The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
587 
resolution images, neighboring pixels in a 3*3 window were 
participated in input vector generation. Furthermore, the 
normalized distances of all 9 pixels in the mentioned window to 
the mean vector of road pixels are added to input parameters. It 
was intended to use more spectral information to the network in 
order to increase the discrimination ability of the network 
between road and background pixels and also to reduce the 
iteration times in learning stage. The proposed distance 
parameter witch is calculated for each pixel as below: 
~R 
~R~ 
G 
5 
G 
B 
i 
B 
d ‘ = 44lW ( *” “ ^ + (°- ~ G >) 2 - B $f ( E S- 
While, \R m G m B m J. is the mean of road pixels in 
training set. The maximum distance in the spectral space, is 
441.673. 
Thus the input layer is consisted of 9 red, 9 green, 9 blue and 
finally 9 normalized distances to the road mean vector which 
means 36 neurons are designed in this layer. Figure 3 shows 
network's structure and Table 2 presents obtained results. 
Figure 3: Network structure when neighbor pixels with their normalized 
distances form input parameters 
Hi.N 
Best 
Iteration 
RCC 
BCC 
RMSE 
Kappa 
Coeff. 
Overall 
Ace. 
5 
1000 
75.60 
93.85 
0.2011 
71.31 
95.11 
10 
1000 
75.20 
94.60 
0.2002 
71.60 
95.15 
15 
1500 
75.51 
95.57 
0.2014 
71.86 
95.15 
20 
1500 
76.60 
95.24 
0.2008 
72.18 
95.17 
Table 2: Spatial information and normalized distance as input 
parameters 
Comparison between Table 1 and 2 shows that designed input 
parameters could improve network's ability in both road and 
background detection. Although input layer enlargement can 
make training stage more time consuming, but this problem is 
compensated to some extent by decrease in requested hidden 
layer size and iteration time. 
Finally, as a comparison with statistical methods, obtained 
results from Maximum-Likelihood classification method and 
best network from Tables 1 and 2 are shown together in Figure 
4 with their accuracy assessment parameters. A detailed 
explanation of the road detection strategy used in this research 
can be found in (Mokhtarzade and Valadan Zoej, 2007). 
3. AUTOMATIC ROAD VECTORIZATION 
In the image shown in Figure 4, which is obtained from 
improved BNN, all pixels on the road surface were detected. 
But what we need for GIS is the vectorized road centerline that 
3)should be extracted from road raster map. 
An algorithm was devised to extract road centerline and 
vectorize it so that it is ready to be entered into GIS as a road 
vector layer as explained in the following sections. 
3.1 Road Thinning and Centerline Extraction 
In order to find road centerline from the available road raster 
map, thinning algorithms could be used to find road skeleton, 
which is a one pixel width road representation. Morphologic 
operators and an iterative boundary erosion algorithm were 
implemented to find the road centerline and omit other road 
pixels from the binary image. The obtained result is shown in 
Figure 5. 
It can be seen that the obtained image shows the main skeleton 
of the roads, but there are two kinds of unwanted pixels that 
should be removed before vectorization: 
Single elements distributed all over the image which are not in 
the road class 
Spur pixels in the road network that are caused by thinning 
algorithm. 
In the following sections the strategies for removing these 
pixels are explained. 
3.1.1 Removing Single Elements 
Single elements that do not belong to road class are distributed 
all over the image as small sets of isolated pixels. The following 
algorithm has been used to remove them. At fist a run length 
encoding algorithm is implemented (Haralick et al., 1992) and 
then contiguous runs are combined to form a single run. Finally, 
consistent runs are encoded as separate objects. Those objects 
that are comprised from less than a predetermined pixel number 
are regarded as unwanted isolated pixels and are removed in the 
next step. 
It should be noted that the success of this method is directly 
influenced by the defined threshold that is a problem dependent 
value. In this research, the optimum threshold value was found 
in an trial-and-error as 20 pixels. The obtained result is shown 
in Figure 6. 
3.1.2 Removing Spur Pixels 
The term "Spur Pixels" is assigned to non-road small linear 
elements that are connected to the main road skeleton at one 
end and the other ends is free. The length of these elements is 
negligible in comparison with the main road network. They are 
not omitted in the previous section as there is a link between 
them and main road pixels. In order to remove them, the