SPECKLE FILTERING FOR JERS-1/SAR IMAGERY 
Masatoshi MORI, Tomonori YOKOYAMA, and Noboru YAMAMOTO} 
Department of Management Engineering, Kinki University, Iizuka City 820, Japan 
jGraduate School of Advanced Technology, Kinki University, 
Iizuka City 820, Japan 
KEYWORDS: JERS-1, SAR, Speckle noise, Filter algorithm, Radar image, Texture 
ABSTRACT: 
Synthetic Aperture Radar (SAR) images by JERS-1 are enhanced by reducing speckle noise. The Gamma 
MAP filter by Lopes et al. has been applied to ERS-1/SAR or C-SAR images, and they successfully obtained 
the speckle reduced images. The speckle is reduced as a function of the coefficient of variation within a local 
window in the Gamma MAP filter, in which the threshold values are used to classify pixels. The theoretical 
estimate of the thresholds by the Gamma MAP filtering algorithm for JERS-1/SAR images is quite different 
from the estimate by the present analysis of the distribution of the coefficient of variation for 1-look, 2-look, 
and 3-look processes. These thresholds for JERS-1/SAR images are determined based on the distribution of 
the observed local coefficient of variation within a window. SAR images with good quality are obtained by the 
present method of estimating the thresholds. 
1. INTRODUCTION 
First Synthetic Aperture Radar (SAR) images by 
satellite sensors had been obtained by Seasat for Earth 
observation by NASA. However, the active interval of 
observation was short due to a satellite malfunction. 
Now, a massive number of SAR images have been 
acquired by ERS-1/SAR and JERS-1/SAR sensors. 
The only defect of these SAR images is coherent noise 
known as speckle noise. So SAR images are granular in 
quality, and are especially remarkable in homogeneous 
areas such as agricultural areas. 'To date, many re- 
searches have been performed to reduce speckle noise 
(Touzi, 1988; Lopes, 1990). Adaptive speckle filters to 
preserve radiometric and textural information in SAR 
images have been developed by Lopes et al (Nezry, 
1991; Lopes, 1993), which were recently refined, culmi- 
nating in the Gamma MAP (Maximum A Posteriori) 
filter. 
Two thresholds in discriminating between a homoge- 
neous class and a heterogeneous class within a local 
window were introduced in the Gamma MAP filter. 
The operation in that filter is performed depending 
on the coefficient of variation. Many SAR images of 
ERS-1 with high quality are obtained by use of the 
Gamma MAP filter. Two threshold values are, how- 
ever, not so easy or clear to determine. 
2. MODEL FOR SAR IMAGES 
Speckle noise is uniformly distributed on SAR images, 
so the value of the observed intensity includes both 
261 
the true signal value from a target and the speckle 
noise with random process. This multiplicative model 
(Lopes, 1993) is formalized as follows: 
I(t) = S(t)- N(t) (1) 
where I(t) is the observed intensity at the position t 
on an image, S(t) is the true signal value, and N(t) is 
the noise intensity with random process, respectively. 
A texture in images within a homogeneous or het- 
erogeneous class is characterized as a function of the 
coefficient of variation C7(to) within a local window, 
which is defined as 
Cr (to) = o1/T(to) (2) 
where o; is a standard deviation of I(t) within the 
window, /(to) is a mean of I(t), and to is the central 
position of the window. 
The filter operation for SAR images is performed de- 
pending on the value of Cr(to). Regions with a small 
Cr(to), such as agricultural areas, are considered to 
be in the homogeneous class. On the other hand, the 
regions with a large Cr(to) may contain edge or linear 
features. 'Thus, the operation in filtering is defined 
based on two threshold values, Cx and Cuz, Where 
Cy is the speckle coefficient of variation, which is de- 
fined as 
Cn.=on/N (3) 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996