The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
607 
*4=1" 
P 
V 255 
(3) 
While traversing with the second operator, it should be noted 
whether the target pixel is a road candidate point. For candidate 
points and non-candidate points, different weights(X) should be 
attached. Thus, we can get detector R for roads’ linear 
characteristic extraction by formula (4). 
R = R { *R 2 »R 2 *R 4 (4) 
Whether a pixel point is a road point or not is decided by a 
threshold value T. if R>T, then it is a road point. 
4. CONNECT THE LINEAR FEATURES 
The road linear features are detected by the feature detector 
devised in the previous section. However, these linear features 
are always short and dispersed road segments. Effective method 
should be found to compose the segments to obtain significant 
road curves. The method in this paper mainly includes three 
steps(Xie, 2007): Firstly, mark every segment of the road. That 
is to mark the points which belong to the same road in a small 
area to obtain a structural body of this road. Secondly, do 
further connection with these segments on the basis of some 
particular rule. Finally, add constraint to look for the best 
connection method. 
Generally speaking, slender segments are more likely to be a 
part of a road, while the isolated points and smallish segments 
are tend to be fake detection or other line-similar 
segments(Zhang, 2007). Also because the road are all 
continuous, near one segment there should be another segments 
which have the same direction with it. According to these two 
features above, applying two measures of the region area and 
the oblateness (T f = 4 • n ■ Area/(perimeter) 2 ) to screen 
the road segments and keep only the reliable ones. Give the area 
thesholding value Ta and the oblateness thesholding value T f . If 
the area of the road segment is smaller than Ta and the 
oblateness is bigger than Tf, and also there are no close 
segments in the same directions around it, then this segment 
should be recognised as non road segment. 
Every segment obtained has its starting point and end point. It is 
defined that the starting point is the most left-superior or left- 
inferior one, while the end point is the most right-superior or 
right-inferior one. If the segment is vertical, then its starting 
point is the lower one and the end point is the upper one. It is a 
natural choice that to connect adjacent segments into a line. It 
improves the description of the linear character of detected 
segments, and simplifies the relationship between them. 
Therefore the straight line segments will be the foundation of 
the following organising work. All the line segments obtained 
in the linear feature detection have their own direction. If two 
adjacent segments have similar direction, then we consider 
these two segments to be on the same line. Thus we shall get 
the linear road by connecting two segments based on this phase 
grouping method. 
5. EXPERIMENT 
In this paper we experimentized a RadarSat image including 
road area taken in Tangshan, Hebei province, size of which is 
512X512. Using the method putting out in this article to extract 
the road from SAR image, and then do erosion, expantion and 
thinning, we can get results shown in Figure 3(b). 
(a) Original 
Figure 3. Experiment images 
Almost all the main roads are extracted form the SAR image 
using our method, especially the straight roads are extracted 
continuously and smoothly. However, there are some road 
fractions because of the lack of information. Based on the result, 
we can get that: 
1) The method can effectively control the influence of the 
multiplicative speckle noise, especially when the background is 
unhomogeneous. 
2) The extraction effect is better in the area in which the road is 
long and not very curving, as well as the linear fractions are less 
than the curving road. 
3) Threshold has played an important role in the extraction. 
Especially to the first operator, the threshold should be large 
enough to reserve the most possible road candidate, however, it 
results in a higher probability of false alarm.