The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
Three critical conclusions have been drawn through analyzing 
in Figure 3. 
1 ) The curve is sensitive to the pixel numbers of the road region 
(Nm, middle region in Figure 1). As the decrease of the number; 
the increase of the curve kurtosis that is the curve is deviating 
from symmetry. For the Kolmogorov-Smimov (K-S) hypothesis 
testing theory, in order to get to an unchanged significance level 
& (correlated with constant false alarm rate), the threshold T 
should be increased otherwise it will lead to the increase of 
false alarm rate. 
2) The pixel numbers of the abuttal (right or left) has little 
effect with the curve of the distribution curve (Figure 3(b)). The 
performance of the extraction is not effect with the change of 
the abuttal, when we determine the number of middle region 
(road) and k, the result would be foreseen. 
3) From Figure 3(c), the value of ^ & ( w here, P is the 
mean value of the local image and ® is the standard deviation) 
has a great impact on the curve. Through analyzing, the 
decrease of k corresponds to the more uneven in the region, and 
the more deviate from the model assumption, which will 
destroy the ideal road model sooner or later. The roads can not 
be extracted under very low value of k. 
3.4 The methods of reducing uncertainty 
Used the rules above, we selected some sets of parameters to 
get the required results. The thresholds were determined by the 
K-S hypothesis testing: for the given significance level a , 
testing hypothesis HO: middle region and left region are the 
same landmark (homogeneous region); HI: middle region has 
lower grey than the left region. When HO is true, the value of 
R = X m /X m 
statistic 11 is not very large; when HI is true, 
R always very large, and middle region is road. So the 
formation of rejection field 
R lm =X l /X m >T&&R rm =X/X m >T 
jg L ,m l m r ,m r m 
Based on the curve of R probability distribution function, we 
used normal distribution to approximate the stochastic variable 
R. According to the normal distribution, as significance 
, , a = 0 1 #(1.28) = 0.1 q ..... 
level ^ 1 , v ' .So we can initiate the 
threshold T=1.28. Based on P{reject HO I HO is true}< =i * , we 
can get the critical Threshold T, and satisfy 
P{R, m >T}<a T .R Im >T .... . 
K a) .If l - m , then middle region has 
lower grey than abuttal, the middle region must be road, 
otherwise they are homogeneous region. The probability of 
misjudging is a . 
The critical points of the procedure are the judgement of 
whether region is homogeneous or not. We adopted VI (section 
3.1) to testify. In practice, we need add a correction value( A )on 
R to compare with threshold the next for the correction 
judgement. 
Based on the analysis of section 3.3, we reduced the errors of 
extraction for three aspects. The pixel numbers of the middle 
region (road) can not be taken small value, for example one 
pixel width and fifteen pixel length (Nm=15) road region, the 
road in reality should be 13-40 meters width; too narrower the 
road can not be extracted exactly. Due to the curvature of the 
roads are small, the appearance of the roads in local should be 
linear. The length of the road model should accord with this 
characteristic, too short to extract other landmark and enhance 
the false alarm rate. Certainly, the length longer than gentle 
curve segment is not advisable. Because ^ P ^ & has t h e 
drastic impact on the results, the more noise in the SAR image 
(correspond to the low value of k, the higher of misextract rate. 
4. EXPERIMENTS AND ANALYZING 
Since single-look amplitude SAR is Rayleigh distribution, only 
if look number greater than 4, the amplitude SAR image data is 
approximated Gaussian distribution [4]. So for the low look 
number SAR image, we utilized simulated annealing algorithm 
to handle 2 times of iteration. Simulated annealing algorithm is 
set on the statistical feature of the SAR data, and at the same 
time the resolution of the image can not be change, thus this 
approach can achieve a better equivalent look and meet the 
hypothesis in section 2.2. 
There may be very large size of image, for the calculation 
amount and speed; we trim the large image down to small ones 
just like Figure 3 512 X 512 size. Furthermore, distributed 
computation on multi-computers is applied to the procedure. 
Because of the random directions of the roads, all directions 
extraction based on local road model should be applied. In 
practice, for the compute speed and efficiency, the local road 
model extractor rotates along the directions every 18° around 
the circle over the selected size. That is, we label the roads 
points in SAR image in 0° ,18° ,36° , •••,144° ,162° 
directions. Setting parameters ST=2, Nm, N1, 
A _ 0.522 
fi 
as rules in section 3.4, we can implement 
the extraction as the whole procedure flowchart shown in 
Figure 4. We used VC++ programming the algorithm; at the 
same time set extracting area and flattening calculated by 
T f = 4- n ■ Areal(perimeter) 2 
formula * threshold of 
extracting region, for eliminating the obvious not-roads 
landmark and noise. In several directions extraction some pixels 
have been labeled as roads points several times, we consider 
these points are mainly roads doubtlessly. After directions 
extraction, the main roads net are obtained and maybe need 
some auxiliary process such as filling the holes in the main road 
segments based on mathematical morphology. And the centre 
lines of the roads can be obtained by Hilditch thinning 
algorithm. 
Maybe connecting the roads segments into the roads net is also 
a main problem in roads extraction [1].With-the point of view 
of author, this problem can be eliminated by improving the 
roads algorithm detected the roads precisely. 
The end of practice, some subjective and objective evaluation 
indexes were introduced for comparing the performance in 
various parameters and whether according with prediction by 
the theories above or not. In subjective, 2/3 part of a road 
extracted is considered the whole extraction of the road. 
Although subjective tests can sometimes be authentic if 
performed correctly, they are inevitable inconvenient, vary with 
each individual, subjective Therefore some useful objective 
indexes were introduced. These indexes are completeness (the 
ratio of points extracted correctly to reference roads points, the 
theoretical value is 1), correctness (the ratio of points extracted 
correctly to the whole number of extracted, the theoretical value 
is 1), quality (weighted value of the completeness and the 
correctness, the optimum value of this is 1), RMS (the mean