ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS", Bangkok, May 23-25, 2001 
340 
Fig. 1 Study area before filtering Fig. 2 After filtering speckle noise Fig. 3 After classification image 
2.THE STUDY SITE AND THE RADAR DATA 
The study site named Port Island located along the coast of Kobe 
City of Japan. The radar data was imaged on July 13,1999 by 
CRL7NASDA airborne SAR. Full polarimetrie data was acquired 
simultaneously at X- (9.55GHz), L- (1.27GHz) band at a mean 
incidence angle of 39 . Radiometric calibration has been made 
with results of calibration experiments by CRL. Fig. 4 shows the 
Port Island SAR data image using the backscatter coefficient o 
of X band HH polarization data (1), (2). Since the radar site covers 
a large-scale area, a part of radar image was selected as our 
study area (Fig. 4 highlight part). To improve the accuracy of 
analysis, X-band HH polarization data was used, since it has high 
spatial resolution with 1.5m. Fig. 1 shows the studied area in 
detail. 
<T°(dB) hh = 10 log(Re 2 + Im 2 )-1.16 (In X-HH image) (1) 
<7°(dB) vv =101og(Re 2 + Im 2 )-5.20 (In L-VV image) (2) 
Fig. 4 SAR image of Kobe Port Island (X-band HH) 
segmentation and classification. The radar echoes reflected from 
individual scatters within a pixel either randomly reinforce one 
another if they happen to be in-phase or reduce the return signal if 
they are out-of-phase (fading). To reduce the speckle noise of 
SAR polarimetric imagery, in this study, speckle filtering is applied 
using an algorithm based on Lee polarimetric SAR speckle 
filtering method, which has the feature of smoothing the image 
data without removing edges or sharp features in the images (Lee 
et al., 1999). The filter operates in a 7 x 7 moving window. The 
filtering process described as follows. 
The multiplicative noise model can be described as 
y-XV (3) 
where y is the center pixel value (intensity or amplitude of a SAR 
image), X is the noise-free pixel value to be estimated, v is 
the noise with mean 1 and variance cr v . <r v is a measure of 
speckle level. For AIRSAR four-look processed intensity imagery, 
a v =0.5. The filter is 
x = y + b(y - y) (4) 
A 
where x is the filtered pixel value, y is the local mean, and 
b is the weighting function having a value between 0 and 1. The 
parameter b is computed by 
b = 
var(jc) 
var(y) 
and var(x) = 
var(y) - y 2 (J v 2 
l + cr v 2 
(5) 
(var(jc) > 0, var(>’) > 0) 
where var ( v) is the local variance and var (*) ¡ s the variance of 
reflectance without speckle noise. Fig. 2 shows the filtered SAR 
3. POLARIMETRIC SAR SPECKLE FILTERING 
In most case, images produced by radar have noise in 
appearance than comparable pictures produced optically due to a 
phenomenon called speckle. Speckle complicates the image 
interpretation problem by reducing the accuracy of the image 
image. 
4. SEGMENTATION OF THE SAR IMAGE 
Bayes maximum likelihood classification algorithm is applied to