The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
553 
produced reference images applied in accuracy assessment 
procedure. 
Figure 6. Pan-sharpened Quick Bird image of Bushehr harbor 
and its manually produced reference image 
Figure 7. Pan-sharpened IKONOS image of Kish Island and its 
manually produced reference image 
In the following sections, the practical results of different step 
of road extraction are presented accentuating on practical 
aspects of the implementation. 
3.1 Implementation Results of Road Detection 
Road detection was performed using an artificial neural 
network consisting of 7 neuron in its input layer in charge of 
receiving 3 spectral values (R, G, B) and 4 textural parameters 
as explained in section 2.1. The hidden layer was made up of 10 
neurons and the output layer, having only one neuron, was 
designed to show the response of neural network. 
Sample #1 
Sample #2 
RCC 
BCC 
RMSE 
RCC 
BCC 
RMSE 
No Texture 
Parameter 
82.36 
93.53 
0.172 
77.05 
90.86 
0.259 
Using 
Texture 
93.54 
96.31 
0.106 
80.77 
96.15 
0.196 
Parameters 
Table 1. Accuracy assessment of road detection procedure 
About 500 road and 500 background pixels were selected from 
each input image to be used in neural network training stage. 
An adaptive strategy was applied for learning rate and 
momentum parameters to stabilize the training stage of the 
neural network. 
In order to evaluate the performance of the road detection 
procedure, three quality control parameters, RCC, BCC and 
RMSE were used. 
RCC and BCC, stand for “Road/Background Detection 
Correctness Coefficient” respectively, are the average of correct 
neural network response for road and background detection by 
comparison the manually produced reference image (Figures 6- 
b and 7b). Regarding the difference between the neural network 
response and its true expected response (0 for background and 1 
for road pixels) as the error values, the Root Mean Square Error 
(RMSE) can be computed as the third accuracy assessment 
parameter. 
Figure 8 show the neural network road detection results for 
input pan-sharpened images of Figures 6a and 7a. These gray 
scale images are produced by multiplying the normalized neural 
network output by 255. 
Figure 8. Neural network road detection results 
In Figure 8, the left side images (8a and 8d) show the obtained 
result of simple neural network where no texture parameter is 
used. Right side images of Figure 8 (8b and 8d) depicts the 
output of the proposed neural network structure where texture 
parameters of the preliminary road raster maps ( Figures 8a and 
8c) are used beside spectral information for neural network 
input parameters set generation. 
Table 1 show the obtained accuracy assessment parameters for 
both cases where the input source image of Figures 6a and 7a 
are called sample# 1 and Sample#2. 
The presented accuracy assessment parameters in Table 1 show 
that both road and background detection ability of the textural 
improved neural network are improved and thus the efficiency 
of the proposed road detection methodology in this research is 
approved. 
3.2 Implementation Results of Road Vectorization 
The obtained results of improved neural networks (Figures 8b 
and 8d) were converted to road raster map putting a threshold 
on the grey scale values. The obtained road raster maps were 
used in the road vectorization process described in section 2.2. 
At the first attempt, genetically guided road key point 
determination was performed on a simulated road raster map. 
Although the obtained result was acceptable, the computation 
time, even for the small size simulated road raster map, was