ınbul 2004 
sis Center 
1 help. We 
nd staft of 
at UNM. 
in Product 
107. 
0, Pattern 
New York, 
h«4o the 
ely sensed 
amm. Eng. 
ification of 
ropagation 
d Remote 
dology for 
ages Using 
'ters, 9, pp. 
e Analysis 
"amm. Eng. 
APPLICATION OF RAJSKI DISTANCE TO LAND-COVER CLASSIFICATION 
USING POLARIMETRIC SAR AMPLITUDE IMAGE DATA 
T. Yamada", T. Hoshi? 
? Dept. of Electrical Engineering,, Fukushima National College of Technology, 30 Kamiarakawa Taira Iwaki, 
Fukushima, 970-8034 Japan - yamada@fukushima-nct.ac.jp 
? Dept. of Computer and Information Science, Ibaraki University, 4-12-1 Nakanarusawa Hitachi, Ibaraki, 316-8511 
Japan - hoshi@cis.ibaraki.ac.jp 
Poster Session, Working Group VII/1 
KEY WORDS: Remote Sensing, Land Cover, Classification, Polarization, Algorithms, SAR 
ABSTRACT: 
Polarimetric SAR can observe scattering matrix for each resolution cell and provide amplitude images that have gray level in 
proportion to amplitude of the conjunction matrix elements. However, these amplitude image data have been used for pseudo color 
synthesize and construction of feature vector for land-cover classification as the vector elements mainly, discussion about features 
derived from amplitude image data was scarcely. In this paper, we consider pattern difference in different polarization amplitude 
image of polarimetric SAR as probability, and discuss about contribution of expanding dimension of feature vector by introducing 
Rajski distance as a features. To calculate Rajski distance, gray level co-occurrence matrix (GLCM) method that has been often used 
for texture analysis was used. In the proposed method, gray levels of the pixel that was located at the same coordinates in two 
different transmit and receive polarization amplitude images are adapted to line and column of GLCM, then co-occurrence 
probability are calculated. From this matrix, joint entropy and conditional entropy are derived and Rajski distance is found. To 
inspect the proposed theory, we investigate the effect of expanding dimension of feature vector for land-cover classification by 
Euclid minimum distance method and maximum likelihood method as well known supervised classification method generally using 
SIR-C polarimetric data obtained in two kinds of scenes including different land-cover objects. As the results, improvement of 
accuracy in the point of classification score and ambiguity scale, so that the effectiveness of introduction of Rajski distance to 
expanding dimension of feature vector for land-cover classification is demonstrated. 
1. INTRODUCTION Rajski distance is used as a distance measure that indicates the 
similarity of two probability phenomenon systems using 
The various land-cover classification algorithms using the data entropy. In this paper, Rajski distance is calculated for each 
of the polarimetric SAR have been proposed until now, and the pixel by making into a probability phenomenon system the 
validity is reported. For example, they are supervised pixel value distribution of the amplitude image data from that 
classification methods based on Bayes theory and using neural the combination of two different polarizations of polarimetric 
network, and unsupervised classification methods based on SAR data: SIR-C (Shuttle Imaging Radar-C) and airborne Pi- 
scattering types and scattering entropy. Each of these SAR (Polarimetric and Interferometric SAR), the distances for 
corresponds to full-polarimetric data that used both amplitude some area are characterized, and the application effect of Rajski 
information and phase information for each element of distance to land-cover classification is reported. 
scattering matrix obtained by polarimetric SAR (Yueh, H. A. et 
al., 1988; Ito, Y. and Omatsu, S., 1997; Zyl, J. J., 1989; Cloude, 
S. R. and Pottier, E., 1997). However, in the case where whole 
information of scattering matrix is acquired, and the situation 
that only amplitude information is acquired, some of these 
2. CALCULATION OF RAJSKI DISTANCE FROM 
POLARIMETRIC SAR IMAGE DATA 
techniques cannot be apply or deteriorate classification 2.1 Overview of Rajski Distance 
accuracy. In order to analyze to image data only for amplitude 
information, the classification technique as a general digital In this section, overview of Rajski distance is described 
image are applied. It could judge visually that the difference has (Isomichi, Y., 1980). Two probability phenomenon systems are 
arisen in the amplitude image data for the different polarization set to X and Y, respectively, and each phenomenon is set to a; (i 
used for observation by polarimetric SAR, when performing — 70. ..., m-1) and b; (j=0, ..., n—1), respectively. The probable 
such a classification. Therefore, it is necessary to extract relevance of both phenomenon systems is determined by the 
quantitatively the difference in the feature by the difference simultaneous occurrence probability p(a; bj) Joint entropy 
kind of polarization that is latent in the amplitude image data of | H(XY) for X and Y is given using simultaneous occurrence 
the polarimetric SAR, and to raise classification accuracy with probability as follows. 
eg icatı ro . 1 ane si tu m-ln-l 
the application of it for dimension extension of the feature H(XY) = ES pia, Pie pie D.) (1) 
vector of land-cover classification for amplitude image. In this oe : 
paper, application of the Rajski distance based on information The conditional information content A(X) and HHX) are 
theory is proposed as the parameter. Although entropy is 
ray fa given as follows 
conventionally used for the texture analysis of a digital image, 
21