The quick bird high resolution remote sensing image of 
Shenyang which is obtained in September, 2006 is used as the 
research data(see Figure l),In this paper, the software ENVI and 
Matlab Basic Mentality is used. First, we make a pre-treatment 
to the remote sensing image; Second we use the supervised 
classification method- Support Vector Machine to classify the 
remote sensing image, convert the classified image into binary 
image; then we use the mathematics morphology method to 
simplify the image data, maintain their basic shape 
characteristics, and except the irrelevant structure characteristic, 
the tiny branch and the noise, extract the road skeleton; at last 
we use the seed growing algorithm method to extract the road 
median line of certain length and direction, and overlay it with 
the original image to assess precision 
2.2 Image Clustering Segmentation based on Support 
Vector Machine 
The Support Vector Machine classification (Support Vector 
Machine is SVM) is one kind of machine learning method that 
is establishment in the foundation of the statistical study 
theory(Statistical Learning Theory or SLT) , with the aid of the 
optimization method to solve machine learning question, its 
main thought aims at two classification questions, seeks a 
optimization classification in the hyper plane which is token 
as the classified plane to guarantee the smallest classification 
error [ 18 \ This method suits the limited sample (small sample) 
question, has solved the problems existing in the traditional 
method(such as neural net) in the great degree , like model 
choice, study on-linearity, multi-dimensional question, partial 
minimum point question and so on. The biggest different 
between the Support Vector Machine(SVM) and the traditional 
statistical pattem is that the Support Vector Machine cannot 
cause the Hughes phenomenon (Hughes phenomenon) - to the 
limited training sample along with the characteristic 
dimension increases the classification precision reduces, and 
when two categories spectrum average values extremely 
approach, SVM also can separate these two categories 
according to these limited samples .But for remote sensing 
image ,characteristic dimension generally is many, moreover 
the category spectrum is quite close in the panchromatic image. 
Appling SVM to remote sensing image of the multi- spectral, 
the high spectrum or the high spatial resolution can obtain good 
effect, also can enhance the remote sensing image classification 
precision. This classified method is insensitive to the noise, 
increases the classification precision, suits the non-linear 
classification, the classification result is neat, suits GIS, 
extremely suits to four wave bands high resolution data. The 
following(see Figure 2) is the classification result 
2. 
Figure 1 The original image 
EXPERIMENTAL RESEARCH OF ROAD 
EXTRACTION 
This method mainly includes the following several steps: The 
image pretreatment; use the Support Vector Machine 
segmentation method to obtain binary image; process the binary 
image by mathematics morphology method, obtain the skeleton 
image; through limiting the length and the direction to reject the 
tiny branch using seed growth to obtain the final road median 
line; finally superpose the road median line with the original 
image to assess precision. 
2.1 Image Pretreating 
The mainly work of the Image pretreatment is image 
enhancement, deleting cloud and mist noise, the aim is to wipe 
off irrelevance noise, enhance image quality, stand out the 
need information ,which is advantageous to interpretation and 
further processes, there are many image enhancement methods, 
such as spatial filter, color transform ,image operation, multi 
spectrum change and so on .But the concrete method must meet 
the concrete experiment image data need, the image in this 
paper hasn’t cloud and mist, is carried on color transform, 
according to the experiment. 
Figure 2 The result of Support Vector Machine Classification 
Stochastically selects 30 ground control points to carry on the 
precision analysis, the reference [19] thought that the extracting 
result quality has very big relation with the initial clustering, 
whether clustering precise is high or not plays a important role 
to the road. This method classification overall precision 
achieves 88.406%, the kappa coefficient is 84.4%.(see Figure3).