119 
reserved in the optical edge image; otherwise, the centred edge 
pixel will be eliminated from the original edge image. 
ording 
airport 
image 
ley are 
obtuse 
1962]) 
to the 
led by 
e (row 
'orm is 
elevant 
orm is 
on of a 
ividing 
At the 
ne, and 
in the 
>re, the 
; points 
n p6 - 
ng and 
etected 
md the 
•logical 
;hat are 
:ient is 
discard 
y holes 
; of the 
digital 
binary 
5 most 
)licity. 
vith the 
get the 
rch the 
e is the 
•undary 
will be 
4. EXPERIMENTS 
Apply the proposed method to the two original remote sensing 
images shown in Section 2. The segmentation algorithm based 
on MRF is realized using simulated annealing method (Chee 
Sun Won, 1987) on the assumption that the SAR image can be 
described by Rayleigh distribution model. 
P&i I x.) = 
y t 
T ex P 
cr(x.) 
(1) 
where, y. is the gray-level value of the pixel, x. is the marker 
of segment result, and 
{ 0 y. is not the target 
1 y. is the target 
(2) 
<j(x.) is the distribution parameter of Rayleigh distribution 
which can be estimated with the mean value ju R (x.) and the 
standard variance a R (x.) of the sampled area from each 
category in the original image by the following two expressions. 
) = v(x i ) 
2 4 — 2T 2 
c R ( x i) = cr (x,) 
2 
(3) 
(4) 
In the simulated annealing algorithm, it is very important to 
choose an appropriate annealing schedule. In the application of 
image segmentation, either an exponential format or an 
logarithmic format is often used. In our experiment, the 
exponential format is adopted for the annealing schedule, i.e. 
(ij 
□ rz 
Figure 3 First-order Neighbourhood System 
and Its Associated Clique Types 
Figure 4 is the segment result of applied MRF-based algorithm 
to the original SAR image shown in Figure 1 (a). It can be seen 
that a lot of regions whose weak backscattering coefficient cr 
is low have been extracted out, and the runway area take on 
good connectivity except for two holes existing within the left 
runway which are caused by two airplanes in the original image. 
Figure 4 Segment Result 
Figure 6 ROI Area Figure 7 Modified ROI Area 
T = T„a‘ (5) 
where, T Q is the initial temperature which has been set to be 4 
for the experiment here, a is the temperature descending index 
which should be a constant within (0, 1), and k is the iteration 
times. In the experiment, a is set to be 0.9 which is experiential 
according to the reference (S. Geman, 1984). And the 
neighbourhood system and its corresponding cliques being used 
here is shown in Figure 3. 
Figure 5 is the edge image of Figure 4 obtained by Roberts 
detector, since the latter is an binary image. Figure 6 is the 
result after removing the irrelevant segmented regions and 
keeping only the runway region according to the geometric 
characteristics of the typical man-made line-type target as 
discussed in Section 3. Hough transform is utilized here to 
extract the region whose edge is characterized by straight and 
long lines and eliminate the others which fail this requirement. 
One can see that the two holes still exist within the runway. 
Filtering the image with morphological operator (opening and 
then closing) can fill the holes and get Figure 7 with better 
integrality. 
The fusion result of the ROI boundary image and the optical 
Canny edge image (Figure 8) is shown in Figure 9. Compared 
with Figure 8, the fake edges existing within the right runway