C

Ta

b

a

ARA

KOATN

1

aaiustMtlTrMifisumu

0d

International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884–893

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

journa l homepage: www.e lsev ier .com/ locate / jag

omparing object-based and pixel-based classifications for mapping savannas

imothy G. Whitesidea,∗, Guy S. Boggsb, Stefan W. Maierb

Faculty of Health, Business and Science, Bachelor Institute of Indigenous Tertiary Education, Batchelor, NT 0845, AustraliaSchool of Environmental and Life Sciences, Charles Darwin University, Darwin, NT 0909, Australia

r t i c l e i n f o

rticle history:eceived 29 November 2010ccepted 28 June 2011

eywords:bject-based image analysisccuracy assessmentropical savannaorthern Australia

a b s t r a c t

The development of robust object-based classification methods suitable for medium to high resolutionsatellite imagery provides a valid alternative to ‘traditional’ pixel-based methods. This paper comparesthe results of an object-based classification to a supervised per-pixel classification for mapping landcover in the tropical north of the Northern Territory of Australia. The object-based approach involvedsegmentation of image data into objects at multiple scale levels. Objects were assigned classes usingtraining objects and the Nearest Neighbour supervised and fuzzy classification algorithm. The supervisedpixel-based classification involved the selection of training areas and a classification using the maximumlikelihood classifier algorithm. Site-specific accuracy assessment using confusion matrices of both classi-fications were undertaken based on 256 reference sites. A comparison of the results shows a statistically

significant higher overall accuracy of the object-based classification over the pixel-based classification.The incorporation of a digital elevation model (DEM) layer and associated class rules into the object-based classification produced slightly higher accuracies overall and for certain classes; however this wasnot statistically significant over the object-based using spectral information solely. The results indicateobject-based analysis has good potential for extracting land cover information from satellite imagerycaptured over spatially heterogeneous land covers of tropical Australia.

. Introduction

The Northern Territory of Australia is characterised by largereas of land and a very small population. This situation is suit-ble for remote sensing data and analysis to map natural resourcenformation. A range of remotely sensed data has previously beensed to map land cover in northern Australia, however prior to thistudy, most studies of land cover classification were based on tradi-ional pixel-based methods (Ahmad et al., 1998; Hayder et al., 1999;

enges et al., 2000), although there has been recent interest withinhe Northern Territory in using object-based approaches to mapand cover (Crase and Hempel, 2005; Whiteside and Ahmad, 2004).he nature of land cover in tropical Australia creates some issueselating to pixel-based methods of classification (Whiteside, 2000).ost of the native vegetation in northern Australia is relatively

ntact and has undergone little modification apart from grazing andre (Wilson et al., 1990). Thus, there are large areas of spectrallyimilar but compositionally different vegetation cover or continu-

ms of cover of varying densities of tree cover (Hayder, 2001) thatay be difficult to differentiate, potentially increasing classification

ncertainty. This discontinuous nature of tree cover within savan-

∗ Corresponding author. Tel.: +61 8 89463831; fax: +61 8 89463833.E-mail address: tim.whiteside@batchelor.edu.au (T.G. Whiteside).

303-2434/$ – see front matter © 2011 Elsevier B.V. All rights reserved.oi:10.1016/j.jag.2011.06.008

© 2011 Elsevier B.V. All rights reserved.

nas also contributes to the heterogeneous appearance of savannaswithin imagery (Hutley and Setterfield, 2008; Pearson, 2002). Sim-ilarly, when undertaking a per-pixel classification for the purposesof mapping vegetation cover at a community scale, the spectralheterogeneity within a particular land cover can lead to spurious(misclassified at that scale) pixels appearing within classes creatinga ‘salt and pepper’ effect (Blaschke et al., 2000). In addition to this,the increased application of higher resolution imagery (pixel sizeless than 5 m) is problematic as it is difficult to map communityscale classes accurately using traditional pixel-based methods. Theincreased spectral heterogeneity within land cover classes oftenleads to an inconsistent classification of pixels.

The development of robust object-based image analysis (OBIA)methods suitable for the classification of medium (pixel size10–30 m) to high (pixel size 2–10 m) spatial resolution satelliteimagery provides a valid alternative to the ‘traditional’ pixel-based(PB) methods of analysing and categorising remotely sensed data(Baatz et al., 2004; Benz et al., 2004). Pixels are grouped togetherinto objects or segments based on some criterion of homogene-ity (either spectral or spatial). One of the benefits of using objectsis that in addition to spectral information (for example, the mean

band values for each object), objects have a number of geograph-ical/geometrical features attributed to them such as shape andlength, and topological entities, such as adjacency and, foundwithin (Baatz et al., 2004). These attributes create a knowledge base

Earth Observation and Geoinformation 13 (2011) 884–893 885

fio

ocmApttte

1

tatmtAfwttasaaidccdMvbe

1

aici2(mpiiiKcravMfBavw

T.G. Whiteside et al. / International Journal of Applied

or the sample objects that is far richer in information than that forndividual pixels, and which can be called upon in the developmentf a rule-based classification process.

This paper describes a comparison between a supervisedbject-based classification process and a supervised per-pixellassification for mapping tropical savanna land cover at a com-unity level in the monsoonal north of the Northern Territory,ustralia using medium spatial resolution ASTER data. The com-arison uses the VNIR bands of the ASTER image. Accuracies forhe classifications are produced and compared using statisticalests. A comparison between the object-based process based onhe spectral ASTER data and another incorporating an ASTER digitallevation model (DEM) is also conducted.

.1. Object-based image classification

Object-based (OB) image classification involves the segmenta-ion of an image into homogeneous objects followed by the analysisnd classification of these objects. One of the advantages of segmen-ation is that it creates objects representing land cover types that

ay be spectrally variable at the pixel level and thus eliminateshe salt and pepper effect associated with per-pixel classification.nother advantage is that OB classification uses non-arbitrary units

or analysis as opposed to pixels; objects can approximate realorld features better than pixels. In addition, the use of shape fea-

ures, the hierarchical structures of objects and classes, and theopological features relating to the objects are other benefits of OBpproaches. Object-based analysis enables the construction of ruleets that can be used across a variety of scenes producing a repeat-ble methodology. One of the disadvantages of OBIA is a requisitepriori knowledge of the area and the types of land cover under

nvestigation, which may not necessarily be available. Anotherisadvantage is that the segmentation process and subsequent cal-ulation of the topological relationships described between objectsan utilise a large amount of computer memory. Further, there is noefinitive algorithm or parameters for the creation of image objects.ost assessment of the suitability of segmentation is undertaken by

isual assessment, although recently the use of local variance haseen applied to determine a suitable segmentation scale (Dragutt al., 2010).

.2. Comparison of OBIA and pixel-based image analysis

While there have been some studies comparing object-basednd pixel-based classification techniques little has been conductedn northern Australia. Many publications claim that object-basedlassification has greater potential for classifying higher resolutionmagery than pixel-based methods (Mansor et al., 2002; Oruc et al.,004; Willhauck et al., 2000). However, Dingle Robertson and King2011) found no statistically significant difference between the two

ethods in their study using McNemar’s non-parametric test forroportional difference (de Leeuw et al., 2006), although visual

nspection indicated the OB approach incurred less significant errorn the larger regions of homogeneous cover and performed bettern temporal analysis of land cover change (Dingle Robertson anding, 2011). Niemeyer and Canty (2003) claim that object-basedlassification has greater potential for detecting change in higheresolution imagery. Castillejo-Gonzalez et al. (2009) found thatn object-based method out-performed five pixel-based super-ised classification algorithms (parallelepiped, minimum distance,ahanalobis Distance Classifier, Spectral Angle Mapper, and MLC)

or mapping crops and agro-environmental associated measures.

ased on their reference data, Gao et al. (2006) achieved far greaterccuracy in mapping 12 land cover classes using an object-basedersus a pixel-based approach (83.25% and 46.48%, respectively),hile the object-based methodology used by Gao and Mas (2008)

Fig. 1. Location of the study site.

outperformed both MLC and NN pixel-based methods in map-ping cover using SPOT 5 (10 m spatial resolution) data. Howeverthe authors noted that after smoothing filters were applied to theimagery, the accuracy of the pixel-based methods increased whileobject-based accuracies decreased.

Jobin et al. (2008) noted that one of the advantages of OBIA is theutility of a knowledge base that is beyond purely spectral informa-tion and includes object-related features such as shape, texture andcontext/relationship, along with the capability to include ancillarydata. Manakos et al. (2000) found that ancillary data utilised withinobject-based classification improved classification accuracy. Simi-larly, Devhari and Heck (2009) found the overall accuracy of theobject-based classification was greater than pixel-based and betterwhen ancillary surface data (DEM and contours) were introducedinto the segmentation. Myint et al. (2011) found that including prin-cipal component images and a Normalized Difference VegetationIndex within an object-based rule-set classification produced sig-nificantly higher accuracy than MLC. Yu et al. (2006) incorporatedancillary and derivative data (IHS transformations, elevation slopeand aspect, and a water channel GIS layer) with 4 band airbornedigital camera imagery to map vegetation in Douglas-fir, CaliforniaBay and Coast live oak communities to 60% overall accuracy.

2. Methods

2.1. Study area

The study area is part of the Florence Creek region of Litch-field National Park, in the northwest of the Northern Territory ofAustralia (Fig. 1). The area covers 1373 ha and is located near twoof the park’s major features, Florence Falls and Buley Rockhole. Theregion’s climate is characteristic of the wet/dry tropics; consistingof a long dry season (May–September) with little to no rainfall, andover 75% of the annual rainfall (1500 mm) occurring in the periodbetween November and March. Maximum daily temperatures varyfrom just under 32 ◦C in June and July to over 36 ◦C in October andNovember. The vegetation within the study area is predominantlyopen forest and savanna woodland with a discontinuous Eucalyp-tus spp. (mostly E. tetradonta and E. miniata) dominated canopy andcontinuous annual grass (Sarga spp.) understorey (Griffiths et al.,1997). Patches of monsoon rain forest are located on springs nearthe base of the escarpment and other areas of permanent water.

Melaleuca spp. forests occur along creek lines and share overlappingspecies with the monsoon rain forest (i.e. Xanthostemon eucalyp-toides and Lophostemon lactifluus) (Lynch and Manning, 1988). Thesouthern portion of the study area is generally plateau surfaces

8 Earth Observation and Geoinformation 13 (2011) 884–893

idwgaa

2

v(wiwgnoTrbi

owagpgdiS1((

2

Dccsc(

2

iimettiwhwvtecacht

86 T.G. Whiteside et al. / International Journal of Applied

ntersected by drainage lines, while low lying areas subject to inun-ation are located to the north. At the northern end of the scene,here drainage lines intersect with the plateau edge, a number of

orges and associated waterfalls occur. This terrain will producereas of shading and slope effects which will influence irradiancelthough the area affected is small.

.2. Remote sensing data

ASTER data for the area were captured on 28 July 2000 pro-iding 15 spectral bands; four in the visible and near infrared0.52–0.86 �m) (VNIR) at 15 m nominal pixel size, six in the short-ave infrared (1.60–2.43 �m) at 30 m and five in the thermal

nfrared at 90 m (8.12–11.65 �m) (Yamaguchi et al., 1998). The dataere delivered processed to level surface reflectance with topo-

raphic corrections. The near infrared (NIR) band 3 is captured atadir (3N) and backwards looking (3B), producing a stereo pairf images that can be used to create a DEM (Hirano et al., 2003).he relative DEM image derived from ASTER bands 3N and 3B wasequested from NASA’s EROS Data Center and acquired in Octo-er 2002. A slope layer (%) was derived from the DEM image for

nclusion.Pre-processing included geometric correction and the creation

f subsets for the study area. Geometric rectification of the imageryas undertaken by the authors using a first order polynomial withnearest neighbour interpolation, incorporating the DEM with 25round control points taken from a 1:100,000 topographic maproducing a RMSE of less than 0.5 pixels (7.5 m). Further topo-raphic adjustments of the spectral imagery were not undertakenue to the relatively small area affected by slope and the unsuitabil-

ty of the ASTER DEM in accurately reflecting such small variations.ubsets for the study area were created from the ASTER VNIR bands, 2 and 3N (green, red and NIR nadir, respectively) and the DEMFig. 2). Pixel dimensions of the subset are 249 (columns) × 245rows) for the 15 m ASTER layer.

.3. Object-based classification

Two object-based classifications were undertaken usingefiniens Developer version 7 software. The first classification wasonducted using only the VNIR bands from the ASTER imagery foromparison with the per-pixel classification, while the second clas-ification incorporated information from the ASTER DEM. Both OBlassifications involved two sub-processes: (i) segmentation andii) classification (Fig. 3).

.3.1. Multi-scale segmentationThe object-based approach first involved the segmentation of

mage data into objects on two scale levels. The subset images werenitially segmented into object primitives or segments using the

ultiresolution segmentation algorithm that follows the fractal netvolution approach (Baatz and Schäpe, 2000). The segmentation ofhe images into object primitives is influenced by three parame-ers: scale, colour and form (Willhauck et al., 2000). The algorithms primarily an iterative bottom-up segmentation method starting

ith individual pixels and merging these pixels based upon pixeleterogeneity and object shape and colour. These are determinedithin the algorithm by two parameters; (i) scale and (ii) colour

ersus form. The scale parameter within the algorithm is set byhe operator and is influenced by the heterogeneity of the pix-ls. The colour parameter balances the homogeneity of an object’solour with the homogeneity of its shape. The form parameter is

balance between the smoothness of a segment’s border and its

ompactness. The weighting of these parameters establishes theomogeneity criterion for the object primitives. Visual inspection ofhe objects resulting from a number of segmentations using varia-

Fig. 2. Subsets of ASTER VNIR bands with RGB displaying NIR, Red and Green, respec-tively, (a) and the ASTER DEM (b). (For interpretation of the references to colour inthis figure legend, the reader is referred to the web version of the article.)

tions in the weightings was used to determine the overall values forthe parameter weighting at each scale level (Table 1). Scale param-eters greater than 10 tended to undersegment the image withnoticeable mixes of land covers. Scale parameters smaller than 5

tended to oversegment the image with many adjacent objects of thesame land cover observed. Due to irregular patterns and boundariesof land covers in the natural landscape within the image, empha-sis was placed on the colour (lack of spectral variability) criterion

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884–893 887

Fig. 3. Data flow for object-based classification. Rectangles represent image layersur

oiwsifisTll

2

btaOsataswb

aotwt

TS

Table 2Classes used in this paper. Note: The last two classes WF, CFl, EW1 and WSare interim classes not used in the final classification and thus have no colourassigned to them. (For interpretation of the references to colour in this table,the reader is referred to the web version of the article.)

Colour code Class code Class name

EOFB Burnt Eucalypt open forest

EOF Eucalypt open forest

CFo Mixed closed forest

MelF Melaleuca riparian forest

EW Eucalypt woodland

EWB Burnt Eucalypt woodland

EWRO Eucalypt woodland with rocky outcrops

OW Open woodland

MW Mixed woodland

GL Grassland

- WF Woodland flats

- CFl Creek flats

- EW1 Eucalypt woodland 1

sed in the procedure; circles represent processing and analysis steps; and ‘clouds’epresent objects.

f an object over its shape (60% to 40%, respectively). An increasen the ratio of colour to shape in the algorithm resulted in objects

ith extremely convoluted boundaries, whereas an increase in thehape parameter saw regular shaped segments not representingdentifiable land cover boundaries. Two levels were chosen as therst level provided objects at a suitable scale for mapping howeverome of the objects still contained noticeable spectral variability.hus, a finer scale was introduced to produce smaller objects ofess variability that could then be combined into suitable classesater.

.3.2. ClassificationA total of ten land cover classes for the study area were identified

ased on the structural formation of the vegetation and characteris-ic taxon (Table 2). Two of these classes were introduced to includereas of the study site that were identified as recently burnt. TheB classification involved supervised classification using objects

elected for training data based on their class as determined bycombination of field observation and aerial photo interpreta-

ion. The mean spectral values of the training objects were useds input into the Nearest Neighbour algorithm which uses featurepace to classify objects based on the closest training examplesith an object being assigned the class of the majority of its neigh-

ours.The OB classification procedure incorporating the DEM involved

combination of a supervised classification and the developmentf class rules based upon the mean spectral signatures, as well as

he DEM and slope values of the objects. Samples for each classere selected from the image objects to act as training areas for

he classification.

able 1egmentation parameters used at both scale levels.

Scale level Scale parameter Shape factor Compactness Smoothness

2 10 0.4 0.7 0.31 5 0.2 0.7 0.3

- WS Woodland slope

Details of the multilevel OB classification are depicted in Fig. 5.Initial classification of the broad scale (Level 2) segmentation wasconducted using the Nearest Neighbour (NN) classification algo-rithm with the relevant training samples shown in Fig. 4. The NNclassification used the object means of all three spectral bands tocreate five broad classes (Riparian, EOF, EW, GL, and MF). The finerscale Level 1 objects were initially categorised based on the classassigned to their Level 2 super-object. Further division of classeswas then undertaken using NN supervised classification based onthe sample objects (Fig. 4). The samples were then used as a basisfor fuzzy classification of the data in Definiens (Baatz et al., 2004).Fuzzy classifications use fuzzy logic to account for the heteroge-neous nature (consisting of multiple land covers) of pixels (or in thiscase objects) in medium spatial resolution imagery as well as thelack of hard, sharp boundaries in nature where classes grade intoone another (Foody, 1992). Where there exists m number of classes,pixels are assigned m class memberships providing the degree oftruth to which pixels (or objects) belong to each class (Jensen,2005). The class with the highest membership (or truth) value isassigned as the best and final class for each object (Baatz et al.,2004). After the Level 1 classification step, the interim WF (‘Wood-land flats’) and CFl (‘Creek flats’) objects were then merged into MW(‘Mixed woodland’) class producing 10 classes for the purpose ofcomparison with the pixel-based classification and reference data(Fig. 5).

2.4. Object-based classification with ancillary DEM layer

A second object-based classification was conducted this time

incorporating the DEM data. The procedure using training sam-ple objects and the NN algorithm was the same as for the initialobject-based classification; however two DEM based class ruleswere introduced for the purpose of classifying the areas of the

888 T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884–893

Fig. 4. Sample objects used for the object-based NN classification at levels 1 and2 draped over the ASTER VNIR imagery (RGB = bands 1, 2, 3N). Colours and landcover codes correspond to those in Table 2. Points are locations of sites used foraccuracy assessment. Circles are field sites, triangles are derived from aerial photoit

swolfsfi

Table 3Reference data details.

Year No. sites

2000 12002 112003 292004 202006 42006 100

sification of the communities. Data included the dominant taxon

nterpretation. (For interpretation of the references to colour in this figure legend,he reader is referred to the web version of the article.)

ubset where irradiance was affected by slope. Level 2 ‘Eucalyptoodland’ objects were classified at level 1 as ‘Open woodland’

bjects if their DEM value was less than 140 m AMSL and as ‘Wood-and slope’ if their slope was greater than 8%. The threshold values

or these class rules distinguished the affected areas. ‘Woodlandlope’ objects were then reclassified as ‘Eucalypt woodland’ for thenal output.

Fig. 5. Classification processes for the object-based lan

Air photo 91

Total 256

2.5. Pixel-based supervised classification

The pixel-based classification used a supervised maximum like-lihood (MLC) algorithm (Jensen, 2005; Lillesand and Kiefer, 2008).50 training areas representative of the ten land cover classes (Fig. 4)were selected to develop class signature files similar to those usedin the object-based classification to ensure consistency. The MLCthen assigned pixels to the class of highest probability producing aland cover map for these 10 classes.

2.6. Accuracy assessment

The accuracy of the fuzzy classification was estimated using themean membership value of the best classification (class with high-est membership value) and the mean stability for each class (Baatzet al., 2004). The mean stability is the difference between an object’sfirst and second class memberships and is expressed as a percent-age – the higher the value the greater the stability and the lessexpected mixing of the classes.

The accuracy of the classified image was also assessed using arange of reference data including field data collected in the studyarea between 2000 and 2006 (Table 3) and interpretation of aerialphotography of the area. Each campaign has collected vegetationdata that have enabled the determination of the structural clas-

(Genus or species), canopy height, stem diameter (at 130 cm abovethe ground), basal area, foliage projective cover, canopy cover andcanopy density. Vegetation structural classes were determined

d cover classifications based on spectral bands.

Earth Observation and Geoinformation 13 (2011) 884–893 889

b2tuo5oecsi

mrais

t1fcCsfic2

bddM(i2

�

wilRtD

3

6(cTsoa

onWc‘ssit

T.G. Whiteside et al. / International Journal of Applied

ased on the standard systems used in Australia (Hnatiuk et al.,009; Specht, 1981). In areas that were inaccessible (due to cul-ural restrictions or terrain), aerial photograph interpretation wasndertaken to delineate the vegetation community. A stereo pairf 1:43,000 colour positive aerial photographs taken at 1500 h onMay 2000 covering the study area were observed under a stere-

scope with 6× magnification. Cover calculations after Fenshamt al. (2002) and dominant taxon were identified and structurallasses determined over 90 points derived from a stratified randomampling over the area. Locations of all reference sites are shownn Fig. 4.

For ease of comparison between classification methods, the-atic accuracy was undertaken using only the point-based

eference data mentioned above. There was no consideration ofny object-based accuracy assessment or accuracy measures relat-ng to the geometric accuracy of the objects (such as location andhape).

Accuracy assessments of all three classifications were under-aken using confusion matrices and Kappa statistics (Congalton,991). Producer and User accuracies (Story and Congalton, 1986)or each class were calculated along with the overall accura-ies and Kappa coefficient statistics (Congalton and Green, 1999).onditional Kappa was calculated for each class. The statisticalignificance of the difference between the two overall Kappa coef-cients for the pixel-based classification and the object-basedlassification was assessed using a Z-test (Congalton and Green,009).

As one set of reference data was used to assess the accuracy ofoth classifications, the two confusion matrices may show depen-ence and thus a further test for statistical significance of theifference between the two classifications was undertaken usingcNemar’s test (de Leeuw et al., 2006). The test is non-parametric

Eq. (1)) assuming that the number of correctly and incorrectlydentified points is equal for both classifications (de Leeuw et al.,006; Dingle Robertson and King, 2011; Gao et al., 2006).

2 = (f12 − f21)2

f12 + f21(1)

here f12 and f21, respectively, are the number of points correctlydentified by one classifier and not the other. The statistic fol-ows a chi-squared distribution with one degree of freedom (Dingleobertson and King, 2011). The first object-based classification washen compared to the object-based classification incorporating theEM data using the same statistical tests.

. Results

Within the object-based analysis, the segmentation provided15 level 2 objects and 1495 level 1 objects for classificationFig. 6a). At Level 2 it is visually apparent that objects of this size stillan contain more than one spectrally distinct land cover (Fig. 6b).he objects within the segmentation at level 1 tend to follow thepectral boundaries more closely. The image resulting from thebject-based classification is shown in Fig. 7a. For comparison, therea (ha) assigned to each class is presented in Table 4.

According to the object-based classification the land cover classccupying the largest area is ‘Eucalypt woodland’ (376 ha) with theumber of objects identified as belonging in that class being 675.ithin the per-pixel classification only 256 ha were assigned to this

lass. The class with the largest area in the per-pixel classification isEucalypt open forest’ with 302 ha (140 ha in the object-based clas-

ification). The land cover with the smallest area within the studyite is ‘Grassland’ occupying only 15 ha and consisting of just 21mage objects in the OB classification and 16 ha in the PB classifica-ion. A visual comparison of the resultant land cover images shows

Fig. 6. A section of the study area showing hierarchical segmentation at scale level2 (a) and level 1 (b).

noticeable differences between the classifications (Fig. 7). Whileboth methods produce aggregations of pixels based on land coverclasses, the object-based classification yields multi-pixel featureswhereas the pixel-based classification contains many small groupsof pixels or individual pixels. This produces classes with mixed clus-ters of pixels as displayed by the heterogeneous nature of the PBclassified image.

The areas of classes with cover that is more spectrally homo-geneous such as ‘Mixed closed forest’, ‘Melaleuca riparian forest’and ‘Grassland’ classes are relatively similar in both OB and PB

classifications (Fig. 7 and Table 4). The area classified as ‘Euca-lypt woodland’ is noticeably lower in the pixel-based classificationcompared to the two object-based classifications. This is supportedby visual comparison with the original image showing ‘Eucalypt

890 T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884–893

F

wc‘‘

t

Table 4Areas of classes determined by object-based and pixel-based classifications.

Class name Area (ha)

Object-based OB with DEM Pixel-based

Burnt Eucalypt open forest 228.38 228.38 188.53Eucalypt open forest 140.54 140.54 302.24Mixed closed forest 19.58 19.58 21.94Melaleuca riparian forest 168.09 168.09 173.68Eucalypt woodland 376.02 368.24 256.10Burnt Eucalypt woodland 114.39 114.39 77.38Eucalypt woodland with

rocky outcrops42.59 42.59 118.19

Open woodland 57.55 65.33 37.59

ig. 7. Resultant images of the object-based (a) and pixel-based (b) classifications.

oodland’ classes apparently under-represented in the pixel-basedlassification, while there is an apparent over-representation ofEucalypt open forest’, ‘Eucalypt woodland with rocky outcrops’ and

Grassland’ classes.

A summary of the accuracy assessment of the fuzzy classifica-ion for both object-based classifications is presented in Table 5.

Mixed woodland 208.84 208.84 180.86Grassland 15.18 15.18 16.11

The Mean p1st column displays the mean best classification ofobjects belonging to that class. Nearly all classes have a meanp1st value over 0.8 suggesting that the degree to which imageobjects were being assigned to the best class was high. Within thefirst object-based classification, the class with the lowest meanbest classification value is ‘Open woodland’ with 0.68. Most ofthe classes display stability in their classification; only the ‘Euca-lypt open forest’, ‘Melaleuca riparian forest’ and ‘Mixed woodland’classes have a mean stability value under 0.1. This suggests uncer-tainty in some much as a number of objects within these classesmay not belong to their assigned class or potentially contain morethan one class.

From the results of the confusion matrices based on the refer-ence data set, the proportions of pixels classified due to chanceagreement (Congalton and Green, 2009) were 13% for the OB clas-sification and 12.5% for the PB classification. The overall accuracyof the object-based classification was higher than for the pixel-based classification, 78.5% versus 69.5%, respectively (Table 6).This was also the case for the overall Kappa statistic: the object-based classification had an overall Kappa of 0.74 with a standarderror of ±0.03 while the pixel-based classification’s overall Kappastatistic was 0.65 with a standard error of ±0.03. The resul-tant Z-value (Table 7) calculated from the comparison of Kappacoefficients from the two classification methods was 2.29, whichis greater than the critical value for Z at the 95% confidencelevel (1.96). This indicates a statistically significant differencebetween the results of the two classifications. The �2 value fromthe McNemar’s test (Table 7) was 8.97 with a p-value of 0.01,strengthening the argument for a statistically significant differ-ence between the two classification methods. In addition, theProducer and User accuracies were greater for the majority of theclasses in the object-based classification. Five classes in the OBclassification had User and Producer accuracies over 70%, whileonly three did for the PB classification. Six of the classes fromthe OB classification had conditional Kappa coefficients of 0.75or greater while the PB classification had four. The land coverclasses that were more accurately classified using the pixel-basedmethod were the heavily treed classes of homogeneous cover,‘Melaleuca riparian forest’ and ‘Mixed closed forest’. The classesthat had poor User’s accuracy in both classifications were ‘Mixedwoodland’ and ‘Grassland’ and may, in part, be attributable tothe small number of reference data points for those classes (19and 4, respectively), as any reference site erroneously includedin these classes will have an effect. Object-based classificationappears to be able to differentiate more accurately the relativelyheterogeneous ‘Eucalypt open forest’ and the several woodlandclasses.

Comparison of the OB and OBDEM classifications shows aslightly higher accuracy for the OBDEM classification. The twoclasses (‘Eucalypt woodland’ and ‘Open woodland’) influenced by

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884–893 891

Table 5Fuzzy classification accuracy assessment of the initial object-based image analysis and with DEM.

Class Class name Object-based Object-based with DEM

Mean p1st Mean stability Mean p1st Mean stability

EOFB Burnt Eucalypt open forest 0.880 0.136 0.880 0.136EOF Eucalypt open forest 0.827 0.092 0.827 0.092MFo Mixed closed forest 0.810 0.362 0.810 0.362MelF Melaleuca riparian forest 0.796 0.048 0.796 0.048EW Eucalypt woodland 0.832 0.145 0.858 0.215EWB Burnt Eucalypt woodland 0.856 0.373 0.856 0.373EWRO Eucalypt woodland with rocky outcrops 0.839 0.362 0.839 0.362OW Open woodland 0.676 0.373 0.756 0.385MW Mixed woodland 0.874 0.048 0.874 0.048GL Grassland 0.840 0.138 0.840 0.138

Table 6Summary of confusion matrices for the accuracy of object-based and pixel-based classifications including the Producer’s accuracy (PA), User’s accuracy (UA) and conditionalKappa (CK) for each class as well as overall accuracy, probability of chance results and Kappa statistic.

Class name Object-based classification Object-based classification with DEM Pixel-based classification

PA (%) UA (%) CK PA (%) UA (%) CK PA (%) UA (%) CK

Burnt Eucalypt open forest 91.00 75.00 0.68 91.00 75.00 0.68 63.64 65.63 0.61Eucalypt open forest 81.00 76.32 0.73 81.00 76.32 0.73 75.00 57.45 0.50Mixed closed forest 100.0 66.67 0.66 100.0 66.67 0.66 91.67 100.0 1.00Melaleuca riparian forest 67.74 80.77 0.75 67.74 80.77 0.75 87.10 90.00 0.89Eucalypt woodland 77.42 90.57 0.88 77.77 92.45 0.89 50.00 75.61 0.68Burnt Eucalypt woodland 85.71 80.00 0.78 85.71 80.00 0.78 84.62 55.55 0.53Eucalypt woodland with rocky outcrops 66.67 100.0 1.00 66.67 100.0 1.00 63.63 87.5 0.85Open woodland 91.67 91.67 0.87 92.31 92.31 0.88 83.33 76.92 0.76Mixed woodland 73.68 50.00 0.43 73.68 50.00 0.43 80.00 51.61 0.48Grassland 50.00 50.00 0.51 50.00 50.00 0.51 75.00 42.86 0.39

tai(

4

ptiMaahi

mlo

TCc

TCa

Overall accuracy 78.51%Chance results 13.15%Kappa 0.7526

he rule set have higher Producer and User accuracies. The over-ll accuracy and Kappa were also higher although the differencen accuracies (based on Kappa) was not statistically significantTable 8).

. Discussion

The object-based image analysis method applied in this paperrovided results with statistically significant higher accuracieshan the pixel-based classification. This is consistent with find-ngs within the literature (Castillejo-Gonzalez et al., 2009; Gao and

as, 2008; Gao et al., 2006). This result suggests that object-basednalysis has potential as an alternative method (over per-pixelpproaches) for extracting land cover information from medium toigh resolution satellite imagery captured over tropical savannas

n Australia.

The improved classification using OBIA can be attributed pri-

arily to its use of objects to reduce the spectral variability inand cover types that are heterogeneous such as savanna. Basedn a community level classification such as used here, pixel-based

able 7omparison statistics Z-test and McNemar’s �2 for pixel-based and object-basedlassifications.

Z-value for Kappas 2.285 p-value 0.05 (p = 0.01)McNemar’s �2 8.966 p-value 0.01 (p = 0.0028)

able 8omparison statistics Z-test and McNemar’s �2 for the Object-based classificationnd Object-based with DEM classification.

Z-value for Kappas 0.2175 p = 0.41McNemar’s �2 0.995 p = 0.35

79.30% 69.53%13.24% 12.5%

0.7611 0.6516

classifications do misclassify pixels, particularly in land covers thatare spectrally heterogeneous in medium (pixels of 30 m or less)and high (pixels of 5 m or less) spatial resolution imagery, suchas tropical savanna. Classifying savanna into discrete communitylevel classes (such as forest or woodland) is difficult with per-pixel approaches due to their sensitivity to the discontinuous andvariable nature of the woody cover within such landscapes. Forexample, within the pixel-based classification many of the pixelsclassified as ‘Eucalypt woodland with rocky outcrops’ are actuallygrass understorey or bare ground between tree crowns where astree crown canopy itself has been classified as ‘Open forest’ as theco-dominants within savanna woodland are trees and grass. Thecombination of the two classes should produce a ‘Woodland’ clas-sification. This would account for the greater confusion found inthe Eucalypt dominant classes as mentioned in the results wherethe ‘Eucalypt woodland’ classes are under-represented while thereis an over-representation of the ‘Eucalypt open forest’, ‘Eucalyptwoodland with rocky outcrops’ and ‘Grassland’ classes. Object-based classification appears to overcome some of the problemsencountered using pixel-based methods to classify communitylevel Eucalypt land cover types and their characteristic spatial het-erogeneity (Pearson, 2002), while it is evident that pixel-basedclassification is still quite successful in classifying land cover of aspectrally homogenous nature (i.e. ‘Mixed closed forest’).

One of the advantages of object-based classification is the abilityto use ancillary data (such as derivative data sets, data from othersensors and existing GIS layers) as additional information layers toassist with land cover mapping is being investigated. Another fea-ture is the utility of contextual information related to objects. One

such data set as used in this study is the incorporation of heightinformation in the form of DEM and derivative slope layer. Suchinformation enables better delineation of land cover types thatoccupy certain locations within the landscape such as on slopes

8 Earth

a(

matwutsasocl(oa

cgacrip(wsuclttmtttc

5

aafRmadcstscc

R

A

B

92 T.G. Whiteside et al. / International Journal of Applied

nd drainage flats. This is consistent with findings of other studiesDevhari and Heck, 2009; Manakos et al., 2000).

Comparing object-based and pixel-based classifications in thisanner presents a number of issues. While, both methods producethematic map, the object-based approach appears to eliminate

he salt and pepper or noise effect by considering mean pixel valuesithin objects as opposed to individual pixel values. However, insing the ‘traditional’ point-based method of accuracy assessment,he comparison between point-based reference data and the clas-ification is based on ‘individual’ or site-specific cases (i.e. pixels inpixel-based classification and specific points within objects). As

uch, information relating to the spatial accuracy of the classifiedbjects is not considered (Lucieer, 2004). For example, is that classonsistent across the entire object? Is the object located in a similarocation to the real-world object it is representing? To what degreearea, length, etc.) is the shape of the object similar to the real-worldbject? This information is important in determining the quality ofn object-based classification.

The tendency to produce spurious or misclassified pixels withinlasses (the so-called ‘salt & pepper’ effect) means that hetero-eneous land covers will nearly invariably have slightly lowerccuracies for pixel-based classifications than object-based usinglasses such as used here. Part of this may attributed to mis-egistration between the imagery and field data. Methods tomprove accuracies for pixel-based classifications include someost-classification editing such as filtering and manual removalJohansen and Phinn, 2006). Potential under-evaluation will occurithin certain classes that are heterogeneous in cover such as

avanna in which cover is co-dominated by grass and discontin-ous and variable woody cover. Trees will be assigned to a forestlass and understorey gaps between trees will be assigned to grass-and class. Thus it may be necessary to redefine classes away fromhe ‘traditional’ land cover or vegetation classes into more contex-ual classes, e.g. canopy versus non-canopy or using quantitative

easures (e.g. % canopy cover). This is where the hierarchical struc-ure of OB classification has potential in enabling the use of theseypes of classes at a particular level and then the proportions ofhese classes in super-objects at higher level to determine level ofanopy cover in those objects.

. Conclusions

The results of this study show a significant difference in theccuracy between a pixel-based maximum likelihood classifiernd object-based hybrid nearest-neighbour/rule-based classifieror mapping land cover from tropical savanna using ASTER data.esultant noise in the pixel-based classification suggests that the-atic mapping using high spatial resolution satellite data requiresnew methodology in land cover classification forgoing the tra-

itional community level classifications for the initial stages oflassification and perhaps focussing on the smaller spatial elementsuch as tree/canopy, grass, and bare ground. These objects couldhen be the basis to develop the community/structural level clas-ification grouping based on the proportional values of the variousomponents. This is where OBIA has great advantage over per-pixellassification methods.

eferences

hmad, W., Hill, G.J.E., O’Grady, A.P., Menges, C., 1998. Use of multispectral scannerdata for the identification and mapping of tropical forests of northern Australia.

Asian-Pacific Remote Sensing and GIS Journal 11, 14–22.

aatz, M., Schäpe, A., 2000. Multiresolution segmentation – an optimizationapproach for high quality multi-scale image segmentation. In: Strobl, J.,Blaschke, T., Griesebner, G. (Eds.), Angewandte Geographische Informationsver-arbeitung XII. Wichmann-Verlag, Heidelberg, pp. 12–23.

Observation and Geoinformation 13 (2011) 884–893

Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., Hofmann, P., Lingenfelder,I., Mimler, M., Sohlbach, M., Weber, M., Willhauck, G., 2004. eCognition Profes-sional: User Guide 4. Definiens-Imaging, Munich.

Benz, U., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-readyinformation. ISPRS Journal of Photogrammetry and Remote Sensing 58,239–258.

Blaschke, T., Lang, S., Lorup, E., Strobl, J., Zeil, P., 2000. Object-oriented image pro-cessing in an integrated GIS/remote sensing environment and perspectives forenvironmental applications. In: Cremers, A., Greve, K. (Eds.), EnvironmentalInformation for Planning, Politics and the Public. Metropolis Verlag, Marburg,pp. 555–570.

Castillejo-Gonzalez, I.L., López-Granados, F., Garcia-Ferrer, A., Pena-Barragan,J.M., Jurado-Expósito, M., Sanchez-de la Orden, M., Gonzalez-Audicana, M.,2009. Object- and pixel-based analysis for mapping crops and their agro-environmental associated measures using QuickBird imagery. Computers andElectronics in Agriculture 68, 207–215.

Congalton, R.G., 1991. A review of assessing the accuracy of classifications ofremotely sensed data. Remote Sensing of Environment 37, 35–46.

Congalton, R.G., Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data:Principles and Practice. Lewis Publishers, New York.

Congalton, R.G., Green, K., 2009. Assessing the Accuracy of Remotely Sensed Data:Principles and Practices, 2nd ed. CRC Press, Boca Raton, FL.

Crase, B., Hempel, C., 2005. Object based land cover mapping for Groote Eylandt: atool for reconnaissance and land-based surveys. In: Proceedings of NARGIS 05,the North Australian Remote Sensing and GIS Conference, Darwin, July 4–7.

de Leeuw, J., Jia, H., Yang, L., Liu, X., Schmidt, K., Skidmore, A.K., 2006. Compar-ing accuracy assessments to infer superiority of image classification methods.International Journal of Remote Sensing 27, 223–232.

Devhari, A., Heck, R.J., 2009. Comparison of object-based and pixel based infraredairborne image classification methods using DEM thematic layer. Journal ofGeography and Regional Planning 2, 86–96.

Dingle Robertson, L., King, D.J., 2011. Comparison of pixel- and object-based clas-sification in land cover mapping. International Journal of Remote Sensing 32,1505–1529.

Dragut, L., Tiede, D., Levick, S.R., 2010. ESP: a tool to estimate scale parameter for mul-tiresolution image segmentation of remotely sensed data. International Journalof Geographical Information Science 24, 859–871.

Fensham, R.J., Fairfax, R.J., Holman, J.E., Whitehead, P.J., 2002. Quantitative assess-ment of vegetation structural attributes from aerial photography. InternationalJournal of Remote Sensing 23, 2293–2317.

Foody, G.M., 1992. A fuzzy sets approach to the representation of vegetation continuafrom remotely sensed data: an example from lowland health. PhotogrammetricEngineering and Remote Sensing 58, 221–225.

Gao, Y., Mas, J.F., 2008. A comparison of the performance of pixel-based andobject-based classifications over images with various spatial resolutions. In:Proceedings of GEOBIA 2008 – Pixels, Objects, Intelligence: Geographic Object-Based Image Analysis for the 21st Century, Calgary, Alberta, August 6–7.

Gao, Y., Mas, J.F., Maathius, B.H.P., Xiangmin, Z., van Dijk, P.M., 2006. Comparison ofpixel-based and object oriented image classification approaches – a case studyof a coal fire area, Wuda, inner Mongolia, China. International Journal of RemoteSensing 27, 4039–4055.

Griffiths, A.D., Woinarski, J.C.Z., Armstrong, M.D., Cowie, I.D., Dunlop, C.R., Horner,P.G., 1997. Biological Survey of Litchfield National Park, Report No. 62. PWCNT,Darwin.

Hayder, K., 2001. Study of remote sensing and GIS for the assessment of their capa-bilities in mapping the vegetation form and structure of tropical savannas inNorthern Australia. Unpublished Ph.D. Thesis. Northern Territory University,Darwin, 315 p.

Hayder, K., Ahmad, W., Williams, R.J., 1999. Use of varied resolution imagery forland cover mapping in a semi-arid tropical savanna at Kidman Springs, NorthernTerritory, Australia. In: Proceedings of NARGIS 99, 4th North Australian RemoteSensing and GIS Conference, Darwin, NT, June 28–30.

Hirano, A., Welch, R., Lang, H., 2003. Mapping from ASTER stereo image data:DEM validation and accuracy assessment. ISPRS Journal of Photogrammetry andRemote Sensing 57, 356–370.

Hnatiuk, R.J., Thackway, R., Walker, J., 2009. Vegetation. Australian Soil and LandSurvey Field Handbook. CSIRO, Melbourne.

Hutley, L.B., Setterfield, S.A., 2008. Savanna. In: Jorgensen, S.E., Fath, B.D. (Eds.),Encyclopedia of Ecology, vol. 4: Ecosystems. Elsevier, Oxford.

Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote Sensing Perspec-tive, 3rd ed. Prentice Hall, Upper Saddle River, NJ.

Jobin, B., Labrecque, S., Grenier, M., Falardeau, G., 2008. Object-based classificationas an alternative approach to the traditional pixel-based classification to identifypotential habitat of the Grasshopper Sparrow. Environmental Management 41,20–31.

Johansen, K., Phinn, S., 2006. Mapping structural parameters and species composi-tion of riparian vegetation using IKONOS and landsat ETM+ data in AustralianTropical Savannahs. Photogrammetric Engineering & Remote Sensing 72,71–80.

Lillesand, T.M., Kiefer, R.W., 2008. Remote Sensing and Image Interpretation, 6th ed.

Wiley, New York/Hoboken.

Lucieer, A., 2004. Uncertainties in segmentation and their visualisation. Unpub-lished Ph.D. Thesis. International Institute for Geo-Information Scienceand Earth Observation (ITC) and the University of Utrecht, Netherlands,177 p.

Earth

L

M

M

M

M

N

O

P

S

T.G. Whiteside et al. / International Journal of Applied

ynch, B.T., Manning, K.M., 1988. Land Resources of Litchfield Park, Report No. 36.Conservation Commission of the Northern Territory, Darwin.

anakos, I., Schneider, T., Ammer, U., 2000. A comparison between the ISODATAand the eCognition classification on basis of field data. In: Proceedings of XIXISPRS Congress, Amsterdam, July 16–22.

ansor, S., Hong, W.T., Shariff, A.R.M., 2002. Object oriented classification for landcover mapping. In: Proceedings of Map Asia 2002, Bangkok, GIS Development,August 7–9.

enges, C.H., Bell, D., van Zyl, J.J., Ahmad, W., Hill, G.J.E., 2000. Image classificationof AIRSAR data to delineate vegetation communities in the tropical savannasof northern Australia. In: Proceedings of 10th Australian Remote Sensing &Photogrammetry Conference, Adelaide, August 21–25, pp. 1286–1298.

yint, S.W., Gober, P., Brazel, A., Grossman-Clarke, S., Weng, Q., 2011. Per-pixelvs. object-based classification of urban land cover extraction using high spatialresolution imagery. Remote Sensing of Environment 115, 1145–1161.

iemeyer, I., Canty, M.J., 2003. Pixel-based and object-oriented change detectionanalysis using high-resolution imagery. In: Proceedings of 25th Symposium onSafeguards and Nuclear Material Management, Stockholm, May 13–15.

ruc, M., Marangoz, A.M., Buyuksalih, G., 2004. Comparison of pixel-based andobject-oriented classification approaches using Landsat-7 ETM spectral bands.In: Proceedings of ISPRS Conference, Istanbul, July 19–23.

earson, D.M., 2002. The application of local measures of spatial autocorrelationfor describing pattern in north Australian landscapes. Journal of EnvironmentalManagement 64, 85–95.

pecht, R.L., 1981. Foliage projective cover and standing biomass. In: Gillison, A.N.,Anderson, D.J. (Eds.), Vegetation Classification in Australia: Proceedings of a

Observation and Geoinformation 13 (2011) 884–893 893

Workshop Sponsored by CSIRO Division of Land Use Research, October 1978.CSIRO and ANU Press, Canberra.

Story, M., Congalton, R.G., 1986. Accuracy assessment: a user’s perspective. Pho-togrammetric Engineering and Remote Sensing 52 (3), 397–399.

Whiteside, T., 2000. Multi temporal land cover change in the Humpty Doo regionNT, 1990–1997. Unpublished Master of Natural Resources Management Thesis.University of Adelaide, Adelaide, 90 p.

Whiteside, T., Ahmad, W., 2004. Object-oriented classification of ASTER imageryfor landcover mapping in monsoonal northern Australia. In: Proceedings of12th Australasian Remote Sensing and Photogrammetry Conference, Fremantle,Spatial Sciences Institute, October 18–22.

Willhauck, G., Schneider, T., De Kok, R., Ammer, U., 2000. Comparison of object-oriented classification techniques and standard image analysis for the use ofchange detection between SPOT multispectral satellite images and aerial photos.In: Proceedings of XIX ISPRS Congress, Amsterdam, July 16–22.

Wilson, B.A., Brocklehurst, P.S., Clark, M., Dickinson, K.J.M., 1990. Vegetation Surveyof the Northern Territory, Australia, Report No. 49. Conservation Commission ofthe Northern Territory, Palmerston.

Yamaguchi, Y., Kahle, A.B., Tsu, H., Kawakami, T., Pniel, M., 1998. Overviewof Advanced Spaceborne Thermal Emission and Reflection Radiometer(ASTER). IEEE Transactions on Geoscience and Remote Sensing 36, 1062–

1071.

Yu, Q., Gong, P., Clinton, N., Biging, G.S., Kelly, M., Schirokauer, D., 2006. Object-based detailed vegetation classification with airborne high spatial resolutionremote sensing imagery. Photogrammetric Engineering and Remote Sensing 72,799–811.