00b4952d40932eaf9b000000

8/9/2019 00b4952d40932eaf9b000000

1/7

Field-based landcover classification using TerraSAR-X texture analysis

Ali Mahmoud a,⇑, Samy Elbialy a, Biswajeet Pradhan a,b,1, Manfred Buchroithner a

a Institute for Cartography, Faculty of Forestry, Geo and Hydro-Science, Dresden University of Technology, 01062 Dresden, Germanyb Institute of Advanced Technology, University Putra Malaysia, 43400 Serdang, Malaysia

Received 20 December 2010; received in revised form 5 April 2011; accepted 6 April 2011Available online 12 April 2011

Abstract

The present study aims to evaluate the field-based approach for the classification of landcover using high-resolution SAR data.TerraSAR-X (TSX) strip mode imagery, coupled with digital ortho-photos (DOPs) with 20 cm spatial resolution was used for landcoverclassification and parcel mapping respectively. Different filtering and analysis techniques were applied to extract textural informationfrom the TSX image in order to assess the enhancement of the classification accuracy. Several attributes of parcels were derived fromthe available TSX images in order to define the most suitable parameters discriminating between different landcover types. Then, theseattributes were further statistically analysed in order to define separability and thresholds between different landcover types. The resultsshowed that textural analysis resulted in high classification accuracy. Hence, this paper confirms that integrated landcover classificationusing the textural information of TerraSAR-X has a high potential for landcover mapping. 2011 COSPAR. Published by Elsevier Ltd. All rights reserved.

Keywords: Landcover classification; TerraSAR-X; Field-based; Texture analysis; Remote sensing

1. Introduction

Landcover classification is probably among the mostprominent applications of remote sensing (cf. i.a. Soergel,2010). Crop mapping at a specific time or growth stage isof high importance for agricultural and economic applica-tions. In some cases it is important to survey the existentcrops in order to manage possible irrigation requirements.In other cases it is inevitable for crop-yield estimation forcash crops (Ren et al., 2008) or for subsidies control (Blaeset al., 2005). Remotely sensed (RS) data plays an important

role in retrieving landcover classes and to discriminatebetween different types of crops. In recent years, theobject-based image analysis (OBIA) proved to be more effi-cient than pixel based classification mainly due to the avail-ability of the high spatial resolution RS data (Al Fugaraet al., 2009; Blaschke et al., 2008).

Historically, Synthetic Aperture Radar (SAR) datawas made available for landcover classification muchlater than optical RS data. As a result, more researchhas been conducted and accumulated in the extractionof features from optical data than from SAR images.However, SAR has many advantages over the opticaldata because of its ability to penetrate cloud cover andits night sensing capabilities. In some cases the informa-tion of interest is better visible at microwave frequenciesrather than at optical ones (Dell’Acqua and Gamba,2010). Radar interacts very differently with surface fea-

tures than optical data, providing information morerelated to shape and structure than surface composition(Herold and Haack, 2004). Thus, for many applicationssuch as disaster management or when data have to beacquired at specific dates within a short period of time radar systems are more valuable (Pradhan, 2010;Pradhan and Shafie, 2009).

Recently, field-based landcover classification using dif-ferent data sources –as an OBIA approach – has beenapplied and improved the classification accuracy (Lu andWeng, 2007). Dean and Smith (2003) used an Airborne

0273-1177/$36.00 2011 COSPAR. Published by Elsevier Ltd. All rights reserved.

doi:10.1016/j.asr.2011.04.005

⇑ Corresponding author. Tel.: +49 351 463 34809; fax: +49 351 46337028.

E-mail address: [email protected] (A. Mahmoud).1 Tel.: +60 3 8946 8466; fax: +60 3 8656 6061.

www.elsevier.com/locate/asr

Available online at www.sciencedirect.com

Advances in Space Research 48 (2011) 799–805

http://dx.doi.org/10.1016/j.asr.2011.04.005mailto:[email protected]://dx.doi.org/10.1016/j.asr.2011.04.005http://dx.doi.org/10.1016/j.asr.2011.04.005mailto:[email protected]://dx.doi.org/10.1016/j.asr.2011.04.005

8/9/2019 00b4952d40932eaf9b000000

2/7

Thematic Mapper (ATM) imagery with 1.25 m spatial res-olution and found that the parcel-based representation wasshown to be most appropriate for mapping agriculturallandcover in comparison to semi-natural areas becauseagricultural landscapes have an inherent parcel structure(Dean and Smith, 2003). There is, however, a limitationto this field-based approach as it is heavily dependent onthe field boundaries which need to be acquired prior tostarting such classification. Some studies used the existingparcels-data (digital or hard copies) or generated it for cur-rent and further studies (Wu et al., 2007) using differenttechniques such as segmentation.

Improvement of classification accuracy has beenachieved for both multispectral optical data (Chauhanet al., 2008) and radar imagery (Blaes et al., 2005; Waskeand Braun, 2009). More recently, TerraSAR-X (TSX)images have been used in various studies for landcoverclassification (Baghdadi et al., 2009; Breidenbach et al.,2010; Burini et al., 2008; Mróz and Mleczko, 2008). In a

recent paper, Breidenbach et al. (2010) stated that the use

of textural parameters (Haralick et al., 1973; Liang, 2008;Lloyd et al., 2004; Tso and Mather, 2009), object-basedclassification approaches and multi-temporal data can sig-nificantly improve the classification result of the TSXimages. In summary, the above body of literature indicates

that high-resolution TSX imagery has not been fullyexploited for landuse/cover classification yet.

Thus, the present study aims to test the field-basedapproach for classifying landcover using the TSX data byemploying texture analysis and various filtering methodsin order to maximise the extracted information from theSAR images. First, the classification was applied to twoTSX scenes using each SAR image as a single image thenboth of the two available images were used jointly. Theadvantage of the analysis of single SAR images (besidestheir costs) is the necessity of rapid mapping, for instancein the case of time critical events (Soergel, 2010). In the

present study, field boundaries were digitised using the dig-ital ortho-photos (DOPs) with 20 cm spatial resolution.Several attributes of parcels were derived from the avail-able TSX imagery, then the separability and threshold(SEaTH) method (Nussbaum and Menz, 2008) was appliedin order to define the most suitable attributes that discrim-inate between different landcover types. Finally, these attri-butes were used in the classification tree.

2. Study area

The study area is located near Pirna, Saxony, Germany

(Fig. 1a). In order to apply the proposed methodology

Fig. 1. (a) Location map of the study area, (b) TSX image of the study area acquired 31.05.2010 with parcel boundaries, and (c) TSX image acquired17.06.2010 with parcel boundaries.

Table 1TerraSAR-X data applied.

Date Sensor Polarisation Pass direction Incidence angle range Look direction Resolution (m)

31.05.2010 StripMap HH Ascending 41.76–43.89 Right 317.06.2010 StripMap HH Ascending 29.66–32.42 Right 3

Table 2Object features used for B and J calculations.

Mean GLCM meanStandard deviation GLCM standard deviationGLCM homogeneity GLCM correlationGLCM contrast GLDV ang. 2nd momentGLCM dissimilarity GLDV entropyGLCM entropy GLDV meanGLCM ang. 2nd moment GLDV contrast

800 A. Mahmoud et al. / Advances in Space Research 48 (2011) 799–805

8/9/2019 00b4952d40932eaf9b000000

3/7

efficiently for landcover (crop) mapping, an agriculturalarea was chosen (Fig. 1b and c, demarcating the parcelsin red). During the acquisition of the TerraSAR-X datathe following landcover classes were mapped in the field:crops, orchards and grass.

3. Methodology

3.1. Data preparation and segmentation

TerraSAR-X images acquired on 31.05.2010 and17.06.2010 (Table 1) were imported into ERDAS Imagineand subsequently filter types with three different kernelsizes (3 3, 5 5 and 7 7) were applied. On the otherhand, the field boundaries were digitised from the digitalortho-photos (2005) and updated from the TerraSARimages to match the current landcover boundaries. Defini-

ens 7 Software supports different segmentation algorithms

which are used to subdivide the entire image represented bythe pixel level domain or specific image objects from otherdomains into smaller image objects. In the current study,the multi-resolution segmentation algorithms were appliedusing the field-boundaries thematic layer coupled with theused images. The image layers were weighted differentlydepending on their importance or suitability for the seg-mentation result. The higher the weight, the more of itsinformation is used during the segmentation process.Therefore, the thematic layer weight was assigned to 1while the image weight was assigned to zero (Definiens,2007).

3.2. Feature extraction and analysis

Representative samples for each landcover type wereselected. Then a total of 62 texture feature values (Table 2)

for each image were exported as

.csv files. This step was

Table 3Separability and thresholds.

Object class combination Separability Omen Threshold

1. Cereals from grass

Gama map 7_17.06.10 GLCM ang. 2nd moment (45) 1.25 Great 0.00241Mean 7_17.06.10 GLCM ang. 2nd moment (45) 1.24 Great 0.00246

2. Cereals from maizeTSX 31.05.10 GLCM contrast (all dir.) 1.99 Small 351.74TSX 31.05.10 Mean 1.92 Small 165.94

3. Cereals from orchards

TSX 170610 GLCM dissimilarity (0) 1.94 Small 12.85Local region 7_17.06.10 GLCM homogeneity (all dir.) 1.95 Great 0.145Mean3_17.06.10 Mean 1.87 Small 160.01

4. Cereals from rape

TSX 17.06.10 GLCM dissimilarity (0) 1.98 Small 13.25Frost 3_17.06.10 GLCM homogeneity (0) 1.98 Great 0.103

5. Grass from maize

TSX 31.05.10 GLCM contrast (90) 1.99 Small 259.66TSX 31.05.10 Mean 1.96 Small 164.16TSX 31.05.10 GLCM homogeneity (135) 2.00 Great 0.0603

6. Grass from orchards

TSX 31.05.10 GLDV ang. 2nd moment (90) 1.96 Small 0.0398TSX 31.05.10 GLCM homogeneity (135) 1.99 Great 0.0620

7. Grass from rape

TSX 31.05.10 Mean 1.99 Great 167.49TSX 31.05.10 GLCM dissimilarity (all dir.) 1.99 Small 13.65Adaptive median 7_17.06.10 GLCM homogeneity (0) 1.77 Great 0.1686

8. Maize from orchards

Median 3_31.05.10 GLCM ang. 2nd moment (0) 1.84 Small 0.000679Median 3_31.05.10 GLCM entropy (0) 1.92 Great 7.699TSX 31.05.10 – TSX 17.06.10 Mean 1.99 Great 160.80

9. Maize from rape

Adaptive median 3_17.06.10 GLCM homogeneity (45) 1.99 Small 0.0855Local region 3_17.06.10 GLDV ang. 2nd moment (45) 1.90 Great 0.0313TSX31.05.10 – TSX17.06.10 Mean 2.00 Great 153.77

10. Orchards from rape

Median 7_31.05.10 GLCM dissimilarity (0) 1.65 Small 3.146Median 7_31.05.10 GLDV ang. 2nd moment (0) 1.64 Great 0.1336TSX 31.05.10 Mean 1.35 Small 178.28

A. Mahmoud et al. / Advances in Space Research 48 (2011) 799–805 801

8/9/2019 00b4952d40932eaf9b000000

4/7

applied to the original and the filtered TerraSAR-Ximages. Additionally, an output layer was created by sub-tracting both images acquired on 31.05.2010 and17.06.2010, and then used as additional layer to separatemaize from other crops. In order to define the proper fea-tures that discriminate between different landcover types,

the separability was calculated for each class-pair of land-cover types. The Bhattacharya Distance B (Eq. (1)) andthe Jeffries–Matusita Distance J (Eq. (2)) were calculated(Nussbaum and Menz, 2008). Then the threshold T foreach pair of the classes was calculated using the Eq. (3).An excel model was built to calculate these parametersautomatically from the *.csv files using Visual Basic Appli-cation (VBA), in order to find the best features showing thehighest J value.

B ¼1

8 m1 m2ð Þ

2 2

r21 þ r

21

þ1

2 ln

r21 þ r

22

2r1r2

; ð1Þ

J ¼ 2ð1 eð BÞÞ; ð2ÞFig. 2. Methodological flowchart adopted in this study.

0

0.5

1

1.5

2

S e p a r a b i l i t y

GLDV Ang. 2nd moment (0°)

TSX_310510

Adaptive Median 3

Local Region 3

Local Region 7

Gamma Map7

Median 7

(a )

0

0.5

1

1.5

2

S e p a r a b i l i t y

GLCM Homogeneity (0°)

TSX image

Adaptive Median 3

Adaptive Median 7

Frost 3

Frost 5

Local Region 7

(b )

Fig. 3. Effect of filter type on separability value: (a) image acquired 31.05.2010; (b) image acquired 17.06.2010.

802 A. Mahmoud et al. / Advances in Space Research 48 (2011) 799–805

8/9/2019 00b4952d40932eaf9b000000

5/7

x1ð2Þ ¼m2r

21 m1r

22 r1r2

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðm1 m2Þ

2 þ 2 Aðr21 r22Þ

q ðr21 r

22Þ

;

ð3Þ

A ¼ log r1

r2

n2

n1

: ð4Þ

3.3. Classification

Depending on the J values (Table 3) the proper imagesand relative texture features for classification were definedfor each class. Then the class description was assignedusing the membership functions that offer a transparentrelationship between feature values and the degree of membership to a class. The process tree employed as aclass-by-class analysis that enabled to exclude the classifiedobjects from the classification process. Generally speakingthe classification algorithm uses class descriptions to clas-sify the image objects by evaluating the class descriptionand determines whether an image object can be a memberof this class or not. Moreover, it allows a fuzzy-logiccombination of different features (Definiens, 2007). Fig. 2shows the methodology applied in this study in the form

of a flow-chart.

4. Results and discussion

The results show that some texture features of the origi-nal TerraSAR-X images (without filtering) produced highvalues of separability (J ) for some class-pairs such ascereals–maize and grass–maize. However, specific filters

have improved the separability for other classes such as

Fig. 4. Landcover map of the study area.

Table 4Overall classification accuracy statistics.

Classified data Reference data

Cereals Maize Rape Grass Orchards Sum.

Cereals 30 0 0 1 1 32Maize 0 5 0 0 0 5Rape 0 0 5 0 0 5Grass 0 0 0 4 0 4Orchards 1 0 0 0 20 21Sum. 31 5 5 5 21 67

Accuracy

Producer 97% 100% 100% 80% 95%

User 94% 100% 100% 100% 95%Kappa 88% 100% 100% 100% 93%

k hat 93.27%Overall accuracy 95.52%


8/9/2019 00b4952d40932eaf9b000000

6/7

8/9/2019 00b4952d40932eaf9b000000

7/7

Ren, J.Q., Chen, Z.X., Zhou, Q.B., Tang, H.J. Regional yield estimationfor winter wheat with MODIS-NDVI data in Shandong, China. Int. J.Appl. Earth Obs. 10 (4), 403–413, 2008.

Soergel, U. Review of radar remote sensing on Urban Areas, in: RadarRemote Sensing of Urban Areas. Springer, New York, 2010.

Tso, B., Mather, P.M. (Eds.), Classification Methods for Remotely SensedData, second ed., CRC Press, Boca Raton, 2009.

Waske, B., Braun, M. Classifier ensembles for land cover mapping usingmultitemporal SAR imagery. ISPRS J. Photogram 64 (5), 450–457,2009.

Wu, S., Silvan-Cardenas, J., Wang, L. Per-field urban land use classifi-cation based on tax parcel boundaries. Int. J. Remote Sens. 28 (12),2777–2800, 2007.


Documents

00b4952d40932eaf9b000000