 -28 Mean -33 -30 -34 -2 -42 -51 -31 STD 39 22 16 31 12 34 12 IQR 48 34 16 57 9 59 13

North IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI

N 24 22 2 20 9 10 1 5 14 11 1 14 Min -142 -100 -11 -277 -142 -100 23 -277 -39 -80 -11 -136 Max 23 92 23 134 19 92 23 134 20 47 -11 60 Median -8 -28 6 -13 -5 -33 23 34 -11 -33 -11 -26 Mean -15 -24 6 -32 -28 -23 23 -2 -10 -26 -11 -41 STD 34 51 24 90 51 58 164 16 49 55 IQR 20 73 34 80 76 78 384 21 76 75 Outliers

removed

N 23 17 Min -78 -136 Max 23 107 Median -6 -10 Mean -10 -22 STD 22 59 IQR 18 58

East IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI IGN

Nebel+

Partner

FOI

outlines FOI

N 24 22 2 20 10 11 1 6 14 11 1 14 Min -60 -114 -76 -114 -60 -114 -76 -98 -50 -97 -17 -114 Max 8 137 -17 177 8 137 -76 177 7 90 -17 77 Median -13 30 -46 4 -24 39 -76 23 -8 27 -17 -7 Mean -17 21 -46 2 -24 27 -76 23 -11 14 -17 -7 STD 20 60 42 72 22 66 100 16 56 58 IQR 24 80 59 114 25 70 146 19 66 90 Outliers

removed

N 13 Min -36 Max 7 Median -6 Mean -8 STD 12 IQR 19

Table 4-12: Amiens building location accuracy in cm.

The results are further illustrated in Figure 4-5.

60

Espoonlahti, all buildings

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Cyb

erCity

Ham

burg

Stuttg

art

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

r

C+B

Tech

nik

FOI o

utlin

es FOI

IQR

[m

]

IQR Height

IQR North

IQR East

Espoonlahti, selected buildings

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Cyb

erCity

Ham

burg

Stuttg

art

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

r

C+B

Tech

nik

FOI o

utlin

es FOI

IQR

[m

]

IQR Height

IQR North

IQR East

Hermanni

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Cyb

erCity

Stu

ttgar

tIG

N

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

Aalbo

rg

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

IQR

[m

]

IQR HeightIQR NorthIQR East

Senaatti

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

Cyb

erCity

Ham

burg

Stuttg

art

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

IQR

[m

]IQR Height

IQR North

IQR East

Amiens

0.0

0.2

0.4

0.6

0.8

1.0

1.2

IGN

Neb

el+P

artn

er

FOI o

utlin

es FOI

IQR

[m

]

IQR HeightIQR NorthIQR East

Figure 4-5: Location accuracy (IQR) of the models with respect to single points.

Figure 4-5 shows the differences between various models with respect to elevation and planimetric errors for single point targets. The errors are given separately for north-south and east-west directions due to the different point spacing in along and across track directions of TopoSys systems.

61

CyberCity achieved a good quality in all three Finnish test sites. However, it should be remembered that measurements in CyberCity are done manually from stereoscopic images and automation is mainly used for connecting points into 3D models. Thus, the manual part of the process is verified in this analysis. The quality improved when image quality was better, but in general the accuracy variation within the test sites was low.

The Hamburg and Stuttgart models were affected more by the site and site-wise data characteristics. Since the Hermanni test site was relatively easy, the Stuttgart model used high automation in Hermanni resulting in decreased performance. With the most difficult test site, Espoonlahti, the Stuttgart model was almost as good with all buildings as the CyberCity model and even better with selected buildings (same building modelled by all partners). It has to be remembered that the Stuttgart process was done by a student and with CyberCity the process was done by an expert. The Hamburg model showed lower performance than these two other photogrammetric techniques even though it was based on a more standard photogrammetric process but with an inexperienced operator.

Hybrid techniques using TerraScan (see ICC laser+aerial) showed a quality almost comparable to the best photogrammetric methods. The same method (i.e. TerraScan) used by Nebel and Partner showed typically lower accuracy, which can be explained by the lower amount of time spent and the lower amount of aerial image information used in the process. In ICC, aerial images and image orientations were used to measure the outlines, and in Nebel and Partner, images were used for visual interpretation purposes. It seems that good quality results can be obtained when building outlines are measured using image information and laser-derived planes.

In Hermanni, where all participants using laser data provided a model, the ICC model (based on TerraScan) showed the best accuracy. Also, the Delft and Aalborg models (the latter developed by undergraduate students within the project) resulted in reasonably good accuracy. Delft, Aalborg and C+B Technik mainly used given ground plans to determine building location, so planimetric accuracy is directly affected by the ground plans which do not include roof eaves. This can also be seen in the results: the planimetric accuracy is almost the same for these three and also point deviations in Hermanni (Appendix 2) seem to have “common” points. It should be noted that the generation of the FOI model was fully automatic and their newest extraction algorithm was based on the use of first and last pulse data, and last pulse data was not provided in the test, since it was not available from all test sites. C+B did exceptionally well in Espoonlahti, which was the most difficult area, but they provided only the buildings where outlines were given. One possible explanation for the variability of the C+B performance in various test sites is that since it is based on TINs, it works better than plane-based models (laser points defining a plane) with more complex and with small structures. Also, the Dresden model performed extremely well in Senaatti, but not in Hermanni. The number of points used in the analysis was extremely low for the Dresden model.

62

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

0 50 100Laser data use %

Bu

ild

ing

ou

tlin

e d

evia

tio

n IQ

R [m

]

EspoonlahtiHermanniSenaattiAmiensLinear (Espoonlahti)Linear (Hermanni)Linear (Senaatti)Linear (Amiens)

Figure 4-6: Building outline deviation, all targets. Laser data use of 0% refers to

photogrammetric methods, 100 % to fully laser scanning based techniques and intermediate

values to hybrid techniques.

In general, photogrammetric methods were more accurate in determining building outlines (mean of IQR North and IQR East), as may be seen in Figure 4-6. Taking all test sites into account the IQR value of photogrammetric methods ranged from 14 to 36 cm (average 21 cm, median 22 cm and std 7.2 cm of IQR values). The corresponding values for aerial image assisted laser scanning ranged from 20 to 76 cm (mean 44 cm, median 46 cm, std 18.5 cm). Laser scanning based building outline errors ranged from 20 to 150 cm (mean 66 cm, median 60 cm, std 33.2 cm).

The effect of the point density on the achieved average accuracy (planimetric, mean of IQR North and IQR East, and height errors) is depicted in Figure 4-7.

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

0 2 4 6 8 10 12 14 16

Laser point density [points per m2]

Ta

rge

t a

ccu

racy IQ

R [m

]

PlaneHeightLinear (Plane)Linear (Height)

Figure 4-7: Average accuracy versus laser point density.

63

Point density, shadowing of trees and complexity of the structure were the major reasons for site wise variation of the laser scanner based results. The lowest accuracy was obtained with the lowest pulse density (Senaatti). Also in Amiens, the complexity deteriorated the performance. It was almost impossible to reveal the transition from one house to another using DSM data in Amiens. The small number of trees, simple building structure and relatively high pulse density resulted in the highest accuracy at the Hermanni test site.

The IQR value for laser scanning height determination ranged from 4 to 153 cm (mean 32 cm, median 22 cm, std 31.5 cm). One fully automatic method caused high errors modifying the mean value. Laser scanning assisted by aerial images resulted in IQR values between 9 and 34 cm (mean 18 cm, median 16.5 cm, std 8.5 cm). Photogrammetric height determination ranged from 14 to 54 cm (mean 33 cm, median 35 cm, std 18 cm). The accuracy of height determination exactly followed the laser scanning point density. With the high-density data in Espoonlahti, all participants were able to provide average heights with a better accuracy than 20 cm IQR value.

The comparison of various laser point densities and methods for height determination and with photogrammetry is given in Figure 4-8. Photogrammetric methods are on the left side (laser data use 0%) and laser scanning methods on the right (laser data use 100%). In the cases of Hermanni, Amiens and Espoonlahti, the regression line is decreasing indicating that on the average height determination is slightly more accurate using laser scanner than photogrammetric techniques. In Senaatti, the opposite conclusion can be drawn due to the low pulse density and to an unsuccessful fully-automatic building extraction case. In Figure 4-9 the target height and plane deviations are shown separately for eaves and ridges for Hermanni and Senaatti. As expected, for laser scanning methods the accuracy is better for ridges, which are determined by the intersection of two planes. In theory, the height determination of ridges is extremely accurate if the number of hits defining the planes is high enough. The planimetric accuracy of ridges and eaves was more accurate with photogrammetry than with laser scanning, Figure 4-9.

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

0 50 100Laser data use %

He

igh

t d

evia

tio

n IQ

R [m

]

EspoonlahtiHermanniSenaattiAmiensLinear (Espoonlahti)Linear (Hermanni)Linear (Senaatti)Linear (Amiens)

Figure 4-8: Target height deviation.

64

Figure 4-9: Target height and plane (mean of north/east) deviations for different targets.

The degree of automation varied significantly among the participants of this test. In general, the laser data allows higher automation in creating the models. Editing of the complex building models is needed slowing down the process. Even though some laser-based processes are relatively automatic, the processes are still under development. In general, the plane target accuracy is affected by the degree of automation (and method, Figure 4-10). The accuracy of low automation methods is about 20-30 cm, while on high automation methods it is about 60-100 cm (IQR). The target height accuracy seems to be almost independent of the degree of automation (Figure 4-11). The degree of automation was estimated as a value from 0 to 10 based on the workflow charts and the procedure descriptions and comments provided by participants.

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

2 4 6 8 10Degree of automation

Pla

ne

accu

racy IQ

R [m

]

EspoonlahtiHermanniSenaattiAmiensLinear (Espoonlahti)Linear (Hermanni)Linear (Senaatti)Linear (Amiens)

Figure 4-10: Obtained planimetric accuracy (mean of north/east) as a function of automation.

Height

0.0

0.5

1.0

1.5

2.0

2.5

0 50 100Laser data use %

IQR

[m

]

Eaves HermanniRidges HermanniEaves SenaattiRidges SenaattiLinear (Eaves Hermanni)Linear (Ridges Hermanni)Linear (Eaves Senaatti)Linear (Ridges Senaatti)

Plane

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

0 50 100Laser data use %

IQR

[m

]

Eaves HermanniRidges HermanniEaves SenaattiRidges SenaattiLinear (Eaves Hermanni)Linear (Ridges Hermanni)Linear (Eaves Senaatti)Linear (Ridges Senaatti)

65

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

2 4 6 8 10Degree of automation

He

igh

t a

ccu

racy IQ

R [m

]

EspoonlahtiHermanniSenaattiAmiensLinear (Espoonlahti)Linear (Hermanni)Linear (Senaatti)Linear (Amiens)

Figure 4-11: Obtained height accuracy as a function of automation.

In general the methods using aerial images and interactive processes are capable of producing more detail in building models, but only some providers modelled more detailed structures such as chimneys and ventilation equipment on roofs.

4.2.2 Building Height

Building height accuracies of different models are shown in Table 4-13 (Espoonlahti), Table 4-14 (Hermanni) and Table 4-15 (Senaatti). No reference data for building height determination was available for Amiens. Models without DEM and models where building vectors do not follow ground height are marked with grey italics. Leaving these models out of the comparison, laser scanning produced the same accuracy as photogrammetric methods (Figure 4-12).

66

All

observations

Cyber-

City Hamburg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner ICC laser

C+B

Technik

FOI

outlines FOI FOI 2005

N 33 23 6 26 26 29 20 4 8 19 21 Min -169 -163 -272 -98 19 15 -65 -26 -39 -58 -48 Max 116 285 311 228 598 598 520 106 770 323 74 Median 6 9 -34 31 178 94 282 -7 4 -1 -6 Mean 3 46 -15 37 231 201 232 17 129 13 -8 STD 47 118 193 80 185 187 182 60 283 81 29 RMSE 46 124 177 86 294 272 292 55 295 79 30 IQR 35 163 161 76 299 311 326 23 40 15 39 R2 0.9980 0.9905 0.4834 0.9952 0.9740 0.9709 0.9793 0.9997 0.9830 0.9990 0.9995 Outliers

removed

N 30 5 24 3 6 12 20 Min -25 -272 -98 -26 -39 -35 -48 Max 73 50 152 -3 16 6 29 Median 7 -35 28 -11 -9 -1 -9 Mean 9 -80 21 -13 -8 -6 -12 STD 22 121 59 12 22 11 23 RMSE 24 135 62 16 21 12 25 IQR 34 239 76 15 39 12 35

Table 4-13: Espoonlahti building height accuracy in cm.

All

observations

Cyber-

City Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 24 17 22 22 23 17 12 15 13 1 14 13 Min -34 -385 -120 42 52 42 -34 -678 -39 574 -34 -86 Max 91 213 511 460 584 463 57 115 105 574 1137 22 Median -5 -8 46 276 253 213 13 -226 35 574 20 -16 Mean -4 -16 79 249 235 226 16 -254 35 574 193 -19 STD 26 169 175 135 138 144 25 243 47 407 34 RMSE 26 164 188 282 270 266 29 345 57 574 437 37 IQR 23 266 69 261 205 311 29 369 58 71 59 R2 0.9953 0.8549 0.9341 0.8837 0.8768 0.8993 0.9977 0.7854 0.9929 0.6694 0.9950Outliers

removed

N 23 19 11Min -34 -120 -34 Max 22 104 62Median -5 28 1Mean -8 15 13STD 16 62 33RMSE 18 62 34IQR 22 57 31

Table 4-14: Hermanni building height accuracy in cm.

67

All

observations

Cyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 13 12 4 12 13 12 12 12 7 0 5 10 Min 3 -348 -392 -29 266 193 251 -141 -55 23 -68 Max 52 110 -126 98 614 473 531 27 131 44 45 Median 23 -63 -209 18 332 319 347 17 13 30 -12 Mean 28 -76 -234 16 368 322 372 -9 11 31 -9 STD 18 119 113 39 111 80 89 59 66 8 32 RMSE 33 137 253 41 383 331 382 57 62 32 31 IQR 32 145 195 47 140 70 71 38 76 11 24 R2 0.9987 0.8962 0.1337 0.9905 0.9591 0.9735 0.9638 0.9784 0.9569 0.9996 0.9942 Outliers

removed

N 11 10 10 10 8 Min -29 193 251 -17 -37 Max 72 353 467 27 28 Median 17 307 337 21 -12 Mean 9 294 341 15 -8 STD 31 51 57 13 19 RMSE 30 298 345 20 20 IQR 43 59 33 9 20

Table 4-15: Senaatti building height accuracy in cm.

Building height, Espoonlahti

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Cyb

erCity

Ham

burg

Stuttg

art

IGN

ICC la

ser+

aeria

l

Neb

el+P

artn

er

ICC la

ser

C+B

Techn

ik

FOI o

utlin

es FOI

FOI2

005

[m]

IQR

RMSE

Building height, Hermanni

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Cyb

erCity

Stuttg

art

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

r

Del

ft

Aalbo

rg

C+B

Techn

ik

Dre

sden

FOI o

utlin

es FOI

[m]

IQR

RMSE

Building height, Senaatti

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Cyb

erCity

Ham

burg

Stu

ttgart

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

C+B

Tech

nik

Dre

sden

FOI o

utline

sFO

I

[m]

IQR

RMSE

Figure 4-12: Building height accuracy. IQR: all observations, RMSE: outliers removed.

68

4.2.3 Building Length

Building length accuracies of different models are shown in Table 4-16 (Espoonlahti), Table 4-17 (Hermanni), Table 4-18 (Senaatti) and Table 4-19 (Amiens).

All

observations

Cyber-

City

Ham-

burg Stuttgart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser

C+B

Technik

FOI

outlines FOI FOI 2005

N 107 98 48 85 97 97 73 21 25 43 45 Min -36 -186 -29 -122 -64 -144 -94 -33 -146 -562 -720 Max 112 125 59 190 30 127 170 101 257 1604 1427 Median 7 -16 5 -2 -7 36 27 1 5 68 27 Mean 8 -15 6 6 -7 29 30 8 23 87 49 STD 17 34 14 47 18 52 48 32 107 283 265 RMSE 19 37 15 47 19 59 56 32 107 292 267 IQR 13 30 12 45 25 64 71 37 161 140 105 R2 0.9992 0.9971 0.9996 0.9925 0.9992 0.9930 0.9919 0.9979 0.8714 0.8440 0.7516 Outliers

removed

N 100 91 45 78 96 95 72 20 39 39 Min -18 -58 -13 -71 -55 -81 -94 -33 -138 -172 Max 29 34 28 81 30 127 124 54 310 203 Median 7 -16 4 -2 -6 36 26 -1 68 27 Mean 7 -14 5 1 -7 33 29 3 65 34 STD 10 20 9 33 17 46 45 24 108 79 RMSE 12 24 10 33 18 56 53 24 125 85 IQR 12 27 11 38 25 60 67 20 126 92

Table 4-16: Espoonlahti building length accuracy in cm.

All

observations

Cyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 40 36 37 40 40 36 21 30 22 6 24 32 Min -28 -37 -184 -42 -72 -92 -182 -182 -182 -123 -167 -204 Max 34 74 156 43 66 65 42 75 7 -44 78 233 Median 6 6 -48 13 21 -13 -34 -19 -33 -94 -38 6 Mean 5 7 -44 11 19 -7 -47 -34 -49 -89 -47 30 STD 14 22 61 19 23 34 58 61 51 33 64 99 RMSE 14 22 75 22 30 34 74 69 70 94 78 102 IQR 10 21 61 19 20 38 72 56 56 65 96 116 R2 0.9998 0.9995 0.9956 0.9996 0.9994 0.9986 0.9969 0.9985 0.9976 0.9968 0.9953 0.9869Outliers

removed I

N 32 34 34 37 38 35 28 20 31 Min -15 -37 -157 -20 -20 -68 -120 -121 -190 Max 24 37 17 43 59 65 75 7 233 Median 6 5 -44 17 21 -11 -15 -31 7 Mean 5 3 -42 15 20 -5 -24 -37 38 STD 8 16 41 13 17 31 50 33 91 RMSE 9 16 58 20 30 31 55 49 97 IQR 8 18 55 19 19 38 54 50 116

Table 4-17: Hermanni building length accuracy in cm.

69

All

observations

Cyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 32 28 14 21 27 27 17 21 8 6 9 21 Min -33 -163 -57 -150 -93 -177 -118 -195 -169 -18 -187 -149 Max 40 40 45 23 123 299 161 107 -39 5 102 336 Median 7 -20 -1 -39 2 10 -13 -37 -83 -13 -43 84 Mean 4 -26 -5 -44 3 41 -13 -34 -98 -10 -49 99 STD 14 45 30 35 36 102 62 86 49 9 91 101 RMSE 14 51 29 56 35 108 61 91 109 13 0.99 139 IQR 16 43 32 33 17 66 69 126 90 11 139 125 R2 0.9999 0.9993 0.9997 0.9996 0.9995 0.9961 0.9987 0.9976 0.9997 1.0000 0.9984 0.9962 Outliers

removed

N 30 26 20 23 21 16 Min -19 -84 -88 -32 -83 -118 Max 20 40 23 25 120 52 Median 7 -19 -39 2 8 -18 Mean 4 -17 -39 -2 11 -24 STD 11 31 26 15 46 44 RMSE 11 35 46 14 46 48 IQR 15 43 32 17 29 65

Table 4-18: Senaatti building length accuracy in cm.

All observations IGN Nebel+ Partner FOI outlines FOI

N 8 4 0 5 Min -7 -9 69 Max 58 99 264 Median 21 57 102 Mean 23 51 130 STD 25 47 80 RMSE 33 65 148 IQR 41 85 69 R2 0.9969 0.9961 na 0.9779 Outliers removed

N 4 Min 69 Max 139 Median 88 Mean 96 STD 32 RMSE 100 IQR 33

Table 4-19: Amiens building length accuracy in cm.

70

Building length, Espoonlahti

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

Cyb

erCity

Ham

burg

Stu

ttgar

tIG

N

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

r

C+B

Tech

nik

FOI o

utlin

es FOI

FOI2

005

[m]

IQR

RMSE

Building length, Hermanni

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Cyb

erCity

Stuttg

art

IGN

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

Aalbo

rg

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

[m]

IQR

RMSE

Building length, Senaatti

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

Cyb

erCity

Ham

burg

Stu

ttgar

tIG

N

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

[m]

IQR

RMSE

Building length, Amiens

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

IGN

Neb

el+P

artn

er

FOI o

utlin

es FOI

[m]

IQR

RMSE

Figure 4-13: Building length accuracy. IQR: all observations, RMSE: outliers removed.

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0 50 100Laser data use %

Bu

ild

ing

le

ng

th d

evia

tio

n R

MS

E [m

]

EspoonlahtiHermanniSenaattiAmiensLinear (Espoonlahti)Linear (Hermanni)Linear (Senaatti)Linear (Amiens)

Figure 4-14: Building length deviation (all observations).

71

In building length determination (Figure 4-14), laser based methods were not as accurate as photogrammetric methods, as could be expected from the above. The photogrammetrically derived lengths varied from 14 to 51 cm (RMSE, mean 26 cm, median 22 cm, std 12.6 cm). Lengths obtained with aerial image assisted laser scanning varied from 19 to 108 cm (mean 59.4 cm, median 57 cm, std 31.2 cm). The laser scanning based lengths varied from 13 to 292 cm (mean 93 cm, median 84.5 cm, std 60.9 cm). With laser scanning the complexity of the buildings was the major cause for site wise variation rather than the point density.

4.2.4 Roof Inclination

Roof inclination accuracies of different models are shown in Table 4-20 (Hermanni), Table 4-21 (Senaatti) and Table 4-22 (Amiens).

All

observations

Cyber-

City

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft Aalborg

C+B

Technik Dresden

FOI

outlines FOI

N 17 17 17 17 17 17 11 16 12 3 12 15 Min -2.41 -2.38 -1.57 -1.49 -2.38 -1.49 -1.15 -0.93 -5.29 0.03 -1.48 -1.67 Max 4.30 5.54 1.78 0.34 0.32 0.41 0.72 0.83 -0.36 0.45 0.52 0.46 Median 1.16 1.09 -0.16 -0.38 -0.33 -0.32 -0.42 -0.18 -2.22 0.32 -0.68 -0.70 Mean 0.95 1.27 -0.07 -0.38 -0.51 -0.34 -0.29 -0.15 -2.58 0.27 -0.52 -0.61 STD 2.02 2.62 1.01 0.52 0.73 0.58 0.60 0.54 1.65 0.22 0.59 0.54 RMSE 2.18 2.84 0.99 0.63 0.88 0.65 0.64 0.55 3.03 0.32 0.77 0.80 IQR 3.01 5.10 1.85 0.71 0.89 1.17 0.62 0.87 3.34 0.29 0.59 0.55 R2 0.7701 0.6559 0.9419 0.9851 0.9714 0.9814 0.9565 0.9805 0.7927 0.8592 0.9728 0.9823Outliers

removed

N 16 0.14 Min -1.95 -1.27 Max 0.32 0.46 Median -0.30 -0.68 Mean -0.40 -0.53 STD 0.57 0.47 RMSE 0.68 0.70 IQR 0.50 0.62

Table 4-20: Hermanni roof inclination accuracy in degrees.

All

observations

Cyber-

City

Ham-

burg

Stutt-

gart IGN

ICC

laser+

aerial

Nebel+

Partner

ICC

laser Delft

C+B

Technik Dresden

FOI

outlines FOI

N 10 9 6 8 9 9 5 4 2 3 2 6 Min -2.00 -7.16 -4.00 -2.42 -2.12 -4.24 -3.10 -1.34 -22.10 -1.02 -2.61 -2.55 Max 1.31 2.65 6.49 2.27 0.46 0.57 0.32 0.42 -4.38 0.36 -1.09 0.03 Median -0.23 -2.29 0.72 -0.22 -0.17 -1.00 -1.99 -0.37 -13.24 0.24 -1.85 -0.72 Mean -0.20 -2.52 0.80 -0.14 -0.61 -1.41 -1.51 -0.41 -13.24 -0.14 -1.85 -0.95 STD 0.98 2.93 3.76 1.95 0.99 1.89 1.37 0.93 12.53 0.76 1.07 1.04 RMSE 0.95 3.74 3.52 1.83 1.12 2.27 1.95 0.91 15.93 0.64 2.00 1.34 IQR 1.44 4.07 5.07 3.28 2.02 4.55 2.54 1.70 17.72 1.25 1.52 1.69 R2 0.9741 0.7699 0.8233 0.9498 0.9765 0.9392 0.9675 0.7406 1.0000 0.9904 1.0000 0.9708

Table 4-21: Senaatti roof inclination accuracy in degrees.

72

All observations IGN Nebel+ Partner FOI outlines FOI

N 5 1 0 2 Min -34.97 -1.19 -4.50 Max 15.93 -1.19 11.90 Median -0.49 -1.19 3.70 Mean -4.80 -1.19 3.70 STD 19.32 11.60 RMSE 17.93 1.19 9.00 IQR 40.53 16.40 R2 0.1306 1.0000

Table 4-22: Amiens roof inclination accuracy in degrees.

Roof inclination, Hermanni

0.0

1.0

2.0

3.0

4.0

5.0

6.0

Cyb

erCity

Stu

ttgar

tIG

N

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

Aalbo

rg

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

[de

gre

es]

IQR

RMSE

Roof inclination, Senaatti

0.0

2.0

4.0

6.0

8.0

10.0

12.0

14.0

16.0

18.0

Cyb

erCity

Ham

burg

Stu

ttgar

tIG

N

ICC

lase

r+ae

rial

Neb

el+P

artn

er

ICC

lase

rDelft

C+B

Tech

nik

Dre

sden

FOI o

utlin

es FOI

[de

gre

es]

IQR

RMSE

Roof inclination, Amiens

0.0

5.0

10.0

15.0

20.0

25.0

30.0

35.0

40.0

IGN

Neb

el+P

artn

er

FOI o

utlin

es FOI

[de

gre

es]

IQR

RMSE

Figure 4-15: Roof inclination accuracy of the models. IQR: all observations, RMSE: outliers

removed.

Figure 4-15 gives the roof inclination accuracy for each model, defined with IQR and RMSE. Roof inclination determination was more accurate when using laser data than photogrammetry, but a large variation in quality exists due to methods and test sites (i.e. complex buildings). The RMSE (all observations included) using laser scanning for roof inclination varied from 0.3 to 15.9 degrees (mean 2.7 degrees, median 0.85 degrees, std 4.4 degrees). The corresponding values for aerial image assisted laser scanning ranged from 0.6 and 2.3 degrees (mean 1.3 degrees, median 1.1 degrees, std 0.6 degrees) and for photogrammetry ranged from 1.0 to 17.9 degrees (mean 5.2 degrees, median 3.2

73

degrees, std 6.3 degrees). In Senaatti and Amiens, the roof inclinations are steep and roofs short, so even small errors in target height determination lead to large errors in inclination angle. The Hermanni test site is relatively easy for both methods. In Hermanni the accuracy of roof inclination determination was about 2.5 degrees for photogrammetric methods and about 1 degree (RMSE, all observations included) for laser based methods.

In Hermanni the best result was obtained with the Dresden model and basically all laser based methods resulted in less than 1 degree error (RMSE). The C+B approach had a larger error than the photogrammetric processes. Most probably the TIN based principle of C+B does not give as good results as using all the points hitting one surface and defining a plane using all of them. In photogrammetry, the roof inclination is obtained from two measurements and therefore, the same accuracy as laser was clearly not achieved. When the building size was smaller and/or the pulse density was lower, this difference between photogrammetry and laser scanning was reduced or even disappeared (Figure 4-15).

4.3 Total Relative Building Area and Shape Dissimilarity

Buildings or parts of buildings are missing in many models. The reasons for this are not always known, but the shadowing of big trees and the complexity of buildings can certainly cause problems. Participants may also strive for different goals in building extraction; some want to model the buildings in as detailed a manner as possible and some are aiming for basic building form only. For automatic methods vegetation also causes problems as trees may be classified as buildings. With other methods the extra buildings are mainly temporary constructions, such as site huts for construction sites, which are usually impossible to separate out from “real” buildings.

Total relative building areas and shape dissimilarities for the whole test site and for modelled buildings at Espoonlahti are shown in Table 4-23 and Figure 4-16 and at Hermanni in Table 4-24 and Figure 4-17. Images showing the differences in 2D between reference raster and delivered models are in Appendix 3.

Cyb

er

City

Ham

burg

Stut

tgar

t

IGN

ICC

lase

r+

aeri

al

Neb

el+

Par

tner

ICC

lase

r

C+

BT

echn

ik

FO

I ou

tl

FO

I

FO

I 20

05

All 98.6 94.2 49.6 97.3 92.3 99.7 89.2 95.2 105.2 108.2 97.6 Total relative

building area Modelled 100.0 96.8 80.7 97.8 94.8 101.8 94.0 97.4 105.2 111.5 101.4

All 7.7 12.4 52.6 10.7 10.2 11.8 17.5 6.9 10.9 29.9 25.8 Total relative shape

dissimilarity Modelled 5.5 10.0 22.8 10.2 7.8 10.1 12.9 4.7 10.9 26.6 22.7

Table 4-23: Total relative building area and shape dissimilarity at Espoonlahti in %.

74

Espoonlahti

0

10

20

30

40

50

60

Cyb

ercity

Ham

burg

Stu

ttgar

tIG

N

ICC la

ser+

aeria

l

Neb

el+Partn

er

ICC la

ser

C+B T

echni

k

FOI o

utlFO

I

FOI2

005T

ota

l re

lative

sh

ap

e d

issim

ila

rity

[%

]

All buildings

Modelled buildings

Figure 4-16: Total relative shape dissimilarity at Espoonlahti.

Cyb

er-C

ity

Stut

tgar

t

IGN

ICC

lase

r+

aeri

al

Neb

el+

Par

tner

ICC

lase

r

Del

ft

Aal

borg

C+

BT

echn

ik

FO

I ou

tl

FO

I

All 108.4 104.1 96.6 99.7 108.3 89.1 97.5 70.9 97.3 101.1 112.7Total relative

building area Modelled 108.2 105.0 97.6 105.4 109.2 101.8 97.8 100.3 97.6 101.1 121.0

All 11.6 10.2 7.2 14.5 13.7 19.9 4.8 34.4 4.9 8.5 52.2 Total relative shape

dissimilarity Modelled 11.0 9.4 6.3 9.6 12.9 8.4 4.5 7.2 4.6 8.5 25.0

Table 4-24: Total relative building area and shape dissimilarity at Hermanni in %.

Hermanni

0

10

20

30

40

50

60

Cyb

ercity

Stu

ttgart

IGN

ICC

lase

r+aer

ial

Neb

el+P

artner

ICC

lase

rDelft

Aalbo

rg

C+B

Techn

ik

FOI o

utl

FOIT

ota

l re

lative

sh

ap

e d

issim

ila

rity

[%

]

All buildings

Modelled buildings

Figure 4-17: Total relative shape dissimilarity at Hermanni.

75

5 Discussion and Conclusions

It can be concluded that, presently, photogrammetric techniques and hybrid techniques provide the highest accuracy and level of detail in 3D city reconstruction. Despite the high amount of research during the last decade the level of automation is still relatively low.

Improvement in automation can be achieved most significantly by utilising the synergy of laser and photogrammetry. The photogrammetric techniques are powerful for visual interpretation of the area, measurement of the building outlines and of small details (e.g. chimneys), whereas laser scanning gives height, roof planes and ridge information at its best. When the advantages of both methods are implemented in a single system well, high accuracy and a relatively high level of automation can be achieved. In this test, there was only one hybrid system.

The quality of the laser scanning data has a significant impact on the accuracy of the fully laser based models. Since pulse density is increasing substantially (even though the novel European Toposys system already had an 83 kHz system in 1997, the PRF of ALS produced by today’s leading manufacturer Optech has almost followed Moore’s Law. In 1995, the PRF was just a few kHz, while today it is 100 kHz) and there are new possibilities to use pulse information (multiple pulse, waveforms), it is expected that the accuracy of the laser based techniques will follow this technical development. The rapid rate of implementation development of laser scanning data can also be seen from the improvement of FOI results during the test. On the other hand, when the study was initiated, the implementation level of the majority of laser-based techniques was low. It can be seen from Table 3-3 (time spent), that thanks to the good implementation of photogrammetric methods, the time spent on these models is not much higher that those of laser scanning.

We see, based on the results and development of the last ten years, that - There is a need to integrate laser scanning and photogrammetric techniques in a more

advanced way. The accuracy of single height measurements with the laser should be used in the photogrammetric process to increase the automation. There are currently some developments in this field.

- The quality improvement of digital camera data will increase the quality of photogrammetric techniques and it will also allow higher automation by using remote sensing methods in the image processing.

- The rapid technological development in laser scanning will allow higher accuracies to be obtained by techniques relying only on laser scanning.

- Laser scanning allows higher automation needed in the building extraction process. - There is a huge need to implement laser-based techniques into practical working processes. - New algorithmic innovations will emerge in all areas, thanks to digital camera data and laser

scanning developments that will increase automation.

In order to be able to separate methodological development from development due to improvement of the data, it is recommended to use the same test sites, with old and new aerial data, for future verifications.

In practise, the availability of the material (aerial image or laser point clouds) determines significantly what kind of techniques can be used. The costs of laser scanning are significantly higher, although in many cities the material has already been collected. In many cities, building outline information exists as well as some kind of existing 3D models. In future, the challenge is to find the answers to: how existing models are to be updated, how changes are verified and how existing information can be used optimally in a more automatic process. In the project, first pulse data was used for the extraction. It is expected that the use of last pulse and combined use of first and last pulse would lead to improvements.

76

Acknowledgments

All participants, i.e. George Vosselman (International Institute for Geo-Information Science and Earth Observation), Alexandra Hofmann (Dresden University of Technology, presently Definiens), Urs Mäder (CyberCity AG), Åsa Persson, Ulf Söderman and Magnus Elmqvist (Swedish Defence Research Agency), Antonio Ruiz and Martina Dragoja (Institut Cartogràfic de Catalunya), Sylvain Airault, David Flamanc and Gregoire Maillet (MATIS IGN), Thomas Kersten (Hamburg University of Applied Sciences), Jennifer Carl and Robert Hau (Nebel + Partner), Emil Wild (C+B Technik GmbH) and Lise Lausten Frederiksen, Jane Holmgaard and Kristian Vester (University of Aalborg) and especially Eberhard Gülch (Stuttgart University of Applied Sciences) are gratefully acknowledged for cooperation in the project. Eberhard contributed remarkably to the planning of the project. Hannu Hyyppä from Helsinki University of Technology and Leena Matikainen from FGI also contributed significantly. Support of MATIS IGN, Espoo and Helsinki City Survey Divisions, FM-Kartta Oy and Terrasolid Oy in providing data for the project is gratefully acknowledged.

77

References

Alharthy A. and Bethe, J, 2004. Detailed Building Reconstruction from Airborne Laser Data Using a Moving Surface Method. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):237-242.

Ameri, B. and Fritsch, D., 2000. Automatic 3D building reconstruction using plane roof structures. In:ASPRS congress, Washington, www.itp.uni-stuttgart.de/publications/2000/pub2000.html

BAE Systems, 2000. SOCET Set – User’s Manual Version 4.3.

Baillard, C. and Maïtre, H., 1999. 3-D reconstruction of urban scenes from aerial stereo imagery: A focusing strategy. Computer Vision and Image Understanding 76(3): 244-258.

Baillard, C. and Dissard, O., 2000. A stereo matching algorithm for urban digital elevation models. PE & RS 66(9):1119-1128.

Baltsavias, E.P., 1999. Airborne laser scanning: existing systems and firms and other resources. ISPRSJournal of Photogrammetry and Remote Sensing 54(1999):164-198.

Baltsavias, E.P., 2004. Object extraction and revision by image analysis using existing geodata and knowledge: current status and steps towards operational systems. ISPRS Journal of Photogrammetry and Remote Sensing 58(3-4):129-151.

Braun, C., Kolbe, T.H., Lang. F. et al., 1995. Models for photogrammetric building reconstruction. Computers & Graphics 19(1):109-118.

Brenner, C., 2001. City models – automation in research and practice. Photogrammetric Week ’01, pp. 149-158.

Brenner, C., 2003. Building reconstruction from laser scanning and images. Proc. ITC Workshop on Data Quality in Earth Observation Techniques, Enschede, The Netherlands, November 2003, 8 p.

Brenner, C., 2005. Building reconstruction from images and laser scanning. International Journal of Applied Earth Observation and Geoinformation 6:186-198.

Brunn, A., Gülch, E. et al., 1998. A hybrid concept for 3D building acquisition. ISPRS Journal of Photogrammetry and Remote Sensing 53(2):119-129.

Brunn, A. and Weidner, U., 1997. Extracting Buildings from Digital Surface Models. IAPRS 32(3-4W2), pp.27-34.

Brunn, A. and Weidner, U. 1998. Hierarchical Bayesian nets for building extraction using dense digital surface models. ISPRS Journal of Photogrammetry and Remote Sensing 53(5):296-307.

Burrough, P. and McDonnell, R., 1998. Principles of Geographical Information Systems. Oxford University Press, ISBN: 0-19-823366-3.

78

Centeno, J.A.S. and Miqueles, J., 2004. Extraction of buildings in Brasilian urban environments using high resolution remote sensing imagery and laser scanner data. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B4), 4 p.

Chen, L.-C., Teo, T.-A., Shao, Y.-C., Lai, Y.-C. and Rau, J-Y. 2004. Fusion of lidar data and optical imagery for building modelling. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B4), 6 p.

Cho, W. et al., 2004. Pseudo-grid based building extraction using airborne lidar data. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):378-381.

Cord, M. and Declerq, D. 2001. Three-dimensional building detection and modeling using a statistical approach. IEEE Transactions on Image Processing 10(5):715-723.

Cord, M., Jordan, M. and Cocquerez, J.P., 2001. Accurate building structure recovery from high-resolution aerial imagery. Computer Vision and Image Understanding 82(2):138-173.

Dash, J., Steinle, E., Singh, R.P. and Bähr, H.P., 2004. Automatic building extraction from laser scanning data: an input tool for disaster management. Advances in Space Research 33:317-322.

Dold C. and Brenner C., 2004. Automatic Matching of Terrestrial Scan Data as a Basis for The Generation of Detailed 3d City Models. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):1091-1096.

Elaksher, A.F. and Bethel, J.S., 2002. Reconstructing 3D buildings from lidar data. IAPRS 34(3A):102-107.

Elaksher, A.F., Bethel, J.S. and Mikhail, E.M., 2003. Roof boundary extraction using multiple images. Photogrammetric Record 18(101):27-40.

Elmqvist, M., 2002. Ground Surface Estimation from Airborne Laser Scanner Data Using Active Shape Models. Proceedings of the ISPRS Technical Commission III Symposium "Photogrammetric Computer Vision" PCV'02, September 2002, Graz, Austria. IAPRS, Vol. XXXIV, part 3A.

Fischer, A., Kolbe, T.H., Lang, F. Cremers, A.B. Förstner, W., Plümer, L. and Steinhage, V.,1998. Extracting buildings from aerial images using hierarchical aggregation in 2D and 3D. Computer Vision and Image Understanding 72(2):185-203.

Flamanc, D., Maillet, G., Jibrini, H., 2003. 3D city models: an operational approach using aerial images and cadastral maps. ISPRS archives, Vol XXXIV, Part 3/W8, Munich, 17.-19. Sept. 2003.

Forlani, G., Nardinocchi, C., Scaioni, M. and Zingaretti, P., 2003. Building reconstruction and visualization from LIDAR data. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXIV (5W12): 151-156.

Forlani, G., Nardinocchi, C., Scaioni, M. and Zingaretti, P., 2006. Complete classification of raw LIDAR data and 3D reconstruction of buildings. Pattern Analysis and Applications, 8(4): 357-374.

79

Fraser, C.S., Baltsavias, E. and Grün, A., 2002. Processing of Ikonos imagery for submetre 3D positioning and building extraction. ISPRS Journal of Photogrammetry and Remote Sensing 56(3):177-194.

Frederiksen, L., Holmgaard, J. and Vester, K., 2004. Automatiseret 3D-bygningsmodellering – fran laserscanning til bygningsmodel. University of Aalborg (in Danish).

Fuchs, C., Gülch, E. and Förstner, W., 1998. OEEPE Survey on 3D-City Models. OEEPE Publication N° 35:9-123. Bundesamt für Kartographie und Geodäsie. Frankfurt.

Fuchs, F. and Le-Men, H., 2000. Efficient subgraph isomorphism with ‘a priori’ knowledge – Application to 3D reconstruction of buildings for cartography. Lecture Notes in Computer Science 1876:427-436.

Fujii, K. and Arikawa, T., 2002. Urban object reconstruction using airborne laser elevation image and aerial image. IEEE Transactions on Geoscience and Remote Sensing 40(10):2234-2240.

Förstner, W., 1999. 3D-City Models: Automatic and Semiautomatic Acquisition Methods. Fritsch/Spiller (eds.), Photogrammetric Week ’99, CD-ROM, pp. 291-303.

Grün, A., 1997. Automation in building extraction. Photogrammetric Week ’97, pp. 175-186.

Grün, A., 1998. TOBAGO – a semi-automated approach for the generation of 3-D building models. ISPRS Journal of Photogrammetry and Remote Sensing 53(2):108-118.

Grün, A. and X. Wang, 1998a. CC-Modeler: a topology generator for 3-D city models. ISPRS Journal of Photogrammetry and Remote Sensing 53(5):286-295

Grün, A. and Wang, X., 1998b. Acquisition and Management of Urban 3D data using CyberCity Modeler, www.cybercity.tv/pub/ccm_3D_data.pdf

Grün, A. and Wang, X., 1999a. CyberCity Modeller, a tool for interactive 3-D city model generation. Photogrammetric Week’99, pp. 317-327.

Grün, A. and X. Wang, 1999b. CyberCity Spatial Information System (CC-SIS): a new concept for the management of 3-D city models in a hybrid GIS. The 20th Asian Conference on Remote Sensing, Nov. 22-25, 1999, Hong Kong, China, 8 p.

Grün, A., Li, Z. and Wang, X. 2003. Generation of 3D city models with linear array CCD-sensors. http://carstad.gsfc.nasa.gov/topics/JBRESEARCH/STARLABO/Opt3D_citymod_new.pdf

Gülch, E. and Müller, H., 2001. New applications of semi-automatic building extraction. In: Proceedings Third International Workshop on Automatic Extraction of Man-Made Objects from Aerial and Space Images. Ascona, Switzerland, 10-15 June 2001. Swets & Zeitlinger Publishers/Balkema, The Netherlands.

Haala, N., Brenner, C. and Anders, K.-H., 1998. 3D urgan GIS from laser altimeter and 2D map data. IAPRS 32(3/1):339-346.

Haala, N. and Brenner, C., 1999. Extraction of buildings and trees in urban environments. ISPRSJournal of Photogrammetry and Remote Sensing 54:123-129.

80

Haithcoat, T.L., Song, W. and Hipple, J.D., 2001. Building footprint extraction and 3-D reconstruction from LIDAR data. IEEE/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, 5 p. CD-ROM.

Henricsson, O. and Baltsavias, E., 1997. 3-D building reconstruction with ARUBA : a qualitative and quantitative evaluation. International Workshop "Automatic Extraction of Man-Made Objects from Aerial and Space Images (II), 5.-9. May [1997], Ascona, Switzerland. http://e-collection.ethbib.ethz.ch/show?type=inkonf&nr=96

Hofmann, A., Maas, H.-G. and Streilein, A., 2001. Knowledge-based building detection based on laser scanner data and topographic map information. ISPRS Commission III, Symposium 2002, Photogrammetric Computer Vision, PCV02, Graz Austria, 6 p.

Hofmann, A., Maas, H.-G. and Streilein, A., 2003. Derivation of Roof Types by Cluster Analysis in Parameter Spaces of Airborne Laser Scanner Point Clouds. Proceedings of the ISPRS working group III/3 workshop ‘3-D reconstruction from airborne laserscanner and InSAR data’,Dresden, Germany, 2003. http://www.isprs.org/commission3/wg3/workshop_laserscanning/

Hofmann, A.D., 2004. Analysis of tin-structure parameter spaces in airborne laser scanner data for 3-D building model generation. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III): 6 p.

Jaynes, C., Riseman, E., and Hanson, A., 2003. Recognition and reconstruction of buildings from multiple aerial images. Computer Vision and Image Understanding 90(1): 68-98.

Juhl, J., 1980. Det analytiske stråleudjævningsprogram sANA. Laboratoriet for fotogrammetri og landmåling, University of Aalborg (in Danish).

Jutzi, B. and Stilla, U., 2004. Extraction of features from objects in urban areas using spacetime analysis of recorded laser pulses. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B2), 6 p.

Jülge K., and Brenner, C., 2004. Object Extraction from Terrestrial Laser Scan Data for a Detailed Building Description. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,XXXV(B3/III):1079-1084.

Kaartinen, H., Hyyppä, J., Gülch, E., Hyyppä, H., Matikainen, L., Vosselman, G., Hofmann, A.D., Mäder, U., Persson, Å., Söderman, U., Elmqvist, M., Ruiz, A., Dragoja, M., Flamanc, D., Maillet, G., Kersten, T., Carl, J., Hau, R., Wild, E., Frederiksen, L., Holmgaard, J., and Vester, K., 2005a. EuroSDR Building Extraction Comparison. ISPRS Hannover Workshop High-Resolution Earth Imaging for Geospatial Information, May 17-20, 2005, CD-ROM, 6 p.

Kaartinen, H., Hyyppä, J., Gülch, E., Vosselman, G., Hyyppä, H., Matikainen, L., Hofmann, A.D., Mäder, U., Persson, Å., Söderman, U., Elmqvist, M., Ruiz, A., Dragoja, M., Flamanc, D., Maillet, G., Kersten, T., Carl, J., Hau, R., Wild, E., Frederiksen, L., Holmgaard, J., and Vester, K., 2005b. Accuracy of 3D city models: EuroSDR comparison. Proceedings of ISPRS Workshop Laser scanning 2005, September 12-14, 2005, Enschede, The Netherlands, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXVI(Part 3/W19): 227-232, CD-ROM.

81

Khoshelham K., 2004. Building Extraction from Multiple Data Sources: A Data Fusion Framework for Reconstruction of Generic Models. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):980-986.

Kim, T. and Muller, J.P., 1998. A technique for 3D building reconstruction. Photogrammetric Engineering and Remote Sensing 64(9): 923-939.

Lee, I. and Choi, Y., 2004 . Fusion of terrestrial laser scanner data and images for building reconstruction. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,XXXV(B5): 6 p.

Li, B.J., Li, Q.Q., Shi, W.Z. and Wu, F.F., 2004. Feature extraction and modeling of urban building from vehicle-borne laser scanning data. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B5): 6 p.

Maas, H.-G. and Vosselman, G., 1999. Two algorithms for extracting building models from raw laser altimetry data. ISPRS Journal of Photogrammetry & Remote Sensing, 54(2-3):153-163.

Maas, H.-G., 2001. The suitability of airborne laser scanner data for automatic 3D object reconstruction. In: Baltsavias, E. (ed.), Automatic Extraction of Man-Made Objects from Aerial and Space Images III, pp. 291-296.

Madhavan, B.B., TachibanaK., Sasagawa T., Okada H. and Shimozuma Y., 2004. Automatic Extraction of Shadow Regions in High-resolution Ads40. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):808-810.

Mayer, H., 1999. Automatic object extraction from aerial imagery – A survey focusing on buildings. Computer Vision and Image Understanding 74(2):138-149.

Morgan, M. and Habib, A., 2001 . 3D TIN for automatic building extraction from airborne laser scanner data. Proceedings of the American Society of Photogrammetry and Remote Sensing (ASPRS), St. Louis, Missouri.

Morgan, M. and Habib, A., 2002. Interpolation of lidar data and automatic building extraction. ACSM-ASPRS Annual Conference proceedings.

Neidhart, H. and Sester, M., 2003. Identifying building types and building clusters using 3D-laser scanning and GIS-data. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B4), 6 p.

Niederöst, M., 2003. Detection and reconstruction of buildings for automated map updating. Diss., Technische Wissenschaften ETH Zurich. (Mitteilungen/Institut für Geodäsie und Photogrammetrie/ETH), 139 p.

Noronha, S. and Nevatia, R., 2001. Detection and modeling of buildings from multiple aerial images. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(5):501-518.

82

Oda, K., Takano, T., Doihara, T. and Shibasaki, R., 2004. Automatic building extraction and 3D city modeling from lidar data based on Hough transformation. IAPRS 35(B3:III): 277-280.

Oriot, H. and Michel, A., 2004. Building extraction from stereoscopic aerial images. Applied Optics43(2):218-226.

Overby, J., Bodum, L., Kjems, E. and Ilsøe, P.M., 2003. Automatic 3D building reconstruction from airborne laser scanning and cadastral data using Hough transform. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3), 6 p.

Paparoditis, N., Cord, M., Jordan, M. and Cocquerez, J.P., 1998. Building detection and reconstruction from mid- and high-resolution aerial imagery. Computer Vision and Image Understanding 72(2):122-142.

Peternell, M. and Steiner, T., 2004. Reconstruction of piecewise planar objects from point clouds. Computer-aided Design 36(4):333-342.

Rottensteiner, F., 2000. Semi-automatic building reconstruction integrated in strict bundle block adjustment. International Archives of Photogrammetry and Remote Sensing,XXXIII(B2/2):461-468.

Rottensteiner, F., 2003. Automatic generation of high-quality building models from lidar data. IEEEComputer Graphics and Applications 23(6): 42-50.

Rottensteiner, F., Trinder, J., Clode, S. and Kubik, K., 2003. Detecting buildings and roof segments by combining LIDAR data and multispectral images. http://sprg.massey.ac.nz/ivcnz/Proceedings/IVCNZ_12.pdf

Rottensteiner, F., Trinder, J., Clode, S. and Kubik, K., 2005. Using the Dempster-Shafer method for the fusion of LIDAR data and multi-spectral images for building detection. Information Fusion,6(4): 283-300.

Rottensteiner, F., Trinder, J., Clode, S. and Kubik, K., 2004. Fusing airborne laser scanner data and aerial imagery for the automatic extraction of buildings in densely built-up areas. IAPRS35(B3:III):512-517.

Rottensteiner, F. and Briese, Ch., 2002. A new method for building extraction in urban areas from high-resolution lidar data. Symposium der ISPRS-Comm. III, Graz, 9-13 September 2002. IAPRS XXXIV(3A):295-301.

Rottensteiner, F. and Jansa, J., 2002. Automatic extraction of buildings from lidar data and aerial images. http://www.ipf.tuwien.ac.at/publications/fr_jj_ottawa_02.pdf

Sahar, L. and Krupnik, A., 1999. Semiautomatic extraction of building outlines from large-scale aerial images. PE & RS 65(4):459-465.

Schwalbe, E., 2004. 3d Building Model Generation from Airborne Laserscanner Data by Straight Line Detection in Specific Orthogonal Projections. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III): 249-254.

83

Sequeira, V., Ng, K., Wolfart, E., Goncalves, J.G.M. and Hogg, D., 1999. Automated reconstruction of 3D models from real environments. ISPRS Journal of Photogrammetry and Remote Sensing 54(1):1-22.

Shufelt, J.A., 1999. Performance evaluation and analysis of monocular building extraction from aerial imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence 21(4):311-326.

Sinning-Meister, M., Grün, A. and Dan, H.,1996. 3D City models for CAAD-supported analysis and design for urban areas. ISPRS Journal of Photogrammetry and Remote Sensing 51:196-208.

Sohn G., 2004. Extraction Of Buildings from High-resolution Satellite Data and Lidar. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B3/III):1036-1042.

Stilla, U. and Jurkiewicz, K., 1999. Reconstruction of building models from maps and laser altimeter data. Lecture Notes on Computer Science 1737:34-46.

Süveg, I. and Vosselman, G., 2004. Reconstruction of 3D building models from aerial images and maps. ISPRS Journal of Photogrammetry and Remote Sensing 58(3-4):202-224.

Söderman, U., Ahlberg, S., Elmqvist, M. and Persson, Å., 2004. Three-dimensional environment models from airborne laser radar data. Proceedings of SPIE Defense and Security Symposium, vol. 5412, 12. p. http://www.sne.foi.se/documents/SNE_SPIE_DS_04.pdf

Taillandier, F. and Deriche, R., 2004. Automatic building extraction from aerial images: a generic Bayesian framework. IAPRS 35(B3:III):343-348.

Takase, Y., Yano, K., Kawahara, N. et al., 2004. Reconstruction and visualization of ”virtual time-space of Kyoto”, A 4D-GIS of the city. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXV(B5): 6 p.

Tan, G. and Shibasaki, R., 2002. A research for the extraction of 3D urban building by using airborne laser scanner data. Proceedings of the 23rd Asian Conference on Remote Sensing 2002.

TerraSolid, 2003. TerraScan User’s Guide.

Tsay, J.-R., 2001. A consept and algorithm for 3D city surface modelling. ISPRS Commission III/WGIII/2, 4 p. http://www.isprs.org/commission3/proceedings02/papers/paper092.pdf

Tseng, Y.H. and Wang, S.D., 2003. Semiautomated building extraction based on CSG model-image fitting. Photogrammetric Engineering and Remote Sensing 69(2):171-180.

Vosselman, G., 2002. Fusion of Laser Scanning Data, Maps, and Aerial Photograps for Building Reconstruction. http://www.geo.tudelft.nl/frs/papers/2002/vosselman.igarss02.pdf

Vosselman, G. and Dijkman, S., 2001. 3D building model reconstruction from point clouds and ground plans. IAPRS 34(3W4):37-43.

84

Vosselman, G. and Süveg, I., 2001. Map based building reconstruction from laser data and images. Automatic Extraction of Man-Made Objects from Aerial and Space Images (III), Ascona, Switzerland, June 11-15, Balkema Publishers, pp. 231-239.

Wang, X. and Grün, A., 2003. A hybrid GIS for 3D city models. IAPRS 33(4/3):1165-1172.

Weidner, U. and Förstner, W., 1995. Towards automatic building extraction from high-resolution digital elevation models. ISPRS Journal of Photogrammetry and Remote Sensing 50(4):38-49.

Zhan, Q., Molenaar, M. and Tempfli, K., 2002. Building extraction from laser data by reasoning on image segments in elevation slices. International Archives of Photogrammetry and Remote Sensing, XXXIV(3,B):305-308.

Zhang, Y. and Wang, R., 2004. Multi-resolution and Multi-spectral Image Fusion for Urban Object Extraction. Proceedings of the XXth ISPRS Congress, July 2004, Istanbul, Turkey, International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,XXXV(B3/III):960-966.

Zhao, H.J. and Shibasaki, R., 2003. Reconstructing a textured CAD model of an urban environment using vehicle-born range scanners and line cameras. Machine Vision and Applications14(1):35-41.

85

Index of Figures

Figure 2-1: Senaatti test site. ................................................................................................................ 14Figure 2-2: Senaatti test site (TopoSys-1). As can be seen, ground plans are shown in advance for

selected buildings. ........................................................................................................... 14Figure 2-3: Hermanni test site. ............................................................................................................. 15Figure 2-4: Hermanni test site (TopEye). ............................................................................................. 15Figure 2-5: Espoonlahti test site. .......................................................................................................... 16Figure 2-6: Espoonlahti test site (TopoSys Falcon).............................................................................. 16Figure 2-7: Amiens test site. ................................................................................................................. 18Figure 2-8: Amiens test site (TopoSys). ............................................................................................... 18Figure 3-1: Work- and Dataflow CC-Modeler™ ................................................................................. 22Figure 3-2: Point definition in CC-Modeler. ........................................................................................ 23Figure 3-3: Workfow of photogrammetric data processing using SOCET Set on the DPW770. ......... 24Figure 3-4: Sequence of manual photogrammetric point measurements for peaked roofs (BAE

Systems, 2000). ............................................................................................................... 25Figure 3-5: Sequence of manual photogrammetric point measurement for gabled roofs (BAE

Systems, 2000). ............................................................................................................... 25Figure 3-6: Breaking down a complex building into simpler components (BAE Systems, 2000). ...... 26Figure 3-7: Geometrical corrections and modelling of photogrammetrically measured buildings

using AutoCAD 2000...................................................................................................... 26 Figure 3-8: Workflow of laser scanning data processing using TerraScan........................................... 27Figure 3-9: Superimposed aerial images and generated roof boundaries. Manual change of the

position of the yellow roof. ............................................................................................. 27Figure 3-10: Building boundary types of roofs..................................................................................... 28Figure 3-11: Parametric building models (here saddleback-roof building). ......................................... 29Figure 3-12: Polyhedral flat-roof building. a) First five corners are measured. b) The whole roof

outline is measured and with one matched ground point the walls are derived. ............. 29Figure 3-13: Building extraction with inJECT with parametric and polyhedral building models and

textured 3D view in a VRML Browser (Student exercise by A. Novacheva and S.H. Foo). ................................................................................................................................ 29

Figure 3-14: (a) Classified building ground plan. (b) DSMzmin.......................................................... 31Figure 3-15: The roof segments............................................................................................................ 32Figure 3-16: Topological points (triangles), intersection lines (thicker) and height jump sections of

the building. .................................................................................................................... 32Figure 3-17: Left: Estimated lines for a jump section. Right: Lines adjusted according the

orientation of the plane.................................................................................................... 32Figure 3-18: Topological points and estimated vertices along height jump sections (triangles). ......... 33Figure 3-19: The 3D model output from the building reconstruction algorithm. ................................. 33Figure 3-20: Work flow used by C+B Technik in building extraction. ................................................ 34Figure 3-21: Types of buildings to be modelled. .................................................................................. 36Figure 3-22: Building model algorithm by Dresden............................................................................. 37Figure 3-23: A building’s point cloud with TIN-structure. .................................................................. 38Figure 3-24: Clusters in the parameter space........................................................................................ 38Figure 3-25: A 3D building model with TIN-structure......................................................................... 39Figure 4-1: Extracted buildings of Espoonlahti. ................................................................................... 45Figure 4-2: Extracted buildings of Hermanni. ...................................................................................... 46Figure 4-3: Extracted buildings of Senaatti. ......................................................................................... 47Figure 4-4: Extracted buildings of Amiens........................................................................................... 48Figure 4-5: Location accuracy (IQR) of the models with respect to single points. .............................. 60

86

Figure 4-6: Building outline deviation, all targets. Laser data use of 0% refers to photogrammetric methods, 100 % to fully laser scanning based techniques and intermediate values to hybrid techniques. ........................................................................................................... 62

Figure 4-7: Average accuracy versus laser point density. .................................................................... 62Figure 4-8: Target height deviation. ..................................................................................................... 63Figure 4-9: Target height and plane (mean of north/east) deviations for different targets. .................. 64Figure 4-10: Obtained planimetric accuracy (mean of north/east) as a function of automation. .......... 64Figure 4-11: Obtained height accuracy as a function of automation. ................................................... 65Figure 4-12: Building height accuracy. IQR: all observations, RMSE: outliers removed.................... 67Figure 4-13: Building length accuracy. IQR: all observations, RMSE: outliers removed.................... 70Figure 4-14: Building length deviation (all observations). ................................................................... 70Figure 4-15: Roof inclination accuracy of the models. IQR: all observations, RMSE: outliers

removed........................................................................................................................... 72Figure 4-16: Total relative shape dissimilarity at Espoonlahti. ............................................................ 74Figure 4-17: Total relative shape dissimilarity at Hermanni. ............................................................... 74

87

Index of Tables

Table 2-1: Aerial images. ..................................................................................................................... 17Table 2-2: Laser scanner data. .............................................................................................................. 17Table 2-3: Participant-generated 3D-models and data formats............................................................. 20Table 3-1: Summary of used data, level of automation and time use of building extraction................ 21Table 3-2: Block diagram of Delft processing strategy ........................................................................ 36Table 3-3: Time used per test site and estimated time needed for extraction of 10 km2 area (hours)... 39Table 3-4: Estimated degree of automation in building extraction*..................................................... 40Table 4-1: Total amount [%] of reference points that could be found from produced 3D-models....... 44Table 4-2: Espoonlahti building location accuracy in cm, all buildings............................................... 49Table 4-3: Espoonlahti building location accuracy in cm, selected buildings. ..................................... 50Table 4-4: Hermanni building location accuracy in cm, all targets. ..................................................... 51Table 4-5: Hermanni building location accuracy in cm, eaves. ............................................................ 52Table 4-6: Hermanni building location accuracy in cm, ridges. ........................................................... 53Table 4-7: Hermanni building location accuracy in cm, other targets. ................................................. 54Table 4-8: Senaatti building location accuracy in cm, all targets. ........................................................ 55Table 4-9: Senaatti building location accuracy in cm, eaves. ............................................................... 56Table 4-10: Senaatti building location accuracy in cm, ridges. ............................................................ 57Table 4-11: Senaatti building location accuracy in cm, other targets. .................................................. 58Table 4-12: Amiens building location accuracy in cm. ........................................................................ 59Table 4-13: Espoonlahti building height accuracy in cm. .................................................................... 66Table 4-14: Hermanni building height accuracy in cm......................................................................... 66Table 4-15: Senaatti building height accuracy in cm............................................................................ 67Table 4-16: Espoonlahti building length accuracy in cm. .................................................................... 68Table 4-17: Hermanni building length accuracy in cm......................................................................... 68Table 4-18: Senaatti building length accuracy in cm............................................................................ 69Table 4-19: Amiens building length accuracy in cm. ........................................................................... 69Table 4-20: Hermanni roof inclination accuracy in degrees................................................................. 71Table 4-21: Senaatti roof inclination accuracy in degrees. ................................................................... 71Table 4-22: Amiens roof inclination accuracy in degrees. ................................................................... 72Table 4-23: Total relative building area and shape dissimilarity at Espoonlahti in %.......................... 73Table 4-24: Total relative building area and shape dissimilarity at Hermanni in %............................. 74

89

Appendix 1: Test Site Visualizations by Participants

Espoonlahti, Senaatti and Hermanni by CyberCity.

90

Senaatti by Delft.

Wireframe models of Espoonlahti and Senaatti by Hamburg.

91

Wireframe models of Espoonlahti, Senaatti, Hermanni and Amiens by Nebel + Partner.

92

Espoonlahti by IGN, with and without texturing from aerial images.

93

Hermanni by IGN, with and without texturing from aerial images.

94

Senaatti by IGN, with and without texturing from aerial images.

95

Amiens by IGN, with and without texturing from aerial images.

Espoonlahti by Stuttgart.

96

Hermanni by Aalborg.

97

Appendix 2: Single Point Deviations

FOI outlines

Hamburg

Nebel+Partner

ICC laser

FOI

ICC laser+aerial IGN

Stuttgart

C+B Technik

CyberCity

Planimetric single point deviations in Espoonlahti

(4m x 4m window, points outside the window: IGN 1, FOI outlines 2, FOI 21).

98

Delft Dresden

Nebel+Partner

IGN

Aalborg

FOI

ICC laser

ICC laser+aerial

FOI outlines

C+B Technik

StuttgartCyberCity

Planimetric single point deviations in Hermanni

(4m x 4m window, points outside the window: ICC laser 1, FOI 1).

99

CyberCity

Nebel+Partner

Delft

Dresden FOI

Hamburg

ICC laser

ICC laser+aerial IGN

FOI outlines

Stuttgart

C+B Technik

Planimetric single point deviations in Senaatti

(4m x 4m window, points outside the window: Nebel+Partner 1, Delft 1, FOI 10).

100

Nebel+Partner

FOI

IGN

FOI outlines

Planimetric single point deviations in Amiens

(4m x 4m window, points outside the window: FOI 1).

101

Vertical single point deviations in test sites (points outside the scale: Espoonlahti IGN 2, ICC

laser 3, Nebel+Partner 1, C+B Technik 2, FOI outlines 5, FOI 1; Hermanni Stuttgart 3, IGN 4,

Delft 2, C+B Technik 2, FOI outlines 6; Senaatti Delft 2, C+B Technik 5, FOI 1; Amiens IGN 2,

FOI 1).

Espoonlahti

Senaatti

Hermanni

Amiens

103

Appendix 3: Difference images

Difference Images, Whole Test Site, Espoonlahti

Aerial images only: CyberCity, Hamburg, Stuttgart (white = extra area, black = missing area).

Laser data and aerial images: IGN, ICC laser+aerial, Nebel+Partner

(white = extra area, black = missing area).

104

Laser data only: ICC laser, FOI (white = extra area, black = missing area).

Laser data and ground plans: C+B Technik, FOI outlines

(white = extra area, black = missing area).

105

Difference Images, Whole Test Site, Hermanni

Laser data only: CyberCity, Stuttgart (white = extra area, black = missing area).

Laser data and aerial images: IGN, ICC laser+aerial, Nebel+Partner

(white = extra area, black = missing area).

106

Laser data only: ICC laser, Aalborg, FOI (white = extra area, black = missing area).

Laser data and ground plans: Delft, C+B Technik, FOI outlines

(white = extra area, black = missing area).

107

Difference Images, Modelled Buildings, Espoonlahti

Aerial images only: CyberCity, Hamburg, Stuttgart (white = extra area, black = missing area).

Laser data and aerial images: IGN, ICC laser+aerial, Nebel+Partner

(white = extra area, black = missing area).

108

Laser data only: ICC laser, FOI (white = extra area, black = missing area).

Laser data and ground plans: C+B Technik, FOI outlines

(white = extra area, black = missing area).

109

Difference Images, Modelled Buildings, Hermanni

Laser data only: CyberCity, Stuttgart (white = extra area, black = missing area).

Laser data and aerial images: IGN, ICC laser+aerial, Nebel+Partner

(white = extra area, black = missing area).

110

Laser data only: ICC laser, Aalborg, FOI (white = extra area, black = missing area).

Laser data and ground plans: Delft, C+B Technik, FOI outlines

(white = extra area, black = missing area).

EuroSDR-Project

Commission 2 “Image analysis and information content”

“Change Detection”

Final Report

Report by Klaus Steinnocher and Florian Kressler

Department Environmental Planning - ARC systems research GmbH, Vienna

113

Abstract

This report describes a methodology employed to classify orthoimages and detect changes by the integration with land use data. The methodology is based on an object-based classification approach which is more adept to deal with the complexity of high-resolution image data than traditional classification methods. Using this method basic land cover classes are identified in the image. These are then compared to the land use data on the basis of plausibility to show where changes have taken place. It is then left to the user to interpret these changes and update the land use data. Four case studies were carried out based on data for Austria, Denmark, Germany and Switzerland. Results show that while the methodology gives consistent results irrespective of the image data, it is very much dependant on the completeness of the land use data.

1 Introduction

In each country in Europe, one or more mapping agencies are concerned with the collection and update of core geo-spatial data. Depending on the different definitions as well as the means of data collection, data sets can vary enormously from country to country. Updates are usually based on visual interpretation of orthophotos, a time consuming and cost intensive task. In order to support this process, automatic and semi-automatic procedures have been examined and some progress in the detection of anthropogenic objects such as buildings and roads has been made. However, due to their complexity and the great variations of spatial data bases, no general solution to identify changes exists.

The necessity of having harmonised spatial data bases, created and kept up-to-date using standardised method-ology, across Europe, has been recognised by the INSPIRE1 initiative. This need for harmonising data and methodologies as well as to support update procedures led to the collaborative project "Change Detection" within the framework of the Science Committee of EuroSDR2. It was initiated by the Austrian Federal Map-ping Agency (Bundesamt für Eich- und Vermessungswesen – BEV) and ARC systems research. Other Euro-pean partners were from Belgium, Denmark, Germany, Finland, Switzerland and Turkey. In international projects EuroSDR functions as a coordination and communication platform by financing travel cost and workshops. In Austria the project was funded by the Ministry of Economics and Labour (Bundesministerium für Wirtschaft und Arbeit – BMWA).

This report describes the methodology employed to classify orthoimages and detect changes by the integration with land use data. The methodology is based on an object-based classification approach which is more adept to deal with the complexity of high-resolution image data than traditional classification methods. Using this method basic land cover classes are identified in the image. These are then compared to the land use data on the basis of plausibility to show where changes have taken place. It is then left to the user to interpret these changes and update the land use data.

Four case studies were carried out based on data for Austria, Denmark, Germany and Switzerland. Results show that while the methodology gives consistent results irrespective of the image data, it is very much dependant on the completeness of the land use data. As the method was developed for data provided by Austria, the best results could be obtained here, which was also confirmed by an evaluation of the responsible mapping agency. The land use data from other countries differs both in content as well as level of generalisa-tion which on the one hand sometimes lowers the utility of the result for the respective user while on the other hand underlines the necessity for more harmonized spatial data as well as update procedures. In addition the suitability of digital scanner data (true colour as well as colour-infrared) for change detection was examined for the Austrian test site.

1 INSPIRE (Infrastructure for Spatial Information in Europe) is a recent initiative launched by the European Commission and developed in collaboration with Member States and accession countries. It aims at

making available relevant, harmonised and quality geographic information to support formulation, implementation, monitoring and evaluation of Community policies with a territorial dimension or impact.

(http://www.ec-gis.org/inspire/ )

2 EuroSDR (European Spatial Data Research) is a research platform for national mapping agencies, academic institutes, the private sector, industry and other groups concerned with European spatial data

infrastructures vital to sustainable spatial planning and development. (http://www.eurosdr.net )

114

2 Project highlights

The project Change Detection has been carried out by ARC systems research in cooperation with the Austrian Federal Mapping Agency and a consortium of international partners. This consortium is composed of the following organisations and their representatives: - BEV, Austria, Michael Franzen

- BKG, Germany, Andreas Busch

- GCM, Turkey, Oktay Aksu

- IGN, Belgium, Ingrid Vanden Berghe

- KMS, Denmark, Brian Pilemann Olsen

- NLS, Finland, Tapio Tuomisto

- SwissTopo, Switzerland, Andre Streilein.

The first meeting of the consortium took place during the EuroSDR meeting in Madrid in October, 2004. There the suggested methodology for change detection was discussed (see annex 1 for project presentation) and the partners agreed to provide data sets. The agreement was presented at the EuroSDR meeting and approved by the science committee.

The data was provided in the end of 2004 and beginning of 2005. The quality of the image data ranges from black and white images with rather low radiometric quality to true-colour and colour-infrared images. The scale of the vector data ranges from 1:1,000 to 1:25,000 with varying levels of generalisation. As the classification procedure requires true-colour or colour-infrared information and minimum information provided with the vector data, the case studies had to be limited to the Austrian, German, Danish and Swiss test sites.

First results were presented at the ISPRS workshop on "High resolution earth imaging for geospatial information" in Hannover, Germany in May, 2005.

In June 2005 a workshop was held at the BEV with participants from nine different countries (see annex 2 for minutes). First results were presented and discussed with the different project partners. Different adjustments were suggested and subsequently incorporated in the procedure. The new results were then conveyed to the partners for evaluation and feedback.

In July 2005 a paper was presented at IGARSS05 conference in Seoul. The paper can be found in annex 3.

The final meeting of the project consortium took place at the EuroSDR meeting in Nicosia in October 2005 (see annex 4 for presentation and annex 5 for minutes). Final feedback was given by the project partners on the change detection results and some of their experiments were presented. In general it was felt that the change map is one possible source for identifying changes but all other available sources must be included as well. Updating data bases can cost up to 40 % of the cost of data base generation and thus more efficient update procedures are desir-able. This coincides with the need to shorten the update cycles (e.g. from ten to three years). There are different priorities for different object groups and the update should be adjusted accordingly. The presentation of the change information must be simple (e.g. traffic light system). If it is too complex the operator would require too much time.

3 Methodology

Automated or semi-automated analysis of high resolution images has been hampered by the high complexity of such images. Traditional pixel-based classifiers such as Maximum Likelihood very often lead to an undesired speckle effect (Willhauck, 2000) as they only look at one pixel at a time without considering its spatial context.

115

Object based classifiers, such as eCognition by Definiens3, deal with this problem by segmenting an image into homogenous segments prior to any classification (Baatz and Schäpe, 2000). Subsequent classification is based on features calculated for each segment. These features not only draw on spectral values but may also relate to size, form, texture, neighbourhood, previous classifications, and so forth. It can be seen that now much more complex classifications can be carried out and are thus better suited to deal with very high resolution data. In the following sections the basic principles of segmentation and object-based classification will be discussed, followed by a description of how the classification results can be compared to land use data and the changes highlighted.

3.1 Segmentation

Segmentation involves grouping neighbouring pixels into homogeneous segments. The degree of homogeneity is governed by the parameters scale, colour, shape, compactness and smoothness (Benz et al. 2004), allowing the procedure to be adjusted to suit different data sets and applications.

Scale determines how heterogeneous the segments may be. A higher value leads to more heterogeneous, and thus larger, segments. The segment size also depends on the complexity of the image data. At a given scale the more homogenous the data are the larger the segments. In order to further influence the shape of the segment the parame-ters colour and shape are used. They sum to one and depending on the values assigned to them colour or form carry more weight. Shape is further divided into smoothness and compactness. They also sum to one and lead either to larger segments with smooth borders or to smaller and very compact segments. The pixel layer can be seen as the lowest segmentation level. New levels are created by grouping neighbouring pixels into segments. New segmenta-tion levels can be created above an existing level by merging neighbouring segments into larger ones. In order to create a level below an existing segmentation, segments are divided into smaller ones. Higher levels can also be created by defining only a parameter for spectral difference. Here all segments which have a spectral difference below the defined threshold are merged. This allows the adjustment of oversegmentation which may take place when very large homogenous objects such as fields or water bodies are split into different segments due to restric-tions introduced by the parameters described above. The important principles of all segmentations are that the borders are preserved, allowing the creation of a segmentation hierarchy (see Figure 3-1). This allows the classifica-tion of an image on more than one level and thus a much more detailed analysis.

Figure 3-1: Segmentation hierarchy Source: eCognition 3.0 User Guide p. 3-28

Any segmentation based classification can only be as good as the underlying segmentation. Inaccuracies encoun-tered here cannot be corrected at a later stage. In order to make the segmentation parameters transferable from one data set to another a trade off has to be made between how many segmentation layers are created and how finely the parameters are tuned to arrive at the desired results.

3.2 Object-oriented classification

Object-oriented classification is based on segments. For each segment different features can be calculated. They may relate to spectral values, shape, texture, hierarchical as well a spatial relations (see Table 3-1) A detailed description of these features is given in Baatz et al. (2003).

3 www.definies-imaging.com

116

Feature Characteristics

Object related Layer values

Shape

Texture

Hierarchy

Thematic Attributes

Class-related Relations to neighbour object

Relations to sub-object

Relations to super-object

Membership to …

Classified as …

Classification value of …

Operators Nearest Neighbour

Similarity to …

Logical operator

Table 3-1: Features used by eCognition for classification

In eCognition classification can be carried out either by a nearest neighbour classifier, using selected segments as training samples to define the different classes, or by fuzzy functions defined for selected features calculated for each segment. The description of these classes is stored in the class hierarchy (see Figure 3-2).

Figure 3-2: Example of class hierarchy

117

Even though a rule based system is more time consuming to create it is given preference over the nearest neighbour approach as it allows more control over the classification process and can be more easily adapted to fit new data. Behind each class in the class hierarchy is the class description in the form of fuzzy functions. An example is shown in Figure 3-3. Here it is determined whether a segment is to be assigned to the class vegetation or not on the basis of two features on the basis of membership. Both class hierarchy as well as class description allow very complex definitions for each class but the basic rule should be: as complex as necessary, as simple as possible. If this is observed then it is easier to understand why segments were assigned to certain classes. In addition when moving to new data sets it facilitates the necessary adjustments of the class description.

Figure 3-3: Example of a class description

In order to remove unwanted artefacts in the initial classification, it is refined on a new segmentation level. This level is created by the so called classification based segmentation. Here the borders of the segments are defined by the classification boundaries of each class. Refinement is then carried out on the basis of neighbourhood, creating the final classification output and the basis for the comparison with the reference data.

3.3 Comparison classification – land use data

The aim of the analysis is to highlight those areas where changes are likely to have taken place. Depending on the number of classes and the aggregation level of the reference data the number of classes can be very different compared to those of the classification. For this reason comparison is carried out on the basis of plausibility as opposed to creating a change/no change map. Depending on which class is present in the reference data and which is in the classification, a segment is either assigned to the change class identical, plausible, questionable or new.

Identical are those class combination which indicate that the reference data and the classification show the same kind of class e.g. red roof in the classification and building in the reference data. Plausible are those combinations that do not directly agree but very likely do not indicate change e.g. meadow in the classification and arable land in the reference data. The term questionable refers to doubts whether, based on the classification, the reference data is correct. An example is grassland in the classification and road in the reference data. As grassland can usually be identified very well the reference data must be put into question. Loss of built up area would also fall into this class.

118

New are those combinations that indicate that building activities have taken place. The information on how each class combination is to be evaluated is laid down in a matrix containing all possible combinations. Through this matrix it is not only possible to accommodate different data sets but also to adjust the results to suit the needs of the operator using the data.

3.4 Contributions of project partners

All project partners provided data in the form of orthophotos and vector data. Four of these data sets could be used for the case studies (see also section four). The results of the change detection were evaluated and commented by each partner for his study site. The most detailed analysis of the proposed methodology was possible with the data provided by the BEV (Austria), both in terms of data provided for the initial cases study as well as reference data used for the evaluation. BKG (Germany) compared the results of the change detection to those derived from its own methodology and provided a detailed evaluation. KMS (Denmark) and SwissTopo (Switzerland) examined and commented the results for their test sites.

4 Case studies

Within the framework of this project raster and vector data were provided by the seven partners. Table 4-1 gives an overview of the characteristics of the raster data and Table 4-2 of the vector data.

Country Number of

images Spectral properties

Spatial resolution (cm)

Size/image (km)

Austria 25 True colour 25 1,25 x 1

Belgium 9 Panchromatic 50 2 x 2

Denmark 2 True colour/colour infrared 50 1 x 1

Finland 2 Panchromatic 50 5 x 5

Germany 1 True colour 40 2 x 2

Switzerland 3 True colour 50 3 x 3

Turkey 2 Panchromatic 100 4 x 5

Table 4-1: Characteristics of available raster data

The number of images ranges from one (Germany) to twentyfive (Austria). Five countries provided true colour images and Denmark also supplied a colour infrared images. The remaining countries provided panchromatic images. The spatial resolution varies from 25 cm (Austria) to 1 m (Turkey) and Image size ranges from 1 x 1 km to 5 x 5 km. All images were ortoimages derived from scanned analogue aerial photographs and subsequently digitally scanned and orthorectified.

The vector data differs very much from country to country in terms of content, accuracy and completeness. The number of classes ranges from seven to twentythree although this also depends partially on the complexity of the study areas. Nevertheless, great differences exist were the aggregation level and data display are concerned. This is important as the methodology improves with the level of detail of the reference data. In Table 4-2 the number of classes, whether single houses are shown, and if and how streets are presented serve as an indicator of level of detail and thus suitability of the vector data for the analysis.

119

Country Number of

classesSingle buildings Streets Format

Austria 23 Yes Yes Polygons

Belgium 20 Yes (large) Yes (lines) Polygons, Lines, Points

Denmark 7 Yes Yes (lines) Polygons, Lines

Finland 11 Yes Yes (lines) Polygons, Lines, Points

Germany 16 No Yes (lines) Polygons, Lines

Switzerland 11 No Yes (lines) Polygons, Lines, Points

Turkey 14 No No Polygons, Lines, Points

Table 4-2: Characteristics of available vector data

In the following sections the case studies carried out using data from Austria, Switzerland, Germany and Denmark will be described. For the other data sets either the image or the vector data were not suitable for the analysis. For each case study the available data, segmentation, classification and comparison will be examined.

4.1 Austria

For Austria the most detailed data sets were available both in terms of images as well as reference data. This made it possible to independently evaluate the results of the change detection technique as well as to examine the benefits and drawbacks of both analogue and digital sensor systems as well as true colour and colour infrared data. In this section the results of the change detection using orthoimages from scanned analogue aerial photographs is pre-sented followed by the independent evaluation of the results by the BEV. The analysis of the images recorded by an airborne scanner is presented in chapter 5.1.

4.1.1 Data and study area

The study area comprises the municipalities of Weinitzen and Wenisbuch near Graz, Austria and covers approxi-mately 17.38 km². Twenty-five real-colour orthoimages, recorded in 1994, were available. Each covers an area of 1.25 x 1 km at a spatial resolution of 25 cm (Figure 4-1). Land use is mainly agricultural with large tracts covered by fields and forest. Some urban areas exist which, except for the city centres, are dominated by single houses with gardens. Outside the cities a large number of individual farms are located. A more detailed view is given in Figure 4-2 showing a subset of the most urbanised part of the study area. The Digital Cadastral Map (DCM) served as reference as well as the basis for the change detection (Figure 4-3). It covers thirtythree classes, nineteen of which are present in the study area. The DCM identifies individual houses and contains a class for roads, making it the most detailed land use data used in this project (see Figure 4-4).

120

Figure 4-1: Orthophotomosaic from 1994 from Austria with outline of study area

Figure 4-2: Subset of orthophotomosaic

121

Figure 4-3: Digital cadastral map of study area in Austria from 1994

Figure 4-4: Subset of cadastral map

122

4.1.2 Segmentation

The first step of the analysis is the segmentation of the image data. Two segmentation levels were created. On the first level a scale of 35 was selected. The parameters colour and shape were given with 0.5 equal weights. Shape is further defined by compactness and smoothness and here more emphasis was put on the former with 0.9, leaving 0.1 for the latter. As the main emphasis is the identification of man-made structures, more compact segments are more appropriate for identifying them. The result of this process can be seen in Figure 4-5. Due to high emphasis on compactness, the segments are all very similar in size and regular in shape. While this is not necessarily a problem in urban areas, it leads to an oversegmentation in forests and especially in homogenous areas covered by meadow as well as in fields. In order to overcome this problem, a second segmentation layer is created. Here only a parame-ter for spectral difference was used. This allows the merging of spectrally similar segments into homogenous areas, while keeping those in heterogeneous areas separate. A spectral difference of 10 was used as the criteria for the segmentation. The result can be seen in Figure 4-6. It is the basis for the classification described in the next section.

Figure 4-5: Segmentation level 1

123

Figure 4-6: Segmentation level 2

4.1.3 Classification

The aim of the classification is to identify basic land cover types. The following classes were defined for the Austrian data set: Vegetation (further divided into forest and meadow), shadow, water, fields, bright objects, red roofs, grey roofs and other urban objects. The features used in the classification were selected in such a way that they can be easily adjusted to fit new data sets. Streets are not an individual class but are classified as one or more of the urban classes. At the lowest level the class hierarchy could be limited to the classes vegetation, shadow, fields, urban and water. Preference, however was given to a more detailed classification, as this allows a subsequent refinement of the classification such as the correction of fields wrongly classified as red roofs.

As described above a rule-based classification using fuzzy functions was selected. Table 4-3 gives an overview of the types of features used for each class. Depending on the type of land cover present in an image the class hierar-chy has to be adjusted e.g. to account for water bodies. The parameters were chosen in such a fashion as to allow an easy adjustment to data provided by different sources. Depending on the radiometric correction performed on the data, no or hardly any adjustments have to be done to classify images from one data set. Most classes are defined by parameters derived from spectral values such as brightness, ratio and standard deviation. Only fields are also defined by other features such as area and number of sub-objects. Streets were not defined as a separate class but are assigned to whatever urban class they fall into depending on the type of surface used. In this strictly hierarchical classification non-urban features such as vegetation and fields are classified first and if the differentiation were only to emphasise urban/non urban objects the classification could be considered finished with the class Not Fields. However, in order to allow refinement of the classification, remove unwanted artefacts and allow a better compari-son with the reference data, the urban classes are differentiated further. The result of the classification for the whole study area can be seen in Figure 4-7. Large parts are covered by meadows (light green) and forest (dark green). A number of bare fields (orange) were identified as well. Of the four urban classes grey objects (cyan) are most abundant. Bright objects (magenta) indicate urban structure with a strong reflectance as is the case for flat roofs. Red roofs (red) are usually single houses and can be identified very well. The remaining segments are assigned to other urban objects (yellow). Shadow is shown in black. In total thirteen features were necessary to arrive at this classification. Many classes could be identified by only one feature. Field is the most complex class requiring four features for identification. A subset (see Figure 4-8) shows the high detail of the classification, going down to single building level. The next step is the comparison with the DCM in order to highlight where changes have taken place.

124

Class Features

Vegetation Ratio green

Brightness

Forest Brightness

Standard deviation green

Meadow Not forest

Shadow Brightness

Field Area

Brightness

Number of sub-objects

Ratio red

Bright Object Brightness

Red Roof Ratio red

Grey Object Brightness

Mean

Other urban object Not grey object

Table 4-3: Features used in classification on level 2

125

Figure 4-7: Classification of 1994 orthoimages

Figure 4-8: Subset of classification of 1994 orthoimages

126

4.1.4 Change map

As the results of the classification can normally not be directly compared to the reference data (here the DCM) an evaluation matrix was created (see Table 4-4). Each possible class combination that may occur has been assigned to one of four plausibility classes (identical in green, plausible in yellow, questionable in blue and new in red). Applying this matrix to the classification and reference data shows where changes might have taken place and give some indication as to their nature (see Figure 4-9).

Classification

DCM

Forest Meadow Field Red roof Bright object

Grey object

Other urban object

Agriculture

Building

Building plot not sealed

Building plot sealed

Extensively used pasture

Fallow land

Field

Forest

Market Garden

Meadow

Orchard

Other

Pasture

Railway

Recreational Area

Road

Water running

Water stagnant

Vineyards identical plausible questionable new

Table 4-4: Evaluation matrix for comparing the classification for Austria with the DCM

It can be seen from the comparison that most areas were considered identical (green). Plausible (yellow) are especially areas that are fields in the DCM but, due to the vegetation cycle, were classified as meadow. Blue (questionable) appears especially along the borders of forests, where either shadows were classified as forest or the forest has increased compared to the reference data. In addition, houses present in the DCM but not present in the image are also classified as questionable.

127

Figure 4-9: Result of evaluation of classification and reference data for Austria

Figure 4-10: Subset of evaluation map for Austria

128

Two types of changes were assigned to new (red). One represents real increases due to new buildings, while the others represents areas which are assigned to building plots not sealed in the DCM but are really driveways (see Figure 4-10).

4.1.5 Evaluation of change map

During the course of the project, the DCM was updated by the BEV land surveying office in Graz based on the same orthophotos used in this analysis together with additional information. This allows a direct comparison between the changes of the DCM using the normal update routine and the changes returned by the proposed change detection procedure. Of course not all updates of the DCM were derived from the visual analysis of orthophotos alone. A number of changes were from other sources and are not visible from the image data, e.g. a change from the class meadow to pasture is a change that cannot be seen in an orthophoto. In order to allow an evaluation of the results, the changes in the DCM were grouped into three classes: increases and decreases in building areas visible in the orthophoto, and changes that are not visible in the image.

Table 4-5 shows how each combination that occurred in the comparison of the two DCMs was assigned. The spatial distribution of the changes is shown in Figure 4-11. Orange indicates changes, that are not visible in the image, magenta those that are visible and represent an increase of built-up areas and blue those that represent a loss of built-up areas. Based on the total area, changes that cannot be seen in the orthophoto are most abundant at approxi-mately 57.7 ha. Visible increases in built-up areas are approximately 13.3 ha and loss of built-up areas 1.2 ha.

Based on the number of changed polygons in the DCM, 709 represent visible increases, ninety-one visible de-creases and 669 other changes (see Table 4-6). As a large number of changes that cannot be derived from an image are related to some sort of agricultural assignment, they tend to larger than those that are due to building activities. Of the 709 increases which should, by definition, be visible, 677 or 95.5 % were identified as change in the classification. The remaining thirtytwo could not be identified which in most cases is due to overhanging vegetation or shadows. Of the ninety-one visible decreases, seventytwo or 79.1 % could be identified, leaving nineteen or 20.9 % undetected. In total of 800 changes visible in the DCM, 749 or 93.5 % could be successfully detected leaving fiftyone undetected. However, based on the change map, sixtyfive additional changes were found.

These changes were evaluated by the BEV land surveying office in Graz. Each change was commented and can be summarised into six categories (see Table 4-7). Thirteen changes concern building sites or other structures where the use is not yet clear and were thus not incorporated in the new DCM. Eightteen changes were objects which were missed during the update and eightteen require further investigation before an update can be carried out. Five are due to land use changes that were missed. Four refer to buildings that are shown in the DCM but not visible in the image. An examination of the orthophotos showed that they are underground structures where the roof is used as a garden. A further seven changes were not incorporated in the provided DCM but had apparently been included before the evaluation had been carried out. In summary, of the sixtyfive changes, thirteen changes were not in the DCM because the final use is not clear and four are underground structures. The remaining fortyeight changes (including the seven already put in the DCM) are shown where some kind of action should be taken, either by updating the DCM or doing field checks.

The comments of the BEV are shown in annex 7. In general it was felt that most additional changes would, without the changes map, only have been found by accident or not at all. If the change map had been available earlier many mistakes would not have happened. The change map together with the DCM and the orthophotos leads to an improvement in the interpretation which cannot be achieved by traditional image interpretation.

129

DCM 2004

DCM 1994 Agr

icul

ture

Bui

ldin

g

Bui

ldin

g pl

ot n

ot s

eale

d

Bui

ldin

g pl

ot s

eale

d

Ext

ensi

vely

use

d pa

stur

e

Fal

low

land

Fie

ld

For

est

Mar

ket G

arde

n

Mea

dow

Orc

hard

Oth

er

Pas

ture

Rai

lway

Rec

reat

iona

l Are

a

Roa

d

Wat

er r

unni

ng

Wat

er s

tagn

ant

Vin

eyar

ds

Agriculture -Building -Building plot not sealed -Building plot sealed -Extensively used pasture -Fallow land -Field -Forest -Market Garden -Meadow -Orchard -Other -Pasture -Railway -Recreational Area -Road -Water running -Water stagnant -Vineyards -

visible increase visible decrease change not visible combination does not occur

Table 4-5: Comparison matrix for land use data

130

Figure 4-11: Evaluation of changes in DCM 1994 - 2004

Number of changes according to DCM

Number of changes shown in change

map Changes missed Additional changes

found

Visible increase 709 677 (95.5 %) 32 (4.5 %) 65

Visible decrease 91 72 (79.1 %) 19 (20.9 %)

Other changes 669

Table 4-6: Evaluation of change map

131

Figure 4-12: Additional changes found using change map

Evaluation Number of changes

Final use not clear 13

Objects missed 18

Objects have to be checked again 18

Land use change has been missed 5

Underground objects 4

Have already been updated in DCM 7

Table 4-7: Summary of evaluation of additional changes not identified on DCM

4.2 Switzerland

4.2.1 Data and Study Area

Three orthophotos were available for the experiments for Switzerland (see Figure 4-13). They were recorded in true colour and each covers 3 x 3 km at a spatial resolution of 50 cm. They show three very different land cover ranging from urban (see Figure 4-13a) and rural (see Figure 4-13b) to an alpine environment (see Figure 4-13c).

132

Figure 4-13: Orthophotos of study areas in Switzerland

Two data sets were used as reference data. One is a land use data base (VECTOR25) covering ten classes (urban, forest, orchard, shrubs, river, lake, wetland, gravel, rock and a class called other which seems to refer to open areas). In addition a road network was available as a vector data set. This was converted into polygon format using data on street width, provided with the data set, and incorporated in the land use data base (Figure 4-14 a-c). The road data was included as otherwise most roads would be shown as change in the change map. When the land use data is compared to the orthophotos, it can be seen that the reference data is very generalised, especially in urban areas.

133

Figure 4-14: Land use data with roads

4.2.2 Segmentation, classification and comparison

The procedure for the analysis follows the same basic steps already employed in the previous case study. Based on a segmentation, the images are classified and the results compared to the reference data to highlight where changes have taken place. For the segmentation on level 1 a scale of 45 was selected. All the other parameters were kept the same as for the Austrian data set (colour/shape 0.5/0.5, compactness/smoothness 0.9/0.1). On Level 2 a segmenta-tion based on spectral difference with a value of 7 was performed.

Classification was carried out on level 2 using the same classes and features as those for the Austrian dataset, only adjusting the parameters to account for radiometric differences. A class for water had to be defined and was included in the class hierarchy. The result for the urban environment (see Figure 4-15a) shows clearly the heavily sealed areas as well as the suburban and vegetated areas. Some misclassification occurs in the river which is due to spectral similarities with urban features. In the image covering the rural environment (see Figure 4-15b) the

134

different villages were identified as well as large parts of the road network. The riverbed on the top left corner was classified as bright object due to the high reflectance of the bare stones. In the alpine scene (see Figure 4-15c) bare rock is classified as bright object as well. Shadows are very prominent in the image and some small patches were classified as fields.

Comparison was carried out based on an evaluation matrix developed for the classes present in the classification as well as in the reference data (see Table 4-8). The results for the urban scene (see Figure 4-16a) show that large parts of the city have been assigned to plausible (yellow) showing parts of the city structure. This is due to generalisation employed in the land use data base and, according to the evaluation matrix, urban areas in the land use data are assigned to plausible if they have been classified as forest or meadow. The river, which has been misclassified, was assigned to questionable (blue) as were some roads in the data base which could not be identified due to tree cover. Some building activities, which have taken place in the study area, are highlighted in red.

Classification

Land use

Forest Meadow Field Red roof

Bright object

Grey object

Otherurbanobject

Water

River

Gravel

Orchard

Lake

Urban

Wetlands

Other

Forest

Rock

Shrub

Road

identical plausible questionable new

Table 4-8: Evaluation matrix for data from Switzerland

135

Figure 4-15: Classification of Swiss data set

The change map for the rural scene (see Figure 4-16b) shows many small villages which are not part of the data base due to generalisation. In addition, many roads are visible as new as well as questionable. This indicates some geometrical inconsistency between the image and the land use data. In the alpine scene many bare rocks, visible in the image but not shown in the land use data are highlighted as new. Rocks covered by vegetation are assigned to plausible.

136

Figure 4-16: Change maps of Swiss data sets

The change maps were evaluated by swisstopo by visual comparison with the new VECTOR25 data. The basic requirements of swisstopo for a change detection system are that the rate of objects wrongly detected as new must be below 25 % while the rate of new object correctly detected must be 100 %. The conclusion of the evaluation were that too many new areas were detected as new even though they were not new, new forest areas were not clearly identified and no optimal results for water could be achieved. In addition it was noted that as long as the swisstopo vector data set is based on map information (two steps of interpretation) accuracy problems (geometry and semantic) will remain (e.g. streets). Possible improvements might be achieved by better tuning between swisstopo categories and detected categories and road buffering. In addition, individual buildings should be used (not available for this study) instead of the data for built up areas employed here.

137

4.3 Germany

4.3.1 Data and study area

For the German study area one real-colour orthophoto covering 2 x 2 km with a spatial resolution of 40 cm was available (see Figure 4-17). It is a predominantly rural area with one very compact town. The reference data is from the ATKIS (Amtliches Topographisch-Kartographisches Informationssystem) database and while being very detailed thematically, the representation of some classes is very generalised. Roads are stored by means of vectors. They were converted into polygons, assigning each road class to a certain street width to mask out road for the comparison.

Figure 4-17: Orthophoto and vector data of study area in Germany © “Hessische Verwaltung für

Bodenmanagement und Geoinformation” (Surveying and Mapping Authority of the Federal State Hessia)

and BKG.

4.3.2 Segmentation, classification and comparison

As before, segmentation was carried out on two levels. On level one a scale of 45 was used with the parameters for shape and colour of 0.5 and compactness and smoothness with 0.9 and 0.1 respectively. Level two was created on the basis of a spectral difference of 7. For this study area the same class hierarchy as that employed for the Austrian data could be used, only adjusting the parameters of some features to account for radiometric differences.

The classification contains 8 classes (see Figure 4-18). Meadows (green) and fields (orange) are very prominent. Due to varying vegetation cover a clear separation is not possible. Forests (dark green) can be found throughout the image. The urban areas are clearly defined. Roads are classified mainly as grey object (cyan) as are some of the buildings. Other buildings were assigned either to bright object (magenta) and red roof (red), depending on the

138

spectral reflectance. Shadows (black) are visible in some parts of the image, especially close to large building blocks and forest edges. The class other urban object (yellow) is here more related to bare fields and, as it is not clearly defined, will be treated as such in the evaluation.

Figure 4-18: Classification of German data set

For the comparison of the classification with the reference data a new evaluation matrix was set up to account for the different classes of the ATKIS data base (see Table 4-9). Each class combination is assigned to one of the four plausibility classes. As described above the combination of other urban object area in the classification and farm-land in the land use data is considered to be identical. The result of the evaluation is shown in Figure 4-19. Very large areas are assigned to questionable. In most cases these are areas which are considered to be grassland in the ATKIS data base but are actually covered by forest in the classification. Some areas where building activities had taken place could be identified and are shown in red. Vegetation in urban areas is highlighted in yellow as are roads along fields. Although most of these roads show up in the initial classification, they were later merged with fields in the refinement of the classification. If these roads were to represent a problem when using the change map, they could be removed by including the first classification stage in the evaluation.

139

Classification

Land use

Forest Meadow Field Red roof

Bright object

Grey object

Otherurbanobject

Allotment

Built up - special feature

Cemetery

Farmland

Grassland

Grove

Industrial areas

Market garden

Mix. of res./ind. areas

Nondescript

Purification plant

Recreational area

Residential area

Specialised cultivation

Sports facilities

Sports field

Road identical plausible questionable new

Table 4-9: Evaluation matrix for data from Germany

140

Figure 4-19: Change map of German data set

4.3.3 A comparison of change detection methods and analysis of results (by Andreas Busch from BKG)

German test data

The German test site selected for this EuroSDR test is located close to Frankfurt am Main. The coloured aerial image defining the test site covers an area of 2km × 2km. The pixel size on the ground is 0.40m. The topographic data set to be checked by means of this orthoimage is the ATKIS Basis DLM. ATKIS is the Authoritative Topog-raphic-Cartographic Information System on the territory of the Federal Republic of Germany. ATKIS is a trademark of the Working Committee of the Surveying Authorities of the States of the Federal Republic of Ger-many (AdV). The most important components of ATKIS are object-based digital landscape models (DLM) encompassing several resolutions and digital topographic maps (DTK). The ATKIS Basis DLM, i.e. the ATKIS data offering highest resolution, is produced by the sixteen surveying authorities of the federal states of Germany and is delivered to the BKG. At the Geodata Centre (GDC) of BKG, the ATKIS Basis DLM is checked with respect to logical consistency, joined to one homogeneous data set for the territory of the Federal Republic of Germany, and stored in a database. Since it is the intention of the test to detect differences of topographic reference datasets and digital images we deliberately selected out-of-date ATKIS data from the archive that do not represent the current state of data acquisition.

Results

The German test site has been analysed by the automated change detection methods from Austria and from Ger-many. The method from Austria is based on the commercial software eCognition and has been developed by the Austrian Research Center (Kressler et al. 2005) whereas the software from Germany has been developed within a common project of BKG and the University of Hannover (Busch et al., 2004, Gerke and Busch, 2005).

The results of the two methods are compared to reference results as obtained by a human operator (Figure 4-20). This is done by means of selected examples and by statistics. As mentioned already out-of-date ATKIS data from

141

the archive have been selected for this test to ensure that there are enough differences between the ATKIS model and reality as given by the orthoimage.

Figure 4-20: Objects of ATKIS Basis DLM where need for update has been marked by a human operator.

Overview

The method from Austria based on eCognition gives diagnostics of the complete scene. It shows mainly five different states, namely “identical”, i.e. the method detected correspondence of model and orthoimage, “new”, i.e. need for update, “questionable”, i.e. the decision of the method is uncertain, “plausible”, i.e. some discrepancies have been detected, but the object as given by the model is still plausible, and “shadow”, i.e. there are shadows in the image which prevent a diagnosis.

While the diagnostics of the Austrian method is based on pixels, the diagnostics of the German method is based on ATKIS objects. The results of the automatic procedure are passed to the human operator in the form of a so-called traffic light diagnostics, i.e. the results are displayed by means of red and green colours. An attribute corresponding to the traffic light diagnostics and indicating the result of the automatic procedure is attached to each inspected object. If the algorithm is able to detect and locate the corresponding image object without observing inconsisten-cies, the ATKIS object is marked with green colour. Otherwise the object is labelled red, i.e. it is not accepted, since the algorithm was not able to establish full correspondence (Figure 4-21). The human operator can access more parameters of the diagnostics whenever necessary. The method from Germany is focused on man-made objects, i.e. roads and built-up areas. One component of the method looks for new buildings and roads by inspecting the whole image and checking agricultural and wooded areas for disturbances caused by these new objects. Another component aims at verifying the roads as they are stored in the ATKIS Basis DLM. Since this component starts from the existing ATKIS roads, it is not able to detect new roads without being guided to a certain area of interest.

142

Figure 4-21: Diagnostics as given by the German method: rejected ATKIS objects are marked by red colour,

cf. text.

Examples

In this section success or failure of the different methods is demonstrated by example. The first example shows a newly constructed built-up area that is not yet registered in the ATKIS data. This has been detected by both methods (Figure 4-22). The method from Germany was not able to verify another built-up area which is stored in ATKIS correctly.

Figure 4-22: ATKIS data: red = built-up area, brown = farmland (left), diagnostics from Austria (middle),

diagnostics from Germany (right), cf. text.

In the second example (Figure 4-23) grassland as given in the ATKIS data should better be stored as grove. Again this has been indicated by both methods.

143

Figure 4-23: ATKIS data: green = grassland, brown = farmland (left), diagnostics from Austria (middle),

diagnostics from Germany (right), cf. text.

The third example (Figure 4-24) shows a situation where the objects as stored in ATKIS have been accepted by a human operator. Here both methods generate “false alarm” because the area does not look homogeneous in the image.

Figure 4-24: ATKIS data: green = grassland, brown = farmland (left), diagnostics from Austria (middle),

diagnostics from Germany (right), cf. text.

Roads and paths are captured in ATKIS by means of linear objects, i.e. they are not represented by areas or regions. Since this is not compatible with the data models used by the Austrian method, the Austrian participants of the test had to apply pre-processing to be able to analyse these ATKIS objects. In the last example (Figure 4-25) paths not stored correctly in ATKIS have been detected by the German method.

Figure 4-25: ATKIS data: green = path, red = road (left), diagnostics from Austria (middle), diagnostics

from Germany (right), cf. text.

After taking into account the complete results of the Austrian method for the German test site in our opinion the method was very successful in detecting newly constructed houses. Agricultural areas are rated as “questionable” if they do not show a homogeneous structure. Trees and rows of trees are not always classified correctly even if they show a similar texture.

Statistics

For rating the performance of change detection methods the results were compared to those obtained by a human operator. The different situations that can occur when comparing decisions from a human operator and diagnostics from automatic procedures are classified in Table 4-10.

144

Automatic

Human Operator

Accepted Rejected

Accepted True Positive False Negative Rejected False Positive True Negative

Table 4-10: Confusion matrix of decisions: human operator vs. automatic procedure, terminology.

The consequences of this approach, when using automatic procedures and leaving the final decision in critical cases to the human operator, are described in Table 4-11. As a consequence the rate of undetected errors should be as low as possible when using this procedure in a production process.

Automatic

Human Operator

Accepted Rejected

Accepted Efficiency Interactive Final Check

Rejected Undetected Errors

Interactive Final Check

Table 4-11: Confusion matrix of decisions: human operator vs. automatic procedure, effects.

The tables below show the statistics based on the German test site for both methods.

Automatic

Human Operator

Accepted Rejected

Accepted 58.0 % 26.0 % Rejected 3.3 % 12.7 %

Table 4-12: Statistics for method from Germany, objects of type region (areas).

Since the method for road extraction from Germany is designed for rural areas, wooded and built-up areas have not been included.

Automatic

Human Operator

Accepted Rejected

Accepted 39.4 % 55.4 % Rejected 1.6 % 3.6 %

Table 4-13: Statistics for method from Germany, roads.

As the results from Austria are pixel-based we have related them to the ATKIS objects. In order to eliminate small clusters of pixels, areas smaller than 300 pixels have not been taken into account. The diagnostics “identical” and “plausible” have been subsumed by the attribute “accepted” whereas “questionable” and “new” have been merged to “rejected”

Automatic

Human Operator

Accepted Rejected

Accepted 57.3 % 29.2 % Rejected 3.4 % 10.1 %

Table 4-14: Statistics for method from Austria.

Conclusions

Table 4-12 and Table 4-14 demonstrate that the quality of the results from Austria and Germany is quite similar. There is no counterpart of Table 4-13 for the Austrian method as the roads from ATKIS do not fit to the model used by the software of the Austrian participants of the test. The concepts of both methods are different, commercial software in the case of the Austrian method and software developed together with a university in the case of the

145

German method. Since eCognition is a product that is well known for its sophisticated classification methods, it would be promising to combine both approaches and to make use of results produced by eCognition within the software developed by BKG and the University of Hannover.

4.4 Denmark

4.4.1 Data and study area

For the study area in Denmark two orthophotos (one true colour and one colour infrared) with a spatial resolution of 50 cm were available. Due to the benefit associated with using colour infrared data, the analysis focuses on that data (see Figure 4-26). The study area covers 1 x 1 km and is located in a suburban area. The available reference data is very detailed, although only some classes are available in polygon format. These are the classes for lake, churchyards, industry, building and forest. Other classes such as roads, streams, railway and so forth are only available in line format and did not have enough information to allow a conversion into polygon format that could be used in the analysis. The building layer is very detailed, showing each individual building.

Figure 4-26: Image and land use data for Danish study area

146

4.4.2 Segmentation, classification and comparison

For the segmentation the colour infrared image is treated in exactly the same way as the real-colour images. Segmentation was carried out on two levels. On Level 1 a scale of 35 with values for shape and colour of 0.5 for each and compactness/smoothness with 0.9/0.1 were applied. The second level was created by using a spectral difference of fiveteen.

For the classification fewer features were needed due to the information provided by the infrared band. Fields were not defined as they are not present in the image. The result (see Figure 4-27) shows that roads and a large number of houses were classified as grey object. A few larger structures were classified as bright object and some objects as red roof, even though they appear in the false-colour composite yellow but, for reasons of consistency, the descrip-tion was kept. Water as well as vegetation could be identified very well as well as some shadow areas.

Figure 4-27: Classification of Danish data set

For the comparison of the classification results with the reference data a downsized approach was chosen. Due to the limitations of reference data only a comparison between the building layer of the reference data and the classification result was carried out (see Table 4-15). The class plausible is not used in the comparison and only those areas were evaluated where buildings are either present in the reference data or in the classification.

147

Classification

Land use

Forest Meadow Field Red roof

Bright object

Grey object

Otherurbanobject

Water

Building

No-Building - - - - identical questionable new - not included in comparison

Table 4-15: Evaluation matrix for data from Denmark

Figure 4-28: Change map of Danish data set

The result shows the problem associated with the lack of road information. Real changes due to building activities cannot be discerned from changes highlighted because of the missing road information. A shift between the reference and the image data can be made out by the large number of small areas classified as questionable along buildings. At this stage the potential of the change detection routine, especially when using colour infrared data, can only be assumed and more reference data would be necessary to make a more thorough test of the methodology.

The results were also evaluated by the National Survey and Cadastre – Denmark and the comments are summarised here. For the classification it was felt that the definition of classes is somehow not "sufficient" and may have to be redefined to suit the "Danish object model". But in general it seems as if it is possible to differentiate man made objects from natural objects. Buildings are differentiated into different classes which also contain other objects especially roads which have to be differentiated from buildings. Dark shadow is detected very well but sometimes man-made objects, covered by "light" shadow are also included in the shadow class. Here the class description needs refinement, maybe by including height information. Water is detected very well and forest and meadow are

148

differentiated very well. The change detection in the present form is not usable due to missing road information. Using cartographic information (road width) may lead to better results since the centre line can be buffered. The shifts of the buildings are obvious and can be handled by transformation of vector data to fit the image data. In addition a problem regarding the "registration description" for the map data base is revealed. Building with atrium gardens are registered as "one" object, without cut out. Some actual new buildings have been detected.

4.5 Additional applications

Within the course of the project, other applications were examined. They all are based on the principle of object-oriented classification but either represent an additional application using the methodology of change detection or address some specific questions of classification. In the first example data from an airborne digital scanner, recorded over the Austrian study area was analysed. This is followed by an application identifying vineyards from real-colour orthophotos. Finally the classification of dwarf pines and other land cover in high alpine areas is examined.

4.5.1 Digital Scanner

4.5.1.1 Data and study area

Real-colour and colour infrared data were recorded with the Ultracam digital scanner over the study area in 2004. Even though the spatial resolution of 0.5 m is lower than that of the optical orthophotos, the availability of a colour infrared band as well as better radiometric quality associated with digital scanner data allowed the prospect of achieving similar results with the change detection process. The orthophotos are in the eastern part of the study area (see Figure 4-29). In the western part of the mosaic there appears to be a cloud-shadow, obscuring parts of the image. A subset of the mosaic is shown in Figure 4-30. When these images are compared to the real-colour images (see Figure 4-31 and Figure 4-32) one can see that the main difference lies in the greater spectral variations of vegetation cover. These variations also allow a better discrimination of vegetation from non-vegetation.

4.5.1.2 Segmentation, classification and comparison

On the first level segmentation parameters were the same for both images with a scale of 12, shape/colour of 0.5 each, compactness of 0.9 and smoothness of 0.1. Spectral difference had for the colour infrared image a value of ten, and for the real-colour image of seven. For both images class hierarchies were created.

The real-colour images could use the same features as those used for the classification of analogue orthophotos, only adjusting the parameters of the fuzzy functions. This meant that the classification could be performed after only a few iterations of adjusting the parameters to arrive at a satisfactory result. The result of the classification (see Figure 4-33 and Figure 4-34) is very similar to that of the analogue orthophotos (see Figure 4-7 and Figure 4-8). The greatest difference is in the assignment of field and meadow due to different vegetation cover. Within urban classes some variation between bright object and grey object is visible, but this is due to radiometric variations and represents no drawback for the comparison with the land use data.

The classification of the colour infrared images could be carried out with fewer features than were necessary for real-colour images. Due to the superior separation of vegetation from non-vegetation, a further refinement in the urban areas was not found to be necessary. This led to the classification of most urban areas to the class other urban object (see Figure 4-35 and Figure 4-36). This speeds up the classification procedures and represents no disadvan-tage for the comparison. The delineation of urban areas is more precise compared to the result using the real-colour images. More areas were classified as shadows in the colour infrared classification (especially prominent is the cloud shadow visible in the mosaic) but given more time this could have easily been adjusted.

149

Figure 4-29: Ultracam false-colour infrared mosaic

Figure 4-30: Subset of false-colour infrared mosaic

150

Figure 4-31: Ultracam real-colour mosaic

Figure 4-32: Subset of real-colour composite

151

Figure 4-33: Classification of real-colour Ultracam mosaic

Figure 4-34: Subset of classification of real-colour Ultracam mosaic

152

Figure 4-35: Classification of colour infrared Ultracam mosaic

Figure 4-36: Subset of classification of colour infrared Ultracam mosaic

153

The classification results were compared to the DCM (see Figure 4-3 and Figure 4-4) using the plausibility matrix shown in Table 4-4.

At this stage an evaluation of the change map can only be based on a visual comparison with the results from the analysis of the analogue orthophotos (see Figure 4-9 and Figure 4-10). The result of the comparison based on the real-colour data (see Figure 4-10 and Figure 4-38) is very similar to that based on the orthophotos. In the subset one can see the same objects were identified as new, as is also the case for most of those identified as questionable.

For the visualisation of the change map based on the classification of the colour infrared data (see Figure 4-39 and Figure 4-40) shadow is not highlighted separately but is displayed in green as is the class identical. This was done to show more details of the relevant changes which otherwise would be obscured in the figure by the profusion of shadow. It can be seen that despite the fewer classes used in the classification, the results of the comparison are very similar to those of the other comparisons. Again basically the same changes were highlighted as in the other change maps, confirming the transferability of the methodology to other sensor systems.

Figure 4-37: Change map from Ultracam real-colour data

154

Figure 4-38: Subset of change map from Ultracam real-colour data

Figure 4-39: Change map from Ultracam colour infrared data

155

Figure 4-40: Subset of change map from Ultracam colour infrared data

156

5 Conclusions and Outlook

In this project a number of applications of a change detection technique were examined based on high-resolution orthophotos using an object-oriented classification approach. The main application combines up-to data orthoi-mages with older land use data to show where changes have taken place and to support the update of the data base. Within the framework of the EuroSDR project data from Austria, Germany, Denmark and Switzerland were examined. The main principle of the methodology is that the images are first segmented and then classified to identify basic land cover types such as meadow, forest, water as well as different urban classes. The next step is to carry out an evaluation based on both the classification as well as the land use data to show where changes have taken place. As the classes employed change from one data base to another and very often cannot be directly derived from a classification of an orthophoto, evaluation of these changes is carried out on the basis of plausibility. Each class combination is assigned to one of four plausibility classes, giving an indication where changes have taken place and the general nature of these changes.

The methodology was designed such that segmentation, classification and comparison can be adapted to both new image as well as land use data. In addition, the level of detail shown in the change map is fully scaleable to suit different applications. The transferability of the methodology was tested on orthophotos from different recording systems (analogue and digital), different spatial scales (0.25 cm to 1 m), different radiometric properties (true-colour and colour-infrared) and different land use data. It could be shown that comparable classification results can be achieved, irrespective of the type of orthophoto used.

The comparison with the land use data was hampered by the fact that in some cases no road information was integrated in the land use data (Switzerland, Germany and Denmark), although for Switzerland and Germany it could be added by converting line information into polygon format using fixed values for street width, depending on the road type. The usability of the results depends very much on the level of generalisation of the data base and in some cases, as for Switzerland, the change map produced too detailed results.

In order to examine the accuracy of the change map and to test its use in a working environment, an update of the Digital Cadastral Map (DCM) was made available for the study area in Austria. The update was based on the same orthophotos used in the classification thus allowing a direct evaluation of the change map. More than 95 % of the visible increases present in the DCM are shown in the change map as well as over 90 % of the visible decrease. In addition, more changes missed in the update of the DCM were found because of the change map (65) than were missed (40). An evaluation of the usability of the change map showed that its integration in the update process improves the quality of the updating results significantly.

Similar results were obtained when classifying imagery from a digital camera, even though the spatial resolution was lower (0.5 cm). This lower spatial resolution has the added advantage of reducing the time of analysis consid-erably without losing too much detail in the change map. Both true-colour as well as colour-infrared data were analysed and from a classification point of view clear preference must be given to the infrared data as the informa-tion contained in the infrared band drastically improves the discrimination of the different land cover types and thus the classification process.

As a consequence of the results of this project the Austrian Federal Mapping Agency is considering the possibility of implementing the change detection method in their updating routine. On a European level the discussion of data harmonisation and improved update procedures to meet the growing demands are ongoing. EuroSDR provides an important platform to channel these discussions and represents a basis for further developments.

157

References

BAATZ, M., SCHÄPE A., 2000: MULTIRESOLUTION SEGMENTATION – AN OPTIMISATION APPROACH FOR HIGH QUALITY

MULTI-SCALE IMAGE SEGMENTATION. IN: STROBL, J. & BLASCHKE, T. (HRSG.), ANGEWANDTE

GEOGRAPHISCHE INFORMATIONSVERARBEITUNG XII. BEITRÄGE ZUM AGIT-SYMPOSIUM SALZBURG 2000.KARLSRUHE. HERBERT WICHMANN VERLAG,12-23

BAATZ, M., BENZ, U., DEHGHANI, S., HEYNEN, M., HÖLTJE, A., HOFMANN, P., LINGENFELDER, I., MIMLER, M., SOHLBACH, M., WEBER, M., WILLHAUCK, G., 2003: ECOGNITION USER GUIDE 3. MÜNCHEN

BENZ, U.C., HOFMANN, P., WILLHAUCK, G., LINGENFELDER, I., HEYNEN, M., 2004: MULTI-RESOLUTION, OBJECT-ORIENTED FUZZY ANALYSIS OF REMOTE SENISNG DATA FOR GIS-READY INFORMATION. ISPRS JOURNAL OF

PHOTOGRAMMETRY & REMOTE SENSING, 58, PP. 239-258

BUSCH, A., GERKE, M., GRÜNREICH, D., HEIPKE, C., LIEDTKE, C. E., MÜLLER, S., 2004: AUTOMATED VERIFICATION

OF A TOPOGRAPHIC REFERENCE DATASET: SYSTEM DESIGN AND PRACTICAL RESULTS. IN: ALTAN, M.O. (ED.),INTERNATIONAL ARCHIVES OF PHOTOGRAMMETRY AND REMOTE SENSING, VOL. 35, PART B2, ISTANBUL, PP.735-740.

GERKE, M., BUSCH, A., 2005: VERIFICATION OF A DIGITAL ROAD DATABASE USING IKONOS IMAGERY. IN: HEIPKE,C., JAKOBSEN, K., GERKE, M., (EDS.), PROCEEDINGS OF THE ISPRS WORKSHOP: HIGH-RESOLUTION EARTH

IMAGING FOR GEOSPATIAL INFORMATION. CDROM, HANNOVER, 2005

KRESSLER, F., FRANZEN, M., STEINNOCHER, K., 2005: SEGMENTATION BASED CLASSIFICATION OF AERIAL IMAGES

AND ITS POTENTIAL TO SUPPORT THE UPDATE OF EXISTING LAND USE DATA BASES. IN: HEIPKE, C., JAKOBSEN, K., GERKE, M., (EDS.), PROCEEDINGS OF THE ISPRS WORKSHOP: HIGH-RESOLUTION EARTH

IMAGING FOR GEOSPATIAL INFORMATION. CDROM, HANNOVER, 2005

STEINNOCHER, K., KRESSLER, F., PETRINI-MONTEFERRI, F., WEICHSELBAUM, J., 2004: CITY-SAT AUSTRIA –ANALYSE VON STADTGEBIETEN UND SIEDLUNGSFLÄCHEN IN ÖSTERREICH MITTELS SATELLITEN- UND LUFT-

BILDGESTÜTZTER ERDBEOBACHTUNG. ARC SYSTEMS RESEARCH REPORT, ARC-S-0225, BV, 80 PP.+ANHANG, JÄNNER 2004

STEINNOCHER, K., KRESSLER, F., 2005: CHANGE DETECTION AUSTRIA – ZWISCHENBERICHT. ARC SYSTEMS

RESEARCH REPORT, ARC-SYS-0052, MAI 2005

WILLHAUCK, G., 2000: COMPARISON OF OBJECT ORIENTED CLASSIFICATION TECHNIQUES AND STANDARD IMAGE

ANALYSIS FOR THE USE OF CHANGE DETECTION BETWEEN SPOT MULTISPECTRAL IMAGES AND AERIAL PHOTO.INTERNATIONAL ARCHIVES OF PHOTOGRAMMETRY AND REMOTE SENSING, XXXIII(PART B3): 35-42.

Annex 1 EuroSDR Change Detection Workshop Vienna, June, 16th

-17th

2005

Minutes

161

Change Detection Workshop

16.-17.6.2005, BEV, Vienna Participants:

Country Name Organisation

Denmark Brian Pileman Olsen KMS

Germany Andreas Busch BKG

Finland Tapio Tuomisto NLS

Ireland Colin Bray OSI

Northern Ireland Aldo Moscato OSNI

Northern Ireland Trevor Steenson OSNI

Switzerland André Streilein SwissTopo

Turkey A. Samil Demirel GCM

UK Malcolm Havercroft OS

UK Paul Marshall OS

Austria Florian Kressler ARC systems research

Austria Klaus Steinnocher ARC systems research

Austria Michael Franzen BEV

Austria Christine Ries BEV

Franzen opens the workshop and presents the agenda

Steinnocher gives a short overview of the project’s background, its objectives and the schedule

the paper presented at the ISPRS Hanover workshop will be distributed to all participants per email

Kressler gives a presentation on the results achieved so far, including:

methodological background (object-oriented classification)

data sets used (from the porject partners)

results for the single test sites

Discussion on:

Definition of segmentation and classification parameters

parameters have to be defined once for images of certain characteristics

no change of parameters is necessary for similar images in terms of radiometry and if similar landscape types are covered

adaptation is required when radiometry or illumination changes (adjustment of spectral values) or when landscape varies (new land cover types have to be added)

How to use linear features (such as road network) if not represented as polygons?

line features in the data base could be buffered according to attribute data

162

What is the impact of changing spatial resolution?

as long as the range lies between 1m and tens of cm no impact is expected

for the test data (25cm – 1m resolution) no adjustments had to be made

Minimum mapping area?

in order to avoid non significant changes a minimum mapping area would be helpful

might vary for different object types

Tuomisto gives a presentation on experiments performed in Finland:

change detection based on image differencing and image composition of multitemporal pan-chromatic Orthophotos

a paper on the results of these experiments is available upon request

Busch gives a presentation on tests of change detection support:

a software developed by the University of Hanover is used to highlight those features of the geo-spatial data base that are not consistent with the corresponding Orthophotos

comparison with the results of the ARC change detection method would be interesting

Steenson gives a presentation on the activities and requirements of the Ordnance Survey of Northern Ireland

General discussion

there is general agreement that „guided“ update of geo-spatial data sets could save time and ef-fort

there might not exist only one method to be used but a combination of several approaches

support comes from local authorities providing information on changes due to constructions (roads, buildings, infrastructure)

requirement is general coverage of country with orthophoto

although satellite based change mapping could contribute to change location on a larger scale no participant knows of any approaches in this direction

Surface model derived from laser scanning is valuable for change detection, but cost ratio ortho-photo – laser scanning 1:6!!!

Evaluation of results

The resulting change maps of the test sites will be made available via secure ftp

Format will be tiff files (shape files in case of CH)

Before distribution of change maps the roads will be considered as cover type:

• linear representation of road network will be buffered and integrated in change map

• for CH and GER vector data of roads will be used (road width via attributes)

• DK will provide street polygons

Evaluation will be based on comparison between change map and independent update of data base

163

Feedback will be given at next workshop (during EuroSDR meeting in Cyprus)

Discussion on implementation

How can change maps be integrated in existing update processes?

Hot spots could be selected in ordered sequence and zoomed to automatically

Ordnance survey does a manual pre-processing of change locations leading to a list of hot spots – interesting to know what the benefit in terms of time/effort is

Problem of inconsistency in the map data due to decreasing concentration of human operators:

< 4 h good

4-6h bad

> 6h nonsense

Any guidance of operator to actual change location would be helpful

Next workshop

Next Workshop will take place in the frame of the 107 EuroSDR meeting in Cyprus

• Date of workshop: Wednesday 26.10.2005

• Exact location and time still to be defined

Focus of workshop will be results of evaluation

Although project will end in Dec. 2005, follow up activities are likely to be needed

Annex 2 Paper presented at IGARSS05, July, 25th

-29th

2005, Seoul

Kressler, F., Franzen, M. and Steinnocher, K.

167

Object-Oriented Classification of Orthophotos to support update of Spatial Databases

F.P. Kressler, K. Steinnocher ARC systems research

A-2444 Seibersdorf, Austria florian.kressler;klaus.steinnocher@arcs.ac.at

M. Franzen Federal Mapping Agency A-1080 Vienna, Austria

michael.franzen@bev.gv.at

Abstract— Interpretation of aerial images is normally carried out

by visual interpretation as traditional classification routines are

too limited in dealing with the complexity of very high resolution

data. Segmentation based classifiers can overcome this limitation

by dividing images into homogenous segments and using them as

basis for further classification procedures. In this paper this

approach is examined in view of its potential to support the

update of existing land use data bases. A workflow was developed

that allows the classification of high-resolution aerial images, the

subsequent comparison with land use data and the assessment of

identified changes. Special emphasis is put on the transferability

of the procedure in terms of study area as well as image and land

use data.

Segmentation, Orthoimage, Classification, Human Settlement,

Semi-automation, Change Detection, Databases

Introduction

Geospatial data such as cadastral maps are usually established and updated from aerial images by visual interpretation, a both time consuming and tedious task. With the advent of digital aerial images and improved analysis methodologies, semi-automated classification procedures can be an improvement both in terms of reliability as well as costs. Within the framework of EuroSDR (Spatial Data Research) the project ‘Change Detection’ was initiated to examine the update of geospatial data on the basis of digital remote sensing images, supported by semi-automated classification methodologies. Partners from Austria, Belgium, Den-mark, Finland, Germany, Ireland, Switzerland and Turkey provided samples of their databases to allow an analysis whether the proposed procedure can fit the different data sets.

The automated or semi-automated analysis of high resolution images has been hampered by the high complexity of such images. Traditional pixel-based classifiers such as Maximum Likelihood very often lead to an undesired speckle effect [1] as they only look at one pixel at a time without considering its spatial context. Object based classifiers deal with this problem by segmenting an image into homogenous segments prior to any classification [2]. Classification is based on

features calculated for each segment. These features not only draw on spectral values but may also be related to size, form, texture, neighborhood, previous classifica-tions, and so forth. It can be seen that now much more complex classifications can be carried out, better suited to deal with very high resolution data. Although a number of object-based classification experiments have been performed (e.g. [3] and [4]), the number of applica-tions using orthophotos is limited.

Aim of this study was to develop a classification proce-dure for very high resolution orthophotos to support the update of existing land use data bases. As land use classes are often defined by their function rather than properties that can be observed in an image, emphasis is put on showing where changes might have taken place, leaving confirmation of these changes, determination of correct boundaries and assignment of appropriate labels to the user. Development of the classification procedure was based on real-color orthophotos from Austria. In a next step the procedure was tested on data from Den-mark, Germany and Switzerland. Although both images as well as land use data are quite different from those used for Austria, the procedure could be adapted very easily to suit the new data sets.

Data sets and study areas

For the development of the methodology 15 real color orthophotos from 2004, recorded over the towns of Weinitzen and Wenisbuch near Graz, Austria were used. Each orthophoto covers 1.25 x 1 km with a spatial resolution of 25 cm. The Digital Cadastral Map was used as reference data. This data set comprises 34 classes, 21 of which were present in the study area. In addition data sets from Switzerland, Germany and Denmark were examined (see Table 1). For Switzerland 3 real color orthophotos (spatial resolution 0.5 m) recorded in 1998 and reference data from 1993 were available, covering 3 x 3 km each. The reference data stems from the Digital Landscape Model of Switzerland. For Germany one real color orthophoto with a spatial resolution of 0.4 m,

168

covering 2 x 2 km with reference data from the ATKIS (Amtliches Topographisch-Kartographisches Informa-tionssystem) data base were provided and from Denmark two orthophotos (one real-color, one with infrared) with a spatial resolution of 0.5 m, covering 1 x 1 km and reference data were made available.

TABLE I. ORTHOPHTOS AVAILABLE FOR EACH COUNTRY

Country Spec-

tral*

Size (km)

per image

Spatial (m) Number of

images

Austria RGB 1.25 x 1 0.25 15 Denmark RGB/IR 1 x 2 0.50 1 Germany RGB 2 x 2 0.40 1 Switzerland RGB 3 x 3 0.50 3

* RGB red, green, blur; IR near infrared

Segmentation, Classification, comparison

The workflow of the analysis can be divided into the three steps: segmentation, classification and comparison. It was developed using the software eCognition from Definiens. Segmentation involves grouping neighboring pixels into homogenous segments. The degree of homogeneity is governed by the parameters scale, color, shape, compactness and smoothness [5], allowing the procedure to be adjusted to fit different data sets and applications. Any segmentation based classification can only be as good as the underlying segmentation. Inaccu-racies encountered here cannot be corrected at a later stage. In order to make the routine transferable from one data set to another a trade off has to be made between how many segmentation layers are created and how finely the parameters are tuned to arrive at the desired results. For the analysis the initial segmentation is carried out on 2 levels (see Table 2). Depending on the spatial, spectral and radiometric resolution different segmentation parameters were applied to each data set. Scale varies on level 1 between 35 and 45. Shape and color have been given equal weights of 0.5. The shape parameter is further defined by compactness and smoothness and here compactness is given with 0.9 considerable more weight than smoothness. Level 2 is created by merging existing segments of level 1 based on the absolute spectral difference. This value varies between 5 and 15, being very dependent on the radio-metric quality of the data. This allows the merging of large homogenous objects, such as fields and grassland, while keeping other objects separate such as houses from the surrounding gardens. A third segmentation takes place after the initial classification. Here the borders of the segments are defined by the classification on level 2.

In eCognition classification can be carried out either by a nearest neighbor classifier or by fuzzy functions defined for selected features calculated for each segment. These features can relate to spectral values, shape, texture,

hierarchical and spatial relations. Even though a rule based system is more time consuming to create it is given preference over the nearest neighbor approach as it allows more control over the classification process and can be more easily adapted to fit new data. During the classification each image segment is assigned to one (or no) class. With the assignment of a class to an image object, the relations to other classes formulated in the specific class description are transferred to the image objects. The result of the classification is a network of classified image objects with corresponding attributes. The following classes were defined for the Austrian data set: vegetation (further divided into forest and meadow), shadow, water, fields, bright objects, red roof, grey roofs and other urban objects. The features were selected in such a way so that they can be easily adjusted to fit new data sets.

TABLE II. SEGMENTATION PARAMETERS FOR LEVELS 1 AND 2

Country Level 1 Level 2 Scale Shape/

Colour

Compactness/

Smoothness

Spectral

Difference

Austria 35 0.5/0.5 0.9/0.1 10 Denmark 35 0.5/0.5 0.9/0.1 15 Germany 45 0.5/0.5 0.9/0.1 7 Switzerland 45 0.5/0.5 0.9/0.1 7

TABLE III. FEATURES USED FOR CLASSIFICATION ON LEVEL 2

Class Features used

Vegetation Ratio Green; Brightness Forest Brightness; Standard Deviation Green Meadow Not ForestShadow Brightness Field Area; Brightness; Number of sub-objects;

Ratio Red Bright Object Brightness Red Roof Ratio Red Grey Object Brightness; Mean Other urban object Not Grey Object

At the lowest level the class hierarchy could be limited to the classes vegetation, shadow, fields, urban and water. But preference was given to a more detailed classification, as it allows a subsequent refinement of the classification such as the correction of fields wrongly classified as red roofs. From this follows that the classi-fication is performed in two stages, one where each segment is assigned to one of the classes, a second where the classification is refined and unwanted artifacts removed. For the second stage a new segmentation is performed on the basis of the initial classification, i.e. the borders of the segments on the new layer are defined by the classification boundaries for each class. Classifi-cation refinement is then carried out on the basis of neighborhood operations. The result is then the basis for the comparison with the reference data.

169

Table 3 gives an overview over the types of features used for each class. Depending on the type of land cover present in an image the class hierarchy has to be adjusted e.g. to account for water bodies. The parameters were chosen in such a fashion as to allow an easy adjustment to data from different sources. Depending on the radio-metric correction performed on the data, no or hardly any adjustments have to be done to classify images from one data set. Most classes are defined by parameters derived from spectral values such as brightness, ratio and standard deviation. Only fields are also defined by other features such as area and number of sub-objects. Streets were not defined as a separate class but are assigned to whatever urban class the fall into depending on the type of surface used. In this strictly hierarchical classification non-urban features such as vegetation and fields are classified first and if the differentiation were only to emphasis urban/non urban objects the classifica-tion could be considered finished with the class NotFields. However, in order to allow refinement of the classification, remove unwanted artifacts and allow a better comparison with the reference data, the urban classes are differentiated further.

Aim of the analysis is to highlight those areas where changes are likely to have taken place. Depending on the number of classes and the spatial aggregation level of the reference data the number of classes can be very differ-ent compared to those of the classification. For this reason comparison is carried out on the basis of plausi-bility as opposed to creating change/no change map. Depending on which class is present in the reference data and which in the classification, a segment is either assigned to the class identical, plausible, questionable or new. Identical are those class combination which indicate that the reference data and the classification show the same kind of class e.g. reed roof in the classifi-cation and building in the reference data. Plausible are those combinations that do not directly agree but very likely do not indicate change e.g. meadow in the classi-fication and arable land the reference data. The term questionable refers to doubts whether, based on the classification, the reference data is correct. An example is grassland in the classification and road in the reference data. As grassland can usually be identified very well the reference data must be put into question. Loss of built up area falls into this class as well. New are all those combinations that are an indication of building activities such as grassland in the reference data and bright object in the classification.

Results

In the following section the results for the Austrian study area will be presented. In addition first experimental results for the other data sets will be discussed. A subset from one orthophoto was selected to allow a more detailed analysis although the procedure was applied to all 15 orthophotos. As described in section III segmenta-tion was performed on two levels. Fig. 1 shows the subset of the Austrian orthophotos with segmentation level 2 overlaid in magenta. The initial classification was performed on level 2 using fuzzy functions drawing on features calculated for each segment as defined for each class within the class hierarchy. As some segments can be wrongly assigned a new segmentation level, based only on the existing classification, was created. Using neighborhood functions unwanted artifacts such as shadows in forest as well as incorrectly assigned seg-ments in bare fields were corrected. The corrected classification for the subset is shown in Fig. 2. Segments assigned to either red roof (red), bright objects (ma-genta), forest or large trees (dark green), grey objects (cyan), meadow (light green), other urban objects(yellow), fields (orange) or shadow (black). When the classification is compared to the original image it can be seen that all urban features have been assigned to an urban class. Depending on the type of roof, houses are either classified as individual objects or blend in with the surrounding area: As the main emphasis is on the differentiation of urban and non-urban objects this does not represent a disadvantage.

Figure 1. Segmentation Level 2

170

ShadowFieldMeadow

Bright object

Other urban objectForest Grey objectRed roof 0 20 40 Meters

Figure 2. Corrected classification of subset

The next step is the comparison with the reference data to highlight where changes have taken place. Fig. 3 shows a subset of the cadastral map. Buildings (red) are depicted as individual houses, unsealed building plots (yellow) are most dominant with very few sealed building plots (magenta). Roads are shown in black, bare land in cyan, agricultural areas in orange, meadows in light green and forests in dark green. A visual compari-son with the orthophoto shows that some building activities have taken place. To show the location and nature of these changes an evaluation matrix was applied to the classification and the reference data. The result is a change map comprising four classes: identical (green), plausible (yellow), questionable (blue) and new (red) (see Fig. 4). A visual comparison shows that all changes were correctly highlighted. The methodology tends to overestimate change, which is often due to the fact that the assignment of an area to a class in the reference data is not only based on land cover but also on historical or legal reasons.

BuildingBuilding plot sealedRoadBuilding plot not sealed

Bare land

RailwayMeadowForest

Agricultural use

0 20 40 Meters

Figure 3. Subset of cadastral map

Figure 4. Evaluation of classification and reference data

The fact that certain areas are assigned to building plot sealed and others to building plot not sealed cannot be derived from the images alone. Discrepancies are highlighted both by the classes new (e.g. sealed areas on unsealed building plots) and questionable (e.g. meadow on sealed building plots). Other questionable segments appear were areas that are supposed to be covered by meadow are covered by forest. Information as shown in Fig. 4 can be the basis for updating the cadastral map both by updating the data base due to actual changes on the ground as well as by correcting inconsistencies.

The procedure described above was not only applied to all 15 orthophotos of the Austrian data set but experi-ments were also carried out on data provided by institu-tions from Denmark, Germany and Switzerland. These

171

data were selected as they contain either real color or infrared orthophotos, which is a prerequisite for the procedure. The class hierarchy and evaluation matrix were adjusted to fit the land cover and classes present in the reference data. First results show that when using either infrared or real color images, even though an infrared band makes it easier to avoid some misclassifi-cations, having a high radiometric resolution is most important. In most cases the same features could for the classification as originally defined and only some parameters had to be adjusted. The strict hierarchical approach makes it easy to adjust them and arrive at a satisfactory result. Differences in spatial resolution (ranging from 0.25 to 0.5 m) were no issue and no consistent changes had to be made to the segmentation parameters, being very much governed by the radiomet-ric quality. Reference data had in general far fewer classes and thus a higher aggregation level than that of the Austrian cadastral map, but as the comparison is based on plausibility this presented no disadvantage.

Conclusion and outlook

The use of high-resolution aerial images for the update of land use data bases is usually limited to visual analy-sis. Automated or semi-automated classification routines are often hampered by the complexity of the data. In this paper a workflow is described that, on the basis of an object-based classifier, allows the analysis of very high resolution orthophotos with the aim to show where changes have taken place. These changes are then evaluated on the basis of plausibility and the update of reference data can concentrate on those areas where building activities or other discrepancies have been highlighted. The classification routine was designed to ensure that very few changes have to be made when moving from one data set to another, either land use data or orthoimagery. This is even true when using data from different sensor types such as real color and infrared imagery. In order to examine the transferability from one data set to another, first experiments were carried out on data sets from four European countries, confirming the adaptability of the procedure. Based on the available results, the potential of using an object based classifica-tion procedure to support the update land use data bases can be considered very high, especially as it can be easily adapted to suit different kinds of orthoimagery as well as land use data bases. Further work will be carried out to test the procedure in a real-working environment and have its usability evaluated by the different institu-tions involved in this project.

References [1] Willhauck, G., "Comparison of object oriented classification techniques

and standard image analysis for the use of change detection between SPOT multispectral satellite images and aerial photos," In: International Archives of Photogrammetry and Remote Sensing Vol. XXXIII, Supplement B3, 2000, pp. 35-42.

[2] Baatz, M. and Schäpe A., “Multiresolution segmentation – an optimization approach for high quality multi-scale image segmentation,” In Strobl, Blaschke & Greisebener (Edts): Angewandte Geographische Informationsverarbeitung XI. Beiträge zum AGIT-Symposium Salzburg 2000, Karlsruhe. Herbert Wichmann Verlag, 2000, pp. 12-23.

[3] Meinel, G., Neubert, M. and Reder, J., "The potential use of very high resolution satellite data for urban areas – first experiences with IKONOS data, their classification and application in urban planning and environmental monitoring," In: Jürgens C. (Eds.): Remote Sensing of Urban Areas/Fernerkundung in urbanen Räumen (=Regensburger Geographische Schriften, Heft 35). Regensburg, 2001, pp. 196-205.

[4] Herold, M., Scepan, J., Müller, A. and Günther, S, "Object-oriented mapping and analysis of urban land use/cover using IKONOS data,". In: Proceedings of 22nd EARSEL Symposium “Geoinformation for European-wide integration. Prague, 2002.

[5] Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen, M., "Multi-resolution, object-oriented fuzzy analysis of remote senisng data for GIS-ready information,” ISPRS Journal of Photogrammetry & Remote Sensing, 58, 2004, pp. 239-258.

Acknowledgments

The authors would like to thank Oktay Aksu from the General Command of Mapping (GCM), Turkey, Ingrid Vanden Berghe from the Institut Geographique National (ING), Belgium, Colin Bray from the Ordnance Survey Ireland (OSI), Andreas Busch from the Federal Agency for Cartography and Geodesy (BKG), Germany, Brian Olsen from National Survey and Cadastre (KMS), Denmark, Andre Streilein from the Swiss Federal Office of Topography (SwissTopo) and Tapio Tuomisto from the National Land Survey (NLS), Finland, for providing orthoimages and reference data in the framework of the EuroSDR project 'Change Detection'.

Annex 3 EuroSDR CD Workshop Nicosia 2005, October, 25th

2005

Minutes

175

Change Detection Workshop

26.10.2005, Hilton Park Hotel, Nicosia

Participants:Country Name Organisation

Austria Florian Kressler ARC systems research

Austria Klaus Steinnocher ARC systems research

Austria Michael Franzen BEV

Belgium Ingrid Van den Berghe NGI

Cyprus Georgia Papathoma DLS

Denmark Brian Pileman Olsen KMS

France Marc Pierrot Deseilligny IGN

France Nicola Champion IGN

Germany Andreas Busch BKG

Germany Christian Heipke IPI Univ Hannover

Switzerland André Streilein SwissTopo

UK Keith Murray OS

Steinnocher opens the workshop and presents the agenda

Participants agree upon agenda

Evaluation of CD results:

Streilein gives a presentation on the CD results of the Swiss test sites

evaluation based on qualitative comparison

requirements:

• detected as “new” but “not new”: < 25% (rule 1)

• detected as “not new” but “new”: 0% (rule 2)

problem is rule 1 – too many changes detected,

caused to some extent by generalisation:

• vector data represent settlement areas, single houses cause wrong changes

• displacements of linear features such as roads cause wrong changes

conclusions:

• there is potential for the future

• better tuning of classes is required

• generalisation effects will be reduced when vector data are derived directly from information source but not from generalised map data (beginning 2007)

Busch gives a presentation on the CD results of the German test site

two methods were compared from ARC systems research (1) and from University of Hannover (2)

176

method 1 indicates land cover change as clusters of pixels

method 2 indicates ATKIS objects that have changed - more generalised approach – operator does not necessarily see reason for change indication

methods are designed for different scales (1:2.000 versus 1:20.000)

requirements:

• detected as “new” but “not new”: < 25%

• detected as “not new” but “new”: 0%

quantitative analysis will be provided for the final report

Steinnocher presents some results from the Austrian test site

first general evaluation results indicate that change map is very helpful for operator despite sceptical attitude of operator at the beginning of the process

a number of actual changes could be detected by means of the change map that were overseen in the first update cycle (without change map)

detailed evaluation will be provided for final report

as for the German test site the two CD methods were compared for the Austrian test site - resulting change maps are rather different, due to different target scale

Olsen gives a summary on the Danish results

tuning of classes is required, e.g. roads not in data base – too many changes

also problem with generalisation effects

0% for no changes is not necessarily required

change map can only be one source for changes, to be combined with other sources

Discussion

change map is one information source for changes, but all sources available should be used

combination of different methods could be helpful

differentiation between information on changes (to be interpreted) and data on changes (to be di-rectly ingested into data base)

update costs come up to 40% of data base generation costs

in general update cycles are shortened (e.g. from 10 years to 3 years), thus speeding up the update process is a must, however

• there are different priorities for different object groups

• update cycles should be linked to life cycles of objects

updating is done partially in 3-D, but change map could be superimposed

Conclusions

any information to speed up update process are welcome

information should be simple (traffic light system), if complex, operator needs more time for inter-pretation

matrix of objects and potential information sources for update should be set up

tested methods tend to overestimate changes - tuning to national requirements is necessary

177

additional tests and/or adjustment to these requirements not possible within this project, but can be done based on individual contracts

topic of CD is relevant for all participants, EU project within 7th FWP is envisaged, potential coordi-nation by Steinnocher

Actions

setting up matrix and sending it out for contributions Steinnocher, Nov 10

Returning matrix to Steinnocher all, Nov 20

Sending contribution for report – Evaluation of German test site Busch, Nov 15

Sending contribution for report – Evaluation of Danish test site Olsen, Nov 15

Sending contribution for report – Evaluation of Austrian test site Franzen, Nov 15

Sending guidelines for EuroSDR report Busch, Nov 15

Distributing draft report (for Austrian ministery) Steinnocher, Dec 15

EuroSDR report Steinnocher, Jan 31

Murray and Heipke will contribute to discussion with JRC regarding 7th FWP

Project building extraction

Deseilligny gives a summary of the proposed building extraction project

focuses on detailed change analysis of building based on optical and lidar data

availability of lidar data very different in different countries, costs very high

not yet generally available, project will have scientific research focus

Action: IGN will look for interested project partners

Annex 4 Comments on change map by BEV

181

Projekt „Change Detection“

Beurteilung der Ergebnisse

Sehr geehrter Herr Hofrat!

Mit diesem Schreiben übermitteln wir unseren Erfahrungsbericht im Zusammenhang mit der „Change Map“. Selten macht die Begegnung mit neuen Techniken so viel Freude. Die im Text zitierten Abbildun-gen finden Sie im Anhang.

Die Luftbildinterpretation im Kataster beschränkte sich lange Zeit auf das Einzeichnen von Auswerteli-nien in Schwarzweißluftbildern nach erfolgtem Feldbegang (klassischer Feldvergleich). Die genaue und manchmal nicht ganz einfache Festlegung schwieriger Abgrenzungen wurde gerne an die Kollegen an den Auswertegeräten weitergegeben. Seit Einführung der DOP werden die BA/NU ohne Feldbegang direkt am Bildschirm aktualisiert. Das kann zu Fehlinterpretationen führen. Durch manchmal mangelnde Bildqualität und die Fülle an („2D“-)Informationen können Objekte falsch abgegrenzt oder übersehen werden.

Unter „Change Map“ konnten wir uns anfangs nicht viel vorstellen. Die Ausführungen Ihres Vortrages an der TU Graz und die Folien am Doku-Server waren die einzigen Grundlagen. Um die Daten in unsere ACAD-Benutzerschale zu laden, mussten die gelieferte DXF-files zuerst in das DWG-Format umgewandelt werden. Mit Unterstützung der Kollegen der IT Graz wurden die Files in mehrere kleinere Portionen unterteilt. Dadurch konnten die Ladezeiten dementsprechend herabgesetzt werden.Die „neuen und fraglichen“ Änderungen allein betrachtet sorgten anfänglich für Verwirrung (Abb. A) aber gemeinsam mit dem DOP wurden die Linien immer verständlicher (Abb. B). Die Bedeutung der „neuen“ Linien war klar und mit dem Laden der DKM wurde der Sinn der „fraglichen“ Linien erkannt (Abb. C).Aber durch das Zu- und Wegschalten des DOP bzw. der einzelnen Layer wurde mit einem Mal ein

so großer Informations- und Interpretationsgewinn sichtbar, den wir in dieser Form nicht für

möglich gehalten hätten (Abb.D).

Die vorangegangene Aktualisierung der BA/NU mit den bisherigen Möglichkeiten war bestimmt gewis-senhaft erfolgt. Durch das Einarbeiten der „addit_changes“ waren eigentlich keine wesentlichen Diffe-

15.11.2005Kern; Nst.221

Projekt „Change Detection“ z.H. HR DI Michael Franzen

Geschäftszahl:

Datum: Rückfragen:

182

renzen mehr zu erwarten. Doch laut „Change Map“ hatten wir wieder etliche Dinge übersehen, bzw. falsch interpretiert.

Die stichprobenartige Überprüfung des Testgebietes mit Hilfe der „Change Map“ lieferte u.a. folgende Beispiele :

Abb.01 : Weg übersehen; erst durch charakteristische Form in „ch_94_04_n1“ aufgefallen.

Abb.02 : Teich aus DOP herkömmlich abgegrenzt; erst durch „ch_94_04_n2“ festgestellt, dass es sich um zwei getrennte Teiche handelt.

Abb.03 : Weg übersehen; erst durch charakteristische Form in „ch_94_04_n1“ aufgefallen. Waldabgrenzung übersehen; erst durch „ch_94_04_n2“ festgestellt.

Abb.04 : Wegverlegung übersehen; erst durch charakteristische Form in „ch_94_04_n1“ aufgefallen.

Abb.05 : Waldabgrenzung übersehen; erst durch „ch_94_04_n2“ festgestellt.

Abb.06 : Fehler in DKM; bei beiden „Bauflächen“, die als „LN“ ausgewiesen sind, sollte der Layer „HG“ gelöscht werden; erst durch „ch_94_04_n2“ festgestellt.

Abb.07 : Waldabgrenzungen übersehen; wesentlicher „Kronenschluß“ erst durch „ch_94_04_n2“ verdeutlicht.

Abb.08 : Wegverlegung; erst durch „ch_94_04_n1“ festgestellt; katastertechnisch aber problematisch

Abb.09 : Gst. 1186 ist eine Verkehrsfläche, die im Kataster (GDB+DKM) irrtümlich als „Baufläche begrünt“ ausgewiesen ist; erst durch „ch_94_04_n1“ aufgezeigt; Fehler wäre wenn nicht durch Zufall, durch keine Prüfroutine entdeckt worden!

Abb.10 : viele Flächen aus „ch_94_04_n1“ und „ch_94_04_n2“, die anfänglich verwirren, keine Relevanz für den Kataster haben, aber das hohe Informationspotential der „Change Map“ dokumentieren.

Durch die Change Map wurden viele zusätzliche Mängel im Testgebiet aufgezeigt, die eine nochmalige flächendeckende Überarbeitung der BA/NU sinnvoll erscheinen lassen. Die meisten Differenzen wären im Nachhinein nur durch Zufall oder gar nie erkannt worden. Wäre die Change Map vorher verfügbar gewesen, wären durch den enormen Interpretationsgewinn viele Fehler gar nicht enstanden.

Durch die Change Map wird in Verbindung mit DKM und DOP ein höchstmöglicher Informations- und Interpretationsgewinn erreicht, der durch herkömmliche (2D-) Luftbildinterpretation nie erlangt werden kann.

Mit freundlichen Grüßen i.A. Karlheinz Kern

EuroSDR-Project

Commission 2 “Image analysis and information content”

In conjunction with IEEE GRSS data fusion

technical committee (DFC)

“Sensor and Data Fusion Contest

Information for Mapping from Airborne SAR

and Optical Imagery”

Final Report on Phase I

Report by Anke Bellmann and Olaf Hellwich

Computer Vision & Remote Sensing - University of Technology, Berlin

185

Abstract

The EuroSDR “Sensor and Data Fusion Contest” is concerned with the question of whether mapping information usually obtained from optical imagery can also be derived from radar data with a compa-rable quality. In the first phase of the contest both optical and SAR data were processed visually in four test regions. To prepare a standardized evaluation procedure, reference data sets were generated for each test region providing the basis for qualitative as well as quantitative data comparison. In the first part of this report the four test site areas are introduced. Subsequently the description of the generation of the reference data is given. The evaluation methods are then explained in detail. Results of the comparison of contest participants' map data with reference data are in the form of percentage values of false positive and false negative extractions with respect to existing objects and a visual interpretation of the map data. Furthermore, taking the individual nature of each participant's results into consideration, we conclude that a pixel accurate evaluation is not possible for all information types. Finally, short conclusions are given summarizing the results of the visual interpretation phase of the contest.

1 Introduction

In the last few years, high quality images of the earth acquired by synthetic aperture radar (SAR) systems carried on a variety of airborne and spaceborne platforms have become increasingly available. During the past couple of decades much research has been performed regarding Synthetic Aperture Radar (SAR) satellite systems such as ERS-1/2, JERS, and RADARSAT-1, demonstrating the potential of SAR for a wide variety of applications, such as deriving DEMs, geomorphological mapping or extracting object of different types.

1.1 Aim of the Project

The aim of the contest was to answer the questions (1) Can state-of-the-art airborne SAR competes with optical sensors in the mapping domain? (2) What can be gained when SAR and optical images are used in conjunction, i.e. when methods for information fusion are applied?

On the one hand there are indications that airborne SAR images will play a major role in topographic mapping in the future, due to its major advantages, namely independence of day light and cloud cover. On the other hand the interpretation of SAR images is difficult for a number of reasons: The geometry and spectral range is different compared to optical imagery and more importantly differs from the imagery perceived through human eyes. In addition, the reflectance properties of objects in the microwave range depend on the frequency band used and may significantly differ from the usual assumption of more or less diffuse reflection at the earths surface. This effect can be particularly strong for buildings or metal surfaces. Speckle, a consequence of coherent radiation needed to exploit the SAR principle and other disturbing factors complicate the interpretation. Therefore, mapping staff, for example photogrammetric operators, often have difficulties interpreting SAR imagery for topog-raphic mapping.

The EuroSDR sensor and data fusion contest was initiated in order to find out to what extent SAR imagery can compete with high resolution optical imagery for topographic mapping applications and to investigate the potential of fusing data of both sensor types.

186

Phase Image Data Linear Objects Areal Objets Participants

I II III SAR Optical roads railways rivers agricult. forest built-up lakes

Budapest University of Technology

and Economics, HUNGARY

Photogrammetry Department

General Command of Mapping, TURKEY

University of Hannover, GERMANY

Institute for Photogrammetry and GeoInformation

University of Rome “La Sapienza”

Department Info-Com, ITALY

Fraunhofer Institute

Information and Data Processing, GERMANY

Technical University Munich Photogrammetry and Remote Sensing, GERMANY

Ness A.T Inc.

ISRAEL

TNO Human Factors

THE NETHERLANDS

Chinese Academy of Surveying

and Mapping, CHINA

1.2 Contest Phases

In phase 1 test participants have visually interpreted the imagery and captured interactively topog-raphic objects as ortho-images of one or more study areas. In the second phase, test participants used automatic methods for object detection and classification. Finally, data fusion will be used in the third phase to derive results using both sensors (Table 1).

Phase I Visual Interpretation

Phase II Automatic Object Extraction and Classification

Phase III Sensor Fusion

Table 1: Contest Phases

In a preliminary phase of the contest, the test site data were organised and prepared for use in the contest. After this work, the first phase started in July 2003. The duration of each contest phase was approximately one year.

1.3 Participants

In total nine institutes are participating in the contest. Table 2 shows a list of these institutes, the phases they are taking part in, and their interests in term of sensor type and object classes.

Table 2: EuroSDR Contest Participants

187

2 Test Sites and Test Data

The test sites investigated in the contest were selected to be neither too complex nor to contain overly simplistic structures. For example, very densely built-up urban regions are hardly to be found. They certainly would have posed a particularly difficult challenge for SAR imaging, which is why it seemed to be appropriate not to investigate them intensively. Therefore, selected test areas include agricultural and forested rural regions as well as industrial and urban areas. Four different test sites with different contents were selected for the contest. The test areas were located in

• Germany – Trudering: Area around the new fair in Munich - industrial and rural regions

• Germany – Oberpfaffenhofen: Campus and runway of DLR - agricultural and airport regions

• Denmark – Copenhagen: City of Copenhagen - residential and industrial areas

• Sweden – Fjärdhundra - rural area, forested and agricultural regions

2.1 Specification of the Test Data

The data used in the contest consisted of optical color as well as black and white photography of photogrammetric quality and state-of-the-art airborne SAR with multiple polarizations and different wavelengths. To meet the objectives only single-pass SAR data is used. In order to account for the large variety of modern SAR systems, different frequencies as well as polarimetric data were selected. Certainly, there are even more parameters strongly influencing the appearance of topographic objects in SAR imagery – such as incidence angle and direction. In spite of their importance regarding object appearance, these parameters could not be considered, as the high demands of the contest regarding resolution, and other more practical matters, such as data costs and copyright, already impose restric-tive conditions. The specifications of the SAR images are shown in Table 3.

Test Site Sensor PolarisationTest Data

ResolutionLooks

Trudering, Germany

AeS-1 None X-Band 1.5 m 16

Oberpfaffenhofen,Germany

E-SAR Lexicographic L-Band 3.0 m 4

Copenhagen, Denmark

EMISAR Pauli C-Band 4.0 m 8

Fjärdhundra,Sweden

EMISAR Pauli C-Band 4.0 m 8

Table 3: SAR images used for different test areas

The resolution of the test data ranges from 1.5 m to 4.0 m and depends on the resolution of the SAR sensors. The corresponding optical images were of higher resolution so that they had to be resampled to the pixel size of the SAR images to enable a fair comparison. Thus, the resolutions of the corre-sponding optical images are the same as those of the SAR images. The following figures (Figure 1 - Figure 4) show the four chosen test data sets. On the left side of the figure the optical image is shown; on the right side the corresponding SAR image. The sections shown are co-registered with pixel accuracy.

188

Figure 1: Data set Copenhagen, urban and suburban areas

Figure 2: Data set Trudering, industrial and rural areas

189

Figure 3: Data set Fjärdhundra, agricultural and forested areas

Figure 4: Data set Oberpfaffenhofen, agricultural and airport areas

190

The pictures show the same display level of detail, but do not possess the same geocoding. To create reference data, an identical geocoding of corresponding images was necessary to display the pictures with pixel accuracy in the same position. This operation was carried out using ERDAS software from Leica Geosystems. It should be noted, however, that the information of interest in the contest is not a correct coordinate interpretation concerning terrestrial coordinates but the correct detection and interpretation of objects and their classification. Therefore a superordinate georeferencing based on terrestrial coordinates was not essential.

SAR data sets from Copenhagen and Fjärdhundra were provided by DDRE, Copenhagen; the data set from Oberpfaffenhofen was provided by DLR-IHR, Oberpfaffenhofen; and the data set from Truder-ing was provided by AeroSensing GmbH (now Intermap). The optical data set form Trudering was provided by Bayerisches Landesvermessungsamt, Munich; from Oberpfaffenhofen by DLR-IHR; from Fjärdhundra by Lantmäteriet, Sweden; and finally the data set from Copenhagen was provided by KAMPSAX, DDOland Denmark.

The acquisition times were different for the four test sites. Fjärdhundra imagery was taken in July 1995, Copenhagen in June 1996, Oberpfaffenhofen in May 1999 and finally the imagery from Trudering was taken in May 2000.

2.2 Additional Data

In addition to SAR and optical data, ground truth data were needed as additional sources of informa-tion about test regions and object types. Therefore, topographic maps for the areas around the test sites were collected. Topographic maps are, in general, approximations in that small objects are broadened but they also include more detailed information about the type of land use and classification. Table 4 shows the topographic maps used to create the reference data for the four test areas.

Test Site Map Name Scale Provider, Publication Date

Copenhagen Danmark, 1514 II SØ, Holte

1:25.000 Kort- & Matrikelstyrelsen, 1995

FjärdhundraGröna kartan, Encöping 11H NV

1:50.000 Lantmäteriverket Gävle, 1993

Trudering München – Trude-ring, 7836

1:25.000 Bayrisches Landesvermessungsamt München, 1997

Oberpfaffenhofen Weßling, 7933 1:25.000 Bayrisches Landesvermessungsamt, 2002

Table 4: Topographic Maps for the test areas

191

The map of Sweden is only available at a scale of 1:50.000 because no larger scale was offered for this area. Furthermore, the maps from Germany were obtainable also as digital maps. The maps from Sweden and Denmark were digitized and georeferenced to be subsequently used in digital form. Figure 5 shows the used topographic maps.

Figure 5: Topographic Maps: a) Trudering, b) Oberpfaffenhofen, c) Copenhagen (scanned), d)

Fjärdhundra (scanned)

a) b)

c) d)

192

3 Reference Data

In order to have a basis for analyzing the interpretations produced as part of the contest, it was necessary to create reference data sets for the different test sites. Reference data contents should ideally include information from optical and SAR imagery, as well as information from topographic maps, accounting for the fact that visual interpretation of image data alone – as conducted by test participants – can only achieve an image-specific interpretation of reality excluding non-image information. Therefore, the outcome of reference data compilation was mapping content considered optimal for the tests purposes from a mapping experts’ point of view – admittedly to a certain degree a subjective view point. This data was used as the basis for comparing participants’ results.

3.1 Generation of Reference Data

The scanned and the digital topographic maps had to be registered with the corresponding optical and SAR images. For reasons of simplicity, geometric reference was taken from the optical images to register the maps. After registration, the maps were geometrically compatible with optical and SAR imagery. However, quantitative (pixel-accurate) interpretations of georeferenced topographic maps were not possible due to inaccuracies of these types of maps. They were only used to examine the content.

Reference data were generated using the vector-based desktop GIS software ArcView of ESRI. Information from SAR and optical data were overlayed and were used to produce an exact reference map; missing information was obtained from the topographic map. Interaction and thorough process-ing guaranteed highly accurate reference data for all data sets (Figure 6). To avoid a bias of the reference maps with respect to optical or SAR imagery the reference information was extracted from both sensor’s images on equal terms.

Figure 6: From three basic information sources to the reference map (left: optical image,

topographic map, SAR image; right: reference map of Copenhagen)

193

3.2 Accounting for Different Acquisition Times

Since data acquisition times were not the same for optical and SAR images, the data sets have marginal discrepancies in some places. Nevertheless, to achieve equal data from both sensor types, masks were created to cut out objects and terrain not clearly defined (Figure 7). Areas covered by the masks were not included in the analysis in order to avoid incorrect interpretation of information content. The same procedure had to be adopted for a part of the airport area of Oberpfaffenhofen.

Figure 7: Areas with different content in SAR and optical images as well as in topographic map

(left); region with different content covered with a mask (right)

In order to achieve quantitative results and due to the fact that data analysis was conducted on raster images, each layer of the reference maps was converted into a raster data set with pixel-size corre-sponding to the imagery. This data conversion was carried out using ERDAS Imagine of Leica Geosystems.

3.3 Reference Maps

As a basis for data evaluation, object classes were categorized into linear and areal (extended) area classes. Object classes used for test sites were:

- Built-up - Railways

- Watercourses - Roads

- Agriculture - Alleys

- Forest

In addition to these general classes, which were present in all data sets, region-specific object classes were defined for each data set considering the specific data content (Table 5). For some areas in the data sets of Copenhagen and Trudering it was not possible to obtain a correct classification. Neither the image content nor the map information allowed the unambiguous definition of a class. These areas were classified as undefined spots and not included in the analysis.

Mask

194

a)

d)

b)

c)

Copenhagen Fjärdhundra Trudering Oberpfaffenhofen

Sport field Sport field Industrial areas Industrial areas Green corridor Green corridor Green corridor

River Excavation area Excavation area

Trench Railway Parking area

Railway

Only in SAR image Airport area

undefined undefined

Table 5: Additional object classes

Figure 8 shows the reference maps compiled manually with a vector based GIS software system. For each object class a separate layer was created. The reference maps are of the same size as the optical and SAR images. To compare the reference data with the participants’ data to pixel accuracy, each layer of the vector-based maps was converted into a raster data layer of the same pixel size as the original images.

Figure 8: Reference Maps, a) Trudering, b) Oberpfaffenhofen, c) Copenhagen, d) Fjärdhundra

195

4 Analysis

4.1 Basic Information

In the initial stages of analyzing the interpreted data, numerous problems became apparent. One problem that needed to be dealt with is the individual interpretation style of each interpreter. As previously shown in ALBERTZ (1970), the quality of image interpretation is very much affected by the level of experience of the interpreter. This is especially true for SAR imagery because understanding of different imaging characteristics is required. A second issue concerned the different programs used by interpreters. Some interpreters use vector-based programs with different object layers for different objects, while others work with raster-based programs and do not separate different layers.

To achieve consistent analysis, the following basic principles of data analysis must be predefined:

• Interpretations which were computed by participants as vector data were converted into raster images, since data analyses are conducted based on raster images.

• Raster images were separated into different layers; one layer for each object class. • Comparisons of the results were conducted for corresponding layers. • Results of the data analysis are expressed in percentage values. Distances and area sizes were

not computed.

4.2 Evaluation Procedure

To evaluate the interpretation results, objects were separated into areal areas and linear object types, since analysis is different for both object types. Image processing for both types of analysis have been carried out with ERDAS Imagine from Leica.

4.2.1 Areal Areas

Comparison of areal objects can take place with pixel accuracy. Each object class was evaluated separately. Therefore, each layer from the reference image was subtracted from the corresponding layer in the participants’ images.

Figure 9: a) Reference data, b) participant’s data, c) result of subtraction

a) b c)

196

Figure 9 shows one example of image processing inputs and result for part of a built-up area analysis. The reference data set is displayed on the left, the participant’s interpretation in the middle, and the result of image subtraction is shown on the right hand side. Different colours in the resulting image are defined as followed:

• Black: Correctly interpreted areas • Orange: False positively interpreted areas • Yellow: False negatively interpreted areas

Correctly interpreted areas are those parts which were located in the interpreter’s image as well as in the reference image. The parts which are defined as false positively interpreted are those areas which were located by the interpreter but do not appear in the reference map of the corresponding object class. And finally, the false negatively interpreted areas were located in the reference map but not detected by the interpreter. Each object class layer of all interpreters was compared with the corre-sponding layer of the reference data in this way.

Figure 10: Image processing procedure

To achieve comparable information for the results, the number of pixels of reference content of each object layer was set to 100%. Thus the number of pixels of false negatively interpreted areas (yellow) and correctly interpreted areas (grey) will add-up to 100% in the resulting image. Those areas which were false positively classified (orange) by the interpreter were added as percentage values above 100% (Figure 10).

Image processing result

W ald - SA R

0

20

40

60

80

100

120

Int 1 Int 3 Int 5 Int 8 Int 10

Referenceimage

Interpreter’simage

Result

197

4.2.2 Linear Objects

Since it is not meaningful and almost impossible to compare one-pixel-wide lines in two images with pixel accuracy in a pixel by pixel comparison, another approach to comparing linear features had to be developed. To compare linear objects, buffer zones were created around reference and interpreted linear objects. The lines were buffered twice. Since an ordinary expressway is of a width of about 25 m to 30 m the first buffer zone was created with a width of 30 m. This is to ensure that marginal discrepancies in interpretation do not directly imply a failure of the interpretation. The second buffer zone was created with twice the width of the first. For a particular comparison, a buffered line and a one-pixel-wide reference line were processed, either stemming from reference and interpretation or – vice versa. If the one-pixel-wide reference line is compared to the buffered interpreter’s line, one obtains the false negative part outside of the buffer. On the other hand, if the one-pixel-wide inter-preter’s line is compared to the buffered reference line, one obtains the false positive part outside of the buffer.

Figure 11: Definition of false positively and false negatively classified linear objects

Using image processing techniques, the image of the one-pixel-wide line and the image of the buffered lines were combined in one image. If the one-pixel-wide line is placed in the first buffer (Figure 11, black line) the line is correctly located. If the one-pixel-wide line is placed outside of the first buffer but within the second buffer (Figure 11, dashed line), the linear object is considered detected, but in a minimally incorrect position. If the line is placed outside of the second buffer (Figure 11, dotted line), it is considered as false positive or false negative. Comparing the one-pixel-wide reference line with the buffered interpreter’s line yields the false negative part and vice versa for the false positive part.

198

Railways & Roads - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6 Int 9

To obtain standardised values for the quality of interpretation, the numbers of pixels of correctly classified objects (from negative and positive analysis) and the numbers of pixels of incorrectly positioned objects were averaged to give AVcor and AVinc respectively. To calculate the total number of pixels, which will represent 100%, the following measure was used:

100% = AVcor + AVinc + fn were fn is the number of false negative pixels (1)

As was the case for areal objects, false positively classified pixels (orange) were added to 100% to display the number of all interpreted pixels. An example of a graph resulting from analysing linear objects is shown in Figure 12.

Whereas the grey part corresponds to correct interpretations, the green part is well classified with a marginally incorrect position. The yellow part represents false negatively interpreted classifications, i.e. not detected, and the orange part represents false positively interpreted classifications.

Figure 12: Example for linear analysis

4.2.3 Qualitative Comparison

In addition to the quantitative analysis for areal and linear objects, a qualitative comparison was conducted to evaluate special objects which were not detected by each interpreter, such as sports fields, bridges or land use boundaries. The qualitative comparison was carried out visually and was necessary since some interpreters provide a more detailed classification than others. Similarly, there is no possibility for evaluating the detection of boundaries, ruins, barracks and parking lots because these detections depend on the interpreter’s classification and can neither be examined pixel-by-pixel nor validated by the topographic maps.

199

A griculture A reas - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

incorrect pos. 7,6 6 ,5 10 ,1

incorrect neg . 14 ,3 2 1,8 14 ,4

correct 8 5,7 78 ,2 8 5,6

Int 2 Int 3 Int 7 Int 8

A griculture A rea - SA R

0,0

20,0

40,0

60,0

80,0

100,0

120,0

incorrect pos. 12 ,5 7,7 9 ,3 4 ,5 9 ,6

incorrect neg . 18 ,5 19 ,0 3 1,9 2 1,9 2 2 ,6

correct 8 1,5 8 1,0 6 8 ,1 78 ,1 77,4

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6

B uilt -Up A reas - o ptical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

i nc or r e c t ne g. 8 , 6 4 , 2 2 , 5

i nc or r e c t pos. 7 , 8 9 , 1 14 , 0

c or r e c t 9 2 , 2 9 0 , 9 8 6 , 0

I nt 2 I nt 3 I nt 7 I nt 8

Built-Up Areas - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

i nc or r e c t pos. 8 , 3 9 , 4 8 , 4 4 , 0 6 , 8

i nc or r e c t ne g. 10 , 4 9 , 9 12 , 9 19 , 3 2 9 , 5

c or r e c t 8 9 , 6 9 0 , 1 8 7 , 1 8 0 , 7 7 0 , 5

I nt 1 I nt 2 I nt 3 I nt 4 I nt 5 I nt 6

5 Results of Phase I

5.1 Test Site Copenhagen

5.1.1 Areal and Linear Objects

The test site data set of Copenhagen contains urban and suburban areas and is of a resolution of 4 m per pixel. The images have a size of 1077 x 729 pixels. The evaluation of the optical image was carried out by four participants and the evaluation of the SAR image by seven participants. The following figures (Figure 13 - Figure 17) show the results of the analysis for areal and linear objects.

Figure 13: Copenhagen Result, Agricultural Area

Figure 14: Copenhagen Result, Built-Up Area

200

F o rest A reas - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

inco rrect po s. 11,6 7,0 2 5,1 7,7

inco rrect neg . 10 ,6 14 ,0 5,2 17,3

correct 8 9 ,4 8 6 ,0 9 4 ,8 8 2 ,7

Int 2 Int 3 Int 7 Int 8

F o rest A rea - SA R

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

i nc or r e c t pos. 10 , 5 9 , 0 15 , 8 5 0 , 6 2 5 , 5 7 , 2

i nc or r e c t ne g. 16 , 2 2 0 , 7 19 , 8 5 , 2 16 , 7 3 2 , 8

c or r e c t 8 3 , 8 7 9 , 3 8 0 , 2 9 4 , 8 8 3 , 3 6 7 , 2

I nt 1 I nt 2 I nt 3 I nt 4 I nt 5 I nt 6

Lakes - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

inco rrect p os. 12 ,9 11,1 13 ,2 7,6

inco rrect neg . 3 5,3 4 0 ,1 3 1,2 4 2 ,9

co rrect 6 4 ,7 59 ,9 6 8 ,8 57,1

Int 2 Int 3 Int 7 Int 8

Lakes - SA R

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

inco rrect p o s. 2 0 ,5 3 0 ,7 2 0 ,8 10 ,7 5,8 2 ,1

inco rrect neg . 4 6 ,2 4 9 ,1 50 ,1 4 6 ,2 72 ,5 8 5,1

co rrect 53 ,8 50 ,9 4 9 ,9 53 ,8 2 7,5 14 ,9

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6

R ailways & R o ads - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

incorr . po sit ive 2 ,7 12 ,7 0 ,7

incorr . negat ive 9 ,6 14 ,6 19 ,1

incorr . po sit ion 4 ,2 3 ,9 3 ,6

correct 8 6 ,2 8 1,4 77,3

Int 2 Int 3 Int 7 Int 8

Railways & Roads - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

inco rr. po sit ive 5,4 10,6 9,2 14,4 8,3 5,8 0,0

inco rr. negative 23,6 20,5 23,0 16,8 28,2 35,5 51,7

inco rr. po sit io n 6,0 7,0 6,0 6,7 8,0 9,3 3,7

co rrect 70,4 72,5 71,0 76,6 63,8 55,2 44,6

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6 Int 9

Figure 15: Copenhagen Result, Forested Area

Figure 16: Copenhagen Result, Lakes

Figure 17: Copenhagen Result, Linear Objects

201

5.1.2 Qualitative Comparison

In addition to the numerical analysis, a qualitative comparison of special objects was performed, because some participants classified more and detailed objects than others. The results of qualitative comparison are shown in Table 6.

SAR image Optical image

Freeway detected detected

Main Road predominantly detected detected

Alleys few to none predominantly detected

Trenches none few to none

Lakes big = good, small = partially big = all, small = predominantly

Trees depends on interpreter, but possible separate ones mostly detected

Buildings big ones detected, smaller ones partially but not always correct

big and small ones detected exemplary

Golf Course only one detection only two detections

Sports Field partially detected nearly all of them

land use boundary, ruins, barracks, parking lots

Table 6: Qualitative Comparison of Copenhagen

5.1.3 Summary – Copenhagen

The results of visual interpretation of Copenhagen show that big and continuous areas like forest, agricultural and built-up areas were detected equally well in SAR and in optical imagery. Smaller areal objects like lakes and separate buildings are more difficult to extract from SAR imagery. Concerning linear objects it can be ascertained that highways and main roads were detected equally well in both types of image. Smaller linear objects, such as alleys, trenches and secondary roads, cannot be detected in SAR but partially in optical imagery. In addition, classification of roads versus alleys appeared to be easier in optical images.

The golf course, which is placed on the left side of the images, was mostly detected as green corridor. Only two interpreters of optical and one interpreter of SAR imagery classified that area as a golf course. The classification of this area in the ground truth (reference) data could only be done well with information from the topographic map.

The visual analysis shows that some interpreters used a more detailed classification than others. They additionally detected land use boundaries, ruins, barracks and parking lots. These objects cannot be approved with certainty, because there is no evidence for them in the available ground truth data.

202

A griculture A reas - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

inco rrect po s. 1,1 4 ,1 3 ,3

inco rrect neg . 7,1 5,6 4 ,0

correct 9 2 ,9 9 4 ,4 9 6 ,0

Int 2 Int 3 Int 7 Int 8

A griculture A rea - SA R

0,0

20,0

40,0

60,0

80,0

100,0

120,0

inco rrect p os. 7,4 4 ,5 6 ,9 5,5 6 ,5

inco rrect neg . 3 ,2 8 ,3 6 ,9 7,7 10 ,5

correct 9 6 ,8 9 1,7 9 3 ,1 9 2 ,3 8 9 ,5

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6

B uilt -Up A reas - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

i nc or r e c t ne g. 11, 6 9 , 5 2 , 1 3 , 5

i nc or r e c t pos. 13 , 2 2 6 , 4 4 0 , 7 6 1, 1

c or r e c t 8 6 , 8 7 3 , 6 5 9 , 3 3 8 , 9

I nt 2 I nt 3 I nt 7 I nt 8

Built-Up Areas - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

i nc or r e c t pos. 10 , 8 7 , 6 16 , 2 19 , 5 11, 4 0 , 9

i nc or r e c t ne g. 4 0 , 3 3 9 , 3 3 8 , 9 5 4 , 3 5 3 , 2 7 6 , 0

c or r e c t 5 9 , 7 6 0 , 7 6 1, 1 4 5 , 7 4 6 , 8 2 4 , 0

I nt 1 I nt 2 I nt 3 I nt 4 I nt 5 I nt 6

5.2 Test Site Fjärdhundra

5.2.1 Areal and Linear Objects

The test site data set of Fjärdhundra contains mainly agricultural and forested areas and has a resolu-tion of 4 m per pixel. The images have a size of 910 x 956 pixels. The evaluation of the optical image was carried out by five participants and the evaluation of the SAR image by seven participants. The following figures (Figure 18 - Figure 22) show the results of analysis for areal and linear objects.

Figure 18: Fjärdhundra Result, Agricultural Area

Figure 19: Fjärdhundra Result, Built-Up Area

203

F o rest A rea - SA R

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

i nc or r e c t pos. 10 , 8 9 , 4 10 , 1 2 5 , 0 7 , 9 8 , 7

i nc or r e c t ne g. 10 , 8 9 , 4 11, 6 9 , 6 18 , 3 18 , 9

c or r e c t 8 9 , 2 9 0 , 6 8 8 , 4 9 0 , 4 8 1, 7 8 1, 1

I nt 1 I nt 2 I nt 3 I nt 4 I nt 5 I nt 6

F o rest A reas - o ptical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

inco rrect p o s. 4 ,5 9 ,8 12 ,2 5,5

inco rrect neg . 5,4 7,5 1,8 10 ,0

co rrect 9 4 ,6 9 2 ,5 9 8 ,2 9 0 ,0

Int 2 Int 3 Int 7 Int 8

R ailways & R o ads - o pt ical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

incorr . po sit ive 2 ,7 12 ,7 0 ,7

incorr . negat ive 9 ,6 14 ,6 19 ,1

incorr . po sit ion 4 ,2 3 ,9 3 ,6

correct 8 6 ,2 8 1,4 77,3

Int 2 Int 3 Int 7 Int 8

Railways & Roads - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

inco rr. po sit ive 5,4 10,6 9,2 14,4 8,3 5,8 0,0

inco rr. negative 23,6 20,5 23,0 16,8 28,2 35,5 51,7

inco rr. po sit io n 6,0 7,0 6,0 6,7 8,0 9,3 3,7

co rrect 70,4 72,5 71,0 76,6 63,8 55,2 44,6

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6 Int 9

Fjärdhundra - River, Optical

0,020,040,060,080,0

100,0120,0140,0160,0180,0200,0220,0

incorr. positive 0,6 15,3 0,6 118,7

incorr. negative 1,2 5,5 1,1 8,6

incorr. position 3,5 3,4 2,1 2,8

correct 95,3 91,1 96,7 88,6

Int 8 Int 9 Int 10 Int 11

Fjärdhundra - River, SAR

-30,0

20,0

70,0

120,0

incorr. positive 7,8 3,1 2,6 19,2 5,2

incorr. negative 20,8 10,1 26,6 30,1 37,6

incorr. position 6,8 6,1 3,8 3,3 8,0

correct 72,3 83,8 69,7 66,6 54,5

Int 1 Int 2 Int 3 Int 4 Int 5 Int 6 Int 7

Figure 20: Fjärdhundra Result, Forested Area

Figure 21: Fjärdhundra Result, Linear Objects: Roads

Figure 22: Fjärdhundra Result, Linear Objects: River

204

well detectedmostly detectedRiver

detectedpredominantly detectedRoads

land use boundary, gardening

Sports Field

Little grown Tree

Areas

Built-Up Areas

Trees

Trenches

Alleys

Main Road

nearly all of thempartially detected

Detection easier -> Trees can be found even if they are of little growing

Detection hardly possible -> don’t looks like Forrest Areas

Big contiguous areas well detected, even small areas can be found

Big contiguous areas mostly well detected, but small areas seemed to

be hardly detected

separate ones mostly detecteddepends on interpreter, but possible

few to nonenone

predominantly detectedfew to none

detecteddetected

Optical imageSAR image

well detectedmostly detectedRiver

detectedpredominantly detectedRoads

land use boundary, gardening

Sports Field

Little grown Tree

Areas

Built-Up Areas

Trees

Trenches

Alleys

Main Road

nearly all of thempartially detected

Detection easier -> Trees can be found even if they are of little growing

Detection hardly possible -> don’t looks like Forrest Areas

Big contiguous areas well detected, even small areas can be found

Big contiguous areas mostly well detected, but small areas seemed to

be hardly detected

separate ones mostly detecteddepends on interpreter, but possible

few to nonenone

predominantly detectedfew to none

detecteddetected

Optical imageSAR image

5.2.2 Qualitative Comparison

Similarly to the analysis of Copenhagen, a qualitative comparison was performed in addition to the main object classes. The results of qualitative comparison are shown in Table 7.

Table 7: Qualitative Comparison of Fjärdhundra

5.2.3 Summary – Fjärdhundra

The results are similar to the results of Copenhagen. Only the continuous and large areas of agricul-ture and forest are detected in SAR as well as in optical imagery. Even though lakes were not located in the reference map (they were not located in the topographic map either), some participants never-theless classified small lakes in the SAR imagery.

Considering the graph of the built-up areas, it is obvious that there are more difficulties in classifying and identifying such areas in SAR images. Some of the built-up areas are located separately and next to the forest areas, so that they can be easily ignored or classified as the objects near by. This also happened in the optical imagery, but not to the same degree.

Regarding the linear objects, large rivers were located in these data sets in addition to the roads. Main roads and the rivers were detected in SAR as successfully as in the optical images, whereas smaller streets were not detected in SAR but in the optical imagery. The smaller streets (alleys, secondary roads) were not included in the quantitative analysis. Concerning the results of river classification, it is obvious that some parts are not detectable in SAR imagery (yellow part of the graph). Considering the result of interpreter 11, a big mistake was made while classifying land use boundaries instead of rivers in the optical image. This leads to the large value (orange part) in the graph.

205

Agriculture - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

Int 1 Int 3 Int 5 Int 8 Int 10

Agriculture - optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

Int 2 Int 4 Int 9

5.3 Test Site Oberpfaffenhofen

5.3.1 Areal and Linear Objects

The data set of the Oberpfaffenhofen test site contains agricultural and industrial areas, especially on the airport area, and is of a resolution of 3 m per pixel. The images have a size of 907 x 658 pixels. The evaluation of the optical image was carried out by four participants and the evaluation of the SAR image by six participants.

The airport area, which covers one third of the image, caused more problems in interpretation than expected. The classification from the participants concerning this region turned out to be extremely diverse and it was impossible to find a way of analyzing them consistently. Hence, a mask was created for this region to avoid quantitative analyses. The result of a visual analysis can be found in the next chapter. Additionally, there were some regions of excavation which were, surprisingly, found by nearly all interpreters. Therefore, an additional class “excavation area” was introduced for quantitative analysis.

The following figures (Figure 23 - Figure 28) show the results of quantitative analysis for areal and linear objects.

Figure 23: Oberpfaffenhofen Result, Agricultural Area

206

Built-Up - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

Int 1 Int 3 Int 5 Int 8 Int 10

Built-Up - optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

Int 2 Int 4 Int 9

Forest - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

Int 1 Int 3 Int 5 Int 8 Int 10

Forest - optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

Int 2 Int 4 Int 9

Waterbodies - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

Int 1 Int 3 Int 5 Int 8

Waterbodies- optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

Int 2 Int 4 Int 9

Figure 24: Oberpfaffenhofen Result, Built-Up Area

Figure 25: Oberpfaffenhofen Result, Forested Area

Figure 26: Oberpfaffenhofen Result, Water bodies

207

Excavation- SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

Int 1 Int 3 Int 5 Int 8 Int 10

Excavation - optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

Int 2 Int 4 Int 9

Roads - optical

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

Int 2 Int 4 Int 6 Int 9

Roads - SAR

0,0

20,0

40,0

60,0

80,0

100,0

120,0

140,0

160,0

Int 1 Int 3 Int 5 Int 7 Int 8 Int 10

Figure 27: Oberpfaffenhofen Result, Excavation Area

Figure 28: Oberpfaffenhofen Result, Roads

5.3.2 Qualitative Comparison

As stated before, the airport area was not analysed with pixel accuracy, because of the different approaches to classification. Nevertheless, a qualitative comparison of these areas will be given. The big runways for planes were detected as well in SAR as in the optical imagery. Even the smaller tarmac covered area could be located, but were not classified correctly. Some smaller objects, which cannot be classified, are located in the airport area. These small objects were located in both sensors’ images completely. Some classified them as hangar or parking areas for airplanes. Even small spots on the green fields were located in SAR images. Concerning the airport area, there are no big differ-ences between SAR and optical image interpretation and classification.

Regarding the interpretations of linear objects, the result is similar to the test sites of Copenhagen and Fjärdhundra. Railways are the exception. They were classified and detected in all cases of optical interpretation but not in SAR images. The lines of railways were detected also by all SAR interpreters but three out of five participants classified the railways as main roads. This leads to the large false positive part (up to 30%) in the results in Figure 28. It shows that the detection of different linear objects is trouble-free, in contrast to the classification.

208

mostly all parking areas were detected big parking areas could be detectedParking

detecteddetectedRunway

all interpreters detected the railways wellpartially detectedRailway

land use boundary, bridges

detectedpredominantly detectedRoads

Built-Up Areas

Alleys

Main Road

well detected but no distinction of industrial or urban areas, separate houses were detected

well, even small ones

well detected but no distinction of industrial or urban areas, separate big

houses were detected well

predominantly detectedfew to none

detecteddetected

Optical imageSAR image

mostly all parking areas were detected big parking areas could be detectedParking

detecteddetectedRunway

all interpreters detected the railways wellpartially detectedRailway

land use boundary, bridges

detectedpredominantly detectedRoads

Built-Up Areas

Alleys

Main Road

well detected but no distinction of industrial or urban areas, separate houses were detected

well, even small ones

well detected but no distinction of industrial or urban areas, separate big

houses were detected well

predominantly detectedfew to none

detecteddetected

Optical imageSAR image

Detection of buildings was not a big problem in SAR and the optical imagery, due to the large size of industrial buildings. Regarding the classification of parking areas, one can say that the big ones were detected and classified in optical and even in SAR images. The smaller ones lead to more problems in SAR imagery. The same is essentially true for bridges and land use boundaries. Table 8 shows an overview of qualitative comparison.

Table 8: Qualitative Comparison of Oberpfaffenhofen

5.3.3 Summary – Oberpfaffenhofen

Although the resolution of sensor data is a bit higher compared to the test sites of Copenhagen and Fjärdhundra (3 m instead of 4 m), there are no significant discrepancies between interpretations of these test sites. Accuracy and quality of interpreting forest, agricultural and built-up areas as well as linear objects show results similar to those in Copenhagen and Fjärdhundra.

Regarding the special areas of this test site, it was surprising that nearly all participants detected and classified the excavation area correctly. More problems occurred in the interpretation of the airport area. Special regions on this site could be detected, but were not classified well. These problems, however, were equal in SAR and optical imagery. Thus, it appears that industrial areas with very special land uses and objects are almost impossible to interpret correctly.

5.4 Test Site Trudering

5.4.1 Areal and Linear Objects

The data set of the Trudering test site mainly contains rural and industrial areas and has a resolution of 1.5 m per pixel, which is the highest resolution among the test imagery. The images have a size of 2813 x 2289 pixels. The evaluation of the optical image was carried out by four participants and the evaluation of the SAR image by seven participants.

This test site predominantly consists of agricultural and built-up areas. The forest areas are small parts with few trees and only cover one percent of the image area. A small lake is located in the north west of the images, which appears different in SAR and optical images due to the exposure date of the images. Therefore, two different reference maps were created for this small area. Two other small parts of the images have different content so they were covered with a small mask to avoid discrepan-

209

Agriculture - sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 1,18 2,12 1,12 1,42 0,44

falsch negativ 9,47 5,32 6,85 7,93 9,33

richtig 90,53 94,68 93,15 92,07 90,67

Interpret 1 Interpret 3 Interpret 5 Interpret 7 Interpret 9

Agriculture - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 1,92 0,47 3,76 0,00

falsch negativ 5,40 9,19 2,67 0,00

richtig 94,60 90,81 97,33 0,00

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Built-Up - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 13,51 12,00 8,58 0,00

falsch negativ 1,81 1,55 4,04 0,00

richtig 98,19 98,45 95,96 0,00

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Built-Up - sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 5,13 11,05 10,39 12,82 19,44

falsch negativ 11,97 3,45 3,46 5,11 5,46

richtig 88,03 96,55 96,54 94,89 94,54

Interpret 1 Interpret 3 Interpret 5 Interpret 7 Interpret 9

Forest - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

160,00

falsch positiv 0,36 3,50 19,03 33,67

falsch negativ 21,58 24,39 54,91 24,72

richtig 78,42 75,61 45,09 75,28

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Forest - sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

160,00

falsch positiv 14,37 1,24 4,57 45,10 48,04

falsch negativ 55,05 39,95 33,52 41,84 49,88

richtig 44,95 60,05 66,48 58,16 50,12

Interpret 1 Interpret 3 Interpret 5 Interpret 7 Interpret 9

cies in analysis. The following figures (Figure 29 - Figure 34) show the results of analyses for areal and linear objects.

Figure 29: Trudering Result, Agricultural Area

Figure 30: Trudering Result, Built-Up Area

Figure 31: Trudering Result, Forested Area

210

Waterbodies - sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 13,28 5,03 8,65 4,61 6,15

falsch negativ 2,97 1,93 1,41 11,44 6,89

richtig 97,03 98,07 98,59 88,56 93,11

Interpret 1 Interpret 3 Interpret 5 Interpret 7 Interpret 9

Waterbodies - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

falsch positiv 0,54 2,05 4,40 4,18

falsch negativ 11,91 4,04 1,28 7,99

richtig 88,09 95,96 98,72 92,01

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Excavation - sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

falsch positiv 2,77 22,11 31,19 15,20 13,75

falsch negativ 10,11 3,72 5,85 7,85 7,38

richtig 89,89 96,28 94,15 92,15 92,62

Interpret 1 Interpret 3 Interpret 5 Interpret 7 Interpret 9

Excavation - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

falsch positiv 2,79 4,03 24,79 0,00

falsch negativ 5,39 24,88 2,77 0,00

richtig 94,61 75,12 97,23 0,00

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Roads - optic

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

160,00

falsch positiv 5,20 1,17 16,02 46,25

falsch negativ 5,31 12,03 4,77 3,91

lagefalsch 5,42 2,93 2,92 5,38

lagerichtig 89,27 85,04 92,31 90,72

Interpret 2 Interpret 4 Interpret 8 Interpret 10

Roads- sar

0,00

20,00

40,00

60,00

80,00

100,00

120,00

140,00

160,00

falsch positiv 7,20 2,60 1,53 1,46 7,67 6,21 1,46

falsch negativ 9,21 8,70 14,86 23,43 29,95 20,35 23,43

lagefalsch 7,30 6,78 4,63 1,67 2,78 2,00 1,67

lagerichtig 83,50 84,52 80,51 74,90 67,27 77,66 74,90

Interpret 1 Interpret 3 Interpret 5 Interpret 6 Interpret 7 Interpret 9Interpret

11

Figure 32: Trudering Result, Water bodies

Figure 33: Trudering Result, Excavation Area

Figure 34: Trudering Result, Roads

211

5.4.2 Qualitative Comparison

The railways in the north of the images were located by all interpreters (just one interpreter only evaluated roads) and classified correctly by all SAR interpreters. Only two optical interpreters classified them as roads. Some SAR and optical interpreters included several houses in the analysis. In several cases, both optical and SAR images allow the detection and location of houses, even the smaller ones. In addition some interpreters, SAR as well as optical, detected land use boundaries, bridges, parking areas, trees and hedges. It was possible to detect them well in optical and in SAR images.

Some undefined lines, which look like historical objects, can be found in a small portion of the SAR image. These lines are only located in the SAR image. Three interpreters of the SAR image located them well; the other classified this area as an undefined area. That is the reason why it is not included in the quantitative analyses.

5.4.3 Summary – Trudering

The figures relating to the quantitative analyses show the best result of the test sites. The continuous and large parts of agriculture and built-up areas were located and classified with best results (mostly up to 90% correct). Even the smaller areas, such as water bodies and excavation areas were identified with a high degree of accuracy. The interpretation of forests – usually conducted very accurately – resulted in more failures. Some interpreters added hedges and several trees to the forest class which might be a reason for misinterpretation. Due to the high resolution of the image data, it was possible to detect several houses and trees comparatively well.

The detection of linear objects shows best results equally well in SAR and in optical imagery. Just the classification of railways might be difficult, surprisingly in optical imagery. The fact, that this test site’s data is of a high resolution leads to the best visual interpretation result among all test sites. The accuracy and quality for SAR and optical imagery is equal, even for smaller objects.

212

6 Conclusion – Phase I

The results of all test sites show a consistent status of visual interpretation. The goal of the first phase was to compare optical versus SAR image interpretation. The question to be answered was whether it would be possible to obtain the same accuracy and quality when interpreting SAR instead of optical image data.

In general all of the linear and areal features could be interpreted quite well in SAR images. The results for the four different test sites look similar. It is obvious that main roads and highways, i.e. big linear objects, can be detected and well located in both sensor imageries. Although it is no problem to locate big linear objects, classification appears to be much more difficult in SAR images, as railways were sometimes classified as streets. Smaller linear objects, such as roads, alleys and little rivers, lead to more errors when locating and classifying them in SAR images.

While comparing areal areas, large contiguous areas are usually well detected while small areas seem to be barely detectable in SAR imagery, whereas they can be found in optical images. Nevertheless, large buildings can also be detected and well located in SAR while, as before, the smaller ones are not always located correctly. The same results are obtained with sport fields, bridges and several trees, only partially detected in SAR but nearly completely detected in optical imagery.

Small forest areas are barely detectable in SAR because they don’t look like forest areas in the images (which may also depend on the experience of the interpreters). Another problem seems to be the detection of water surfaces: large ones can be located and classified well in SAR images, but smaller ones can not be found, unlike in optical images.

Concerning the optical data there is no difference in performance between panchromatic and color images. This can be inferred from the results of Fjärdhundra and Copenhagen test sites, where the data are of the same resolution, whereas Fjärdhundra is panchromatic and Copenhagen color image data. Regarding the results for test site Trudering, it becomes obvious that a higher resolution leads to better results for interpretations of SAR images. Consequently, discrepancies between SAR and optical sensing are shrinking with increasing resolution of images.

To answer one of the main questions of this contest, it seems that SAR performs as well as optical sensing for large areal and linear objects but is not satisfactory with respect to smaller objects. Even though optical imagery was resampled to the pixel size of SAR sensing, it allows much more detailed analysis if the resolution is about three or four meters. This changes when using SAR data with higher resolutions, at least as far as the test site of Trudering is concerned. Results could be even better when using modern SAR data with resolution in the range of decimeters, as then speckle filtering could be done intensively without losing the interpretability of the data.

Without training the interpreters in finding small or thin objects of a certain kind they are not familiar with in SAR imagery, it is nearly impossible for them to extract those objects correctly. Due to the similarities between the human visual system and optical imaging, this is much easier when using optical images. The analysis shows that under the conditions set by the test regarding resolution of the data and experience of the interpreters SAR sensing alone is not satisfactory for mapping small objects of the size of a few pixels in one or two dimensions.

213

Acknowledgments

We thank all participants of the contest for contributing their work and analyses. Further thanks belong to Gert König, Hartmut Lehmann and Andreas Reigber who helped to initiate the contest, as well as Sübeyde Demirok and Andreas Krause who finished their seminar thesis analysing a part of the data.

Support of several authorities in providing data for the fusion contest, especially Intermap Corp. and Bayrisches Landesvermessungsamt München, IHR/DLR (E-SAR data), Aerosensing GmbH (AeS-1 data) as well as the DDRE (EmiSAR data) and KAMPSAX is gratefully acknowledged.

References

Albertz, J., 1970: Sehen und Wahrnehmen bei der Luftbildinterpretation, Bildmessung und Luftbild-wesen 38, (1970), Karlsruhe.

Bellmann, A., Hellwich, O., 2005: Sensor and Data Fusion Contest: Comparison of Visual Analysis between SAR- and Optical Sensors, in Publikationen der DGPF, Band 14, 2005, pp. 217-223

Hellwich, O., Heipke, C., Wessel, B., 2001: Sensor and Data Fusion Contest: Information for Mapping from Airborne SAR and Optical Imagery, in International Geoscience and Remote Sensing Symposium 01, Sydney. IEEE, 2001, vol. VI, pp. 2793-2795.

Hellwich, O., Reigber, A., Lehmann, H., 2002: OEEPE Sensor and Data Fusion Contest: Test Imagery to Compare and Combine Airborne SAR and Optical Sensors for Mapping, in Publikationen der DGPF, Band 11, 2002, pp. 279-283.

Lohmann, P., Jacobsen, K., Pakzad, K., Koch, A., 2004: Comparative Information Extraction from SAR and optical Imagery: IntArchPhRS. Band XXXV, Part B3. Istanbul, 2004, pp. 535-540.

214

Index of Figures

Figure 1: Data set Copenhagen, urban and suburban areas ..................................................188Figure 2: Data set Trudering, industrial and rural areas .......................................................188Figure 3: Data set Fjärdhundra, agricultural and forested areas ...........................................189Figure 4: Data set Oberpfaffenhofen, agricultural and airport areas ....................................189Figure 5: Topographic Maps: a) Trudering, b) Oberpfaffenhofen, c) Copenhagen (scanned),

d) Fjärdhundra (scanned)..............................................................................................191Figure 6: From three basic information sources to the reference map (left: optical image,

topographic map, SAR image; right: reference map of Copenhagen)..........................192Figure 7: Areas with different content in SAR and optical images as well as in topographic

map (left); region with different content covered with a mask (right) .........................193Figure 8: Reference Maps, a) Trudering, b) Oberpfaffenhofen, c) Copenhagen,

d) Fjärdhundra ..............................................................................................................194Figure 9: a) Reference data, b) participant’s data, c) result of subtraction ...........................195Figure 10: Image processing procedure................................................................................196Figure 11: Definition of false positively and false negatively classified linear objects .......197Figure 12: Example for linear analysis .................................................................................198Figure 13: Copenhagen Result, Agricultural Area ...............................................................199Figure 14: Copenhagen Result, Built-Up Area.....................................................................199Figure 15: Copenhagen Result, Forested Area .....................................................................200Figure 16: Copenhagen Result, Lakes ..................................................................................200Figure 17: Copenhagen Result, Linear Objects ....................................................................200Figure 18: Fjärdhundra Result, Agricultural Area................................................................202Figure 19: Fjärdhundra Result, Built-Up Area .....................................................................202Figure 20: Fjärdhundra Result, Forested Area......................................................................203Figure 21: Fjärdhundra Result, Linear Objects: Roads ........................................................203Figure 22: Fjärdhundra Result, Linear Objects: River .........................................................203Figure 23: Oberpfaffenhofen Result, Agricultural Area.......................................................205Figure 24: Oberpfaffenhofen Result, Built-Up Area ............................................................206Figure 25: Oberpfaffenhofen Result, Forested Area ............................................................206Figure 26: Oberpfaffenhofen Result, Water bodies..............................................................206Figure 27: Oberpfaffenhofen Result, Excavation Area ........................................................207Figure 28: Oberpfaffenhofen Result, Roads .........................................................................207Figure 29: Trudering Result, Agricultural Area ...................................................................209Figure 30: Trudering Result, Built-Up Area.........................................................................209Figure 31: Trudering Result, Forested Area .........................................................................209Figure 32: Trudering Result, Water bodies ..........................................................................210Figure 33: Trudering Result, Excavation Area.....................................................................210Figure 34: Trudering Result, Roads......................................................................................210

215

Index of Tables

Table 1: Contest Phases ........................................................................................................186Table 2: EuroSDR Contest Participants ...............................................................................186Table 3: SAR images used for different test areas................................................................187Table 4: Topographic Maps for the test areas.......................................................................190Table 5: Additional object classes ........................................................................................194Table 6: Qualitative Comparison of Copenhagen.................................................................201Table 7: Qualitative Comparison of Fjärdhundra .................................................................204Table 8: Qualitative Comparison of Oberpfaffenhofen........................................................208

EuroSDR-Project

Commission 2 “Image analysis and information extraction”

“Automated Extraction, Refinement, and Update of Road

Databases from Imagery and Other Data”

Final Report

Report by Helmut Mayer1, Emmanuel Baltsavias

2, and Uwe Bacher

1

1Institute of Photogrammetry and Cartography – Bundeswehr University Munich 1Institute of Geodesy and Photogrammetry – ETH Zürich

219

Abstract

Roads are important objects for many applications of topographic data such as car navigation systems. They are often acquired manually and as this entails significant effort, automation is desirable. Deficits in automation, particularly linked to human interaction have led to the formation of this EuroSDR working group on the automated extraction, refinement, and update of road databases from imagery and other data. In the introductory part of this report we give a short account of the working group's motivation and the actual implementation in terms of actions. After showing why and how we limit the scope of the report we present an analysis of questionnaires to producers of road data as well as researchers and manufacturers. Based on it, the main focus has been the setup of a test comparing different approaches for automatic road extraction. After describing the data and the evaluation criteria used, we briefly present the approaches of a number of groups which have submitted results and give the outcome of the evaluation of these results. Finally, after thoroughly analyzing the evaluation results, we present conclusions and recommendations.

1 Motivation and Implementation

The original proposal for the project leading to the working group including an exhaustive account of background, motivation, and planned implementation can be found in Appendix 1. Here we give a short motivation and list the most important points reached in the actual implementation before finally limiting the scope of the report.

1.1 Motivation

The need for accurate, up-to-date and increasingly detailed information for roads has rapidly increased. They are needed for a variety of applications ranging from provision of basic topographic infrastructure, over transportation planning, traffic and fleet management, car navigation systems, location based services (LBS), tourism, to web-based applications. While road extraction has been sometimes performed by digitizing maps, road update and refinement can mainly be performed only from aerial imagery or high resolution satellite imagery such as Ikonos or Quickbird. Additionally, terrestrial methods, particularly mobile mapping are also of significant importance.

Because road extraction from imagery means large efforts in terms of time and money, automation of the extraction is of high potential interest. Full automation of the extraction of topographic objects is currently practically impossible for almost all applications and thus a combination with human interaction is necessary. An important factor hindering the practical use of automated procedures is the lack of reliable measures indicating the quality and accuracy of the results, making manual editing lengthy and cumbersome. Manufacturers of commercial systems have developed very few tools for semi-automated extraction and their cooperation with academia has been minimal. Thus, users and producers of such data, including national mapping and cadastral agencies (NMCAs) and large private photogrammetric firms, have been left with many wishes to be fulfilled.

NMCAs in Europe increasingly plan to update their data in shorter cycles. Their customers have increasing demands regarding the level of accuracy and object modeling detailedness, and often request additional attributes for the objects, e.g., the number of lanes for roads. The insufficient research output and the increasing user needs, necessitate appropriate actions. Practically oriented research, e.g., the ATOMI project at the ETH Zurich (Zhang 2004), has shown that an automation of road extraction and update is feasible to an extent that is practically very relevant. Companies that

220

have developed semi-automated tools for building extraction and other firms too, could very well offer similar tools for roads. Thus, the central ideas of the proposed working group are:

• To survey the data specification needs of important data producers, mainly NMCAs, and their customers.

• To thoroughly evaluate the current status of research (including models, strategies, methods and used data), to test and compare existing semi- or fully automated methods using various data sets and high quality reference data, to identify weak points and, finally, to propose strategies and methods that lead to a fast implementation of operational procedures for road extraction, update, and refinement. Here, refinement means the improvement of planimetric accuracy, but also the addition of new information such as height and road attributes. (Originally, we intended to focus on update and refinement. However, we got almost no feedback on these issues and thus decided to limit the scope to road extraction.)

1.2 Implementation

A EuroSDR working group (WG) chaired by Helmut Mayer, Bundeswehr University (UniBw) München, and Emmanuel Baltsavias, ETH Zürich, was formed in 2002. It has now significantly exceeded its proposed life-time of one and a half to two years. Actions so far comprise:

• Establishment of a web-page. The WG was officially announced and presented (kick-off) at the workshop "Incremental Updating", October 2002, Frankfurt, Germany.

• The distribution of the two extensive questionnaires to Researchers and Manufacturers as well as Data Producers was done at the end of the year 2002

• In spite of considerable advertising efforts the response to the questionnaire for Data Producers was as low as one reply. For the Researchers and Manufacturers the feedback was better with ten replies, though none from industry. Because of the weak response from the side of the producers and because the reply from the researchers pointed mostly to an interest in a comparison of the results of the individual approaches on the same data sets, we decided to postpone the reporting of the outcome of the questionnaires to this report. Instead, we focused on generating data sets consisting of images together with ground truth and the preparation of software making it possible to evaluate the individual results based on quantitative evaluation criteria.

• Preparation and distribution of test data sets including documentation took much longer than expected, particularly as we wanted to include Leica ADS 40 digital aerial image data, which was only delivered in the middle of 2004. Data has kindly been provided by Swiss Federal Office of Topography, Bern, Switzerland, Leica Geosystems, Heerbrugg, Switzerland, and the Bundeswehr Geoinformation Office (AGeoBw), Euskirchen, Germany. The data was distributed following the ISPRS Istanbul congress in July 2004.

• Some first results were submitted and evaluated by the end of 2004. By mid 2005 the situation was very disappointing, however, more results were received by the end of 2005 and the beginning of 2006.

We are finally in a position to give a meaningful account of the results. Unfortunately, due to the long duration of the WG many things including the scientific focus of the WG chair and the availability of personnel working on road extraction has changed. Thus, we are not in a position to give a broad

221

overview over the whole area. We rather limit the scope of this report to the analysis of the main outcome of the WG, namely the questionnaires received from the researchers and particularly the results and their evaluations for the test data.

2 Questionnaires

We developed two questionnaires, one for data producers and another for researchers and manufacturers. They were posted on the web and advertised on many occasions and with a larger number of personal Emails. Both questionnaires, including a summary of the answers received, can be found in full as Appendices 2 and 3.

2.1 Summary of responses to the Questionnaire for Producers of Road Data

The response to this questionnaire (cf. Appendix 2) consisted of one completed questionnaire even though we had advertised the questionnaire several times at EuroSDR meetings and via Email. We finally came to the conclusion that we requested too many answers, meaning too great an effort to complete the questionnaire, and also requested information which are seen as trade secrets. Additionally to the questionnaire we got some insights and opinions from insiders of the road data business. All this is summarized in the following, though the information was mainly obtained in 2003 and might be slightly out of date.

The one answer came from a state survey in Germany, which spent about fifty man years on road extraction over three years. Roads are obtained from GPS vans, digitizing from digital cadastral maps, as well as from governmental road construction agencies and municipalities. Roads are acquired from data which is a couple of months old and the average age of road data is just one year. Concerning accuracy, it was estimated to be in the 1 m range on average for the individual road segment, with the maximum error below 3 m. Yet, besides these good numbers one has to note that the data is cartographically generalized as the main application is topographic maps at a scale of 1:25000. Roads are acquired together with road class, number of lanes, and road width, all as defined in the (German) object class catalog ATKIS. For the latter, a detailed description of the topology for intersections is restricted to highways, but bridges, underpasses, and tunnels are modeled with some detail. As ATKIS defines a fully-fledged topographic information system, it also contains information about forest, water bodies, and settlements / buildings. The updating for roads takes approximately half an hour per square km and is done in an object oriented Intergraph environment. Right now they are switching to an update procedure based on information from governmental road construction agencies and local topographers. The group of customers comprises rescue services (fire brigades, police, ambulances – navigation systems), mapping (city plans), and engineering companies (urban planning, etc.). All location based services, e.g. insurance companies, parcel services, and geo-marketing are seen as potential customers.

Insiders of the road data business have told us that the requirements on road data are going to increase, to support assistance and control functions for cars. First, more detailed models can be used to control lighting or warn drivers if bends are too sharp. There is even the idea of obtaining 3D models for roads, e.g., from laser scanning, and use them to warn the driver or even brake the car if it is too fast. Information about individual lanes is also of interest for some applications. The requirements are under discussion and they are standardized in GDF (geographic data format) under the umbrella of official standardization bodies (CEN, ISO). Proposals range from cm accuracy (probably too high) up to the 5 m in use today.

222

GDF includes topology, road classes, crossings without lanes but with forbidden turns or one way streets, generalization levels, class codes for bridges and tunnels, and lots of other objects besides roads. The producers employ internal data bases and distribute via GDF. There is a quality code which is mostly not used. The geometry is modeled as polygons, i.e., as straight lines. There is a data field for height information, but it is not used. Whereas the width of the road, the number of lanes, the widths of the lanes, or the material of the road surface are usually not modeled, roundabouts and exits are.

For actual use in navigation systems the GDF data is usually generalized and stored in proprietary formats on four different levels, the lowest containing all roads and the highest only the highways. Generalization comprises geometry as well as attributes. There are often different data for visualization and for routing.

We were also informed that only about 20% of the effort goes into geometry. As, additionally, digitization of roads is nowadays done mostly in India and China, automation is not of first interest. Accuracy is in the range of 5 to 15 m which is enough in combination with map matching. Attributes mostly have to be acquired from ground level. It is done by regional bureaus supported by information from mail or parcel services. The acquisition of attributes, e.g., in the form of traffic signs, means lots of effort which has to pay off.

In summary, road data from governmental organizations could be of interest at least concerning its geometrical quality and information such as the road width and number of lanes for commercial data providers. As the geometry is only a part of the business, the question is whether the advantages are significant enough compared to the efforts needed for integration, let alone financial and legal issues.

2.2 Summary of responses to the Questionnaire for Researchers and Manufactures

Overall ten questionnaires were returned by May 2003. With the exception of one from the photogrammetry department of a mapping agency all answers came from researchers from universities. In the majority of the groups between two and four people are involved in road extraction. The average manpower available for road extraction over the last three years has been about three man years, with one group spending about eight man years. Thirty years / man years both came from the photogrammetry department of the mapping agency and it seems likely that these numbers actually describe production.

The completed questionnaires, a summary of which can be found in Appendix 3, show that many groups concentrate their work on the more simple kind of roads: Eight groups model narrow rural roads, six urban streets, and five dirt roads. Large roads or freeways with complex intersections or crossings are only dealt with by two groups each. With the exception of one, all groups can work with black and white images, though seven can also use color or multispectral images. Satellite images can be processed by half of the approaches, the answers at this time saying that there is a slight preference for panchromatic over multispectral. (Please note that for the actual test two years later – see Section 4 below – most approaches made heavy use of multispectral information.) Height information in the form of a Digital Surface Model (DSM) derived by Airborne Laser Scanning (ALS) is used by six groups and three also make use of Digital Terrain Models (DTM) derived from ALS. Comments point on pan-sharpened Ikonos data, that has actually been used in the test below, SAR images, as well as information about the date and time of image acquisition to allow for a prediction of shadows. A ground resolution of 0.6 m and 2 m is preferred by a majority of seven groups. Higher resolution images are used mainly in approaches originally developed for aerial imagery, with two groups focusing on urban streets going down to 10 cm and below. The coarse resolution from satellites of 10

223

m and above no longer plays an important role in road extraction, with newer, high resolution sensors available.

In terms of local road characteristics parallel roadsides / perceptual grouping and homogeneity in the direction of the road are both employed by a majority of seven groups. Scale-space behavior and the spectral characteristics of the road are used by three groups, whereas markings and cars on the road are only important for urban streets. (Again, the practical tests in Section 4 below showed that for Ikonos data most groups made use of the spectral information.) The comments show a vast variety of ideas for road extraction, comprising line extraction, texture analysis, homogeneous ribbons, machine learning, verification based on profiles, as well of grouping in DSM. The smoothness of road trajectories is modeled by eight of the groups by a variety of techniques consisting of splines, (zip-lock ribbon) snakes, Kalman-filters, piecewise parabolas, polynomial adjustment, Bezier curves, weighting by mean curvature, and weighting using fuzzy values for straightness. All groups except one employ the network characteristics of roads. Six do this by bridging local gaps and five by global optimization of the network. Techniques used comprise graph data structures, global path search, and local context. Gaps are closed according to length and collinearity, but also checking homogeneity, the existence of predicted shadows or by finding cars.

The width of the road is the attribute computed by almost all of the groups. Four groups determine the road class and three groups the number of lanes. Other attributes calculated by one or two groups are the width of lanes, road markings, surface material, height profiles, and internal quality measures. Eight groups model simple crossings while complex crossings and roundabouts are only dealt with by two groups. Highway exits are modeled by one group and highway intersections by none. Seven groups make use of contextual information, most of them by different modeling in urban, rural, and forest areas. Shadows and occlusions are also taken into account in four approaches. The comments show that for detailed contextual modeling information from DSM, possibly from LIDAR, is used to obtain hypotheses for buildings, i.e., high objects where there cannot be a road, or cars on the road as well as for predicting / explaining shadows and occlusions.

Geographic information system (GIS) data is not used at all by seven groups. However, two groups use it to refine the search space and one even to actively verify roads. For the latter, attributes and also the context of roads are employed to find evidence for roads, possibly adapting parameters or selecting specific algorithms - no evidence found meaning rejection. GIS data is also used for supervised learning and might be useful to guide an operator. On the other hand, three groups note that the extraction should be done independently of the GIS data to avoid bias. 3D information is employed by six groups, mostly in the form of a DSM generated by image matching or ALS. It is used mainly to verify that roads are smooth in 3D and have a limited steepness, but also to determine shadow regions, to segment urban and rural areas, as well as to fuse results from different aerial images. Three groups carry out precision estimation for the results and five label the results with a reliability flag, only three assuming that it is actually reliable. The evaluation is done by many groups according to Wiedemann et al. (1998) in comparison to ground truth data. Yet, there is also internal evaluation based on the network characteristics and one particular approach uses independent features evaluated by fuzzy values. For learning based approaches results are evaluated at each stage statistically. This information is then used to tune the algorithms.

Object oriented programming often with C++ is the basis of almost all groups. This is complemented for three groups by modeling uncertainty using the Theory of Evidence and fuzzy logic. Two groups use explicit reasoning and one automatic learning. Most approaches run under Windows, many under LINUX, and three groups use UNIX on different hardware architectures.

Most groups work on automatic solutions, only three groups focus additionally or exclusively on semiautomatic approaches. Therefore, the answers given for this issue have to be taken more

224

cautiously than the rest. Two groups have developed an approach based on tracking, with one giving the points sequentially, guiding the tracking. Only two groups have yet devised an approach for editing automatically computed results, both restricting the automated processing by given regions and attributes defined by an operator or given by a GIS. One group labels the results with a (reliable) reliability flag.

The general characteristics of the approaches for which the corresponding questions have been answered are given in Appendix 3. They comprise key features, strong points, weaknesses or limitations, how the approach could be optimally exploited in a practical application, and most important features for an ideal system for practical, i.e., semi-automated, road extraction. For the latter, most groups propose a combination of reliable results of automatic road extraction with a user-friendly interface. If automation is used during the interactive phase, its speed has to be high enough.

Concerning data for the envisaged test, there was a slight preference for original image data over orthoimages. One comment said that mostly orthoimages show strong distortions in urban areas. Two thirds of the responding researchers prefer color or multispectral imagery, one commenting that infrared data would be helpful. GIS data is needed, e.g. for verification, or important for half of the groups. Seven of the groups wish that orthoimages be generated, two from DTM and four from DSM. Only two groups find ALS data important for the test. There are comments saying that DSM is important for urban areas, but could also come from a source other than ALS. The vector formats most groups can handle are ARC/INFO coverage or Microstation DGN files. Besides this there is a big variety of formats employed. Seven groups are interested in digital aerial line scanner data, e.g., from ADS40, but most of them comment that it is interesting for tests but of low priority. Finally, people commented that they find orthoimages in RGB, DSM, as well as lots of labeled images for evaluation purposes important.

All researches see road centerlines as the correct basic information for deriving statistics for road extraction. As parameters for an evaluation besides completeness, reliability and accuracy (RMS), correctness and the number of connected lanes for crossings as well as the number of lanes for roads are proposed. Half of the groups see an additional qualitative visual inspection as meaningful, yet commenting that this can just give a rough idea and that the eye tends to group together lines closing existing gaps. For semi automatic approaches the proposed idea was to define a quality level which is checked automatically and then to measure interaction time for comparison. This was accepted as a good idea by six groups, one noting that not only the time, but also the number of breaks for the operator is of importance. Finally, it is noted that for supervised methods it might be interesting to work with the same training data to obtain a better defined comparison.

In summary, the low number of groups having devised a semi automatic approach shows the focus on research and not on practice of the groups that sent in answers for the questionnaires. From our experience one of the key problems is that the interaction for a semi-automatic approach needs to be highly optimized for the current state of the approach at hand which implies considerable effort which is only meaningful if it is used in production. As there is still too much to be improved for the automatic approaches and production is shifted towards countries with low wages, there is a tendency to focus on the improvement of the automatic approaches where more is seen to be gained for the moment.

Therefore, we decided to focus the envisaged test on automatic approaches. To still get a practical flavor into the test, images of a size too small for practice, but challenging for the approaches, were used. According to the answer we got we devised a test data set with aerial and high resolution satellite data, all in full color, the digitally acquired data also including infrared. The data and the evaluation criteria used are detailed next.

225

3 Test Data and Evaluation Criteria

Initially, eight test images were prepared from different sensors (for details see the README file in Appendix 5):

• 3 scanned aerial images from the Federal Office of Topography, Bern, Switzerland • Aerial1: suburban area in hilly terrain• Aerial2: rural scene with medium complexity in hilly terrain • Aerial3: rural scene with low complexity in hilly terrain

• 2 Leica ADS40 images from Leica Geosystems, Heerbrugg, Switzerland • ADS40_1 and _2: rural area with medium complexity in flat terrain.

• 3 IKONOS images from Kosovo, provided by Bundeswehr Geoinformation Office (AGeoBw), Euskirchen, Germany, given as pan sharpened images in red, green, blue, and infrared

• Ikonos1: urban/suburban area in hilly terrain • Ikonos2: rural scene with medium complexity in hilly terrain • Ikonos3: rural scene with medium complexity in hilly terrain

All images have a size of at least 4,000 by 4,000 pixels. While we initially thought that this size should be a challenge that people should be able to overcome during the duration of the test, it was finally found to be insurmountable by nearly all approaches even two years after the envisaged end of the test and, therefore, the limiting factor of the test. We, therefore, decided eventually to cut out three smaller parts with 1,600 by 1,600 pixels of the Ikonos images, one from Ikonos1 and two from Ikonos3, as there seemed to be most interest in this area.

For evaluation we use criteria put forward by Wiedemann et al. (1998). The basic assumption is that reference data is available in the form of the center lines of the roads. Additionally, it is assumed that only roads within a buffer of a certain width, usually the average width of the roads, around the road are correct. The extracted roads which are inside the buffer of the given reference roads and vice versa

are determined via matching of the respective vector data. The most important criteria defined by Wiedemann et al. (1998) based on these matching results to which we have restricted the analysis are:

Completeness: This is the percentage of the reference data which is explained by the extracted data, i.e., the part (percentage) of the reference network which lies within the buffer around the extracted data. The optimum value for the completeness is 1.

Correctness: It represents the percentage of correctly extracted road data, i.e., the percentage of the extracted data which lie within the buffer around the reference network. The optimum value for the correctness is 1.

RMS (root mean square): The RMS error expresses the geometrical accuracy of the extracted road data around the reference network. In the given evaluation framework its value depends on the buffer width. If an equal distribution of the extracted road data within the buffer around the reference network is assumed, it can be shown that RMS equals one divided by square root 3 of the buffer width. The optimum value for RMS is 0. As RMS mainly depends on the resolution of the image, it is given in pixels in this report.

The reference data has an estimated precision of half a pixel. It comprises major and secondary roads, but no paths or short driveways. The reference data has not been made available to the participants. The participants usually asked only once or twice for an evaluation, i.e., no optimization in terms of the reference data was pursued. Opposed to (Scharstein and Szeliski, 2002) we allowed people to

226

optimize their parameters for each and every image, as constant parameters were seen as too challenging.

4 Approaches and Evaluation of Results

We have finally obtained for every data set at least one result. Results that could be evaluated have been sent in by six groups. We will shortly introduce the groups and their approaches (alphabetical ordering according to corresponding author given in bold). Details for some of the approaches can be found in Appendices 6 to 8. Selected results and their analysis are presented in (Mayer et al. 2006).

The approaches of Bacher, Gerke, and Hedman are all three based on the work of TU München of Wiedemann and Hinz (1999) and partially Baumgartner et al. (1999a, 1999b). Hedman mostly optimizes the line extraction of the former for the data at hand by employing the blue channel as well as the NDVI. Opposed to this, Bacher and Gerke also make use of the work of Baumgartner et al. (1999a, 1999b). While Bacher has devised an approach automatically generating training data for a multispectral classification and employs snakes on the classification result to bridge gaps, Gerke uses the original approach of Wiedemann and Hinz (1999) with customized parameter settings (Gerke_W) as well as the original approach of Baumgartner et al. (1999a, 1999b) with the improved basic line detection of Wiedemann and Hinz (1999 – Gerke_WB).

• Uwe Bacher, Institute for Photogrammetry and Cartography, Bundeswehr University Munich, Germany: The approach is only suitable for the Ikonos images. It is based on earlier work from TU München of Wiedemann and Hinz (1999) and partially Baumgartner et al. (1999a, 1999b). The approach of Wiedemann and Hinz (1999) starts with line extraction in all spectral bands using the sub-pixel precise Steger line extractor (Steger 1998) based on differential geometry and scale-space including a thorough analysis and linking of the topology at intersections. The lines are smoothed and split at high-curvature points. The resulting line segments are evaluated according to their width, length, curvature, etc. Lines from different channels or extracted with different scales, i.e., line widths, are then fused on a best first basis. From the remaining lines a graph is constructed, supplemented by hypotheses bridging gaps. After defining seed lines in the form of the most highly evaluated lines, optimal paths are computed in the graph and from it gaps to be closed are derived. Bacher has extended this by several means (Bacher and Mayer 2004, 2005). The central idea is to take into account the spectral information by means of a (fuzzy) classification approach based on fully automatically created training areas. For the latter parallel edges are extracted in the spirit of (Baumgartner et al. 1999a, 1999b) in a buffer around the lines and checked if the area in-between them is homogeneous. The information from the classification approach is used to evaluate the lines. Additionally, it is the image information when optimizing snakes to obtain a more geometrically precise, but also more reliable basis for bridging larger gaps in the network, which is another new feature of Bacher's approach.

• Charles Beumier and Vinciane Lacroix, Signal and Image Center, Royal Military Academy, Brussels, Belgium (also see Appendix 6): Their approach for Ikonos images rests on the gradient line detector of Lacroix and Acheroy (1998) which assumes that the gradient vectors on both sides of a line are pointing in opposite directions. Bright lines are extracted from the green channel with a slight Gaussian smoothing employing non-maximum suppression. Lines are tracked limiting the direction difference until a minimum strength is reached. Lines are only kept if they are at least 30 pixels long and are straight enough when checked via the square root of the inertial moment. For each of the line points the Normalized Difference Vegetation Index (NDVI) is computed from the red and the infrared channel and if it is below zero, the point is supposed to be vegetation and is rejected. Finally, the rest of the points are again tracked and checked to see if they are still long

227

and straight enough.

• Markus Gerke, Institute for Photogrammetry and Geoinformation (IPI), Hannover University, Germany: Gerke uses two approaches, suitable for the aerial images as well as for the Ikonos data. Gerke_W is the approach of Wiedemann and Hinz (1999) – see Bacher above. Gerke_WB consists of a combination of Gerke_W with the approach of Baumgartner et al. (1999a, 1999b). The latter is based on extracting parallel edges with an area homogeneous in the direction of the road in between in the original high resolution image and fusing this information with lines extracted at a lower resolution, thereby combining the high reliability of high resolution with the robustness against disturbances, particularly for the topology, of the lower resolution. Baumgartner et al. (1999a, 1999b) then construct quadrangles and from them longer road objects also taking into account local context information. Gerke_WB in essence substitutes the Steger line extractor of the original Baumgartner et al. (1999a, 1999b) approach by the fully-fledged Wiedemann and Hinz (1999) approach and additionally puts less weight on the homogeneity in the direction of the road. Gerke notes that there is still room for improvement as he has not at all optimized the snakes used to bridge gaps.

• Jose Malpica, Subdirección de Geodesia y Cartografía, Escuela Politécnica, Campus Universitario, Alcalá de Henares, Spain: The approach of Malpica and Mena (2003, 2005) makes heavy use of the spectral and color characteristics of roads learned from training data. The latter is usually generated based on (possibly outdated) GIS data from the given image data. The basic image analysis for road extraction is done on three statistical levels. In the first level, only color information is employed using Mahalonobis distance. On the so-called “one and a half order” statistical level, the color distribution is determined for a pixel and its 5 x 5 neighborhood and compared to the learned distribution via Bhattacharyya distance. Bhattacharyya distance is finally also used on the “second order” statistical level where, for six different cross-sections of a 3 x 3 neighborhood of a pixel, co-occurrence matrices and from them 24 Haralick features are computed. The three statistical levels are normalized and combined employing the Dempster-Shafer Theory of Evidence. After thresholding and cleaning, the derived plausibility image for roads is the basis for deriving the main axes of the roads. A standard skeleton showing all, including the usual unwanted details, is combined with a coarse skeleton to obtain a graph with precise road segments without too many wrong short road segments. The segments in the graph are finally subject to a geometrical as well as topological adjustment.

• Karin Hedman and Stefan Hinz, Chair for Photogrammetry and Remote Sensing, Technische Universität München, Germany (see also Appendix 7): Like Bacher's and Gerke's approaches this again rests on (Wiedemann and Hinz 1999). It has only been tested for the smaller pieces cut from the Ikonos images. As Hedman and Hinz found that the line extraction is the critical point, they have optimized it: First, they noted that the blue channel gives the best results, with the NDVI adding little, but complimentary, information particularly for the rural areas. For Ikonos3_sub1 they found that it was advantageous to use two different scales for line extraction in the blue channel.

• Qiaoping Zhang and Isabelle Couloigner, Department of Geomatics Engineering, University of Calgary, Canada (see also Appendix 8): The approach is used with minor modifications for all test images (they produced results for all images besides the larger Ikonos images). For the more high resolution test data the images were rescaled by a factor of two (Aerial) or four (ADS40). At the core of the approach of Zhang and Couloigner is the K-means clustering algorithm with the number of classes set to an empirically found value of six. For most of the images three channels were used. The infrared channel was only employed for Ikonos1_sub1; for the other two Ikonos images it was regarded as too noisy. From one or more clusters the road cluster is constructed by a fuzzy logic classifier with predefined membership functions (Zhang and Couloigner 2006a). The

228

road cluster is refined by removing big open areas, i.e., buildings, parking lots, fields, etc., again by means of a fuzzy classification based on a shape descriptor using the Angular Texture Signature (Zhang and Couloigner 2005a, 2006b). Road segments are found from the refined clusters via a localized and iterative Radon transform with window size 31 by 31 pixels with improved peak selection for thick lines (Zhang and Couloigner 2005b, 2005c). The segments are grouped bridging gaps smaller than five pixels and forming intersections. Finally, only segments longer than twenty pixels are retained.

The results of the evaluation are summarized in Table 1. It is ordered in the first instance according to the test areas (from analog to digital aerial to satellite data) and in the second instance alphabetically according to the group and possibly its approaches. For each test area where more than one result is available, the best result in terms of the geometric mean of completeness and correctness is marked in bold. In addition, all values for completeness or correctness, which are beyond a value of 0.6 or 0.75 respectively, are marked in bold. These numbers can be seen as a lowest needed limit so that the results become practically useful. The value for correctness was set to a higher value as experience shows that it is much harder to manually improve given faulty results than to acquire roads from scratch. To be of real practical importance, in many cases both values probably need to be even higher, e.g., for correctness around 0.85 and for completeness around 0.7, but we have chosen the lower values, to distinguish ‘the probably useful’ for the obtained results from the rest.

In Figures 1 to 10 we give one result for every image, besides the details from Ikonos3. Only for the latter have we obtained a larger number of results in the range beyond 0.85 for correctness and 0.7 for completeness deemed suitable for practical applications. Also because of that, we show some of the results for Ikonos3 and its details in Figures 5, 7, and 10 using a larger size. As the result for Beumier for Ikonos1_Sub1 is very interesting and the result of Gerke_WB for Aerial3 is definitely above 0.6 and 0.75 for completeness and correctness, we also present them larger in size in Figures 6 and 2.

Figure 1: Aerial1 – Zhang (left), Aerial2 – Gerke_WB (right); correctly extracted roads are

given in green, incorrectly extracted roads in blue, and red marks the part of the ground truth

for roads which could not be extracted.

229

Name (best) Test area Completeness ( 0.6) Correctness ( 0.75) RMS [pixels]

Gerke_W Aerial1 0.46 0.47 3.74

Gerke_WB 0.31 0.56 1.53

Zhang (Fig. 1) 0.51 0.49 1.92

Gerke_W Aerial2 0.76 0.66 2.87

Gerke_WB (Fig. 1) 0.65 0.82 1.14

Zhang 0.67 0.49 1.72

Gerke_W Aerial3 0.81 0.63 3.14

Gerke_WB (Fig. 2) 0.72 0.77 1.3

Zhang 0.72 0.63 1.66

Zhang (Fig. 3) ADS40_1 0.56 0.48 2.80

Zhang (Fig. 3) ADS40_2 0.45 0.3 2.58

Gerke_W (Fig. 4) Ikonos1 0.49 0.36 1.83

Gerke_W (Fig. 4) Ikonos2 0.59 0.10 1.95

Bacher (Fig. 5) Ikonos3 0.55 0.35 1.56

Gerke_W 0.62 0.16 1.57

Bacher Ikonos1_Sub1 0.34 0.66 1.29

Beumier (Fig. 6) 0.48 0.69 1.3

Gerke_W 0.27 0.41 1.89

Gerke_WB 0.19 0.49 1.91

Hedman 0,31 0,51 1,25

Malpica 0.25 0.74 1.13

Zhang 0.56 0.41 1.52

Bacher (Fig. 7) Ikonos3_Sub1 0.81 0.87 0.97

Gerke_W 0.8 0.65 1.53

Gerke_WB 0.68 0.75 1.99

Hedman (Fig. 8) 0.77 0.78 1.16

Malpica (Fig. 8) 0.6 0.79 1.41

Zhang 0.72 0.35 1.22

Bacher (Fig. 9) Ikonos3_Sub2 0.86 0.89 1.

Gerke_W 0.75 0.52 1.35

Gerke_WB (Fig. 9) 0.71 0.84 1.7

Hedman (Fig. 10) 0.85 0.91 1.19

Malpica 0.6 0.89 1.59

Zhang 0.7 0.34 1.18

Table 1: Results of the evaluation. Bold names represent the best result in terms of the

geometric mean of completeness and correctness for a test area where more than one result is

available. Bold numbers are beyond 0.6 or 0.75 for completeness or correctness, respectively.

230

5 Analysis of Results

We focus the analysis on the details from the Ikonos images, as it is only for these smaller images that we have received a larger number of results. We comment on the images, discuss the individual approaches, and after giving important overall findings we end up in a short summary.

For the different images, we observed the following:

• Aerial1-3: All three images have only been processed by Gerke and Zhang. The latter performs best for Aerial1 (Figure 1), which is the most difficult of the three images showing a suburban area. It seems that for it the loss of information by down sampling by a factor of two by Zhang is more than made up by employing color information via classification, a feature Gerke is lacking. Gerke_WB gives the best results in terms of completeness and correctness for images 2 (Figure 1) and 3 (Figure 2) showing rural areas for which it was designed. Particularly the result for Aerial3 shown in Figure 3 is on a level which could be a viable basis for a practical application. Finally, comparing the results for Gerke_W and Gerke_WB one can see nicely how for Gerke_WB completeness is still sacrificed for correctness even when introducing additional information in the form of the homogeneity in the road direction for the original high resolution imagery.

Figure 2: Aerial3 – Gerke_WB; colors cf. Fig. 1

231

• ADS40_1 and 2: Results for both images (Figure 3) have only been produced by Zhang after down sampling them by a factor of four, thus not making full use of the available information. The results for both are not practically relevant, but one can observe that the performance for the easier rural image 1 is better than for the suburban image 2 where less than one third of the roads are correct and less than half of the roads have been found.

Figure 3: ADS40_1 – Zhang (left), ADS40_2 – Zhang (right); colors cf. Fig. 1

• Ikonos1, 2, and 3: Gerke_W was the only approach for which we received results for all three Ikonos images. Bacher who, in this case, basically uses the same approach as Gerke_W, but with different parameter settings, only processed Ikonos3. The results for Ikonos 1 (Figure 4) and 3 have been approximately as expected. Completeness and correctness for the suburban, i.e., more difficult Ikonos1 are lower than for the rural Ikonos3, with Bacher (Figure 5) achieving a more complete, but also more incorrect result. Ikonos 2 has proved more difficult than expected as demonstrated by the results (Figure 4). It is a rural scene with medium complexity, but a combination of different complications such as trees alongside the roads and small and elongated fields together with the low quality of the image possibly due to haze in the atmosphere resulted in a very high number of false hypotheses for roads, bringing down the correctness to a mere 10%.

Reasons for an inability to process the larger Ikonos scenes were apparently twofold. First, because of missing functionality for processing the whole image in patches which are then combined into one solution, intermediate results just exceeded the available memory. Second, even if this had not been the case, the time it takes to process the image together with the need to adapt the parameters to all variations in the larger scenes, meant these images required too much effort for most people. One gets a hint for the latter when comparing the results for the whole scenes Ikonos1 and 3 with the much better results for the relatively representative sub-areas, described below.

232

Figure 4: Ikonos1 – Gerke_W (left), Ikonos2 – Gerke_W (right); colors cf. Fig. 1

Figure 5: Ikonos3 – Bacher; colors cf. Fig. 1

233

• Ikonos1_Sub1: This image shows an urban/suburban scene and has been processed by six approaches, none giving a practically useful result. It seems to be too hard a challenge for the current approaches, although one has to notice that the results are notably better than for the whole scene, i.e., Ikonos1, from which it is a representative part. Beumier has only submitted this one result, but it is the best for this scene (Figure 6). It shows that a good line extractor combined with spectral information (NDVI) and well chosen constraints on the geometry can produce a pretty good result. Multiplying completeness with correctness gives a similarly high value for the approaches of Bacher, Malpica, and Zhang. Looking at the individual results, however, one can very well see the different trade-offs one can make particularly for difficult images between completeness and correctness. Yet for practical applications, the high correctness values for Bacher and Malpica would probably be preferred.

Figure 6: Ikonos1_Sub1 – Beumier; colors cf. Fig. 1

234

• Ikonos3_Sub1 and Sub2: These two images depicting rural scenes of medium complexity are the only ones for which a larger number of approaches, namely six, was applied and for which at least some of the results are in a range, which could be suitable for a practical application. This is particularly true for the approaches of Hedman and Bacher. Both rest on the approach of Wiedemann and Hinz (1999) and both make use of color information, Bacher in a more sophisticated way than Hedman. While Bacher has a clear edge for Sub1 (Figure 7), Hedman is slightly better on Sub2 (Figure 10 – please find the other images in Figures 8 and 9). Gerke's approaches also rest on (Wiedemann and Hinz 1999), but are less sophisticated in the way they make use of the color information, which seems to be a clear disadvantage here. While for Sub1 Gerke_W and WB perform very similarly, Gerke_WB taking into account the homogeneity of the road in the original resolution is markedly better on Sub2 (Figure 9). For the approaches based on color and texture Malpica achieves a higher quality particularly in terms of correctness than Zhang. This is especially true for Sub1 (Figure 8). Finally, a comparison of the results for Bacher and Malpica in Figures 7 and 8 shows the benefits of global network optimization inherent in all approaches based on (Wiedemann and Hinz 1999) together with snakes for bridging gaps. It is clearly visible that in Bacher's result many smaller gaps are bridged meaningfully.

We next comment on distinct characteristics of the individual approaches, if they have not been discussed already with the images.

• Bacher, Gerke_W and _WB, and Hedman: All three follow Wiedemann and Hinz (1999), the difference being which additional information is used. Bacher, with classification based on automatically generated training data is the most sophisticated and achieves the best results, but also Hedman's suitable selection of channels and scales as well as the use of the NDVI is sufficient to outperform Gerke_W and _WB, which do not make explicit use of color.

• Gerke_W versus WB: Gerke_WB can be seen as an extension of Gerke_W, taking into account higher resolution information in the form of parallel edges enclosing a region homogeneous in the direction of the road. Gerke_WB enforces more detailed constraints and, thus, as expected, the results for it show a lower completeness, but a higher correctness. This is true in all cases, but Gerke_WB seems to be particularly well suited for open rural scenes, where the roads mostly match its model of homogeneous areas.

• Malpica and Zhang: Both employ a classification approach using color or multispectral information, though in a different way. While Malpica also includes textures and learns the characteristics from given GIS data, Zhang uses an unsupervised classification. Malpica outperforms Zhang for the Ikonos data, but Zhang is more flexible with the unsupervised classification producing results for most images.

235

Figure 7: Ikonos3_Sub1 – Bacher; colors cf. Fig. 1

Figure 8: Ikonos3_Sub1 – Hedman (left), Ikonos3_Sub1 – Malpica (right); colors cf. Fig. 1

236

Figure 9: Ikonos3_Sub2 – Bacher (left), Ikonos3_Sub2 – Gerke_WB (right); colors cf. Fig. 1

Figure 10: Ikonos3_Sub2 – Hedman; colors cf. Fig. 1

237

Finally, we want to note some important overall findings:

• The approaches based on line extraction, i.e., Bacher, Gerke_W, Gerke_WB and Hedman based on the Steger extractor as well as Beumier built on top of Lacroix's work give better results for the more line-like high resolution Ikonos data than approaches based on pixel-wise or local classification, i.e., Zhang and Malpica. It would be interesting to see how the latter perform on higher resolution aerial images where the line structure of the image is less marked and the spectral information should be of higher quality. Please note that the Ikonos data are pan-sharpened with a physical resolution for the color of only 4 m. One issue still to be investigated is the use of the original low resolution images.

• There is a trend to use color / multispectral information particularly for high resolution satellite data. This is done either as simple as the NDVI (Beumier and Hedman) or based on a more or less sophisticated classification (Bacher, Malpica, and Zhang). For the latter, training is done using given GIS data (Malpica), training areas are automatically generated from characteristic homogeneous road parts with parallel road sides (Bacher), or the classification is done unsupervised (Zhang).

• Network optimization and bridging gaps, e.g., by means of snakes, only seems to become important when a certain level of quality has been reached as, for example, by Bacher for Ikonos_Sub1 and 2.

In summary, the results show that it is possible to extract roads with a quality in terms of completeness and correctness which should be useful for practical applications, although only for scenes with limited complexity, namely up to medium complex rural scenes. This is true for aerial as well as high resolution satellite data. The test has also demonstrated that most approaches cannot deal with images larger than about 2,000 by 2,000 pixels at the moment. This is probably due to missing functionality to process images in patches and shows that the approaches focus on furthering the understanding of the basic problems rather than on practical development, where robustness to all possible situations would be the central issue.

With the advent of digital aerial cameras and high resolution satellite data making high quality color and spectral information available, there is a recent focus in research on road extraction to employ this information and the results show its usefulness. However, particularly for the high resolution data such as from the ADS40, there is still much to be done, e.g., by making use of the link to the typical materials used to construct roads or by employing line-finders that make full use of color information.

Additional remarks: One of the reviewers of this report suggested a cross-analysis of the data, i.e., results of the extraction. The idea would be to analyze in detail where the approaches produce the same or similar results and where and how they differ, to find complementary strengths. We had this in mind when starting the test, but finally decided not to analyze the data on this level. The main reason for this is that from what we know the different groups put very different effort into generating the data. Thus, while the data can be used for an overview, we think that analyzing it in depth would lead to an over-interpretation. To still give the possibility for a more in-depth analysis, we have published all data visually on our web-page.

6 Conclusions and Recommendations

The original proposal of the working group was much more focused on practical aspects. We not only wanted to make available a priori data to be used in the test, but also wanted to test semi-automated

238

approaches. Unfortunately just obtaining the data actually used in the test proved a major issue taking more than two years, tens of Emails and phone calls, as well as lots of efforts by people at ETH Zurich, University of the Bundeswehr, and Leica Geosystems. Concerning a priori data we found already in the questionnaire that it is not of major interest for the mostly research oriented approaches. Additionally, opposed to image data which anybody can handle, a priori data is mostly vector data which comes in many formats with none being easily readable by everybody. Because of these two reasons we decided not to make available a priori data, although it exists for some of the datasets.

Similar concerns are valid for semi-automated approaches or automated approaches including pre- and post-processing. First, the interest from the participants was not high. Then, it is not clear what to test. From a practical perspective one wants to achieve a certain quality – which one might be able to define in terms of completeness, correctness, and RMS – and then the main issue is to acquire road data as fast and particularly cheap as possible. For this case one would need to do the test in production environments with experienced operators to obtain a reliable estimate of the ultimate performance. These environments do not exist from what we know. Additionally, they would be often specialized to certain types of landscapes existing in the respective country and thus it might be difficult to transfer results and conclusions. Because all of this we did not test semi-automated approaches.

Both, the use of a priori data as well as for practical semi-automated or automated approaches with pre- and post-processing, will become important (possibly even in the near future), but to define tests for them which give results that are mostly unbiased and give a clear message to the user we still see a couple of issues on the research side.

We planned to publish the findings in conferences or in a theme issue of an international journal. Due to the changed situation of the authors focusing much more on close range and computer vision (Helmut Mayer) or even leaving the University (Uwe Bacher), we are glad having at least managed to write a paper summarizing the main results of the test (Mayer et al. 2006).

The test has shown once again that the focus of the research is mostly on theoretic issues, often neglecting practical issues. We, therefore, had originally envisaged that the above activities and international cooperation would help to create a nucleus of interested researchers who, with the cooperation of NMCAs and, if possible, manufacturers, could form a well-coordinated and focused research network with milestones, commitments, and – most important – the necessary resources which can speed up the development of operational (or quasi-operational) systems for road extraction. We at Bundeswehr University Munich hope that other people will take up this idea and make it finally work. We do not have the resources to do so at the moment.

We finally learned the very hard way that it is extremely important to see as very long term what emerged as the core of our work, namely the evaluation of the results for different approaches based on the same data. The initial setup of the working group was very naive in this respect. Experience for similar tests, such as the highly successful 3D reconstruction test of Scharstein and Szeliski (2002) only gained momentum after some time. The goal has to be that, after a while, papers proposing a new approach only get accepted for higher level conferences when they show comparable or improved results on the test data compared to the state of the art.

It is, thus, very important to continue the test, even though we do not have the resources any more. We, therefore, hope that we can transfer our test. We have started discussing this issue with ISPRS working group III/5 “Road Extraction and Traffic Monitoring” chaired by Uwe Stilla, TU München, and Chunsun Zhang, South Dakota State University, USA. As the recent situation where Uwe Bacher runs the scripts manually and sends out the results via Email by hand is neither good in terms of work

239

load nor response time for anybody interested in the evaluation results, the idea is to set up a web service to which one sends the results and gets back the evaluation.

Finally, we see promising directions for future research such as appearance based approaches, e.g, (Leibe and Schiele 2004), which might not be so easily adapted for roads themselves, but which have shown good results for cars, defining roads particularly in urban areas. Even more interesting might be statistical generative modeling. A particularly impressive example is (Dick et al. 2004) for buildings, but Stoica et al. (2004) have employed it for roads. To our knowledge this is the first time that the natural variability of the road network has been modeled in a realistic way. This is of big importance as the modeling of the network topology including the detailed characteristics of crossings is an area where not much research has been conducted.

Additional remarks: One of the reviewers of the report sees a trend for the NMCAs in the acquisition of very high resolution aerial imagery in the decimeter range. From it, highly detailed information including separate lanes and road marks could be acquired with very short update cycles. For this, a high degree of automation would be the only way to proceed. We agree with this, but note that we are not aware of any research on road extraction focused on this type of imagery at the moment.

A reviewer of (Mayer et al. 2006) pointed on several issues not discussed so far which should be taken into account for a follow up of the test. One issue is certainly how much parameters are allowed to be tuned. Particularly for smaller image sizes optimal sets of parameters might be more important than the approach applied. On the other hand, the images are fairly different and a single set of parameters might lead to very unfavorable results. An issues closely linked to this is if the ground truth is given out to the participants, possibly together with the code used for evaluation. This can on the one hand be used by the participants to find out about good parameter settings but also strategies by running multiple trials. On the other hand it can lead to approaches that just perform well for the given couple of images. Depending on the state of the field both issues have their pros and cons. Concerning the ranking it was proposed to have one official criterion (the geometric mean of completeness and correctness is one, although it needs to be discussed further) for overall ranking, but to also give the user the possibility to see the ranking according to individual criteria, such as completeness.

7 Acknowledgments

We want to thank Stefan Hinz as well as the EuroSDR reviewers Nicolas Paparoditis and Kevin Mooney for their helpful comments and for improving the English of this report. We are grateful to the reviewers of (Mayer et al. 2006) for valuable improvements of a shortened version of this report. We finally would like to thank the Swiss Federal Office of Topography, Bern, Switzerland, Leica Geosystems, Heerbrugg, Switzerland, and the Bundeswehr Geoinformation Office (AGeoBw), Euskirchen, Germany, for making available the data for this test.

References

Bacher, U. and Mayer, H. (2004): Automatic Road Extraction from IRS Satellite Images in Agricultural and Desert Areas, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences (35) B3, 1055-1060.

240

Bacher, U. and Mayer, H. (2005): Automatic Road Extraction from Multispectral High Resolution Satellite Images, The International Archives of the Photogrammetry, Remote Sensing and Spatial

Information Sciences (36) 3/W24, 29-34.

Baumgartner, A., Eckstein, W., Heipke, C., Hinz, S., Mayer, H., Radig, B., Steger, C. and Wiedemann, C. (1999a): T-REX: TUM Research on Road Extraction, Festschrift fur Prof. Dr.-Ing. Heinrich Ebner zum 60. Geburtstag, Lehrstuhl für Photogrammetrie und Fernerkundung der Techni-

schen Universität München, 43-64.

Baumgartner, A., Steger, C., Mayer, H., Eckstein, W., and Ebner, H. (1999b): Automatic Road Extraction Based on Multi-Scale, Grouping, and Context, Photogrammetric Engineering & Remote Sensing (65) 7: 777-785.

Dick, A., Torr, P., and Cipolla, R. (2004): Modelling and Interpretation of Architecture from Several Images, International Journal of Computer Vision (60) 2: 111-134.

Lacroix, V. and Acheroy, M. (1998): Feature-extraction Using the Constrained Gradient, ISPRS

Journal of Photogrammetry and Remote Sensing (53), 85-94.

Leibe, B. and Schiele, B. (2004): Combined Object Categorization and Segmentation with an Implicit Shape Model, ECCV’04 Workshop on Statistical Learning in Computer Vision, 1-15.

Mayer, H., Hinz, S., Bacher, U., and Baltsavias, E. (2006): A Test of Automatic Road Extraction Approaches. The International Archives of the Photogrammetry, Remote Sensing and Spatial

Information Sciences (34) B3.

Mena, J.B and Malpica, J.A. (2003): Color Image Segmentation Using the Dempster-Shafer Theory of Evidence for the Fusion of Texture, The International Archives of the Photogrammetry, Remote

Sensing and Spatial Information Sciences (34) 3/W8, 139-144.

Mena, J.B and Malpica, J.A. (2005): An Automatic Method for Road Extraction in Rural and Semi-Urban Areas Starting from High Resolution Satellite Imagery, Pattern Recognition Letters (26), 1201-1220.

Scharstein, D. and Szeliski, R. (2002): A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms, International Journal of Computer Vision (47) 1: 7-42.

Steger, C. (1998): An Unbiased Extractor of Curvilinear Structures, IEEE Transactions on Pattern Analysis and Machine Intelligence (20): 113-125.

Stoica, R., Descombes, X., and Zerubia, J. (2004): A Gibbs Point Process for Road Extraction from Remotely Sensed Images, International Journal of Computer Vision (57) 2: 121-136.

Wiedemann, C., Heipke, C., Mayer, H., and Jamet, O. (1998): Empirical Evaluation of Automatically Extracted Road Axes, CVPR Workshop on Empirical Evaluation Methods in Computer Vision, 172-187.

Wiedemann, C. and Hinz, S. (1999): Automatic Extraction and Evaluation of Road Networks from Satellite Imagery, International Archives of Photogrammetry and Remote Sensing (32) 3-2W5, 95-100.

241

Zhang, C., 2004. Towards an Operational System for Automated Updating of Road Databases by Integration of Imagery and Geodata, ISPRS Journal of Photogrammetry and Remote Sensing (58): 166–186.

Zhang, Q. and Couloigner, I. (2005a): Road Identification and Refinement on Multispectral Imagery Based on Angular Texture Signature, 25th EARSeL Symposium.

Zhang, Q. and Couloigner, I. (2005b): Accurate Centerline Detection and Line Width Estimation of Thick Lines using the Radon Transform, Accepted for Publication in IEEE Transactions on Image

Processing.

Zhang, Q. and Couloigner, I. (2005c): Iterative and Localized Radon Transform for Accurate Road Centerline Detection from Classified Imagery, Submitted to Photogrammetric Engineering & Remote Sensing.

Zhang, Q. and Couloigner, I. (2006a): Automated Road Network Extraction from High Resolution Multi-Spectral Imagery, ASPRS 2006 Annual Conference, Reno, Nevada, 10 pages.

Zhang, Q. and Couloigner, I. (2006b): Benefit of the Angular Texture Signature for the Separation of Parking Lots and Roads on High Resolution Multi-spectral Imagery, Pattern Recognition Letters (27) 9: 937-946.

243

Appendices

Appendix 1: Project Proposal

Proposed OEEPE Project

Automated extraction, refinement, and update of road databases from imagery and other data

Proposal for an OEEPE test of Commission II "Image analysis and information extraction"

Background and motivation

Extraction of topographic objects from aerial and satellite imagery has been the topic of intense research over the last decade. The main objects to be extracted include: digital surface models (DSM) and digital terrain models (DTM), buildings, roads, and other thematic objects, such as agricultural fields. The data used are mainly aerial imagery. Mostly film is utilized, but more and more also digital cameras are employed, with increasing use of color and color infrared (CIR). Additionally, new sensors such as airborne laser scanning (ALS), airborne SAR, and to a lesser degree hyperspectral and thermal sensors, as well as high-resolution satellite imagery play an important role. This is in contrast to classical spaceborne sensors, which are used more on the extraction of broader classes and landcover/landuse. Additional data, such as a priori information from maps, GIS and cadaster, have been used to support the extraction process, albeit in a limited fashion, by restricting the search space. Regarding methodology, most research efforts are aimed at automating the extraction of the above objects.

Major research tendencies for automated object extraction include: use of more than two images, early transition to 3D, use and combination of multiple cues and multiple algorithms for the detection and extraction of objects, object models at various levels of detail and generality, use of a priori

information in the form of existing data, utilization of extended context and constraints, use of multi (- sensor, -temporal, -spectral, -scale) approaches. The various research groups involved can be divided into three classes: (a) those coming from geomatics, geodesy, and photogrammetry, (b) computer vision groups, often working for military projects, but without much practical experience, (c) Remote Sensing (RS) oriented groups, employing mainly spaceborne data for various applications. In spite of intense research efforts, very few results have reached an operational stage and have been employed in practice, with the exception of advances in DSM/DTM extraction as well as landcover/landuse classification, and the development of useful tools for semi-automated building extraction. Full automation is currently practically impossible for almost all applications. However, very few researchers realized this and focused on performance improvement via genuine semi-automated methods. Another important factor hindering the practical use of automated procedures is the lack of reliable measures indicating the quality and accuracy of the results, thus necessitating lengthy and cumbersome manual editing.

For roads, although they are much better structured and more homogeneous in their appearance than buildings, no practically useful semi-automated tools for their extraction have yet been developed. Lack of focused international cooperation has led to wide repetition of efforts, development of theoretical frameworks that do not work, and small advances in practice. Manufacturers of commercial systems have developed very few tools for the semi-automated extraction of the above

244

objects, and their cooperation with academia has been minimal. Thus, users and producers of such data, including national mapping agencies (NMCAs) and large private photogrammetric firms, have been left with many wishes to be fulfilled.

On the other hand, the need for accurate, up-to-date, and increasingly detailed information (including various attributes) for many topographic objects, primarily buildings and roads, has rapidly increased. Roads in particular are needed for a variety of applications ranging from provision of basic topographic infrastructure, over transportation planning, traffic and fleet management and optimization, car navigation systems, location based services (LBS), tourism, to web-based applications and virtual environments, etc. Roads also constitute an important component of the increasingly requested and used 3D City Models. NMCAs in Europe increasingly plan to update their maps in shorter cycles. Many aim at the development of a landscape/topographical database from which cartographic data will be derived, and plan to add 3D information for important objects such as roads and buildings. Their customers have increasing demands regarding the level of accuracy and object modeling completeness, and often request additional attributes for the objects, e.g., the number of lanes for roads. That the interest of NMCAs and large firms in automated tools is high has been exemplified by the Swiss Federal Office of Topography (L+T), the Belgian national geodetic institute (NGI), and the Japanese firm PASCO regarding the system developed within the ATOMI project of ETH Zürich and L+T. Other projects are performed by the University of Hannover for the German Bundesamt für Kartographie und Geodäsie (BKG), or Technische Universität München (TUM) and the University of the Bundeswehr (UniBw) Munich for the German military survey. While road extraction has sometimes been performed by digitizing maps, road update and refinement usually can only be performed from aerial imagery, while some ground-based methods including mobile GPS are also important.

The insufficient and inefficient research output and increasing user needs, necessitate appropriate actions. Practically oriented research, e.g., ATOMI, has shown that an automation of road extraction and update is feasible to an extent that is practically very relevant. Companies that have developed semi-automated tools for building extraction and other firms too could very well offer similar tools for roads, even more as semi-automated road extraction is much easier and faster than building extraction. Thus, the central ideas of the proposed project are:

To survey the data specification needs of important data producers, mainly NMCAs, and their customers.

To thoroughly evaluate the current status of research (including models, strategies, methods and used data), to test and compare existing semi- or fully automated methods using various data sets and high quality reference data, to identify weak points and to propose strategies and methods that lead to a fast implementation of operational procedures for road extraction, update, and refinement. Although these latter three tasks are strongly interrelated and have many commonalities in used methods and data, the focus will be on update and refinement. Here, refinement means the improvement of planimetric accuracy, but also the addition of new information such as height and road attributes.

It is envisaged that the above activities and international cooperation will help create a nucleus of interested researchers who, with the cooperation of NMCAs and, if possible, manufacturers, can form a well-coordinated and focused research network which can speed up the development of operational (or quasi-operational) systems for road extraction. These systems could be integrated in commercial photogrammetric, RS, or GIS systems. Funding of such a research network could be from the 6th EU FP or other sources such as, for example, EuroGeographics.

245

Implementation

The project should be implemented by the formation of a working group (WG), with a proposed life-time of one and a half to two years from its constitution. It is proposed that the Chair/Co-Chair of the WG are respectively: Helmut Mayer, UniBW München and Emmanuel Baltsavias, ETH Zürich. The WG will recruit members among researchers active in the field, data producers and users, as well as commercial vendors. A list of potential interested groups has already being compiled. These groups will be informed of the aims and plans of the WG and will be asked for input to refine the aims and finalize activities. Concrete plans and activities include:

1. Establishment of a web-page with full and up-to-date information. The WG aims could be officially announced and presented (kick-off) at a planned workshop "Incremental Updating" before Intergeo, 14-15 October 2002, Frankfurt.

2. Distribution of a questionnaire to data producers and users mainly including NMCAs and firms producing digital road maps, especially for car navigation.

3. Distribution of a questionnaire to researchers in the field, including groups outside Europe.

4. Preparation of a report summarizing the outcome of points 2 and 3 and giving an overview and classification of existing methods and the state-of-the-art in the field, including an extensive literature review, for which much material has been already collected. Road modeling aspects as well as data structures, e.g., hierarchical, object-oriented, play a significant role for successful extraction and will be treated accordingly. Attribute information, although currently mainly neglected, may be of importance for generation of a rich and useful road database. Attributes could include width, number of lanes, road class, (coarse) surface material, road markings, etc. Attention will also be paid to related tasks such as automatic generalization and data mining. Knowledge about automated generalization may lead to a refined search space when starting from given generalized map data, while data mining or machine learning can discover and formulate properties hidden in existing road databases, which can be favorably exploited in the update and refinement. The results of this point will be preferably presented within a larger, already planned event such as the Joint ISPRS Conference "Photogrammetric Image Analysis", 17-19 September 2003, Munich.

5. Preparation and distribution of test data sets, including extensive documentation. At this stage, evaluation procedures for the results and quantitative quality measures should be elaborated and finalized before point 6 below.

The test and related data will be designed in a way that it will allow the participation of groups using various methods and data: single/stereo/multiple images, black and white (B/W)/color/CIR imagery, DSMs/DTMs stemming from photogrammetry or ALS, 2D/3D extraction, semi-/fully automated methods, etc. However, the test will clearly focus towards (3D) approaches that aim at operational procedures and the extraction of a large part of the road network and not just a few (isolated) road sides/edges. The input and output data requirements will basically comply with the needs of European NMCAs. For example, aerial color stereo film imagery of scale 1:15,000-1:30,000 will mainly be used together with existing DTMs and vector GIS/map data. To reduce the amount of data and simplify their processing and analysis, as well as for better focus, uncommon data, such as CIR, may be made available at a limited extent. The data will focus on low to medium difficulty cases, i.e., flat to hilly terrain, rural and suburban areas. A limited amount of urban data may be made available to demonstrate the limitations of some recent approaches. Accurate reference data, possibly with some attributes, like width, will be provided for the major part of the data set. If data from the new digital photogrammetric cameras is available sufficiently early, it will be integrated. For the provision of test and reference data, the support of NMCAs is a must. Such data already exist in Switzerland and

246

Belgium, provided by their NMCAs. High-resolution satellite data from Ikonos or Quickbird is considered at a considerably lower priority, as these data are not extensively available, their cost is high, their resolution often insufficient, and research on road extraction from these data limited.

6. Analysis of the submitted results investigating different aspects including: completeness, reliability, accuracy, main advantages/disadvantages of each approach, input data requirements and other conditions, especially the sensitivity of the results to the selection of the parameters, delivery of attribute information, again including completeness, reliability, and accuracy. In addition, qualitative analysis of the results, including stereo superimposition of derived vectors, will be performed. The contribution of NMCAs in this process is necessary. The coordinating center for the analysis will be UniBw München.

7. Presentation of the main WG work with a workshop towards the end of its life in a practically oriented event maybe at Institut Géographique National (IGN), Paris, or Institut Cartographic de Catalunya (ICC), Barcelona. The WG work can also be presented and combined with other similar work at other appropriate events, such as by dedicated technical sessions. Cooperation with ISPRS, especially WGs IC II/IV, III/3,4,6,7,8, IV/7 is absolutely necessary, while ties with IEEE GRSS Data Fusion Committee, and Earsel SIG Data Fusion will be exploited with lower priority.

8. Publication of a final report within the OEEPE official publications.

9. Possible coordinated publication of several papers or a theme issue in an international journal.

Time Plan

If the project and WG is approved at the OEEPE June meeting, the first five action points above should be finalized by the end of 2002. The performance and analysis of test results should be concluded by August 2003. The project work and WG could be concluded by the end of 2003.

Support and financing

The support of NMCAs for provision of data and evaluation of test results, especially manual control, are an important prerequisite. To cover expenses related mainly to organization of events and use of part-time personnel (especially students) for data preparation and analysis, an amount of 3,000 Euro is proposed.

Helmut Mayer and Emmanuel Baltsavias, June 4, 2002

Appendix 1.A. Possible interested groups

(incomplete, to be refined after project and WG establishment)

1. Academia/Research

Active or previously active in road extraction: University College London (Dowman, Muller), École National Supérior de Télécommunications (Maître, Roux), Univ. of Alcalá (Malpica), ETH Zürich (Baltsavias, van Gool), Univ. of Karlsruhe (Bähr), Research Inst. for Optronics and Pattern Recognition (FOM) (Stilla), Univ. Stuttgart (Fritsch, Haala, Walter), Univ. Hannover (Heipke, Liedtke, Sester), TUM (Ebner, Baumgartner, Wiedemann, Hinz), UniBw München (Mayer), Univ. Bonn (Förstner), TU Dresden (Fuchs), Univ. Haifa (Peled, Barzohar), TU Delft (Vosselman), Université des Sciences et Technologies de Lille (Jedynak), Royal Institute of Technology, Stockholm (Klang), Carnegie Mellon University (McKeown), University of Southern California (Nevatia, Price),

247

Brown University (Cooper), Univ. of Main (Agouris), Univ. of New South Wales (Trinder, Sowmya), Sao Paulo State University, Brazil (Aluir Porfirio Dal Poz)

Other: Finish Geodetic Institute (HyppÄ, Sarjakoski), TU Helsinki (Haggren), TU Wien (Kraus), TU Dresden (Maas), Joanneum Research Center (Paar, Schardt), Univ. Massachusetts (Hanson, Riseman), DLR (Lehmann), Purdue Univ. (Bethel), OSU (Schenk), Austrian Research Centers, Seibersdorf (Steinnocher)

2. NMCAs

Switzerland (Käser), Belgium, IGN (Jung, Airault, Vignilo), ICC (Colomer, Torre), Ordnance Survey (Murray, Holland), Ireland, Sweden (Talts), Rikjswaterstaat, BKG (Grünreich), State Surveying Offices in Germany (Baden-Württemberg, Bayern, Niedersachsen, Nordrhein-Westfalen), Austria, Finland

3. System manufacturers and data producers

LHS (Miller, Stegmeier), Z/I Imaging (Dörstel), Inpho (Gülch), ESG (Ohlhof), Supresoft (Zhang), Racurs, Matra Systems, CyberCity AG (Steidler), ESPA (Lammi), Erdas, PCI, eCognition, ESRI, Intergraph, Space Imaging (Dial, Gibson), Earth Watch, ImageSat Int'l, Spot Image

4. Private companies and other

Car navigation data: TeleAtlas, Robert Bosch, Navtech

Large photogrammetric and engineering companies (focus mainly in Europe): EuroSense, Aerodata (Belgium), ISTAR (France), TerraImaging, KLM Karto, Fugro-Inpark (NL), Kampsax (Denmark), Hansa Luftbild (Germany), Swissphoto (Switzerland), Simmons Mapping, Aerofilms (UK), Finnmap (Finland), Lantmaeteriet/Metria (Sweden), Fotonor (Norway), Pasco, Kokusai, Asia Air Survey, Aero Asahi, Kimoto (Japan), ASI, Earth Data Int'l, Woolpert, Merrick, 3001 (USA), North West, Intera, Nortech (Canada), AAM Geoscan (Australia)

Appendix 1.B. Test and reference data

(incomplete, to be refined after project and WG establishment)

Aerial film imagery

Scale: 1:15,000, maybe also 1:30,000

Focal length: 30 cm, maybe also 15 cm

Color; B/W can be derived from color images; maybe CIR in suburban areas

Terrain covered: flat to hilly, rural, forest borders, small villages, suburban, maybe

denser city areas although with not very high buildings. Land cover should include water bodies.

Scanning: high quality geometry, scan pixel size 10-15 microns, good radiometry and color rendition

Camera calibration protocol or interior orientation parameters.

Exterior orientation, preferably from bundle adjustment, however without use of additional parameters.

248

Possibly only parts of images will be distributed to reduce data size.

Some areas should include fourfold image overlap.

Date and time of image acquisition for shadow analysis.

DTM: As typically existing in Europe. Possibly also higher quality derived by ALS.

DSM: Derived from matching and/or ALS.

Orthoimages: It is an open question whether orthoimages should be generated using DTM or DSM.

Optionally associated topographic maps in digital raster form for overview and other purposes, if possible divided in separate color layers, and at scale ca. 1:25,000

Vector GIS or map information comparable to a scale of 1:25,000 including roads, forest boundaries, water body boundaries, (possibly administrative) built-up area boundaries. Data for roads should be topologically structured and include additional information, e.g., road classes and attributes such as width, existence, and type/color of road marks, etc. For road vector data, error statistics should be given, e.g., mean, RMS, maximum absolute error. For geometric attributes, such as width, mean, RMS, and range might be useful. Road data should preferably include medial axes and not the road sides/edges. If possible height information should be made available for the road points, e.g., by means of interpolation using a DTM or DSM. Railway lines, and small features like paths, small bridges, etc., should be excluded.

Formats for distribution of data should be common. For raster, simple and tiled TIFF, for vector, ASCII, DXF, DGN, Arc/Shape, etc., will be used. Details will be specified at a later stage. Multiple (2-3) formats for each data set may be used.

249

Appendix 2: Questionnaire for Producers of Road Data

This appendix gives the original questionnaire together with the single submitted answer (comments

in Times font and selections as - for affiliation we give the respondent’s institution at the time of the completion of the questionnaire):

• Wolfgang Stößel, Bayerisches Landesvermessungsamt, Munich, Germany

This questionnaire is intended for producers of road data in large quantities, such as National Mapping and Cadastral Agencies (NMCAs), National Road Administrations (NRAs), large private firms generating road databases for navigation, etc. Some of these institutions not only produce data, but also use them for own products, e.g., maps, or they get the data from other sources, e.g., outsourcing. Though, they all specify the requirements that this data have to fulfill. Often they also sell the data to external customers, or receive data requests from them.

The response of the questionnaire will support the work of the WG as defined under "Motivation/Objectives" in the above WEB page, and in particular will help to determine (a) how road data are produced today and how they are planned to be produced in the near future; (b) what the specifications are for the road data incl. attribution; (c) what the road data requirements of various users and for various applications are. The results of the questionnaire will be analyzed, synthesized and made publicly available through various means, including an EuroSDR Official Report.

We plan to make the answers publicly available on our website. If you cannot accept this policy, please tick the field below.

General information

Corresponding Person

(First name, Surname):

Email:

Name of your institution:

Full address:

Country:

Phone number:

Fax number:

WWW address:

250

How many people are employed in your organization? 650

How many people of your organization are currently involved in road extraction? Give these in full persons, e.g., if two persons are involved 50%, give 1. 50

How much manpower was available over the last three years for road extraction (in man years)? 50

How many public organizations in your country are involved in acquisition of road data incl. their attributes?

Is there a cooperation and data exchange between these organizations and to what extent?yes

Please tick, if you do not want your answers to be published on the web

A. How are road data produced by your organization today?

1. From aerial imagery (tick all that apply)

Digitized aerial film Digital cameras

Orthoimages Stereo images More than 2 images Scale of aerial imageryFocal length of the camera mm

Pixel ground resolution (pixel footprint) for digitized film, digital images, or orthoimage m

Black/white RGB False color infrared More than 3 spectral channels (specify how many channels)

Data used for first acquisition of the data: yes no

Data used for data update: yes no Which system is used (give product name, e.g., Zeiss Planicomp analytical plotter, LH Systems DPW770 digital photogrammetric station, etc.)? List all systems used:

Is the processing:

Manual: yes no

Semiautomatic (specify which system is used): yes no

Automatic with possible subsequent editing (specify which system is used): yes no

2. From satellite imagery List all sensors used:

Orthoimages Stereo images Pixel ground resolution (pixel footprint) m

251

All spectral channels used: yes no

Data used for first acquisition of the data: yes no

Data used for data update: yes no

Which system is used (give product name, e.g., Erdas Imagine, LH Systems DPW770 digital photogrammetric station, etc.)? List all systems used:

Is the processing:

Manual: yes no

Semiautomatic (specify which system is used): yes no

Automatic with possible subsequent editing (specify which system is used): yes no

3. By terrestrial means Specify here the system and methodology used, e.g., by using GPS van in open areas.

Kinematic GPS in a van

4. By digitizing maps or other plans:

Are paper maps/plans used and/or digital ones

Specify the type of maps/plans used, e.g., topographic map, cadastral plan, etc.

DFK (digitale Flurkarte – cadastral map with some topographical contents)

Map/plan is B/W and/or Color

Give the scale of the maps/plans

Is the digitizing:

Manual: yes no

Semiautomatic (specify which system is used): yes no

Automatic with possible subsequent editing (specify which system is used): yes no

5. By making use of existing road data from other databases Please specify the type of road data used, and if known how they were produced.

(Governmental) Road construction agencies, municipalities

6. Please specify any other data not listed above, that are used for road extraction, e.g., airborne laser scanner with intensity, airborne radar, digital terrain model (DTM), digital surface model (DSM) including trees and buildings, etc. Give important specifications of the data and how road extraction is performed (e.g., from airborne radar data using the Star-3i system of Intermap, manual extraction using orthoimages with pixel footprint 0.5 m, etc.).

252

7. What is the average time difference of the time of the acquisition of the primary data from which the road data is generated to the time of road data acquisition? E.g., road data acquired in 2001 from aerial imagery flown in 1998 would have a time difference of 3 years.

3 to 6 months

8. What is the average age of the road data compared to now? E.g., for 2002 and the above example, it would be 4 years.

about 1 year

9. Roads are extracted as:

road edges or directly as road centerlines including width?

10. Is height information included? yes no If yes, how is height determined?

Directly measured, e.g., in stereo images

Derived from DTM

Derived from DSM

11. Estimated geometric accuracy Planimetry:

Mean error 1 m RMS error 1 m Maximum absolute error < 3 mHeight (if applicable): Mean error – RMS error – Maximum absolute errorHow are the above accuracy measures derived? E.g., from sample statistics using reference data, based on experience, etc.

estimated, based on experience

12. List below all additional information (attributes) derived, e.g., road class, number of lanes, road width, surface material, etc. For geometric attributes, e.g., road width, give the geometric accuracy with which this attributes are extracted and specify what the accuracy measures represent, e.g., RMS. For each attribute give its importance from a user/application point of view as low, medium, high.

road class (high), number of lanes (high), road width (high)

13. Is the road data topologically connected? yes no If yes, how is the topology established?

Automatically

Manually through post-editing

253

14. How are road classes defined?

Based on a map object legend, e.g., highways, 1st class, 2nd class, etc.

Indirectly, by using road width or other attributes. Please specify which. List all road classes used, including dirt agricultural roads, paths, bridges, etc., but excluding railway lines and waterways.

according to ATKIS object catalog

15. What information is stored for road intersections? E.g., number of road segments meeting at intersection, detailed topology for instance of individual lanes.

Intersections only for highways

16. Is the road data cartographically generalized?

yes no

17. Are bridges a separate class with start/end points? yes no

Are underpasses a separate class with start/end points? yes no

Are tunnels a separate class with start/end points? yes no

18. Is time information stored for the road data? yes no (approximately)

Is this the time of primary data acquisition, e.g., of the aerial imagery and/or time of road extraction

19. In some cases of automatic and semi-automatic road extraction it is useful to extract roads with different approaches depending on the land cover. Is the following data available at your organization to determine the land cover?

- forest boundaries yes no

- boundaries of water surfaces yes no

- coastlines yes no

- boundaries of settlements (villages, towns) yes no

- outlines of buildings (possibly generalized) yes no

20. If you use semi-automated or automated methods for road extraction:

- do these methods provide accuracy measures? yes no - do these methods provide reliability measures (e.g., road definitely correct, definitely

wrong, uncertain needing manual control)? yes no If yes, how is the reliability derived?

How reliable are these reliability measures?

254

- what is the degree of completeness of road extraction (incl. possible errors) and for what land cover, e.g., 0% in forests, 90% in open rural areas, 50% in suburban areas with detached houses, 20% in dense urban areas

- what is the percentage of manual pre- and post-editing required, e.g., 70% automatic computer processing, 30% manual editing

21. Independently of whether road extraction is manual, or (semi)automated:

- Are the accuracy measures valid global or for individual road segments?

- What is the acquisition time incl. both automated and manual processes per square km on a state-of-the-art computer for a region of average road density (e.g., mixed rural and urban)?

updating road data: approximately 0.5 hours per square km

- What computer operating system is used in the processes involved in road extraction?

UNIX Linux Windows

22. What data formats are used for the road data incl. their attributes?

Internally: Intergraph; for exchange: NFF, shape, EDBS, DXF

23. What system is used for the storage and management of the road data?

Intergraph MG Dynamo

Is the system?

Relational

Object-relational

Object-oriented

B. How does your organization plan to extract road data in the near/middle future?

If you plan any changes for the points listed under A, please explain them here.

Information provided by (governmental) road construction agencies; local

topographers

When do you plan to implement the above changes? E.g., in 1-2 years.

at the moment (2002)

255

C. Customers and applications

List below all applications for which your organization is using the road data internally, i.e., excluding selling them to third parties.

Production of DTK 25 (digital topographic map 1 : 25 000)

List below all your current groups of customers and for each list the application and briefly describe the required road data. E.g., list geometric accuracy and attributes needed, how up-to-date the data should be, etc. For each required road attribute give its importance from a user/application point of view as low, medium, high. If, e.g., a group of customers are large transportation companies, the application is fleet management and optimization, and the requirements for the road data are: not older than 1 year, planimetric geometric accuracy 5 m RMS, no height information necessary. Attributes needed are road width (high), number of lanes (high), and surface material (medium).

Rescue services (fire brigades, police, ambulances) – navigation systems; mapping (city plans); engineering companies (urban planning, etc.)

List below all your potential groups of customers (only those that are safely potential customers, avoiding speculations) and for each customer, list the application and briefly describe the required road data. E.g., list geometric accuracy and attributes needed, how up-to-date the data should be, etc. For each needed road attribute give its importance from a user/application point of view as low, medium, high.

location based services (insurance companies, parcel service) geo marketing

D. Requirements of your organization on the data

Here, the desired requirements of your organizations should be listed if they differ from the current status listed under A. For each required road attribute give its importance from a user/application point of view as low, medium, high.

Please, add in the field below any additional remarks

Thank you for your collaboration

257

Appendix 3: Questionnaire for Researchers and Manufacturers

This appendix gives the original questionnaire together with a summary of the answers given (numbers on the right hand side, comments in Times font). The detailed answers on the general characteristics of the approaches can be found in Appendix 4.

The questionnaire was completed by the following people whom we want to thank very much (for affiliation we give the respondent’s institution at the time of completion of the questionnaire):

• Chunsun Zhang, Institute of Geodesy and Photogrammetry, ETH Zurich, Switzerland

• Jason Hu, Department of Earth and Atmospheric Science, York University, Toronto, Canada

• Oktay Eker, Photogrammetry Department, GCM. Ankara, Turkey

• Markus Gerke, Institute for Photogrammetry and Geoinformation (IPI), Hannover University, Germany

• Jose Malpica, Subdirección de Geodesia y Cartografía. Escuela Politécnica, Campus Universitario, Alcalá de Henares, Spain

• Christian Wiedemann, Chair for Photogrammetry and Remote Sensing, Technische Universität München, Germany

• Albert Baumgartner, cf. Wiedemann

• Stefan Hinz, cf. Wiedemann

• Arcot Sowmya, School of Computer Science and Engineering, University of New South Wales, Sydney, Australia

• Michel Roux, Ecole Nationale Superieure Des Telecommunications; Departement Traitement du Signal et des Images, Paris, France

Objectives and publication policy

The information gathered by means of this questionnaire will first of all be used to make available test data which optimally fits the interests of the participants. Yet, the result of the test is not only meant to show the strengths of the individual approaches. Together with the information collected from this questionnaire it will help to analyze the benefits of different models and strategies especially for different scene contents such as urban, suburban, as well as rural with concrete roads and rural with dirt roads.

We do not only intend to summarize our findings in an EuroSDR report, but we will also make available all results and answers to the questionnaires on the web page of the working group, if not stated otherwise by the participant.

Instructions for filling out the questionnairePlease tick or answer as appropriate.

258

General information

Corresponding Person:

Email:

Your institution:

Full address:

Country:

Phone number:

Fax number:

WWW address:

How many people of your group are involved in road extraction?

1, 1, (30), 4, 2, 4, 4, 4, 2, 3

How much manpower was available over the last three years for road extraction (in man years)?

3, 1, (30), 3, 3, 8, 8, 8, 2, 2

Please tick, if you do not want your answers to be published on the web

Road modeling

1. What kind of roads are modeled?

Dirt roads [5]

Rural narrow roads [8]

Urban streets [6]

Large Roads with complex crossings [2]

Freeways including highly complex intersections [2]

Others [0]

259

If others please specify:

2. What data is used?

Black and white images [9]

Color imagery [7]

Color infrared (CIR) imagery [1]

Multispectral aerial imagery [3]

Panchromatic satellite imagery [5]

Multispectral satellite imagery [4]

Digital terrain models (DTM) from image matching or airborne laser scanning (ALS) [3]

Digital surface models (DSM) from image matching or airborne laser scanning (ALS) [6]

Other [2]

If others please specify:Pan-sharpened Ikonos; date and daytime of image acquisition; SAR images

3. What ground resolution is basically employed?

Less than 10 cm [2]

10 cm to 60 cm [3]

60 cm to 2 m [7]

2 m to 10 m [2]

More than 10 m [2]

4. What local road characteristics are used?

Roadsides are parallel edges; (perceptual) grouping [7]

Homogeneity in the direction of the road; tracking [7]

Scale-space behavior or roads; combination of lines in coarse scales and roadsides in fine scale [3]

Markings and cars on the road [2]

Cars on the road [1]

Spectral characteristics of the road surface material [3]

Other [1]

Please describe road modeling in more detail:

roads modeled as Steger lines; texture analysis; homogeneous ribbons with constraints on length, curvature and width; machine learning; verification based on profiles; grouping of hypotheses in DSM based on graph representation.

260

5. How is the smoothness of road trajectories modeled?

Not at all [2]

Splines (a) [2]

Snakes (b) [2]

Kalman-filter (c) [1]

otherwise (d) [7]

If a) - d), please specify:

piecewise parabola; evaluation by fuzzy values based on length, straightness, constancy width, constancy gray value; robust polynomial adjustment; weighting of lines according to smoothness and collinearity during grouping; zip-lock ribbon snake, Bezier curve; smoothness via weighting by mean curvature; piecewise linear

6. How is the network character of roads modeled?

Not at all [1]

Local gaps are bridged. (a) [6]

Global optimization of the network (b) [5]

otherwise (c) [4]

If a) - c), please describe your modeling in more detail:

decision for grouping based on gap length and direction difference; graph data structure; global path search incl. analysis of path lengths; based on grouping criteria and local context – network enforced by simple crossing extraction; generation and weighting of connection hypotheses is based on local criteria (length, collinearity) – graph constructed from road segments and connection hypotheses – promising hypotheses from globally optimum paths – verification by checking homogeneity, shadow, or cars; network broken down in components on different levels, e.g. pairs of edges, junctions – compositional: add up to full network

7. What attributes are determined?

Road class [4]

Road width [9]

Number of lanes [3]

Width of lanes [2]

Road markings [2]

Surface material [2]

Other (please list) [1] height profile and intern quality measure

8. What types of crossings are modeled (in which way)?

Simple crossings (3-4 road segments meeting at the intersection) [8]

261

Complex crossings (more than 4 road segments meeting at the intersection) [2]

Roundabouts [2]

Highway exits [1]

Highway intersections including over- and underpasses [0]

Please characterize type and modeling more in detail:

crossing = junction point and parts of neighboring road axes; modeled similarly as connection hypotheses – derived from graph representation

9. Is context made use of?

Not at all [3]

Different modeling of roads in urban, rural, and forest areas [6]

Modeling of shadows and occlusions [4]

Modeling of other objects or the whole scene (global or holistic modeling) [1]

Please explain in more detail:

mainly from buildings obtained by LIDAR data; adapt parameters to these three context classes; detailed model for rural areas only; potential occlusions and shadows from DSM via ray-tracing based on date / daytime information – shadows refined by extracting dark, homogeneous regions – DSM is also used to exclude locally high blobs and to limit search space for car detection (cars in DSM valleys)

10. How is GIS data employed?

Not at all [7]

To refine the search space [2]

To actively verify roads. The reliance on the outdated road data is high. [1]

Please characterize the chosen policy in more detail:

road extraction in area of road objects using attributes and contextual information from the database – goal is to find evidence (if none is found, it is rejected); vectorial data from GIS used in supervised recognition on segmentation level; verification and update of GIS-data should be based on extraction independent of prior GIS data (noted three times) – GIS data mainly as context for adapting parameters and selecting specialized algorithms – could also be used to guide operator through result

11. How is 3D information used?

Not at all [4]

Road has smooth 3D surface and limited steepness. [4]

Information from DSM generated by image matching or ALS is used. [5]

Road features are actually matched in two or more images. [0]

Please explain in more detail:

262

DSM from LIDAR for segmentation of open and building areas; only for determination of height for objects; weighting of extracted lines; detection of approximate shadow regions; to derive context information and to fuse lane segments extracted in multiple overlapping images (accuracy requirements low)

12. Precision and Reliability

Is the precision of the result estimated? [3]

Are the results labeled with a reliability flag and what criteria are used to derive this flag? [5]

Is the reliability flag reliable and to what extent? [3]

Please explain in more detail:

no internal evaluation – comparison with objects from database; evaluation according to Wiedemann 1998; analysis of larger network parts according to smoothness etc.;each object attached with fuzzy quality value computed from features not used before, i.e., it is independent – experience: makes system less sensitive against improper parameter settings; standard statistical measures used to evaluate results at every stage and to tune the learning algorithms

13. Basic modeling and software tools used

Object oriented programing (e.g., C++) [8]

Explicit reasoning (e.g., semantic network or rule based system) [2]

Automatic learning [1]

Modeling of uncertainty (e.g., Bayes network) [3]

Please explain in more detail:

C++, delphi, etc.; modeling of uncertainty by Theory of Evidence; uncertainty: fuzzy logic; Semantic network is used to make knowledge explicit – concepts and relations are mapped to C++ class hierarchy – specialized extraction strategy including where to look for – uncertainty by fuzzy sets

14. Computer environment

Windows NT, 2000, XP [6]

LINUX [4]

UNIX (please specify) [3]

SGI (IRIX), Solaris

Semi-automatic road extraction (only if applicable)

1. How could the basic strategy be categorized?

Road tracking / following starting from a given point in a given direction (a) [2]

Automatic determination of the road trajectory from manually given end points and possibly iteratively refined intermediate points (b) [0]

Editing of the results of automatic extraction (c) [1]

263

Combination of a) or b) with c) [0]

Please explain in more detail:

Seed points are sequentially given and extraction results emerge

2. If results of automatic road extraction are used:

Is the automated processing restricted by given regions and attributes defined by the operator or derived from GIS data? [2]

Are the results labeled with a reliability flag? [1]

Is the reliability flag reliable? [1]

Please describe in more detail:

General characteristics of the approach (answers cf. Appendix 4)

1. Please summarize the key features of the approach:

2. What are the strong points of the approach?

3. What are the weaknesses or limitations of the approach?

4. How could the approach be optimally exploited in a practical application? How would the envisaged practical application look like?

5. What features do you see most important for an ideal system for practical (semi-automated) road extraction?

Test and reference data

Basic test data will consist of

-Aerial film imagery; scale: 1:15,000, maybe also 1:30,000; focal length: 30 cm, maybe also 15 cm; color - B/W can be derived from color images; maybe CIR in suburban areas; high quality scanning geometry; scan pixel size 10-15 microns; good radiometry and color rendition.

- High resolution satellite imagery: Pan-sharpened multispectral data from IKONOS and/or QuickBird.

- Terrain covered: flat to hilly, rural, forest borders, small villages, suburban, maybe denser city areas although not with very high buildings. Land cover should include water bodies. For less developed areas also dirt roads should be covered.

- Camera calibration protocol or interior orientation parameters

- Exterior orientation, preferably from bundle adjustment, however without use of additional parameters

- Some areas should include fourfold image overlap.

- Date and time of image acquisition for shadow analysis.

264

- DTM as typically existing in Europe. Possibly also DSM derived from matching and/or higher quality from ALS.

- Optionally associated topographic maps in digital raster form for overview and other purposes, if possible divided in separate color layers, and at about scale 1:25,000.

- Vector GIS or map information comparable to a scale of 1:25,000 including roads, forest boundaries, water body boundaries, (possibly administrative) built-up area boundaries. Data for roads should be topologically structured and include additional information, e.g., road classes and attributes such as width, existence, and type/color of road marks, etc.

- For road vector data, error statistics should be given, e.g., mean, RMS, maximum absolute error. For geometric attributes, such as width, mean, RMS, and range might be useful. Road data should preferably include medial axes and not the road sides/edges. If possible height information should be made available for the road points, e.g., by means of interpolation using a DTM or DSM. Railway lines and small features such as paths and small bridges should be excluded.

- Formats for distribution of data should be common. For raster, simple and tiled TIFF, for vector, ASCII, DXF, DGN, Arc/Shape, etc., will be used. Multiple (2-3) formats for each data set may be used.

Questions:

1. Are orthoimages or the original aerial/satellite imagery with multiple overlap including calibration and orientation information preferred?

Orthoimages [4]

Original Data [6] Remark: orthoimages show strong distortions in urban areas

2. Is color or multispectral imagery needed or at least preferred?

yes [7]

no [3]

Remark: infrared would be helpful

3. Are GIS data needed or of importance?

yes [5]

no [5]Remark: needed to verify results

4. Should orthoimages be generated?

yes [7]

no [3]

265

If yes, using:

DTM and / or [2]

DSM [4]Remark: both; no ortho but DSM needed; for experiments and extending feature set

5. Is ALS information considered important for the test?

yes [2]

no [7]Remark: yes, for urban areas; or DSM from other source; only for DSM; not really needed but interesting

6. Which vector formats can you handle?

dxf, Arc/Info shape file and others; Arc Info; Microstation DGN (IDGS) or kdms vector formats; Arc/Info coverage or ungenerate files; DGN, ARC-INFO generate; ArcInfo and home-made ASCII; ArcInfo; standard ArcInfo

7. There is a possibility to provide data from the new digital photogrammetric cameras using linear CCDs (e.g., ADS40). Such data can be provided only as orthophotos (no stereo or multiple images) but in several channels (panchromatic, RGB, NIR). Would you be interested in processing such data and to what extent?

yes [7]

no [3]Remark: interested (PAN, RGB); test would be interesting, but low priority; willing to experiment

8. What else is important?

Orthoimages in RGB; DSM; lots of labeled/marked images for evaluation purposes

Test statistics

1. Are the road centerlines possibly augmented by the width the right basic information for deriving statistics for road extraction?

yes [9]

no [0]Remark: we use centerlines from extraction algorithm for comparison with roads from database

2. What else besides completeness, reliability, accuracy, and parameters describing the global network topology could/should be quantitatively evaluated automatically?

crossings – completeness, correctness, RMS, # of connected roads); # of lanes

3. Is a qualitative visual inspection meaningful besides the automatic quantitative evaluation?

yes [4]

266

no [4]Remark: could be useful to get rough idea of individual strengths and weaknesses, but not useful to derive meaningful ranking of different results; not enough detail to discriminate – e.g., eye joins broken lines

4. For semi-automatic approaches we propose to define the expected quality, i.e., level of detail, precision, and completeness, and then compare the timings for operator interaction and computing time. The achieved results would be evaluated automatically to make sure that the demanded quality is actually reached. Does this make sense?

yes [6]

no [0]Remark: not only total interaction time is relevant, but also number and duration of breaks between operator's activities; in addition the time duration of interactions and automatic processing should be recorded in a table

Please, add in the field below any additional remarks

In order to compare different supervised methods, the same images should be used with the same training areas; when will the data be ready

Thank you for your collaboration

267

Appendix 4: General Characteristics of the Approaches

Chunsun Zhang

Key features: the approach extracts 3-D road network from stereo aerial images by integrating knowledge processing of color image data and existing digital geo-database. It uses existing knowledge, image context, rules and models to restrict search space, treat each road subclass differently, check the plausibility of multiple possible hypotheses, and derive reliable criteria, therefore provides reliable results

Strong points: 1. fusion of information from different cues 2. works directly in object space 3. different classes of roads are treated using different features 4. multiple cues are used to create redundancy to reduce the complexity of image processing, to account for errors, and to increase success rate and the reliability of the results 5. junctions are generated and modeled 6. almost all the roads in rural areas are correctly and reliably extracted

Weaknesses or limitations: the performance is poor in urban areas

Optimal exploitations in practical applications: the developed system will soon be used for 3-D road production in rural areas.

Most important features for practical road extraction: completeness and reliability

Jason Hu

Key features: 1. multi resolution method is used 2. a fast and template matching based algorithm is used - 3. hierarchical perceptual grouping approach

Strong points: before the grouping for segment linking, evaluating the saliency of the primitives and using a sequential grouping method make it more reliable

Weaknesses or limitations: need to integrating more informations on complicated texture and contextual for road extraction from dense urban area

Optimal exploitations in practical applications: it should be fused into a practical system seamlessly. easy-to-use and flexible.

Most important features for practical road extraction: to meet the demands in accuracy, robustness and interactivity so the efficiency. it should be integrated into a practical system (for digitizing etc.)

Markus Gerke

Key features: Verification/Falsification of roads in a database (no generalized data) - Making use of additional knowledge (context classes, attributes of roads)

Strong points: reliable results for rural areas -supports an operator: he/she has to focus attention just on objects where the automated process did not find a road

Weaknesses or limitations: success of extraction algorithm in dense urban areas or in forest areas depends on degree of occlusions -extraction of "new" roads reliable only in open landscape

268

Optimal exploitations in practical applications: It is a practical application (prototype running at the BKG (Federal Agency for Cartography and Geodesy, Germany)): It supports an operator in the verification process: he/she has to focus attention just on objects where the automated process did not find a road

Most important features for practical road extraction: Depending on the resolution of the target database:-reliable extraction of primitives -sophisticated grouping algorithms, but with an easy possibility of user interaction/post editing

Additional remarks: We are using the road extraction software from Christian Wiedemann, TU Munich, extended for our purposes. This software uses the so called "Steger algorithm" for line extraction. Rather than extracting roads from a "blank" image our task is the update of an existing GIS database (ATKIS DLM Basis).

Jose Malpica

Key features: Binary segmentation through the Dempster Shafer Theory of Evidence, followed by the conversion raster to vector and the evaluation of the results

Strong points: Texture color analysis in the segmentation phase. Mathematics definition and determination of the topology of the graphics elements.

Weaknesses or limitations: It's convenient to edit bigs gaps (the small ones are bridged)

Optimal exploitations in practical applications: Run the application automatically to extract all the roads, followed by manual edition of the results.

Most important features for practical road extraction: Our system provides in detail the general road network structure using only the RGB imagery

Christian Wiedemann

Key features: local, regional and global modeling, global grouping, network analysis, modeling of simple crossings

Strong points: global grouping, network analysis

Weaknesses or limitations: selection of seeds for path search

Optimal exploitations in practical applications: unknown

Most important features for practical road extraction: fast during interactive phase, user interface

Albert Baumgartner

Key features: local grouping, use of scale-space behavior of roads, local context, rural areas, panchromatic images, ground resolution: 02.-0.5 m

Strong points: scale-space, context

Weaknesses or limitations: weak global grouping, acceptable results only for rural areas

Optimal exploitations in practical applications: no reliable statement possible; maybe intermediate results could be used to speed up interactive extraction; the operator could e.g., decide which road segments should be connected; alternative: post editing (efficient tools required to edit typical errors)

269

Most important features for practical road extraction: reliability of automatic parts; convenient user interface (to control/guide automatic tools and for post-editing of results); fast enough to avoid inactive time for operator ('user real time') or - if not fast enough - breaks can be used for other tasks

Stefan Hinz

Key features: designed for complex urban areas; needs high resolution overlapping images; extracts detailed road information (lanes, cars); is still basic research

Strong points: relies on detailed, scale-dependent, explicit road and context model ;makes heavy use of context to control and complete the extraction; incorporates self;diagnosis systematically.

Weaknesses or limitations: weak model for junctions; occluding dense vegetation; shadowed, very narrow (one-lane) roads; car detection is not reliable in case of traffic jams

Optimal exploitations in practical applications: Aside from gathering detailed road and updating GIS-road axes it would be interesting if this approach could be combined with close range or mobile mapping applications (e.g. for the creation of virtual cities). But, for this, there is still a long way to go

Most important features for practical road extraction: convenient user interface; strategy of necessary interactions / corrections fits into daily work flow

Arcot Sowmya

Key features: Our approach machine learning based, hierarchical, compositional and reuses learned concepts at every stage.

Strong points: Recent innovations include inductive clustering which is able to automatically select a good recognition algorithm, tune its parameters and also pick the best class out of the results. Exciting new classification techniques such as Support Vector Machines and Genetic Algorithms have been applied. A strong reference model and evaluation procedure have been developed.

Weaknesses or limitations: The approach requires labeled images, which are difficult to obtain.

Optimal exploitations in practical applications: The training and recognition phases are separate. Training on labeled images can be performed off-line in an automated setting. Actual recognition should be very fast on the trained system.

Most important features for practical road extraction: The ability to make on-line corrections, which should be automatically included into the system for future images. This is still in the future, though.

271

Appendix 5: README File

EuroSDR Test on "Automated Extraction, Refinement, and Update of Road Databases from Imagery and Other Data"

Objectives of the Test

To thoroughly evaluate the current status of research including models, strategies, methods and used data, to test and compare existing semi- or fully automated methods using various data sets and high quality reference data, to identify weak points, and to propose strategies and methods that lead to an implementation of operational procedures for road extraction, update, and refinement.

Test data

8 Test images from different sensors:

-Aerial Images (3) -Leica ADS40 (2) -IKONOS (3)

Aerial Images

The given ortho images have a ground resolution of 0.5 m and a size of 4000 by 4000 pixels. The original images have an image scale of 1:16000, the principal distance of the camera was 0.3 m (normal angle lens), and they were scanned with a Zeiss SCAI scanner with 14 micrometer pixel size. The ortho-images were generated using an enclosed (ASCII-format) digital surface model (DSM) with 2 m grid size generated by image matching (Leica Photogrammetric Suite including manual elimination of blunders). The images cover an area close to the city of Thun, Switzerland.

-aerial1 contains a suburban area in hilly terrain -aerial2 contains a rural scene with medium complexity in hilly terrain -aerial3 contains a rural scene with low complexity in hilly terrain

Leica ADS40

The Leica ADS40 is a digital photogrammetric camera with 3 panchromatic and 4 spectral linear CCDs, each with 12,000 pixels and 6.5 microns pixel size. The focal length was 62.5 mm. The images cover an area close to Waldkirch, Switzerland. The images are orthoimages (generated with an enclosed, ASCII-format, DTM), have a ground resolution of 0.2 m and the size of the images is:

-ads40_1: 5800 x 5765 pixels -ads40_2: 4880 x 5290 pixels

They are an 8 bit version and the original 16 bit version available. ads40_1 and ads40_2 comprise both rural areas with medium complexity in flat terrain.

272

IKONOS

The images come from the SI product GEO and have a ground resolution of 1 m and an image size of 4000 by 4000 pixels. There are an 8 bit version and the original 11 bit data. The 4 spectral channels of IKONOS are made available as pan-sharpened images. The images cover areas located in the Kosovo.

-ikonos1 contains a urban/suburban area in hilly terrain -ikonos2 contains a rural scene with with medium complexity in hilly terrain -ikonos3 contains a rural scene with with medium complexity in hilly terrain

Evaluation

The evaluation is done against manually digitized ground truth using the same input images?. For this, the test participants should send the resulting data to a web-server which computes the results via a CGI script. In the first stage of the test only the road centerline will be evaluated. The data has to be sent in form of an ASCII file. The format is (col and row in sub-pixel image/pixel coordinates):

col1 row1 col2 row2 ... first line coln rown -99col1 row1 col2 row2 ... second line coln rown -99etc.

A line is defined as a poly-line between two intersections or between an intersection and a dead-end. The center of the upper left pixel is the origin of the coordinate system. The y-axis is pointing downwards and the x-axis is pointing to the right.

Test participants should also send via e-mail a text file (ASCII or Word format) with a description of the algorithms used, the parameter settings for the tests, a personal evaluation of the results (this can be verbal/qualitative mentioning particular difficulties etc.), and a bibliography with relevant own publications and other comments relevant for the test. The description of the algorithms should be kept short and when possible publications describing the details of the used methods should be send as PDF files. Please use for the file names as suffix .txt, .doc, .pdf and as suffix the name of the organization or the participant.

Evaluation Results

The results consist of:

- percentage completeness

273

- percentage correctness - geometric accuracy in pixels - topological correctness - an image containing the correct parts of the network in green, the incorrect parts in red and the missing parts in blue.

The evaluation is always done for the whole image.

Restrictions of Use and Copyright

The use of the data is currently restricted to the EuroSDR test on road extraction. It is not allowed to use for other purposes or distribute the original data without permission of the data providers. If data provided for this test is made publicly available in digital or printed form (e.g. in publications, WEB sites etc.), the data source and a respective copyright statement has to be made:

aerial images: Copyright Swiss Federal Office of Topography, Bern, Switzerland ads_40 images: Copyright Leica Geosystems, Heerbrugg, Switzerland ikonos images: Copyright Bundeswehr Geoinformation Office (AGeoBw), Euskirchen, Germany

Acknowledgments

We would like to thank the Swiss Federal Office of Topography, Bern, Switzerland, Leica Geosystems, Heerbrugg, Switzerland, and the Bundeswehr Geoinformation Office (AGeoBw), Euskirchen, Germany, for making available the data for this test.

275

Appendix 6: Documentation by Beumier and Lacroix

Note: we have edited the documentation by changing the references to the literature according to our nomenclature.

Algorithm

a) Bright line detection (raster output)

A detection of bright lines is performed using the Gradient Line Detector described in (Lacroix and Acheroy 1998). The process is based on the fact that, at each side of a crest or valley line, the gradient vectors point in opposite directions. The filtering process, applied to the green channel, does not require any parameter setting unless the value of the smoothing Gaussian (1.2 in this experiment) and the type of object , i.e., crest or valley (crest in this case). The filtering process generates a crest/valley norm and a discretized direction (0-3) which is locally perpendicular to the local crest/valley direction. In order to obtain a one-pixel wide line detection, a non-maximum deletion is then performed, using the discretized direction. The output is an image with high values where bright lines are likely to be present.

b) Segment analysis (vector output)

b1) Linear Segments were extracted from the output image of a). For this, maximum line values are followed in predefined directions with a tolerance of + and – 45°. A segment search is stopped when no neighboring points has a minimal magnitude (here 4, non critical). The segment is rejected if it does not have enough pixels (here 30) or if the deviation from a straight line is too high. The deviation lin_dev is computed as the square root of the inertial moment (1.0 pixel in this case).

b2) For each segment point, NDVI is computed (red and infrared channels). Points with NDVI superior to 0 are rejected.

b3) Finally, linear segments are searched in the remaining points with the procedure explained in b1) and satisfying segments are saved to file.

Discussion

This simple procedure delivers bright roads (as appearing in this image). Typical false alarms are aligned bright roofs, with a width similar to roads. Missing segments are darker, hidden or shadowed parts of roads.

Parameter about contrast is not critical and can be set from histogram. The lin_dev value is not critical (especially if there were a linking procedure) but helps rejecting spurious lines. The minimum length parameter is a compromise between rejecting spurious lines and keeping short segments (linking in specific direction would help).

No “High-Level” processing has been applied (segment linking and completion) and would be the major improvement.

277

Appendix 7: Documentation by Hedman and Hinz

Note: we have edited the documentation by changing the references to the literature according to our nomenclature

Algorithm

The extraction of roads is performed with the TUM road extraction approach (Wiedemann and Hinz 1999). The first step consists of line extraction using Steger’s differential geometry approach (Steger 1998), which is followed by a smoothing and splitting step. By applying explicit knowledge about roads, the line segments are evaluated according to their width, length, curvature, etc. Afterwards, overlapping lines are fused using a “best-first” strategy based on the previous evaluation. During this fusion step, there is a possibility to fuse the extracted segments from different channels and/or extraction with different line width. A weighted graph of evaluated road segments is constructed. For the extraction of the roads from the graph supplementary road segments are introduced and seed points are defined. Best-valued road segments serve as seed points. They are connected by an optimal path search through the graph.

Parameter Settings for each test area

Three smaller IKONOS images were analyzed.

Image Channel Line width Higher contrast Lower contrast Bright/dark lines IKONOS1_sub1 Blue 9 30 15 Bright IKONOS3_sub1 Blue 14 30 10 Bright Blue 10 30 10 Bright NDVI 10 30 10 Dark IKONOS3_sub2 Blue 10 30 10 Bright NDVI 12 30 10 Dark

Personal Evaluation

The result obtained from the automatic road extraction is highly dependent on the parameters. Especially the set up for the line extraction is a critical step. Therefore much effort has been put into the selection of these parameters. Since each scene shows different scenarios and landscapes (hilly, rural/urban), the careful choice of the parameters for each scene alone and not a common parameter choice for the total collection of the images is justified. Each channel (red, green, blue, NIR, NDVI) was analyzed separately. The best results were obtained for the blue channel. Even though the results from NDVI were poor, the extraction turned out to be complimentary compared to the other channels. A fusion of blue and NDVI turned out to be the best option for the rural scenes (IKONOS3_sub1 and

IKONOS3_sub2). In one of the scenes (IKONOS3_sub1), the road width differs depending on whether the surrounding area is rural or urban. To get a satisfactory result, a second line extraction from the blue channel using wider line width was carried out.

The automatic road extraction procedure is originally developed for rural areas, which explains the poor result of the urban scene (IKONOS1_sub1).

279

Appendix 8: Documentation by Zhang and Couloigner

Note: we have edited the documentation by changing the references to the literature according to our nomenclature and by deleting copied parts.

Algorithm

The main steps of our methodology and their corresponding algorithms are briefly described below.

1) Image resampling (only for Aerial and ADS40)

The original images, are resized to half (Aerial) / one fourth (ADS40) of the original size using an image viewer free-ware.

2) Image segmentation

The K-means clustering algorithm is used for this purpose.

Aerial and ADS40: All the three bands are used. In all the cases, the number of clusters is set to six and two of them have been identified as possible roads. These two clusters are used to create a road cluster.

Ikonos: All the four bands are used for the Ikonos1_sub1 image, while only the RGB bands are used for the Ikonos3_sub1 and Ikonos3_2 images. The NIR band is discarded for the latter two images because of its low quality. We used the 11-bits images. In all the cases, the number of clusters is set to six and two of them have been identified as possible roads. These two clusters are used to create a road cluster.

3) Road cluster refinement

The road cluster is refined by removing the big open areas, which are usually spectrally similar land covers such as buildings, parking lots, crop fields, and so on. The refinement is based on our shape descriptors of the Angular Texture Signature in combination of a fuzzy classifier. Details can be found in (Zhang and Couloigner, 2005a, 2005b).

4) Road centerline segment detection

The road centerline segments are detected from the refined road pixels by a localized and iterative Radon transform. A gliding window algorithm achieves the localization of the Radon transform. The window size used is 31 by 31 pixels. The iterative Radon transform is then performed locally on each window. We have improved the peak selection techniques so that we can accurately estimate the line parameters of thick lines in the Radon transform. Details can be found in (Zhang and Couloigner 2005c, 2005d).

5) Perceptual grouping

The road segments are further grouped to form the road network. A gap will be bridged if it is shorter than 5 pixels along the road direction. Possible road intersections are detected and used in the road

280

polyline formation process. Only the road polylines that are longer than 20 pixels are written to the output file.

6) Coordinates rescaling (only for Aerial and ADS40 – cf. 1)

The final coordinates are rescaled by a factor of two (Aerial) / four (ADS40) to match the original image coordinate system.

Personal Evaluation

Personally we think that our results are very satisfactory. Most of the roads are correctly extracted with a good positional accuracy. This is impressive for a fully-automated approach without any post manual editing. Our main concerns are:

1) Problems with wide roads (Aerial and ADS40)

Although the proposed methodologies work quite well in finding small roads, wide roads have not been extracted successfully. The reason might be that the road width is still too larger for our methods to work well. Further reducing the spatial resolution might be necessary for this kind of wide roads;

2) Problems in perceptual grouping (Aerial and ADS40)

Our perceptual grouping algorithm is still incomplete and under development. It creates some topological problems especially in road intersections.

3) Problems with shadowed areas (Aerial and ADS40)

As common for a fully automated approach, we have problems in extracting roads in shadowed areas.

4) Omission errors (Aerial and Ikonos)

In our road cluster refinement step, we still have problems with some wide roads, major road intersections, and roads adjacent to a building, parking lot or crop field. Those roads are misclassified as non-roads and will harm the topology of the whole road network. We are still working on it.

5) Large commission errors (Ikonos only)

Due to the usage of the spectral clustering approach for image segmentation, many spectrally similar land covers are misclassified as roads. This mainly accounts for the large commission errors.

6) Problems with dense residential areas (Ikonos only)

In dense residential areas, we have problems in both the image segmentation and the road refinement steps due to the large volumes of noises, artifacts, and spectrally similar objects. It is difficult to automatically (even manually) extract a complete road network in this kind of areas.

LIST OF OEEPE/EuroSDR OFFICIAL PUBLICATIONS

State – November 2006

1 Trombetti, C.: „Activité de la Commission A de l’OEEPE de 1960 à 1964“ – Cunietti, M.: „Activité de la Commission B de l’OEEPE pendant la période septembre 1960 – janvier 1964“ – Förstner, R.: „Rapport sur les travaux et les résultats de la Commission C de l’OEEPE (1960–1964)“ – Neumaier, K.: „Rapport de la Commission E pour Lisbonne“ – Weele, A. J. v. d.: „Report of Commission F.“ – Frankfurt a. M. 1964, 50 pages with 7 tables and 9 annexes.

2 Neumaier, K.: „Essais d’interprétation de »Bedford« et de »Waterbury«. Rapport commun établi par les Centres de la Commission E de l’OEEPE ayant participé aux tests“ – „The Interpretation Tests of »Bedford« and »Waterbury«. Common Report Established by all Participating Centres of Commission E of OEEPE“ – „Essais de restitution »Bloc Suisse«. Rapport commun établi par les Centres de la Commission E de l’OEEPE ayant participé aux tests“ – „Test »Schweizer Block«. Joint Report of all Centres of Commission E of OEEPE.“ – Frankfurt a. M. 1966, 60 pages with 44 annexes.

3 Cunietti, M.: „Emploi des blocs de bandes pour la cartographie à grande échelle – Résultats des recherches expérimentales organisées par la Commission B de l’O.E.E.P.E. au cours de la période 1959–1966“ – „Use of Strips Connected to Blocks for Large Scale Mapping – Results of Experimental Research Organized by Commission B of the O.E.E.P.E. from 1959 through 1966.“ – Frankfurt a. M. 1968, 157 pages with 50 figures and 24 tables.

4 Förstner, R.: „Sur la précision de mesures photogrammétriques de coordonnées en terrain montagneux. Rapport sur les résultats de l’essai de Reichenbach de la Commission C de l’OEEPE“ – „The Accuracy of Photogrammetric Co-ordinate Measurements in Mountainous Terrain. Report on the Results of the Reichenbach Test Commission C of the OEEPE.“ – Frankfurt a. M. 1968, Part I: 145 pages with 9 figures; Part II: 23 pages with 65 tables.

5 Trombetti, C.: „Les recherches expérimentales exécutées sur de longues bandes par la Commission A de l’OEEPE.“ – Frankfurt a. M. 1972, 41 pages with 1 figure, 2 tables, 96 annexes and 19 plates.

6 Neumaier, K.: „Essai d’interprétation. Rapports des Centres de la Commission E de l’OEEPE.“ – Frankfurt a. M. 1972, 38 pages with 12 tables and 5 annexes.

7 Wiser, P.: „Etude expérimentale de l’aérotiangulation semi-analytique. Rapport sur l’essai »Gramastetten«.“ – Frankfurt a. M. 1972, 36 pages with 6 figures and 8 tables.

8 „Proceedings of the OEEPE Symposium on Experimental Research on Accuracy of Aerial Triangulation (Results of Oberschwaben Tests)“ Ackermann, F.: „On Statistical Investigation into the Accuracy of Aerial Triangulation. The Test Project Oberschwaben“ – „Recherches statistiques sur la précision de l’aérotriangulation. Le champ d’essai Oberschwaben“– Belzner,H.: „The Planning. Establishing and Flying of the Test Field Oberschwaben“ – Stark, E.: Testblock Oberschwaben, Programme I. Results of Strip Adjustments“ – Ackermann, F.: „Testblock Oberschwaben, Program I. Results of Block-Adjustment by Independent Models“ – Ebner, H.: Comparison of Different Methods of Block Adjustment“ – Wiser, P.: „Propositionspour le traitement des erreurs non-accidentelles“– Camps, F.: „Résultats obtenus dans le cadre du project Oberschwaben 2A“ – Cunietti, M.; Vanossi, A.: „Etude statistique expérimentale des erreurs d’enchaînement des photogrammes“– Kupfer, G.: „Image Geometry as Obtained from Rheidt Test Area Photography“ – Förstner, R.: „The Signal-Field of Baustetten. A Short Report“ – Visser, J.; Leberl, F.; Kure, J.: „OEEPE Oberschwaben Reseau Investigations“ – Bauer, H.: „Compensation of Systematic Errors by Analytical Block Adjustment with Common Image Deformation Parameters.“ – Frankfurt a. M. 1973, 350 pages with 119 figures, 68 tables and 1 annex.

9 Beck, W.: „The Production of Topographic Maps at 1 : 10,000 by Photogrammetric Methods. – With statistical evaluations, reproductions, style sheet and sample fragments by

Landesvermessungsamt Baden-Württemberg Stuttgart.“ – Frankfurt a. M. 1976, 89 pages with 10 figures, 20 tables and 20 annexes.

10 „Résultats complémentaires de l’essai d’«Oberriet» of the Commission C de l’OEEPE – Further Results of the Photogrammetric Tests of «Oberriet» of the Commission C of the OEEPE“ Hárry, H.: „Mesure de points de terrain non signalisés dans le champ d’essai d’«Oberriet» – Measurements of Non-Signalized Points in the Test Field «Oberriet» (Abstract)“ – Stickler, A.; Waldhäusl, P.: „Restitution graphique des points et des lignes non signalisés et leur comparaison avec des résultats de mesures sur le terrain dans le champ d’essai d’«Oberriet» – Graphical Plotting of Non-Signalized Points and Lines, and Comparison with Terrestrial Surveys in the Test Field «Oberriet»“ – Förstner, R.: „Résultats complémentaires des transformations de coordonnées de l’essai d’«Oberriet» de la Commission C de l’OEEPE – Further Results from Co-ordinate Transformations of the Test «Oberriet» of Commission C of the OEEPE“ – Schürer,K.: „Comparaison des distances d’«Oberriet» – Comparison of Distances of «Oberriet» (Abstract).“ – Frankfurt a. M. 1975, 158 pages with 22 figures and 26 tables.

11 „25 années de l’OEEPE“ Verlaine, R.: „25 années d’activité de l’OEEPE“ – „25 Years of OEEPE (Summary)“ – Baarda,W.: „Mathematical Models.“ – Frankfurt a. M. 1979, 104 pages with 22 figures.

12 Spiess, E.: „Revision of 1 : 25,000 Topographic Maps by Photogrammetric Methods.“ – Frankfurt a. M. 1985, 228 pages with 102 figures and 30 tables.

13 Timmerman, J.; Roos, P. A.; Schürer, K.; Förstner, R.: On the Accuracy of Photogrammetric Measurements of Buildings – Report on the Results of the Test “Dordrecht”, Carried out by Commission C of the OEEPE. – Frankfurt a. M. 1982, 144 pages with 14 figures and 36 tables.

14 Thompson C. N.: Test of Digitising Methods. – Frankfurt a. M. 1984, 120 pages with 38 figures and 18 tables.

15 Jaakkola, M.; Brindöpke, W.; Kölbl, O.; Noukka, P.: Optimal Emulsions for Large-Scale Mapping – Test of “Steinwedel” – Commission C of the OEEPE 1981–84. – Frankfurt a. M. 1985, 102 pages with 53 figures.

16 Waldhäusl, P.: Results of the Vienna Test of OEEPE Commission C. – Kölbl, O.: Photogrammetric Versus Terrestrial Town Survey. – Frankfurt a. M. 1986, 57 pages with 16 figures, 10 tables and 7 annexes.

17 Commission E of the OEEPE: Influences of Reproduction Techniques on the Identification of Topographic Details on Orthophotomaps. – Frankfurt a. M. 1986, 138 pages with 51 figures, 25 tables and 6 appendices.

18 Förstner, W.: Final Report on the Joint Test on Gross Error Detection of OEEPE and ISP WG III/1. – Frankfurt a. M. 1986, 97 pages with 27 tables and 20 figures.

19 Dowman, I. J.; Ducher, G.: Spacelab Metric Camera Experiment – Test of Image Accuracy. – Frankfurt a. M. 1987, 112 pages with 13 figures, 25 tables and 7 appendices.

20 Eichhorn, G.: Summary of Replies to Questionnaire on Land Information Systems – Commission V – Land Information Systems. – Frankfurt a. M. 1988, 129 pages with 49 tables and 1 annex.

21 Kölbl, O.: Proceedings of the Workshop on Cadastral Renovation – Ecole polytechnique fédérale, Lausanne, 9–11 September, 1987. – Frankfurt a. M. 1988, 337 pages with figures, tables and appendices.

22 Rollin, J.; Dowman, I. J.: Map Compilation and Revision in Developing Areas – Test of Large Format Camera Imagery. – Frankfurt a. M. 1988, 35 pages with 3 figures, 9 tables and 3 appendices.

23 Drummond, J. (ed.): Automatic Digitizing – A Report Submitted by a Working Group of Commission D (Photogrammetry and Cartography). – Frankfurt a. M. 1990, 224 pages with 85 figures, 6 tables and 6 appendices.

24 Ahokas, E.; Jaakkola, J.; Sotkas, P.: Interpretability of SPOT data for General Mapping. – Frankfurt a. M. 1990, 120 pages with 11 figures, 7 tables and 10 appendices.

25 Ducher, G.: Test on Orthophoto and Stereo-Orthophoto Accuracy. – Frankfurt a. M. 1991, 227 pages with 16 figures and 44 tables.

26 Dowman, I. J. (ed.): Test of Triangulation of SPOT Data – Frankfurt a. M. 1991, 206 pages with 67 figures, 52 tables and 3 appendices.

27 Newby, P. R. T.; Thompson, C. N. (ed.): Proceedings of the ISPRS and OEEPE Joint Workshop on Updating Digital Data by Photogrammetric Methods. – Frankfurt a. M. 1992, 278 pages with 79 figures, 10 tables and 2 appendices.

28 Koen, L. A.; Kölbl, O. (ed.): Proceedings of the OEEPE-Workshop on Data Quality in Land Information Systems, Apeldoorn, Netherlands, 4–6 September 1991. – Frankfurt a. M. 1992, 243 pages with 62 figures, 14 tables and 2 appendices.

29 Burman, H.; Torlegård, K.: Empirical Results of GPS – Supported Block Triangulation. – Frankfurt a. M. 1994, 86 pages with 5 figures, 3 tables and 8 appendices.

30 Gray, S. (ed.): Updating of Complex Topographic Databases. – Frankfurt a. M. 1995, 133 pages with 2 figures and 12 appendices.

31 Jaakkola, J.; Sarjakoski, T.: Experimental Test on Digital Aerial Triangulation. – Frankfurt a. M. 1996, 155 pages with 24 figures, 7 tables and 2 appendices.

32 Dowman, I. J.: The OEEPE GEOSAR Test of Geocoding ERS-1 SAR Data. – Frankfurt a. M. 1996, 126 pages with 5 figures, 2 tables and 2 appendices.

33 Kölbl, O.: Proceedings of the OEEPE-Workshop on Application of Digital Photogrammetric Workstations. – Frankfurt a. M. 1996, 453 pages with numerous figures and tables.

34 Blau, E.; Boochs, F.; Schulz, B.-S.: Digital Landscape Model for Europe (DLME). – Frankfurt a. M. 1997, 72 pages with 21 figures, 9 tables, 4 diagrams and 15 appendices.

35 Fuchs, C.; Gülch, E.; Förstner, W.: OEEPE Survey on 3D-City Models. Heipke, C.; Eder, K.: Performance of Tie-Point Extraction in Automatic Aerial Triangulation. – Frankfurt a. M. 1998, 185 pages with 42 figures, 27 tables and 15 appendices.

36 Kirby, R. P.: Revision Measurement of Large Scale Topographic Data. Höhle, J.: Automatic Orientation of Aerial Images on Database Information. Dequal, S.; Koen, L. A.; Rinaudo, F.: Comparison of National Guidelines for Technical and Cadastral Mapping in Europe (“Ferrara Test”) – Frankfurt a. M. 1999, 273 pages with 26 figures, 42 tables, 7 special contributions and 9 appendices.

37 Koelbl, O. (ed.): Proceedings of the OEEPE – Workshop on Automation in Digital Photogrammetric Production. – Frankfurt a. M. 1999, 475 pages with numerous figures and tables.

38 Gower, R.: Workshop on National Mapping Agencies and the Internet. Flotron, A.; Koelbl, O.:Precision Terrain Model for Civil Engineering. – Frankfurt a. M. 2000, 140 pages with numerous figures, tables and a CD.

39 Ruas, A.: Automatic Generalisation Project: Learning Process from Interactive Generalisation. – Frankfurt a. M. 2001, 98 pages with 43 figures, 46 tables and 1 appendix.

40 Torlegård, K.; Jonas, N.: OEEPE workshop on Airborne Laserscanning and Interferometric SAR for Detailed Digital Elevation Models. – Frankfurt a. M. 2001, CD: 299 pages with 132 figures, 26 tables, 5 presentations and 2 videos.

41 Radwan, M.; Onchaga, R.; Morales, J.: A Structural Approach to the Management and Optimization of Geoinformation Processes. – Frankfurt a. M. 2001, 174 pages with 74 figures, 63 tables and 1 CD.

42 Heipke, C.; Sester, M.; Willrich, F. (eds.): Joint OEEPE/ISPRS Workshop – From 2D to 3D – Establishment and maintenance of national core geospatial databases. Woodsford, P. (ed.):OEEPE Commission 5 Workshop: Use of XML/GML. – Frankfurt a. M. 2002, CD.

43 Heipke, C.; Jacobsen, K.; Wegmann, H.: Integrated Sensor Orientation – Test Report and Workshop Proceedings. – Frankfurt a. M. 2002, 302 pages with 215 figures, 139 tables and 2 appendices.

44 Holland, D.; Guilford, B.; Murray, K.: Topographic Mapping from High Resolution Space Sensors. – Frankfurt a. M. 2002, 155 pages with numerous figures, tables and 7 appendices.

45 Murray, K. (ed.): OEEPE Workshop on Next Generation Spatial Database – 2005. Altan, M. O.; Tastan, H. (eds.): OEEPE/ISPRS Joint Workshop on Spatial Data Quality Management. 2003, CD.

46 Heipke, C.; Kuittinen, R.; Nagel, G. (eds.): From OEEPE to EuroSDR: 50 years of European Spatial Data Research and beyond – Seminar of Honour. 2003, 103 pages and CD.

47 Woodsford, P.; Kraak, M.; Murray, K.; Chapman, D. (eds.): Visualisation and Rendering – Proceedings EuroSDR Commission 5 Workshop. 2003, CD.

48 Woodsford, P. (ed.): Ontologies & Schema Translation – 2004. Bray, C. (ed.): Positional Accuracy Improvement – 2004. Woodsford, P. (ed.): E-delivery – 2005. Workshops. 2005, CD.

49 Bray, C.; Rönsdorf, C. (eds.): Achieving Geometric Interoperability of Spatial Data, Workshop – 2005. Kolbe, Th. H.; Gröger, G. (eds.): International Workshop on Next Generation 3D City Models – 2005. Woodsford, P. (ed.): Workshop on Feature/Object Data Models. 2006, CD.

The publications can be ordered using the electronic orderform of the EuroSDR website www.eurosdr.net