This article was downloaded by: [Beijing Normal University]On: 13 January 2014, At: 21:55Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Remote Sensing LettersPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/trsl20

Multi-scale segmentation approach forobject-based land-cover classificationusing high-resolution imageryLei Zhanga, Kun Jiab, Xiaosong Lia, Quanzhi Yuana & Xinfeng Zhaoa

a Key Laboratory of Digital Earth Science, Institute of RemoteSensing and Digital Earth, Chinese Academy of Sciences, Beijing,Chinab College of Global Change and Earth System Science, BeijingNormal University, Beijing, ChinaPublished online: 06 Jan 2014.

To cite this article: Lei Zhang, Kun Jia, Xiaosong Li, Quanzhi Yuan & Xinfeng Zhao (2014) Multi-scale segmentation approach for object-based land-cover classification using high-resolutionimagery, Remote Sensing Letters, 5:1, 73-82

To link to this article: http://dx.doi.org/10.1080/2150704X.2013.875235

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &

Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

Multi-scale segmentation approach for object-based land-coverclassification using high-resolution imagery

Lei Zhanga, Kun Jiab*, Xiaosong Lia, Quanzhi Yuana, and Xinfeng Zhaoa

aKey Laboratory of Digital Earth Science, Institute of Remote Sensing and Digital Earth, ChineseAcademy of Sciences, Beijing, China; bCollege of Global Change and Earth System Science, Beijing

Normal University, Beijing, China

(Received 30 October 2013; accepted 8 December 2013)

Image segmentation is a basic and important procedure in object-based classification ofremote-sensing data. This study presents an approach to multi-scale optimal segmenta-tion (OS), given that single-scale segmentation may not be the most suitable approachto map a variety of land-cover types characterized by various spatial structures; itobjectively measures the appropriate segmentation scale for each object at variousscales and projects them onto a single layer. A 1.8 m spatial resolution Worldview-2image was used to perform successive multi-scale segmentations. The pixel standarddeviation of an object was used to measure the optimal scale that occurred on thelongest, feature unchanged scale range during multi-scale segmentation. Results indi-cate that the classification of multi-scale object OS can improve the overall accuracyby five percentage points compared to traditional single segmentation.

1. Introduction

As very high spatial resolution (VHR) remote-sensing images become increasingly avail-able, high intra-class spectral variability or spectral heterogeneity in such high-resolutionimages severely limits the application of traditional pixel-based approaches in which theso-called salt-and-pepper effect and the limitation to spectral information are handicapsfor the recognition of semantic classes (Yu et al. 2006). Geographic object-based imageanalysis (GEOBIA) was developed through the realization that image-objects hold morereal-world value than pixels alone (Smith et al. 2007). Object-based techniques identifyland-cover elements based on the spectral characteristics of object features rather thanpixel features. Such techniques utilize spatial concepts that involve geometric features,spatial relation and the scale topology relation of upscale and downscale inheritances(Burnett and Blaschke 2003). Image-objects often can be readily related to landscapeelements; semantic features enable object-based classification more accurately than thepixel-based approach, object analysis is particularly appropriate for categorizing land useand land cover using high spatial resolution imagery (Cleve et al. 2008).

The first critical step in object-based classification is the creation of image-objects withwithin-object homogeneity. Land-cover characteristics are sensitive to and dependent on thescales of image or segmented image, significant objects of many land-cover types, even forthe same land-cover type, do not appear at the same segmentation level (Arbiol, Zhang, andPalá 2006; Trias-Sanz, Stamon, and Louchet 2008; Anders, Seijmonsbergen, and Bouten2011), the selection of an appropriate segmentation scale is important and is currently a hot

*Corresponding author. Email: jiakun@bnu.edu.cn

Remote Sensing Letters, 2014Vol. 5, No. 1, 73–82, http://dx.doi.org/10.1080/2150704X.2013.875235

© 2014 Taylor & Francis

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

research topic in object-based classification (Hay et al. 2001; Powers, Hay, and Chen 2012).Image-objects should ideally be of roughly the same size as or slightly smaller than thefeature to be classified (Anders, Seijmonsbergen, and Bouten 2011). Defining the mostappropriate scale for image segmentation is currently problematic, because no single(spatial) scale is optimal for characterizing the multitude of different scene components.Segmentation does not automatically lead to meaningful objects in the real world, but to‘relatively homogeneous areas.’ The latter term better reflects the ‘near-decomposability’ ofnatural systems (Walker and Blaschke 2008). This current study addresses that a standar-dized or widely accepted guideline to optimize the homogeneity setting (scale parameter) ofthe segmentation is not available (Kim et al. 2011). A unique scale is not suitable for theanalysis of geographic elements in remote sensing (Ardila et al. 2012), and scale analysis isa potential limitation for object-based classification (Desheng and Fan 2010).

Multi-scale analysis and representation require many appropriate scales to account forvarious levels of organization in a landscape structure (Drăguţ, Tiede, and Levick 2010).Previous studies are concentrated on optimum of segmentation parameters (scale, shape,compactness and smoothness) by segmentation assessment (Tian and Chen 2007), otherstudies are to measure the optimal scales to match various land covers using trial-and-errormethods (Kim et al. 2011), multi-scale analysis is time-consuming when it applies tocomplex landscape and increases data processing and storage demands. Therefore, it isdesirable to identify the minimum number and configuration of object scales (hierarchicallevels) that maximize the potential final land-cover classification accuracy when combinedwhile minimizing excess data processing (Smith 2010). Nevertheless, multi-scale analysiscould not effectively meet the demand of all land-cover types, the reason is that the extentand rate of each object change of land cover across segmentation scales are not equal.

This study proposes a multiple scale optimal measurement for the spatial domain ofeach object rather than for an entire scene, and then projects each optimal object frommultiple scales onto a single segmentation layer. This process optimizes each object ofsegmentation level and improves classification accuracy.

2. Data and methods

2.1. Data and study area

A VHR Worldview-2 image with 1.8 m spatial resolution and multispectral mode of themiddle section of Hanjiang, a tributary of Yangze River in China, was acquired for thisstudy. This image was acquired on 29 April 2010 when it was in the crop seeding season,and the feature of some croplands (such as fallow field) was mixed with grey impervioussurfaces which were difficult identified by pixel-based approach. This option was to attemptthe identification of the geometric object rather than spectral features alone by objectapproach. The main land-cover types in this region included broad leaf woodland, needleleaf woodland, grassland, cropland, fallow field, water, residential dwelling and road. Imagepre-processing was primarily performed using Erdas 9.0. The procedures included imagegeo-registration, radiation calibration, atmospheric correction and illuminance correctioncaused by terrain (Degui and Alan 1998), and created a registered image in 16-bit type.

2.2. Hypothesis for the optimal segmentation scale selection

The distinction between two classes of land cover is based on their larger variation between-class than within-class in terms of texture or spectral features. Multiple-scale segmentation is aprocess of pixel/object merging from a part of one class, entire class to grouping of many

74 L. Zhang et al.

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

classes.We hypothesized the following fundamental idea for optimal segmentation (OS) scaleselection: when objects are merged using an equal scale feature threshold, although objectfeatures do not always change, these features may gradually change up to the point that theobject sizematches target real world, and it will keep stable or unchanged for a range of scales,the larger two classes’ difference is, the wider the stable scale range is. When segmentationscale is then increased, the object size is larger than that in the real world and the objects of twoland-cover classes are merged, and the object features continue to change. Thus, these stablerange scales of features for each object over multi-scale segmentation are considered asoptimal scales and objects on these scales are selected for further classification.

2.3. Selection of OS scale for each land-cover patch

We proposed a method of multi-scale OS for each object boundary to match the targeted realworld using multiple scale segmentation, and had the following four steps (Figure 1). (1) AVHR imagewas segmented at a wide range ofmultiple scales, startingwith over-segmentationand gradually increasing the size of the objects towards under-segmentation. (2) Objects wereintersected at the spatial boundary of each scale level based on the most over-segmented level(but kept the attributes of the original features); this process is to produce the same object sizeacross segmentation levels for feature comparison. The feature (e.g. mean, standard deviation(SD) of pixel values in an object) variation of the object across scaling was analysed and theoptimal scale (stable-range scales) of each object was determined. In this step, we used aparameter called scaling stable indicator for measurement (Equation (1)). (3) These optimalobjects on various scale levels were projected onto a single layer. (4) The object boundary wasmerged based on the attributes of the object features.

Multiresolution segmentation using Definiens Developer is a bottom-up region-mergingtechnique, starting with one-pixel objects (Benz et al. 2004). In numerous iterative steps,smaller image-objects are merged into larger ones, which are controlled by the heterogeneitythreshold, and a multi-scale segmentation follows the hierarchical process, boundaries ofover-segmented level preserve in under-segmented level across segmentation levels. Manythresholds produce the multiple scale segmentation levels, the increment and range of thresh-old determine the detail changes in segmentation. We performed a multi-scale segmentationon an equal scale increment of 5 from scale parameters that ranged from 10 to 120. A scale

(a) (b)

Pixe

ls M

ulti-

scal

e ob

ject

seg

men

tatio

n

(c) (d)

Figure 1. Flow chart of multi-scale OS. Bottom blocks represent image pixels, upper levels aredenoted as merged objects and coloured blocks represent the object features of spectrum andgeometry. The four steps in this procedure are as follows: (a) Step 1: multi-scale segmentation;(b) Step 2: objects intersecting based on lowest object level; (c) Step 3: objects selection of optimalscale and projection onto the lowest object level; and (d) Step 4: object merging.

Multi-scale optimal segmentation approach 75

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

parameter less than 10 and more than 120 could result in over-segmentation and under-segmentation for this Worldview image, respectively. A scale parameter range of 5 had a littlechange of features across a multi-scale process, and was sufficiently narrow for the measure-ment of a range of stable scales. Each object SD was analysed for stable-scale measurement.The object SD refers to all-band Euclidean distance (square root of the sum of object SDsquare of each band), SD was more sensitive to scale increment (SD usually increased) thanthe mean of the object (usually fluctuated) or size of the object (weak relevant to classes). Forthe spatial comparison of objects of multi-scale levels, we output each of multi-scalesegmentation produced from the Definiens software (Trimble Navigation Limited,Sunnyvale, CA, USA) and input to the ARCGIS software (ESRI Company, Redlands, CA,USA). All the hierarchical scale layers were intersected to the most over-segmented scalelayer. The output showed spatial boundaries of the most over-segmented scale layer and theirattribute table included the multi-scale SD of the original object. This process ensured that thesubsequent object projections did not generate overlays and gaps between objects. The multi-scale object SDs in the attribute table were then measured for the optimal scale.

To assess the dynamics of SD scaling from an object level to another, a scaling stableindicator (Si) is used:

Si ¼ Fiþ1 � Fi (1)

where i refers to scale level, Fi is the SD of object at i scale segmentation level; andFiþ1 isSD value of object at i + 1 scale segmentation level (plus scale increment).

The attribute table of the object SD was input to the MATLAB software, where the datawere ranked in an increasing order of multi-scale segmentation for each object. Si (Equation(1)) was calculated for the objects of each scale level, the layer of Si was equal to zero or evensuccessive occurrences of zeroes represented a stable or OS scale. In this study, we auto-matically selected a middle scale on a longest scale range of successive Si = 0 from theMATLAB and regarded this scale as the optimal scale. These SDs of the objects with optimalscales were linked to the intersected vector to generate a new attribute. The objects were thenmerged, based on the new SD attributes, to create the OS boundary, which was finally input tothe Definiens software to perform image segmentation based on this boundary.

Classification was performed using support vector machine (SVM) classifiers (Pal2008), which can efficiently perform a non-linear classification using the kernel trick,implicitly mapping their inputs into high-dimensional feature spaces. The object fea-tures of four band reflectances, area, rectangular fit, and shape index, were used in theclassification, and the introduction of the object geometric features may increase theidentification of targets (Myint et al. 2011). Randomly sampling approach was used forsample selection, homogeneous objects of OS composed basic units of the totalpopulation. A total of 147 training and 243 validation object samples on screeninterpretation were selected for SVM classification based on both OS and single scale(SS) segmentation. SS segmentation scale was selected using trial-and-error methods.To obtain an appropriate scale with accurate measurement, we selected three better andsuccessive scales of 50, 55 and 60 for the evaluation. A confusion matrix was thengenerated to assess classification accuracy.

2.4. Assessment of OS scale approach

The proposed approach using scaling stable indicator for optimal scale selection wasevaluated. A total of 80 objects of eight land-cover types (10 for each) were randomly

76 L. Zhang et al.

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

collected from the object image of a scale parameter of 10 (most over-segmented scale).The optimal scale was then selected across multi-scale segmentation (23 scales) usingtrial-and-error methods. The optimal scale was considered as the appropriate boundarythat corresponded to the targeted real-world. The object SD gradually increased acrossmulti-scale segmentation because of object merging. The three types of SD changes are asfollows: successive change (unstable, Si > 0), no changes over non-longest range scale(relatively stable, Si = 0, there are many ranges of successive no SD change, appropriatescale matches scale range of no SD change but not a longest range), and no changes overa longest range scale (most stable, successive Si = 0), then, we analyse the proportion ofoptimal object scale matched to above three types of SD change.

3. Results and discussions

3.1. Optimal scale selection analyses

The most stable, relatively stable and unstable objects that matched the optimal scalesaccounted for 76, 21 and 3%, respectively (Figure 2). Water and broad leaf woodlandbetter matched the optimal scale of the most stable SD, and they had more within-classhomogeneity and large feature difference than the other land covers. A relatively stableSD over multi-scale change of these land covers exhibited variations in within-classstructures. The road object had uncertainty in terms of the optimal scale match. Nearlyhalf of the objects with optimal scales matched with a relatively stable range (short range).

80 15

10

5

0

15

10

5

0

15

10

5

0

15

5

10

0 0

4

1

10

5

0

15

10 8

7

3

0

2

0

5

0

15

10

5

0

15

10

5

0

15

10

7

3

0

5

0

60

61

7

8

2

0

21

Total

Grassland

Residential Road Water

Cropland Fallow field

Most stable Relativelystable

Relativelystable

Relativelystable

Relativelystable

Relativelystable

Relativelystable

Relativelystable

Relativelystable

Relativelystable

Objects Objects Objects

ObjectsObjectsObjects

Objects Objects Objects

Unstable

Most stable Unstable

Most stable Unstable Most stable Unstable Most stable Unstable

Most stable Unstable Most stable Unstable

Most stable Unstable Most stable Unstable

Broad laaf woodland Needle leaf woodland

17

9

10

2

40

20

0

Figure 2. Statistics on three change characteristics of the object SD are compared with the objectoptimal scales for various land-cover types. Values on bars are object numbers of optimal scale.

Multi-scale optimal segmentation approach 77

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

Roads were usually and spatially closed to residential and fallow fields, and these roadshad characteristics of spectral similarity. These two reasons led to some optimal scales ona short range of feature-stable values. For most land covers and regions, the longest rangeof scale stability could be an indicator for the object optimal scale selection.

3.2. Comparison between classification accuracy of OS and SS

Classification results show that the OS produced a higher overall accuracy (89%)(Table 1) than SS segmentation at scale 55 (84%) (Table 2), the classification using SSapproach of scale 50, 55 and 60 produced accuracies of 83.5, 84.4 and 84%, respectively.Broad leaf woodland, needle leaf woodland and water resulted in high accuracy, and roadobtained a relatively low accuracy using the OS approach. Unlike the SS55 classification,needle leaf woodland, grassland, residential and fallow field clearly achieved accuracyimprovement using the OS approach.

The accuracies obtained for broad leaf woodland, cropland, water and road had nosubstantial differences between the two approaches. The reason for little difference wasthat the study area was mostly covered by broad leaf woodland and cropland, the scale ofwhich were usually suitable to SS segmentation which is an average appropriate scale,thus, using the SS approach also better classified these classes, but SS approach cannotbalance and efficiently classify the other classes whose scale far away from the appro-priate scale. Water had within-homogeneity; its classification was not obviously relevantto scale change and thus, obtained a high accuracy using both approaches. Road identi-fication using the OS method was not quite effective because of the spectral mixture of theroad with other impervious surface.

3.3. Approach comparison between OS and SS

OS produced 7273 objects that covered the study area, a little bit fewer than SSsegmentation of 8324 objects. These object numbers obtained using the two approacheswere not substantially different, but the size and shape of each object from each approachwere different. The optimal objects occurred at various segmentation levels, even for thesame land-cover type. Thus, traditional single segmentation had difficulty extracting alloptimal objects that can cover an entire region. OS considers object optimum rather thanscale-level optimum performed in SS across multi-scale segmentation. OS improved theinappropriate segmentation of some objects, divided the objects of under-segmentationand merged the objects of over-segmentation that occurred in SS segmentation (Figure 3).Some fallow field objects were segmented in a lower scale using the SS approach (Figure3(a)), and some grassland objects were merged in a higher scale using the SS approach(Figure 3(b)). Changes in the object size and feature led to classification change. Over-segmentation produced detailed features of the object that represented a part of thecharacteristics of land cover rather than the composition and structure of the entire landcover, which could produce many land-cover types when classifying one type. Bycontrast, under-segmentation mixed the objects of several land-cover types that causedsome patches of land covers to be misclassified.

The segmentation scale selection for the entire image in traditional segmentation doesnot consider the feature variety of land-cover types across segmentation. Therefore, under-or over-segmentation for each object inevitably occurs when using SS segmentation. Multi-scale analysis improves classification with multiple scales. However, the segmentation ofeach object is not quite suitable to that in the real world, because even the same class may

78 L. Zhang et al.

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

Table

1.Con

fusion

matrixforclassificatio

nusingOSapproach.

Classes

Broad

leaf

woo

dland

Needleleaf

woo

dland

Grassland

Cropland

Fallow

field

Water

ResidentialRoad

TotalUseraccuracy

(%)

Broad

leaf

woo

dland

301

132

94

Needleleaf

woo

dland

241

144

93

Grassland

122

1486

Cropland

530

3586

Fallow

field

11

344

4085

Water

1717

100

Residential

232

337

86

Road

13

2024

83

Total

3742

1432

3917

3923

243

Produ

ceraccuracy

(%)

8198

8694

8710

082

8789

Multi-scale optimal segmentation approach 79

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

Table

2.Con

fusion

matrixforclassificatio

nusingSSapproach

atscale55

.

Classes

Broad

leaf

woo

dland

Needleleaf

woo

dland

Grassland

Cropland

Fallow

field

Water

Residential

Road

Total

Useraccuracy

(%)

Broad

leafwoo

dland

301

132

94

Needleleaf

woo

dland

139

444

89

Grassland

11

1214

86

Cropland

41

3035

86

Fallow

field

31

2610

4065

Water

1717

100

Residential

332

237

86

Road

519

2479

Total

3641

2131

2917

4721

Produ

ceraccuracy

(%)

8395

5797

9010

068

9084

80 L. Zhang et al.

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

appear on various appropriate scale levels. The OS approach takes advantage of each objectOS scale and objectively produces a segmentation layer where classification can be easilyperformed. The weakness of OS segmentation is its dependence on land-cover types andtheir spatial patterns. In the situation of heavy heterogeneity of within-classes or similaritybetween classes, if these classes were spatial neighbours, the OS approach does noteffectively measure the correct optimal scale when using a scaling stable indicator.

For future studies, we will compare the feature sensibility of stable status to objectoptimal scale, improve the detection of the OS scale and analyse their adaptations forvarious land cover and various types of remote-sensing data.

4. Conclusions

Segmentation is a basic but important procedure in object-based classification. Thetraditional SS segmentation by trial-and-error exploration is not effective for land-covermapping, apart from being costly and time-consuming. This study proposes an approachto optimal multiple scale segmentation that objectively selects the appropriate segmenta-tion scale for each object and projects them to a single layer. The experiment revealed thatan optimal segment scale occurs on an object feature with a stable status across multi-scale segmentation, where object SD does not change in the maximum range. The objectfeature of SD for optimal scale selection improved the classification accuracy by fivepercentage points compared to traditional SS segmentation. The approach has advantagesof accurate selection of appropriate scale for each object. The proposed method is suitablefor application in multiple scale characteristics of land covers, and commits less under-and over-segmentation errors.

FundingThis research was funded by the project of National Natural Science Foundation of China [41371361];the ‘Strategic Priority Research Program – Climate Change: Carbon Budget and Related Issues’ of theChinese Academy of Sciences [XDA05050109]; the Inventory and Assessment of National EcologicalEnvironment 10 year Changes (2000–2010) using remote sensing.

(a)

OS segmentation OS segmentationOS classification OS classification

SS segmentation SS segmentationSS classification SS classification

Broad leaf woodland

Needle leaf woodland

Grassland

Cropland

Fallow field

Road

Residential

Water

N

0 15 30 60m

(b)

Figure 3. Approach comparison between OS and SS segmentation. Black arrows indicate differ-ences. (a) SS approach produces over-segmented objects (centre coordinates, E107.145°,N33.043°); (b) SS approach produces under-segmented objects (centre coordinates, E107.141°,N33.049°); the Worldview images are displayed in Band 432.

Multi-scale optimal segmentation approach 81

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014

ReferencesAnders, N. S., A. C. Seijmonsbergen, and W. Bouten. 2011. “Segmentation Optimization and

Stratified Object-Based Analysis for Semi-Automated Geomorphological Mapping.” RemoteSensing of Environment 115: 2976–2985.

Arbiol, R., Y. Zhang, and V. Palá. 2006. “Advanced Classification Techniques: A Review.” ISPRSCommission VIIMid-term Symposium “From Pixel to Processes”, Enschede, NL,May 8–11, 2006.

Ardila, J. P., W. Bijker, V. A. Tolpekin, and A. Stein. 2012. “Context-Sensitive Extraction of TreeCrown Objects in Urban Areas Using VHR Satellite Images.” International Journal of AppliedEarth Observation and Geoinformation 15: 57–69.

Benz, U. C., P. Hofmann, G. Willhauck, I. Lingenfelder, and M. Heynen. 2004. “Multiresolution,Object-Oriented Fuzzy Analysis of Remote Sensing Data for GIS-Ready Information.” ISPRSJournal of Photogrammetry and Remote Sensing 58: 239–258.

Burnett, C., and T. Blaschke. 2003. “A Multi-Scale Segmentation/Object Relationship ModelingMethodology for Landscape Analysis.” Ecological Modelling 168: 233–249.

Cleve, C., M. Kelly, F. R. Kearns, and M. Mortiz. 2008. “Classification of the Wildland-UrbanInterface: A Comparison of Pixel- and Object-Based Classifications Using High-ResolutionAerial Photography.” Computers. Environmental and Urban Systems 32: 317–326.

Degui, G., and G. Alan. 1998. “Topographic Normalization of Landsat TM Images of Forest Basedon Subpixel Sun–Canopy–Sensor Geometry.” Remote Sensing of Environment 64: 166–175.

Desheng, L., and X. Fan. 2010. “Assessing Object-Based Classification: Advantages andLimitations.” Remote Sensing Letters 1 (4): 187–194.

Drăguţ, L., D. Tiede, and S. R. Levick. 2010. “ESP: A Tool to Estimate Scale Parameter forMultiresolution Image Segmentation of Remotely Sensed Data.” International Journal ofGeographical Information Science 24 (6): 859–871.

Hay, G. J., D. J. Marceau, P. Dube, and A. Bouchard. 2001. “A Multiscale Framework for LandscapeAnalysis: Object-Specific Analysis and Upscaling.” Landscape Ecology 16 (6): 471–490.

Kim, M., T. Warner, M. Madden, and D. S. Atkinson. 2011. “Multi-Scale GEOBIA with Very HighSpatial Resolution Digital Aerial Imagery: Scale, Texture and Image Objects.” InternationalJournal of Remote Sensing 32: 2825–2850.

Myint, S. W., P. Gober, A. Brazel, S. Grossman-Clarke, and Q. Weng. 2011. “Per-Pixel vs. Object-Based Classification of Urban Land Cover Extraction Using High Spatial Resolution Imagery.”Remote Sensing of Environment 115: 1145–1161.

Pal, M. 2008. “Ensemble of Support Vector Machines For Land Cover Classification.” InternationalJournal of Remote Sensing 29 (10): 3043–3049.

Powers, R. P., G. J. Hay, and G. Chen. 2012. “How Wetland Type and Area Differ Through Scale: AGEOBIA Case Study in Alberta’s Boreal Plains.” Remote Sensing of Environment 117: 135–145.

Smith, A. 2010. “Image Segmentation Scale Parameter Optimization and Land Cover ClassificationUsing the Random Forest Algorithm.” Journal of Spatial Science 55: 69–79.

Smith, G., M. Beare, M. Boyd, T. Downs, M. Gregory, D. Morton, N. Brown, and A. Thomson.2007. “UK Land Cover Map Production Through the Generalisation of OS MasterMap.”Cartographic Journal 44: 276–283.

Tian, J., and D.-M. Chen. 2007. “Optimization in Multi‐Scale Segmentation of High‐ResolutionSatellite Images for Artificial Feature Recognition.” International Journal of Remote Sensing 28(20): 4625–4644.

Trias-Sanz, R., G. Stamon, and J. Louchet. 2008. “Using Colour, Texture, and HierarchialSegmentation for High-Resolution Remote Sensing.” ISPRS Journal of Photogrammetry &Remote Sensing 63: 156–168.

Walker, J. S., and T. Blaschke. 2008. “Object‐Based Land‐Cover Classification for the PhoenixMetropolitan Area: Optimization vs. Transportability.” International Journal of Remote Sensing29 (7): 2021–2040.

Yu, Q., P. Gong, N. Clinton, G. Biging, M. Kelly, and D. Schirokauer. 2006. “Object Based DetailedVegetation Classification with Airborne High Spatial Resolution Remote Sensing Imagery.”Photogrammetric Engineering and Remote Sensing 72 (7): 799–811.

82 L. Zhang et al.

Dow

nloa

ded

by [

Bei

jing

Nor

mal

Uni

vers

ity]

at 2

1:55

13

Janu

ary

2014