Developing decision tree classification system based on discrimination of spectral and textural characteristics to identify land use using ALOS AVNIR-2

Developing decision tree classification system based on discrimination of

spectral and textural characteristics to identify land use

using ALOS A VNIR-2

S. Darmawan, I. H. Ismullah, K. Wikantika, A. Bodi Harto [ 1]. Center for Remote Sensing

[2]. Remote Sensing and Geographical Information Science Research Division Institute of Technology Bandung

JI. Ganesha 10 Bandung 40132, Indonesia Email : soni ~ [email protected]

Abstract

There are many classification strategies to identify and monitor land use in remote sensing. Further study to develop classification technique using hierarchy classification by decision tree model that parameters are based on spectral and textural characteristics. The use of both spectral and textural characteristics in remotely sensed image data has been proven promising for land use and land c~wer classification. The main objective of the research is to develop decision tree model based on spectral and textural characteristics of the ALOS A VNIR-2. The study includes image preprocessing, calculation of vegetation index, calculation of textural measures, evaluation separability analysis and develop tree model. Image preprocessing focuses on image registration while vegetation index generate NDVI, RNDVI, GNDVI, NDRGI, GRVI and MPRI, respectively. Finally, textural measures are calculated based on the gray level co-occurrence matrix (GLCM). The results indicated that decision tree model which is based on combination of band spectral, normalized difference vegetation index (NDVI), Modified Photochemical reflectance Index (MPRI) and -imn" texture measure were accurate for discrimination of different land uses over the study area.

Keywords: spectral, textural, ALOSA VNIR-2, discrimination, land use, decision tree

1.

Introduction

In recent years, decision tree classifiers have been successfully used for land cover classification from remote sensing data. The advantages of using a multistage or tree approach to classification include that different data sources, different sets of features, and even different algorithms can be used at each decision stage. (Richard & Jia, 2006). Many researcher have investigated decision tree classification for vegetation cover mapping (Simrad et al., 2000), forest mapping (Huang & Yang, 2001) urban landscape dynamics (Pavuluri et al., 2002), national park vegetation mapping (Colstoun et al., 2003), habitat and agricultural mapping (Lucas, 2007), habitat classification and change detection (Sensie et al., 2008), updating land cover (Raclot et. al., 2005 and Wentz et. al., 2008).

However land use classification using decision tree based on single features of

spectral characteristic have limitation (Pal & Mather, 2003), another limitation decision tree classification are instability of tree (Miller & Franklin, 2002) and requiring a large number of training samples for tree construction (Joy et al., 2003).

Basically, remote sensing technology can detect reflectance energy from land use and it also provides spectral response in image. Separability analysis indicated that spectral ALOS A VNIR-2 data provided adequate spectral discrimination of land use. Other features which can discriminate land use are using textural characteristics. Texture can improve classification result. Classification of Mediterranean land cover from Landsat TM imagery, texture information was found beneficial for certain land covers (Berberoglu et al., 2007) and combination of spectral and textural aspects significantly improved the classification accuracy compared with

classification with only pure spectral features (Wikantika et al., 2004).

The main objective of this research is to construct a decision tree ( expert · system) for land use classification by determining the optimal spectral and textural measures. The research focuses on investigation of spectral and texhiral characteristics of land uses in study area, investigation of separation value for discriminating different land uses, and construction of a decision tree.

Indonesian archipelago

West Java

2. Methodology 2.1 Study area and data

Location of study area is in Southern part of Bandung, West Java, Indonesia, roughly between latitude 6 ° 5 9' - 7 ° 04' S and longitudes 107 ° 34' - 107 ° 41' E. ALOS A VNIR-2 collected in Juni 2007 and total of training site distributed in study area is 65 samples.

10,C 40'0"E .

Figure 1. Study area

Basically land use classification systems level II used in this study based on Anderson

(1976) and Indonesian's land use

classification systems such as industrial area, water body, forest, mixed plantation, agricultural, grassland, wet paddy field, dry

paddy field, minning area, fallow land, shrubland, field area, and urban area

2.2 Method a. Preprocessing

ALOS A VNIR-2 has been systematically geometric corrected but must be registered to reference data. In this study reference data used was SPOT 5 with spatial resolution 2.5 rn. Distribution of ground control point for image registration is nine control points distributed within study area. Selection of ground control point was based on geographic characteristics such as branch of rivers and roads. To wrap the satellite image, polynomial orde-1 was used and resampling method of nearest neighbor was applied.

b. Calculation of spectral, texture and separability analysis

The study used several vegetation index as following below:

Normalized Difference Vegetation Index (NDVI) (Rouse et al. 1974):

NDVI= NIR-Red ... (1)

NIR+Red

Ratio Normalized Difference Vegetation Index (RNDVI) (Gong et al. 2003)

RNDVI = lNIR -red)*lNIR) ... (2)

NIR +red red

Green Normalized Difference Vegetation Index (GNDVI) (Gitelson et. al 1996)

GNDVI = NIR - green

NIR+green

... .. (3)

Normalized Difference Red Green Index (NDRGI) (Yang et. al 2008)

red-green

NDRGI = _ ___c::..__

red+ green

... (4)

Green Ratio Vegetation Index ( GRVI)

GRVI = NIR

green

... ... (5)

Modified Photochemical Reflectance Index (MPRI) Yang et. al (2008)

MPRI = green-red ... . (6)

green-red

For separation not only based on spectral characteristic but also used texture information. Five textural measurement such as mean, homogeneity, correlation and contrast are proposed by haralick et. al. ( 1973 ). All profiles are based on land use in the research area.

N-1

IPij

Mean texture: i,j=O (7)

µij=y··· N-1 p. Homogeneity:

L

1 '1 2 ... (8) i,j=O}

+

(1- J) N-1 Contrast:

L~,/i-

j)

2 •••••••••••••••••••.. (9) i,j=O

Correlation:

.Il(i-

µ;U-t)J

· · · ...

(10)

,,;=o ~(a; )(a₁)

To separate different land use needed threshold value. Threshold values were derived from training site. This study collected 65 training sites. Threshold values were determined based on median values both mean and standard deviation values for each class. Threshold (Th) value can be as follows:

Th

)xi

±O")+(xj ±O") ... (11) 2

3. Results and Discussions

s\LOS A VNIR-2 has 4 band they are blue (0.42 - 0.50 µm), green (0.52 - 0.60 µm), red (0.61 0.69 µm) and near infrared (0.76 -0.89 µm). The result of ·calculation of NDVI,

255

I

73 band 1 255

I

9 band4 0.45

I

-0.8 GNDVI 0.4

I

-0.3 MPRI

RNDVI, GNDVI, NDRGI, GRVI, and MPRI and textural measures such as mean, varian, homogeneity, contrast and correlation, can be shown by figure 2, respectively.

255

I

42 band2 band 3 0.6

I

-0.7 NDVI RNDVI 0.3

I

-0.4 NDRGI ERVI 63

I

0 texture

Figure 2. Spectral and textural characteristics

Figure 3 shows spectral characteristic for each land use class in study area. According to figure 3, industrial zone has higher spectral performance in band 1, band 2 and band 3, respectively. Vegetation land use such as agricultural land, mixed plantation land, forest, grassland, shrub land, and dry paddy, increase in band NIR but. another land use

Vegetation index characteristic for each land use can be shown by figure 4. Based on figure 4, forest, shrub land, mixed plantation and grassland have higher NDVI values than others but in NDRVI forest, shrub land, mixed plantation and grassland have lower values. Non vegetated land use such as water, industrial zone, mining zone, rural and urban

255

I

25 2.9

I

-0.1 2.6

I

0.08

land use such. as forest, shrub land, mixed plantation and grassland have different values. Water and wet paddy have lower values in

GNDVI but higher values in NDRVI and MPRI, respectively.

1

Figure 3. Spectral characteristic for each land use

,

___

ti_~

-·lt.tdlJild ~-Mi•~

-·~

_-~··~

ww>!lb1ullfd '~"'°'YINMt -1i.sw.ww .icmt ,,Mi***klkit!t ,,,.,fiUwli!'d ,,,!kW,, ··W\lt4t ·lltlt'lp.ffdt -~~!IN -F¼llf~ ,,,,.t,ii•~~ -t<Wtlt

-~

···~fd ·-1:ky ... -1Mi1,-rid Nl'lf ,,f,ltlllli._J.Ont -hll-llllMI ---~n '·Wfflfffi '-IN#l~

Figure 4. Vegetation index characteristic for each land use

Textural characteristic for each land use can be shown by figure 5. Based on figure 5, water and wet paddy have lower values while shrub land has higher values than other land , uses. Based on spectral and textural characteristics for each land use and separability analysis therefore it can be discriminated each land uses consisted .of some steps as follows :

Step 1. Mean texture can discriminate water body, wet paddy and industrial zone, minning zone, fallow area, urban, agricultural, field, grassland, forest, mixed plantation, shrubland.

Step 2.1. band 2 can discriminate water body and wet paddy

Step 2.2 NDVI can discriminate industrial zone, minning zone, fallow land, rural, urban area and agricultural, field, grassland, forest, mixed plantation, shrubland _

Step 3.1 band can discriminate industrial zone, minning zone and fallow land, rural, urban

Wetpaddy

Step 3.2 NDVI can discriminate agricultural, field, grassland and forest, mixed plantation, shrubland

Step 4.1 NIR can discriminate industrial zone and minning area

Step 4.2 band 3 can discriminate fallow land, and urban

Step 4.3 MPRI can discriminate agriculture, field and grassland, dry paddy

Step 4.4 band 1 can discriminate forest and mixed plantation, shrubland

Step 5.1 band 3 can discriminate grassland and dry paddy

Step 5.2 band 1 can discriminate agriculture and field

Step 5.3 NDVI can discriminate mixed plantation and shrubland

Decision tree analysis derived from discrimination of land uses based on spectral and textural characteristic as shown by figure 6. Land use classification .. using decision tree analysis can see in figure 7.

Fallow land

Grassland Dry paddy Agriculture Field land

4¢p~t4

Land Cover Map

South ~ati...i1Jng; W,u;t ,i®V* ft1drw0t'1!,io - A,w,dhd • '"'"'*

- """*"- -

_f1*h/4iJ

foy>''t<¾/ • t:.,.,~n,l

- '""-·J

- h;iiw,:ip,d 8 it.14Jud bJJW.;lli,,j

--l'if@liffl

-··"""

- 1\ll ... t.m, Figure 7. Land Use and Land Cover classification result by decision tree analysis

4. Conclusions

It has been shown that use of spectral and textural measures can identify and separate land use and land cover in level II classification systems. Separation of each land use with decision tree analysis can use band 1, band 2, band 3, band 4, NDVI, MPRI and textural information derived from ALOS A VNIR-2 image.

Acknowledgments

· Firstly, I would like to thank Directorate Higher Education, Ministry of National Education, The Republic of Indonesia for the fellowship. I would also like to thank Prof. Ryutaro Tateishi of Center for Environmental Remote Sensing (CEReS), Chiba University for supporting laboratory.

References

1. Berberoglu S., Curran P.J, Lloyd C.D, Atkinson, P.M., 2007, Texture Classification of Mediterranean Land Cover, International Journal of Applied Earth Observation and Geoinformation 9, 322-334

2. Colstoun, E., C., B., Story, M., H., Thompson, C., Commisso, K., Smith, T., G., dan Iron, J., R., 2003 --National Park vegetation mapping using multitemporal Landsat 7 data and a decision tree classifier", Remote Sensing of Environment 85, 316-327

3. Gitelson, Anatoly A., 2004. Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation, J Plant Physiology, No. 161, pp. 165-173.

4. Haralick R.M., et al., 1973, Tekstural Feature for · Citra Classification, IEEE Transaction on systems,

Man and Cybernetics Vol. SMC-3 No. 6, 610-621.

5. Huang C., and Yang L., 2001, Synergistic Use of FIA Data and Landsat 7 ETM+ Images for Large Area Forest Mapping. The thirty-fifth annual Midwest forest Mensurationists and Th Third Annual FIA Symposium

6. Joy S.M., Reich R.M and Reynolds R.T., 2003 A Non-parametric Supervised Classifications of Vegetation Types on The Kaibab National Forest Using Decision Trees. International Journal of Remote Sensing, 24, 1835 - 1852.

7. Miller J., and Franklin J., 2002, Modeling the Distribution of Four Vegetation Alliances Using· Generalized Linear Model and Classification Trees with Spatial Dependence. Ecological Modeling, 157, 227-247

8. Lucas R., et al., 2007, Rule-based Classification of Multi-temporal Satellite Imagery for Habitat and Agricultural Land Cover Mapping, International Society for Photogrammetry and Remote Sensing (ISPRS), Journal of Photogrammetry & Remote Sensing 62, 165-185. 9. Pavuluri · M.K., Ramanathan S., and Daniel Z.

2001 A Rule-based Classifier Using Classification and Regression Tree (CART) Approach for Urban Landscape Dynamics. International Geosciences and Remote Sensing Symposium vol. 2 (pp.24-28)

10. Pal M., and Mather P.M., 2003, An Assessment of the Effectiveness of Decision Tree Method of Land Cover Classification. Remote Sensing of Environment 86, 554-565

11. Raclot D., F. Colin, C. Puech, 2005. Updating land cover classification using a rule-based decision system, International Journal of Remote

· Sensing, 26:7, 1309 - 1321

12. Richard A.R., and Jia. X., 2006. "Remote Sensing Digital Image Analysis" Springer.

13. Rouse, J., W., R. H. Haas, J. A. Schell and D.W. Deering, 1974, -Monitoring Vegetation Systems in Great Plains with ERTS", Proceedings, Third Earth Resources Technology satellite-I Symposium, Greenbelt:NASA SP-351, 3010-317 14. Simrad M., Saatchi S.S., and Grandi G.D., 2000,

The Use of Decision Tree and Multiscale Texture for Classification of JERS-1 SAR Data Over Tropical Forest, IEEE Transaction on Geosciences and Remote Sensing, 38(5), 2310-2321

15. Sesnie S. E., et al., 2008, Integrating Landsat TM and SRTM-DEM Derived Variable with Decision Tree for Habitat Classification and Change Detection in Complex Neotropical Environment, Remote Sensing of Environment 112, 2145-2159 16. Wentz, Elizabeth A., Nelson, David, Rahman,

Atiqur, Stefanov, William L. and Roy, Shoursaseni Sen (2008) 'Expert system classification of urban land use/cover for Delhi, India', International Journal of Remote Sensing, 29:15, 4405 -4427

17. Wikantika K., 2004, -Spectral and Textural Aspect of Multisensor and Multitemporal Satellite Data for Land Use/Land Cover Mapping in A Tropical Area", Thesis Doctoral, Chiba University, Japan

18. Yang, Z., Willis, P ., Mueller, R., 2008, -Impact of band ratio enhanced awifs image to crop classification accuracy, Proceeding, Denver Colorado