605 
ROAD EXTRACTION FROM SAR IMAGE 
USING AN IMPROVED STATISTICAL ALGORITHM 
M. Cheng 3 ’*, Q. Ye a,b 
a Department of Surveying and Geo-informatics, Tongji University, Shanghai, 200092, China - cmyxl@163.com 
b The Center of R&S and Spatial Information Technical Research, Tongji University, Shanghai, 200092, China- 
yeqin@mail.tongji.edu.cn 
Commission DI, WG IÏÏ/5 
KEY WORDS: Road extraction, Feature extraction, SAR image, Speckle statistics, Phase grouping 
ABSTRACT: 
An efficient statistical algorithm of automatic extracting road from SAR images is devised in this paper. First, an feature detecting 
operator is used to find the road candidate. Then another smaller operator which calculating the homogenous statistic of locate 
region is applied to reduce the false alarm and smooth the road edge. Finally, the road linear features are extracted by fusing the 
results of these two operators, and a phase grouping method is used to combine the road linear features. We apply this algorithm to a 
RadarSat image to illustrate the accuracy and efficiency, and the performance is satisfied. 
1. INTRODUCTION 
The technique of extracting road from SAR images is widely 
applied in many fields, such as road positioning and 
transportation planning. The road detection requires an 
algorithm with high accuracy and efficiency. The traditional 
artificial recognition method has been limited to its very low 
efficiency, although the road network can be extracted with 
high accuracy. Automatic or semi-automatic extraction has 
been a research subject these years, many efficient approaches 
have been proposed to deal with the road detection from SAR 
images. 
Generally speaking, these road detecting methods can be 
divided into two steps: Firstly, using the edge-detecting 
operators such as difference operator, Canny operator etc to 
calculating the intensity of the neighborhood area of the target 
pixel. Secondly, a global method about the prior knowledge of 
the large range structure is introduced to build up a large span 
linear structure with the linear segments calculated by the first 
step(Chen, 2003). However, the disadvantages such as false 
alarm and linear fractions appear after the extraction due to the 
presence of speckle and the non-stationarity of the image data. 
In this paper, an almost automatic algorithm including two 
operators based on the statistical character of the road area is 
devised to detecting the road linear features. The contrast and 
homogeneity of the road area and the background are taken into 
consideration. The influence of the multiplicative speckle noise 
is effectively controlled and this algorithm can detect the road 
linear features with high accuracy and efficiency. 
2. STATISTICAL PROPERTY 
The road extraction from SAR image is subject to the 
multiplicative speckle noise because of the coherent nature of 
the radar. The speckle noise in SAR images complicates the 
character of the histogram and makes automatic road detecting 
by threshold segmentation difficult(Lee, 1989). Thus the 
statistical property of road area is usually been taken account 
into the extraction, and we shall discuss it next. 
Figure 1. A image with a straight road in the center 
A small image including an almost straight road and its 
background area is intercepted from an integrated SAR image, 
and its width and height are 100 and 20, respectively. The road 
is located in the center of this image, parallel to the horizontal 
direction. Calculating the pixel gray value summation of every 
column, the relationship of abscissa and its corresponding 
summation is described in Figure 2(a). The relationship of 
ordinate and its corresponding summation is described in Figure 
2(b). 
In the horizontal direction which is parallel to the road direction, 
the summation is mainly distributed between 60 and 80. There 
exists some peaks and troughs as a result of the speckle noise 
effect. In the vertical direction which is perpendicular to the 
road direction, the summation of the center road area is lower 
than that of the around. Based on this two drawings, we can 
reach the conclusion that the pixel gray value of the road is 
lower than that of its two neighborhood. It indicates that the 
road is more black than the background. 
To investigate the differences between the road area and its 
background, five images whose size are 5><5 is intercepted from 
the area completely belonging to the road, and another five are 
from the background. The expectation and variance of each 
* Cheng Mingyue. EMailxmyxl@mail.tongji.edu.cn .Tel: 86-21-65981710