flood extent mapping based on terrasar-x data · tarian,social,economic, andpoliticalinterest. it...

PFG 2010 / 6, 475–488 ArticleStuttgart, Dezember 2010

© 2010 E. Schweizerbart'sche Verlagsbuchhandlung, Stuttgart, Germany www.schweizerbart.deDOI: 10.1127/1432-8364/2010/0069 1432-8364/10/0069 $ 3.50

Flood Extent Mapping Based on TerraSAR-X Data

Virginia Herrera Cruz, MarC Müller, Friedrichshafen & CHristian Weise,München

Keywords: TerraSAR-X, radar, flood mapping, classification, object-based, pixel-based

Summary: The use of Earth observation data andtechniques in the context of disaster managementsupport has received growing importance in recentyears. Due to global climate change the scientificcommunity foresees an increase in both the inten-sity and frequency of extreme weather events. Thehigh resolution, multi-polarization and multi-inci-dence angle capabilities of the radar satellite Ter-raSAR-X, its quick site access and receiving timesopen interesting perspectives for flood mappingand subsequent assessment of damages. Further-more, the nearly all-weather capacity of SAR dataconstitutes its main advantage above optical sys-tems for mapping of flood events. The work pre-sented in this paper focuses on the development ofan object-based classification approach for opera-tional flood extent mapping in near-real time condi-tions using TerraSAR-X data. The object-basedmethodology is assisted by pixel-based operationsand both are implemented in a solution within theeCognition software environment. The final objec-tive is to develop a tool which can be easily used byan interpreter without the need of a profoundknowledge of image processing. For this purposean eCognition based semi-automated solution – aGraphical User Interface with a rule set behind – isset up. This application is tested in various scenari-os and is further developed in parallel to these tests.The overall study clearly demonstrates the poten-tial of TerraSAR-X data and object oriented meth-ods for a rapid delineation of flood extent in an op-erational environment.

Zusammenfassung: Verwendung von TerraSAR-X-Daten zur Kartierung überschwemmter Berei-che. In den letzten Jahren hat die Anwendung vonFernerkundungstechniken sowie der Einsatz vonErdbeobachtungsdaten im Zusammenhang mit Ka-tastrophenmanagement an Bedeutung gewonnen.Bedingt durch den globalen Klimawandel erwartetdie Wissenschaftsgemeinde eine Zunahme an ext-remen Wetterereignissen sowohl bezüglich Aus-maß als auch Häufigkeit. Die Fähigkeit des Terra-SAR-X-Satelliten, hochauflösende und multi-pola-risierte Satellitenbilder unter verschiedenen Ein-fallswinkeln aufzunehmen sowie seine Tasking-Flexibilität und der schnelle Datenerhalt bieten er-weiterte Perspektiven für die Hochwasserkartie-rung bzw. die Schadenabschätzung. Außerdemstellt das nahezu allwettertaugliche Aufnahmever-fahren von SAR-Daten einen bedeutenden Vorteilgegenüber optischen Daten für die Kartierung vonHochwasserereignissen dar. Die hier vorgestelltenArbeiten konzentrieren sich auf die Entwicklungeines objektbasierten Klassifikationsansatzes fürdas operationelle Kartieren von Hochwasser unternahe Echtzeit-Bedingung und unter Verwendungvon TerraSAR-X-Daten. Die objektbasierte Metho-de wird unterstützt durch eine pixelbasierte Metho-de, beide Ansätze sind in einer eCognition-Soft-wareumgebung implementiert. Ziel ist die Erstel-lung eines Software-Werkzeuges, das relativ ein-fach und ohne fundierte Kenntnisse der Bildbear-beitung eingesetzt werden kann. Zu diesem Zweckewurde eine auf eCognition basierende semi-auto-matische Lösung entwickelt, die eine graphischeBenutzeroberfläche inklusive Regelsatz im Hinter-grund als wesentlichen Bestandteil hat. Diese Lö-sung wurde in verschiedenen Szenarien getestetund parallel dazu weiterentwickelt. Insgesamt zei-gen die Ergebnisse deutlich das Potential von Ter-raSAR-X-Daten und objektorientierten Methodenfür eine schnelle Kartierung der Hochwasseraus-dehnung unter operationellen Bedingungen.

476 Photogrammetrie • Fernerkundung • Geoinformation 6/2010

(e. g., Ramesh et al. 2010) and space-borne –has been used for this purpose. Using opticaldata (especially its red to SWIR spectrum) isoften referred to as a most straightforwardway to extract flood extent from Earth obser-vation data (Ji et al. 2009). huang et al. (2009)– analysing Spot-5 data – and iRimescu et al.(2010) – using MODIS TERRA and AQUAdata – are among the related works. Other sat-ellites used are, e. g., Landsat and the Ad-vanced Very High Resolution Radiometer (Ji

et al. 2009). iRimescu et al. (2010) propose theuse of Normalised Difference Vegetation In-dex (NDVI) for water detection while Ji et al.(2009) defends when possible the use of NDWIcalculated from green-SWIR / green+SWIR.Nevertheless, as flooding is often associatedwith extreme weather events, optical data ac-quisition encounters limitations due to cloudcoverage. Contrarily, radar data presents theadvantage of its near all weather operating ca-pability allowing to acquire imagery of af-fected areas also during night time. Thereforeradar is seen as the only secure alternative formapping flood events in its peak (schumann etal. 2009). For a long time its limited resolutionhas been seen as a hinder for its use in map-ping of flood events. In 2006 schumann et al.(2006) reported SAR data as having mediumspatial resolution, often limited to single fre-quency, difficult to georectify accurately andgeometrically and radiometrically distorted.But the new SAR satellites generation over-come these limitations by providing a widecatalogue of resolutions (down to 1meter) andswaths with a very appropriate relation forflood evaluation. Discussion about the appro-priate spatial resolution for flood mapping canbe found in schumann et al. (2009). Further-more, today’s SAR data – and especially Ter-raSAR-X data – is highly precise in terms ofgeocoding, positional accuracy and absoluteradiometric calibration (ageR & BResnahan

2009).The examples presented here focus on three

flood events of high intensities that occurredin June 2008 in a highly populated area of theUSA and in February 2009 and April 2009 intwo savannah areas of Australia and Namibia,respectively.The first area of interest is located close to

the St. Louis County (Missouri). The flood

1 Introduction

Flood is one of the most widespread naturaldisasters and the one which affects the highestnumber of people in average (CRED 2009). Itregularly causes large numbers of casualtieswith increasing economic loss. Annually floodevents result in estimates of up to 25,000deaths worldwide, extensive homelessness,disaster-induced disease, crop and livestockdamage and other serious harm (UNU 2010).Even the most advanced nations are severelyaffected: The seasonal floods of the Missis-sippi River have averaged 25 deaths annuallysince the 1980s (ibid.). In Europe, the 2002floods killed roughly 100 people, affected450,000 people and left $ 20 billion in dam-ages (ibid.). Already in the first half of 2010,floods caused by the storm Xynthia have af-fected all Atlantic countries from Portugal toGermany, hitting especially France. Also, inMarch 2010 the continuing rain was inundat-ing the South of Spain, leaving hundreds ofpeople homeless. Only two to three monthlater heavy rains caused floods in Poland,Hungary, Slovakia and the Czech Republic –let alone the devastating flood events in Chinaand above all in Pakistan in August of 2010.The latter affected almost 14.5million peoplewith an official death toll risen to 1,384 andwith 1,680 people reported as injured. Over722,000 houses have been either damaged ordestroyed (UN-SPIDER 2010). Consequently,disaster management is of particular humani-tarian, social, economic, and political interest.It involves assessments of vulnerability andrisk, the monitoring of hazard prone zones, theplanning and management of rescue opera-tions and post disaster damage assessment tobe finally better prepared for the next time of adisastrous event.Using remote sensing data allows for map-

ping inundated areas for a more accurate as-sessment of the disaster extent – especially forareas difficult to access – and for a more large-scale and cost-effective way than ground ob-servation (WaisuRasingha et al. 2008, OkamO-tO et al. 1998). Remote sensing is also an im-portant information source for the calibrationof hydrological models which are used, e. g.,in forecasting (ZWenZneR & VOigt 2009). Op-tical data as well as radar data – both airborne

Flood Extent Mapping 477

Neural Networks (kussul et al. 2008) or Ac-tive Contour Models – in its statistical varia-tion (hORRitt et al. 2001, masOn et al. 2010) orin its geometric variant (huang et al. 2005) socalled level-sets – have been applied to seg-ment SAR data in the context of flood map-ping both for whole scenes and region-based(silVeiRa & helenO 2009). But Active Con-tour Models require enormous data prepara-tion which today prevents their application fornear real time mapping. A common approachfor segmentation, more simple to apply are al-gorithms using the Fractal Net Evolution Ap-proach (FNEA), commercially introduced byBaatZ & schäpe (1999) and implemented inthe eCognition software. neuBeRt et al. (2008)and meinel & neuBeRt (2004) show in an ex-tensive review and comparison of segmenta-tion methods best performances for theFNEA.To profit from the simplicity of thresholding

and the additional information saved in thepixel objects and the description power of thedecision trees, different combined approachescould be considered.maRtinis et al. (2009) forexample combine a histogram thresholdingapproach and a segmentation based classifica-tion for flood detection in high resolution Ter-raSAR-X data. The threshold is automaticallycalculated based in local thresholding analysisof selected areas or splits. The threshold is af-terwards integrated into a multi-scale segmen-tation process. They proved the better per-formance of object-based thresholding overpixel-based one. tOWnsend (2001) and pieRdic-ca et al. (2009) combine multiple thresholdswith decision trees. Another possible combi-nation is the supervised and unsupervisedmethod: For instance, at the Center for Satel-lite Based Crisis Information of the GermanAerospace Center an unsupervised clusteringalgorithm is applied to images enhanced bylocal image statistics (minimum / maximum,mean and variance) followed by interactivelyidentification of water clusters (hahmann etal. 2008).Considering multi-temporal analysis as su-

perior to single data approaches (hahmann etal. 2008) change detection methods to detectvariations of backscatter or on the coherenceamong SAR image pairs (BOni et al. 2007,pieRdicca et al. 2008) or by using ratios (an-

event that occurred from May to August 2008was caused by heavy rains. It affected highpopulated areas of the States of Illinois, Indi-ana, Iowa, Michigan, Minnesota, Missouri,and Wisconsin.The second area is the Gulf Country region

in North Queensland. It is a large savannaharea with a monsoonal climate characterisedby the alteration of dry winters and wet sum-mers. It has a well known dynamic of floodevents. Populations in the area are used tohave three to four weeks of isolation duringthe wet season. Nevertheless, at the beginningof 2009, the unexpected volume of water keptthe two towns of the area much longer isolat-ed.Finally, the Caprivi region is a narrow strip

of land compressed between the arms of theZambezi and the Kwando rivers and borderingwith the countries Zambia, Botswana andZimbabwe. The area is the wettest of Namibia,with a seasonal tropical climate with the rainyseason lasting from December to March.

2 SAR Based Flood ExtentMapping

In this paper main emphasis is put on the de-velopment of a SAR-based flood mappingworkflow suited for an operational implemen-tation. In this domain quite a lot of researchhas been done in recent years (maRtinis et al.2009). Often used is the thresholding of theimage histogram (maRtinis et al. 2009, geB-haRdt et al. 2008, henRy et al. 2006, schu-mann et al. 2006) or of an arithmetic image(ratio or difference usually, BRakenRidge et al.2003) with different methods of interactive orautomatic detection of the threshold values.Thresholding is a simple and resources eco-nomical method, but when applying it to pix-els it requires a previous speckle filtering andagain post-processing filtering and it is onlyeffective in well contrasted images (maRtinis

et al. 2009).Segmentation approaches play then an im-

portant role. The mapping process benefits,this way, not only from the absence of the saltand pepper effect, but also from the texturaland contextual information contained in theobjects. Sophisticated methods as Artificial


tails on the different SAR signal interactionswith the Earth’s surface it is referred to Raneyin hendeRsOn & leWis (1998).The work presented here focuses on the de-

velopment of a semi-automated procedure forflood extent mapping under near-real time(NRT) conditions. NRT is here understood asa combination of delivering TerraSAR-X im-ages to the operator within 2 hours after imageacquisition plus producing and delivering theflood extent map to the user within 2 hours ontop.The whole flood mapping procedure is im-

plemented within the eCognition softwaresuite and developed in the Cognition NetworkLanguage (CNL). CNL is an object-based pro-cedural computer language which is speciallydesigned to enable the analysis of complex,interactive, context-dependent image analysistasks, like flood mapping. Each language ele-ment representing the dynamic of the analysisis called “process” and organized in a processtree hierarchy (rule set). Especially interactiveprocesses allow interactions with users in animage analysis workflow. The execution envi-ronment of the software suite uses a work-space concept in which the user of a workflowmay process many images offline and – ifneeded – in parallel on a computer cluster. It isthus possible to give a step-wise structure tothe flood mapping analysis, to separate mod-ules with different degrees of automation inthe overall workflow and to analyze images ina production mode.The outputs of the tests using eCognition

confirmed the potential of TerraSAR-X dataand the object-oriented methods for a rapid de-

dReOli & yesOu 2007) are also widely investi-gated. One advantage of those methods is thatthey profit by comparison possibilities in manycases of uncertainties. The disadvantage isthat these methods require much data prepara-tion (andReOli & yesOu 2007) and stronglydepend on available pre-event data which isnot always at hand in a real emergency situa-tion.Also, real emergency events require simple

tools which can provide a rapid and accurateoverview of the situation. Such a tool shouldbe time-effective, reliable, robust and flexibleenough to adapt to the circumstances in rela-tion to data and staff available. This kind of asimplified tool also offers the benefit directlyto be used by local operators. For all these rea-sons the creation of a robust rule set and thecreation of a user interface for operators hav-ing local knowledge but not being remotesensing experts was the guiding idea duringthe development of the method right from thebeginning of the DeSecure project (desecuRe

2010).The mapping of water with SAR data is

based on the specular behaviour of the micro-wave signal when encountering a smooth sur-face. For instance, in the case of calm water,the SAR signal is scattered away and the sur-face appears dark in the satellite image. How-ever, in case of a flood event, when extremewinds and rain occur at the time of image ac-quisition, the induced changes in the smoothsurface of the water can increase the intensityof the response signal and produce artefactswhich will complicate a straightforward clas-sification of the flooded surface (kussul et al.2008). Further difficulties are flooded vegeta-tion and / or man-made structures. Differentstructures poking out of the water can producea enhancement of the signal due to a doublebounce effect in the different SAR bands. Thiseffect depends on the relation between the sizeand spacing of the structures and the chosenwavelength. As an example, changes fromflood to not flooded in forested areas can beobserved in C-band (hahmann et al. 2008)and with L-band (tOWnsend 2001), becausethe signal is able to penetrate the forest cano-py. Concerning X-band the shorter wavelengthdoes not allow the penetration of the forestcanopy (hahmann et al. 2008). For further de- Fig. 1: Flood mapping – general workflow.


based on the size of the area affected. Strip-Map was chosen to cover the area selected forthe Mississippi event while ScanSAR sceneswere acquired for the Gulf Country and theCaprivi region, respectively. Additionally,SpotLight data of Caprivi were also used for amore detailed mapping. Compared to VV po-larization, the HH polarized backscatter coef-ficient generally presents higher contrast be-tween water and land surfaces (henRy et al.2006). The use of dual polarization data re-sults in a reduced coverage and resolution,which needs to be traded off against the in-creased information content. Here, HH polari-sation was used for all areas of the study. Shal-low incidence angles are preferred for floodmapping as steep incidence angles result instronger backscatter for open water and reducethe contrast to land surfaces. In an emergency,however, acquiring the first possible scene ofthe area affected clearly takes priority overconsiderations on the choice of a proper inci-dence angle. For mapping the flood extent ge-ocoded data sets are required (EEC products).Concerning the resolution mode, preference isgiven to RE products (radiometrically en-hanced) as they already comprise a specklereduction and are therefore more suited to seg-mentation processing.In parallel to the acquisition planning, an

archive search on existing pre-event and aux-iliary data is required. TerraSAR-X data arepreferred as pre-event data, since scenes ac-quired by the same sensor can be directlycompared to each other. However, if no Ter-raSAR-X pre-event acquisitions are available,data from any other available source are used(e. g., optical, airborne data). In case of theMississippi flooding, no TerraSAR-X archivedata of the region was available. Thus, a Land-sat 7 image was deployed for pre-event infor-mation. Additionally, approximately fourmonths after the flood a new TerraSAR-XStripMap scene of the area was acquired andused as pre-event data, to simulate a real sce-nario for which TerraSAR-X archive datawould be available. Also in Namibia, data ofthe flood recession period had to be used assimulated pre-event data. However, by thetime the floods in Queensland occurred, theTerraSAR-X Background Mission was able tosupply pre-event data from archive for that

lineation of flood extent in an operational en-vironment. The resulting mapping chain iscompleted by building a Graphical User Inter-face (GUI). Three modules sequentially acti-vated are distinguished in the developed rulessystem: Segmentation with automated AoI(Area of Interest) selection, semi-automatedextraction of water masks (classification) andautomated final refinement (Fig. 1).After introducing the satellite data used

here including a description of the data char-acteristics and the pre-processing steps (Sec-tion 3) each of the three modules is explainedseparately in the Sections 4–6.

3 Data Set Selection andPreparation

TerraSAR-X is a German radar satellite oper-ating in X-Band (9.65 Gigahertz). In the stand-ard operating mode image data can be ac-quired in three different imaging modes:ScanSAR, StripMap and SpotLight modeswhich differ in their spatial coverage and spa-tial resolution. The ScanSAR mode providesthe largest possible coverage with a singlescene of 100 km (range) × 150 km (azimuth,extension up to 1,650 km possible) and a cor-responding ground resolution of 18m. Strip-Map scenes cover a 30 km wide swath with atypical length of 50 km (also extendable to1,650 km), the ground resolution improves to3m. Finally, the SpotLight mode allows theobservation of relatively small areas (10 ×10 km²) at up to 1m resolution.The high resolution, multi-polarization and

multi-incidence angle capabilities of TerraS-AR-X data acquisition open very interestingperspectives for flood mapping and damageassessment. Further, the system capabilities ofTerraSAR-X render the satellite as a valuableimagery source for disaster management andsite monitoring, e. g., due to the satellite’squick site access time of 2.5 days to any pointon Earth at 95% probability.The programming of TerraSAR-X acquisi-

tions for the affected regions requires a numberof decisions regarding acquisition parameters(for a detailed description of products and pa-rameters see FRitZ & einedeR 2009). The se-lection of the best suitable acquisition mode is


object-oriented classification, the final aim ofthe segmentation is to delineate the meaning-ful regions which represent the objects thatneed to be classified. Despite the popularity ofsegmentation processes for remote sensingimages in the last years, segmentation is by nomeans a new field of research. On the contrary,segmentation was already in the 1970’s a topicof interest (cf. kettig & landgReBe 1976).And so far, about one thousand kinds of seg-mentation approaches have already been de-veloped (chen & luO 2003) within disciplineslike artificial intelligence, signal processing,computer vision etc. with successful applica-tions in medicine, telecommunications, engi-neering or neuro-informatics (schieWe et al.2001). There is a high variety of segmentationalgorithms already available for remote sens-ing images. A good review and comparisoncan be found in neuBeRt et al. (2008) and mei-nel & neuBeRt (2004).Referring back to Section 2, FNEA is one of

the most commonly used approaches. It is abottom-up region-growing technique startingwith one-pixel objects. In many subsequentsteps smaller image objects are merged into

area, and therefore a complete ScanSAR dataset (pre and post event) could be analysed inthat case. Detailed information on the imageparameters is shown in Tabs. 1–3.Auxiliary data sets are also of great im-

portance. Any supplementary informationabout the test site support and accelerate theevent interpretation. It also reduces the ef-fort for the refinement step which is laterdescribed. In addition to the image data sets,a visual interpretation of the environmentalconditions was performed for the Australianflood event. This interpretation map, chieflybased on manual work, was used as refer-ence.

4 Automated Segmentation

Segmentation is the process of completelypartitioning a scene (e. g., a remote sensingimage) into non-overlapping regions (seg-ments) in scene space (e. g., image spaceschieWe 2002), based on homogeneity andheterogeneity criteria, respectively (haRalick

& shapiRO 1985 in neuBeRt et al. 2006). In the

Tab. 1: Mississippi area.

Sensor Centre (lon/lat) ImagingMode

Product Acquisition date

TerraSAR-X −90.75° / 38.96° SM EEC/RE 23 June 2008

−90.61° / 39.15° SM EEC/RE 17 October 2008

Landsat7 39.3°/ 90.7° - - 10 January 2008

Tab. 2: Gulf Country area.



TerraSAR-X −17.15°/ 141.00° SC EEC/RE 3 June 2008

−17.25/ 141.23° SC EEC/RE 11 February 2009

Tab. 3: Caprivi area.



TerraSAR-XTerraSAR-X

24.76°/−17.68° SC EEC/RE 11 April 2009

24.72°/−17.59° SC EEC/RE 1 September 2009

24.78°/−17.88° SL EEC/RE 6 April 2009

24.71°/−17.92° SL EEC/RE 7 September 2009


can be batched to analyse as many images asnecessary.The activation of the automated segmenta-

tion step is initiated with an automated revi-sion of the project parameters for the numberof layers and the calculation of scene statistics(mean, standard deviation and extreme pixelvalues). Then – after tiling the image – theparallel processing begins for each tile with achessboard segmentation. By this means thewhole tile is segmented into squares of 60 × 60pixels. Based on the scene statistics previouslycalculated and on empirical analysis of a largecollection of SAR data, different thresholdvalues are identified for units not containingwater. Those values will be included as previ-ous condition for the segmentation. If thesquare object in question is within this thresh-old or in contact to an object which is in it, itwill be segmented.The segmentation of the units containing

water is then made using the MRS algorithm.Considering objects not as a static unit, but asa dynamic concept in which objects grow,shrink, and become re-segmented andsmoothed etc., there exists a wide range ofscale parameters which could fit the segmen-tation purposes here. Hence, an ideal parame-ter set for a perfect segmentation can never beselected.Pragmatically, the necessary parameter ad-

justments depend on the size, shape and back-scatter of the expected objects and the speci-

larger ones. The growing decision is based onlocal homogeneity criteria describing the sim-ilarity of adjacent image objects in terms ofsize, distance, texture, spectral similarity andform. The scale at which this segmentation isdone can and has to be chosen in relation tothe characteristics of the data set and the ob-jective of the work. The aim of the segmenta-tion is to delineate the meaningful regionswhich represent the objects that will be classi-fied. For this reason it is mandatory for thequality of the further classification step thatthe borders of the water bodies (permanent ornot) are clearly highlighted.The segmentation module of the process de-

veloped here is a rule set for automated analy-sis which basically combines cycles of MRSand classification and is assisted by other moresimple segmentations not based on the imagecontent (chessboard segmentation). To allowthe processing of large data sets the rule setincluded the tiling, parallel processing and fi-nal stitching of the region. In order to acceler-ate the segmentation, which is the most timeconsuming process of the overall flood map-ping processing chain, only regions potentiallycontaining water are automatically selectedand segmented. Thus, the amount of segments(objects) and the respective processing time isdrastically reduced. The result is an automatedand target-centred segmentation “image” inwhich the information of both dates, pre andpost event is taken into account. The process

Fig. 2: MRS result using Landsat 7 and TerraSARX data – Overlaid on TerraSAR-X postevent scene.

Fig. 3: MRS result using Landsat 7 and TerraSARX data – Overlaid on Landsat 7 preeventscene.


dark in colour-infrared images. The character-istic spectral reflectance curve shows a reduc-tion of intensity with increasing wavelength.In the near infrared the reflectance of deep,clear water is virtually zero, for this reason thechannel 4 of Landsat 7 is used for the discrim-ination of water / non-water; in case of SARX-band, under low to moderate wind condi-tions, backscatter from water surfaces is low.Thus, a water body appears dark compared tothe backscatter of other land cover features;based on this water surface behaviour, a rule-based classification system with tuneable vari-ables and a user-friendly interface is con-structed for the effective extraction of the wa-ter extent.The first classification step is based on an

object-based thresholding of the intensity val-ues, controlled by texture, shape and contextfeatures. A tool supports the operator allow-ing him to manipulate the different features sothat the water mask extent can be changed on-line. However, the resulting water cover is notalways perfect, as can be seen in Fig. 4, whichshows the permanent water body extracted us-ing channel 4 of the Landsat scene (a) and theTerraSAR-X scene (b), respectively. Main dis-crepancies in these two water masks resultfrom the diverse image resolutions and fromthe different acquisition dates. It can be no-ticed that in the two cases areas affected byshadows are classified as water since theypresent a low reflectivity. These errors that ap-pear in more in the Landsat classification arecorrected with further widgets that can be op-tionally used. Two types of corrections or in-teractive refinements are implemented withthose widgets. The first one is an objects-basedreclassification of the “water” segments, basedon the previously defined water class. The sec-ond one is the pixel-based thresholding of thewater borders and neighbouring objects. As insome occasions a certain water course is inter-rupted because few water pixels dropped in amainly non-water polygon. Subsequent to thepixel-based correction the water course willbe connected again.For the post-event data the semi-automated

extraction of the water mask is based on thesame methodology. The obtained results areshown again in Fig. 4 for the Mississippi Rivercase (c and d). The water masks look similar

fied Minimum Mapping Unit (MMU). In thisstudy – depending on the availability of one ortwo scenes as input data – a mono-temporalsegmentation or a bi-temporal segmentation,respectively, has been automatically per-formed. Both processes will produce only onesegmentation layer, but for the bi-temporalsegmentation a segmentation of the pre-eventdataset is produced inside the segments of thepost-event data set. Finally, all tiles are stitchedtogether. A re-segmentation of the tile bordersis then conducted to harmonize the segmentsin the whole scene and to avoid false objectscreated by the tiles’ borders.Figs. 2 and 3 display an example of the seg-

ments obtained using a combination of Ter-raSAR-X and Landsat data for the Mississippiscenario. The same segments are displayedonce over the TerraSAR-X post-event sceneand once over the Landsat 7 scene. As it can beseen the automated segmentation process usesinformation of both scenes. When comparingthese segmentation results to those obtainedusing only TerraSAR-X data the object shapesdiffer between the two scenarios due to thedifferent resolution of the two pre-event layers(3m, TerraSAR-X image vs. 30m [resampledto 3m], Landsat 7 image) and the different na-ture of radar and optical imaging techniques.Specifically, the speckle of the SAR datastrongly influences the object shapes andsizes.

5 Semi-automated Extraction ofWater Mask (Classification)

Unlike pixel-based techniques which only usethe pixel values, the object-based techniquescan also use texture, shape and context infor-mation of a scene. The optimal features, i. e.,measurable properties of the segments, forwater detection are identified and constitutethe basis for the “water” / ”non water” classifi-cation of the pre-event and post-event scenes.The similar spatial behaviour of water in

SAR images and in the NIR channel of opticaldata is used here as common ground for aflood mapping workflow considering not onlyTerraSAR-X data, but also optical archivedata: In optical images, water absorbs most ofthe incident radiation and therefore appears


a b

c

d

Fig. 4: a) Landsat 7based pre-event water mask b) TerraSARXbased pre-event water maskc) TerraSARXbased post-event water mask d) Permanent water / flood mask derived from thecombination of Landsat 7 and TerraSAR-X data.

Fig. 5: Intermediate step – result of the semiautomated classification (blue: permanent water body, cyan: flooded area).

Fig. 6: Intermediate step – result of the semiautomated classification (colours: see left).


6 Automated Refinements

The final step in the flood mask production isthe generation of two water masks in order toseparate the flooded areas from the permanentwater bodies such as rivers, lakes and ponds.The automated refinement step consists in

the improvement of the classifications madeon the pre-event and post-event scenes. It in-cludes the automated extraction of the floodand permanent water masks as well as the ap-plication of final refinements such as smooth-ing and minimum mapping filters. A finalMMU is also defined depending on the quan-tity of detail desired in the final water maskand the precision of the final map, respective-ly. Finally, the borders of the polygons aresmoothed using growing and shrinking pixelbased algorithms.The results obtained for the Australia case

are very satisfying. Fig. 7 shows a detail of thearea considered for the analysis. A referencemap (Fig. 8) was manually delineated for part

despite the fact that two different layers wereused as pre-event data for the original seg-mentation. It is worth mentioning that the se-lection of the pre-event layer had a direct im-pact on the shape and size of the objects andon the specific value needed for the features tobe selected within the classification process.The need for refinements of the water masks

typically depends on the relief and on the en-vironmental conditions. For example, windhas a particular influence on SAR-based clas-sification results as higher wind speeds couldresult in larger errors. In the Caprivi scenario,the strong winds affecting the ScanSAR sceneprevented the detection of the whole mass ofwater (Fig. 5). In comparison, the SpotLightscene (Fig. 6), recorded at a different date andwith a shallower angle shows a very clear dis-tinction of water and therefore needs nearly norefinement steps. Worth mentioning are alsothe double-bounce effect that can be observedto the sides of the rivers (Fig. 5).

Fig. 7: Detail of ScanSAR post-event scene over the Gulf Country affected area. Low backscattering of the radar signal (dark) indicates permanent water and flooded areas, respectively.

Fig. 8: Manually delineated water extent map (reference), Gulf Country (cyan: permanent waterbody, blue: flooded area), superimposed on ScanSAR post-event scene (colours in the background due to display).


Altogether, TerraSAR data demonstratedits high capacity for the mapping of floodevents. Furthermore, the high resolution, themulti-polarization and multi-incidence anglecapability of TerraSAR-X, its quick site accessand receiving times open very interesting per-spectives for flood mapping and the assess-ment of damages after a natural disaster. Con-cerning an automation of mapping processes,the degree to which it can be applied is direct-ly related to the quality of the received SARsignal in the available TerraSAR-X images forthe respective flood events. Strong wind orrain and some landscape conditions might leadto certain noise or other effects which producean atypical backscatter signal of water. Thiseffect is intensified by a shallow incidence an-gle of the acquired scene. As near-real timeconditions for flood mapping are also ad-dressed, it is not possible to change the envi-ronmental condition if they are not favourable.However, the image acquisition capabilities(e. g., combining descending and ascendingorbits) and the quick tasking of the satellitecan provide within relatively short time moreimages to counteract such disturbances.Nevertheless, object-basedmethods –which

allow the use of context information – and thestepwise approach with its different levels ofautomation and possibility for parameter tun-ing showed to be rather adjustable to these cir-cumstances.The ScanSAR post event scene for the Na-

mibia scenario was a good example of unfa-vourable conditions: On the one hand low in-cidence angle and wind, on the other handwater covered areas under vegetation provok-ing double-bounce effects changing the typi-

of the ScanSAR image. This map allowed apreliminary qualitative evaluation of the re-sults (Fig.9) of the developed flood mappingprocessing chain. Further qualitative andquantitative evaluations are currently underway.It is worth mentioning that the manual de-

lineation required more time for a third of thetotal scene than time needed by the semi-auto-mated process for the whole scene. This meansthat even if no further process was implement-ed for the correction of the shadows in floodplain areas, it could be assumed that deselect-ing areas of shadows by manual editing toolsmight be still more time-effective than visualinterpretation.

7 Conclusion and Outlook

This paper gives an overview of the develop-ment and implementation of a TerraSAR-Xbased semi-automated workflow for flood ex-tent mapping performed in the framework ofthe DeSecure project.For this purpose we combined a series of

existing techniques for pixel- and object-basedanalysis to produce a user friendly tool forflood mapping. The developed semi-automat-ed workflow was tested in a variety of scenar-ios and a preliminary validation of the resultshas been made confronting them with floodextent maps generated manually by visual in-terpretation. Finally, the workflow was imple-mented into a pre-operational test using twodifferent TerraSAR-X products: A bi-temporalScanSAR data set and a bi-temporal SpotLightdata set.

Fig. 9: Final flood mask product from the semi-automated procedure, Gulf Country (colours andbackground as above).


when processes such as rising or decliningwater levels should be monitored. Neverthe-less, one should not forget the additional effortrequired for co-registration when workingwith different sensors.Regarding the validation of the flood map-

ping for the Australian Gulf Country test siteit has been noticed that, first of all, the process-ing of the whole scene with the developed pro-cedure required approximately less than 5%of the working time than the manual delinea-tion of the flood extent. Secondly, the flooddelineation results of either method were iden-tical. Finally, this comparison has also demon-strated the effectiveness of the workflow: Themask generated with the semi-automatedprocess shows a level of detail that an inter-preter by means of digitising is not able toreach in reasonable time. Concerning furtherqualitative and quantitative evaluations of theresults the analysis has only recently been ini-tiated and could not be completed within theterm of the DeSecure project. Nor was it pos-sible to perform any field work or to collectfield data and / or additional reference data.Using SpotLight scenes showed to be the

ideal condition for an automated mapping of aflood event. It was concluded that within oneto two hours after image availability at the op-erator it is possible to produce a rapid floodextent map (best case scenario for two TerraS-AR-X SpotLight acquisitions or subsets incase of larger TerraSAR-X data sets). Fromthis point in time, additional time needs to beconsidered depending on the need of pre-processing non TerraSAR-X data, the numberof pixels to processed, the level of detail re-quired, the level of accuracy and smoothnessrequired for the delineation of the boundarybetween “water” and “no water” and the levelof uncertainty given by the environmentalconditions during image acquisitions.

ReferencesageR, t.p. & BResnahan, p.c., 2009: Geometricprecision in space radar imaging: Results fromTerraSAR-X – ASPRS Annual Conference: onCD.

andReOli, R. & yesOu, h., 2007: Change detectionanalysis dedicated to flood monitoring using En-visat wide swath mode data. – Envisat symposi-um: on CD.

cal low backscattering characteristic of water.As a consequence for the mapping the Scan-SAR-based map required more post-process-ing than the SpotLight-based one, which onlyrequired minor corrections.In terms of workflow the main findings and

framework conditions to be put here can besummarised as follows:Large images were successfully processed

(30 × 50 km2 [StripMap], 100 × 150 km2 [Scan-SAR]) and the TerraSAR-X RE products couldbe used without previous pre-processing orneed of speckle filtering:The adaptability and flexibility of the work-

flow was demonstrated applying it to satellitescenes representing different environmentalconditions. Concerning the mapping chain’srobustness it has been tested using differentTerraSAR-X imaging modes and, as in case ofthe Mississippi River flooding, different sen-sors (TerraSAR-X data [3m res], Landsat 7ETM+ [30m res]). Nevertheless, to draw con-clusions about the rule set’s performance us-ing optical data, further tests are required.Regarding the result’s uncertainties caused bystrong winds and / or rain during image acqui-sition, additional manual editing may be re-quired while in some special cases the mete-orological conditions may not allow the appli-cation of the rule set in all. However, in orderto minimise wind and rain effects in the satel-lite image, the incidence angle of the TerraS-AR-X data acquisition should be as shallow aspossible. As to the homogeneity level of thesegments this can be adjusted by the operator,reflecting the system’s adaptation possibili-ties, especially when the environmental con-ditions during image acquisition are not fa-vourable (inundations in denselywooded areasare unfavourable but have not yet been ana-lysed). Under favourable conditions though,the manual editing is minimised but cannot beeliminated completely.The enhancements of the semi-automated

system as presented here towards a multi-mis-sion suited system, has several advantages fora NRT damage assessment: Highly accurateflood extents could be delineated even if noTerraSAR-X pre-event data are available.And, the revisit time over the affected regionis increased by the use of multi source remotesensing data which is an important factor


es. – International Journal of Remote Sensing 26(6): 1145–1156.

huang, m., gOng, J., shi, Z. & Zhang, l., 2009:River bed identification for check-dam engineer-ing using SPOT-5 image in the HongShiMao wa-tershed of the Loess Plateu, China. – Interna-tional Journal of Remote Sensing 30 (8): 1853–1865.

iRimescu, a., cRaciunescu, V., stancalie, g. &neRtan, a., 2010: Remote sensing and GIS tech-niques for flood monitoring and damage assess-ment. – Study case in Romania, BALWOIDS2010, Ohrid, Republic of Macedonia: on CD.

Ji, l., Zhang, l. & Wylie, B., 2009: Analysis ofdynamic thresholds for the Normalized Differ-ence Water Index. – Photogrammetric Engineer-ing and Remote Sensing 75 (11): 1307–1317.

kettig, R.l. & landgReBe, d.a., 1976: Classifica-tion of Multispectral Image Data by Extractionand Classification of Homogeneous Objects. –IEEE Transactions on Geoscience Electronics14 (1): 19–26.

kussul, n., shelestOV, a. & skakun, s., 2008: Gridsystem for flood extraction from satellite images.– Earth Science Informatics 1: 105–117.

maRtinis, s., tWele, a. & VOigt, s., 2009: Towardsoperational near real-time flood detection usinga split-based automatic thresholding procedureon high resolution TerraSAR-X data. – NaturalHazards and Earth System Science 9: 303–314.

masOn, d.c., speck, R., deVeReux, B., schumann,g.J.p., neal, g.c. & Bates, p., 2010: Flood de-tection in urban areas using TerraSAR-X. – IEEETransactions on Geoscience and Remote Sens-ing 48 (2): 882–894.

meinel, g. & neuBeRt, m., 2004: A Comparison ofsegmentation programs for high resolution re-mote sensing data. – XXth ISPRS Congress: onCD.

neuBeRt, m., heROld, h. & meinel, g., 2008: As-sessment of remote sensing image segmentationquality. – GEOBIA: on CD.

OkamOtO, k., yamakaWa, s. & kaWashiWa, h.,1998: Estimation of flood damage to rice produc-tion in North Korea in 1995. – InternationalJournal of Remote Sensing 19: 365–271.

pieRdicca, n., chini, m., pulViRenti, l. & macina,F., 2008: Integrating physical and topographicinformation into a fuzzy scheme to map floodedarea by SAR. – Sensors 8: 4151–4164.

Ramesh, a., glade, t., kappes, m. & malet, J.p.,2010: Delineation of Flood inundated areas us-ing aereal photpinterpretation and GIS-basedhydrological modelling – an application in Bar-celonnette, France. – Geophysical Research Ab-stracts 12, EGU General Assembly 2010: onCD.

BaatZ, m. & schäpe, a., 1999: Multiresolution seg-mentation – an optimization approach for highquality multiscale image segmentation. – Ange-wandte Geographische Informationsverarbei-tung.

BOni, g., castelli, F., FeRRaRis, l., pieRdicca, n.,seRpicO, s. & siccaRdi, F., 2007: High resolutionCOSMO/SkyMed SAR data analysis for civilprotection from flooding events. – IGARSS: onCD.

BRakenRidge, g.R., andeRsOn, e., nghiem, s.V.,caquaRd, s. & shaBaneh, t., 2003: Flood warn-ings, flood disaster assessments, and flood haz-ard reduction: the roles of orbital remote sens-ing. – 30th International Symposium on RemoteSensing of the Environment: on CD.

chen, q. & luO, J., 2003: A geostatistic based seg-mentation approach for remotely sensed images.– www.definiens.de.

CRED (Centre for Research on the Epidemiology),2009: Disasters in numbers. – New release.www.unisdr.org/preventionweb/files/12470_2009disasterfigures.pdf (15th June2010).

desecuRe, 2010: DeSecure project homepage. –www.desecure.info/intro_de.html (15th June2010).

geBhaRdt, s., künZeR, c., WehRmann, t., gstaigeR,V., huth, J., schettleR, h. & schmidt, m., 2008:The WISDOM Project – A water-related Infor-mation System for the Mekong Delta, Vietnam:first results of remote sensing. – World watercongress: on CD.

FRitZ, t. & einedeR, m. (Eds.), 2009: TerraSAR-XGround Segment, Basic Product SpecificationDocument, Issue 1.6.

hahmann, t., ROth, a., maRtinis s., tWele, a. &gRuBeR, a., 2008: Automatic extraction of waterbodies from TerraSAR-X data. – IGARSS 2008:103–106.

haRalick, R. & shapiRO, l., 1985: Image segmen-tation techniques. – Computer Vision, Graphics,and Image Processing 29: 100–132.

hendeRsOn, F.m. & leWis, a.J. (Eds.), 1998: Princi-ples and applications of imaging radar. – Manualof Remote Sensing 2, John Wiley & Sons, Inc..

henRy, J., chastanet, p., Fellah, k. & desnOs, y.,2006: Envisat multipolarized ASAR data forflood mapping. – International Journal of Re-mote Sensing 27 (10): 1921–1929.

hORRitt, m.s., masOn, d.c. & luckman, a.J.,2001: Flood boundary delineation from Synthet-ic Aperture Radar imagery using statistical ac-tive contour model. – International Journal ofRemote Sensing 22 (13): 2489–2507.

huang, B., li, h. & huang, x., 2005: A level setmethod for oil slick segmentation in SAR imag-


ment and Emergency Response. – www.un-spi-der.org/disaster/3665/2010-07-22/pakistan/flood-pakistan (6th September 2010).

WaisuRasingha, c., aniya, m., hiRanO, a. & sOm-mut, W., 2008: Use of RADARSAT-1 data anddigital elevation model to access flood damageand improve rice production in the lower part ofthe Chi River Basin, Thailand. – InternationalJournal of Remote Sensing 29 (19–20): 5837–5850.

ZWenZneR, h. & VOigt, s., 2009: Improved estima-tion of flood parameters by combining spacebased SAR data with very high resolution digitalelevation data. – Hydrology and Earth SystemSciences 13: 567–576.

Addresses of the Authors:

ViRginia heRReRa cRuZ & Dr. maRc mülleR, Info-terra GmbH, D-88039 Friedrichshafen, Tel.: +49-7545-8-3382, -8439, e-mail: [email protected], [email protected]

chRistian Weise, DefiniensAG,D-80339München,Trappentreustrasse 1, Tel.: +49-89-231180-0, e-mail: [email protected]

Manuskript eingereicht: Juni 2010Angenommen: September 2010

schieWe, J., tuFte, g.l. & ehleRs, m., 2001: Poten-tial and problems of multi-scale segmentationmethods in remote sensing. – GIS – Geographis-che Informationssysteme 6: 34–39.

schieWe, J., 2002: Segmentation of High-Resolu-tion Remotely Sensed Data – Concepts, Applica-tions and Problems. – International Archives ofPhotogrammetry Remote Sensing and SpatialInformation Sciences 34 (4): 380–385.

schumann, g., Bates, p.d., hORRitt, m.s. & mat-gen, p., 2009: Progress in integration of remotesensing-derived flood extent and stage data andhydraulic models. – Reviews of Geophysics 47,RG4001.

schumann, g., matgen, p., pappenBeRgeR, F.,Black, a., cutleR, m., hOFFmann, l. & pFisteR,l., 2006: The REFIX model: Remote Sensingbased flood modelling”. – International Archivesof Photogrammetry, Remote Sensing and SpatialInformation Sciences 36 (7): on CD.

silVeiRa, m. & helenO, s., 2009: Separation be-tween Water and Land using Region Based Lev-el Sets. – IEEE Geoscience and Remote SensingLetters 6 (3): 471–475.

tOWnsend, p.a., 2001: Mapping seasonal floodingin forested wetlands using multi-temporal radar-sat SAR. – Photogrammetric Engineering andRemote Sensing 67 (7): 857–864.

UNU, 2010: United Nations University. – www.unu.edu/ (15th June 2010).

UN-SPIDER, 2010: United Nations Platform forSpace-based Information for Disaster Manage-

flood extent mapping based on terrasar-x data · tarian,social,economic, andpoliticalinterest. it...

Documents