The detailed explanation about this subject can be found in 
(Mokhtarzade, et al, 2007). 
Road vectorization 
The vectorization methodology implemented in this research is 
based on the idea of road raster map clustering, first introduced 
by (Doucette et al., 2001). This process can be divided into two 
main steps as “road key point identification” and “road key 
point connection”. 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
A detailed survey on road detection using artificial neural 
networks can be found in (Mokhtarzade and Valadan, 2007). 
Beside spectral values, textural behaviour of road pixels as 
being homogeneous areas is the most outstanding road pixel 
property in the high resolution satellite images. Hence, 
incorporating spectral and textural parameters for road feature 
vector generation is going to be implemented in this research. 
The input image was converted to different levels of intensity 
values and a variety of co-occurrence texture parameters, 
consisting of Energy, Entropy, Contrast, Homogeneity, were 
extracted from different window sizes. 
Except the source satellite image, the preliminary road raster 
map generated from a simple neural network (trained only with 
spectral information) was also used as the input image for 
texture analysis. 
Different combinations of texture and spectral parameters were 
put in the road feature vector and the functionality of the neural 
network was evaluated comparing the road raster map and the 
reference manually determined road pixels. 
It was determined that using all four texture parameters, 
extracted from the preliminary road raster map, accompanied 
by the spectral information of the source image can make the 
optimum road feature vector. This road feature vector could 
improve both road and background detection ability of neural 
network. 
Road Key Point Identification 
In the first step of road key point identification, the road raster 
map, obtained from the road detection process, is segmented 
into different adjacent road patches based on image space 
clustering algorithm. When the clustering is performed, the 
centroid of each road patch is regarded as a road key point. 
In (Doucette et al., 2001), a K-mean crisp clustering algorithm 
with user defined cluster number was applied on high resolution 
road raster map using a uniform distribution of cluster centres. 
This traditional method can produce acceptable results provided 
that the available roads in the raster map share rather the same 
distribution in the image and have similar widths. Furthermore, 
the initial number of clusters, determined by trial and error, has 
a major influence on the success of this method.