R image near 
in noticeable 
| geocoding, a 
image is then 
“the same ter- 
i-to-SAR rela- 
ic transforma- 
e is remapped, 
cted image. 
data such as 
nerated (these 
SAR viewing 
on). 
ata and result 
ENT 
nakes images 
e order of the 
ear as speckle 
m distribution 
| in the SAR 
mage reduces 
also degrades 
des a number 
s well as other 
hanced using 
ion as well as 
|A P-Refined. 
  
  
Figure 3: A subregion of the DTED corresponds to 
the SAR image shown in Fig. 2. Data is sampled in a 
lat-long grid with spacing of 3' and 6', respectively. 
  
Figure 4: The simulated SAR image generated from 
the DTED shown in Fig. 3 and the known SAR 
platform parameters. 
Figure 5: The geocoded image of Fig. 2 to the same 
lat./long. grid as the DTED image. ((OESA 1991) 
Speckle can be modelled as a multiplicative noise 
process. Both Frost [1] and Lee [5] devised linear 
adaptive filters which incorporate multiplicative noise 
statistical properties. Both filters employ a minimum 
mean square error (MMSE) approach and are compu- 
tationally efficient. The results of filtering a SAR image 
(Fig. 6) are shown in Fig. 7 (Frost) and Fig. 8 (Lee). 
These filters exhibit good speckle reduction with min- 
imal loss of sharpness. 
Both the Frost and Lee filters do not assume an 
explicit model for the underlying (speckle-noiseless) 
signal and incorporate only its local mean and variance. 
As well, these MMSE filters assume that the speckle 
is everywhere fully developed and are optimal only if 
both the received and underlying signals are gaussian. 
These shortcomings are addressed using à maximum 
a posteriori (MAP) filter considered by Kuan et al. 
[4] and modified by Lopes et al. [6]. The MAP 
filter implemented in EV-SAR allows modelling of the 
underlying signal using either a symmetrical Beta or 
Gamma distribution. It locally determines thresholds, 
above and below which filtering is not applied, instead 
retaining the pixel value in question or replacing it by 
the local mean. The result of applying the MAP filter 
is shown in Fig. 9. A refinement can also be done 
by utilizing edge and line ratio detectors to separate 
texturally different regions in the local application 
of the MAP filter. Figure 10 shows the results of 
the MAP-refined filter. The MAP and MAP-refined 
filters provide excellent results, suitable for subsequent 
classification of the image contents. 
301