The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008 
Learning (training) a network is the process of adapting or 
modifying the connection weights between neurons so that the 
network can fulfil a specific task. Back Propagation is the most 
common learning algorithm which is an iterative gradient 
decent algorithm designed to minimize error function expressed 
in equation 1: 
manually produced reference map which is used for accuracy 
assessment. 
In the first try, the spectral information for each pixel, after 
normalizing the RGB values between 0 and 1, is simply entered 
to the network as its input parameters. 
(Eq.l) 
Where d j and cr represent desired output and current 
response of the neuron “j” in the output layer respectively and 
“L” is the number of neurons in the output layer. In an iterative 
method, corrections to weight parameters are computed and 
added to previous values as below: 
< 
(Eq2; 
Where W- ■ is weight parameter between neuron i and j, TJ is a 
positive constant that controls the amount of adjustment and is 
called learning rate, OC is a momentum factor that can take 
values between 0 and 1 and “t” denotes the iteration number. 
The parameter ^ smoothes rapid changes between the weights. 
Learning rate and momentum parameter have a major influence 
on the success of learning process. 
2.1 Network Structure 
In order to use neural networks in road detection, input layer is 
consisted of neurons the same number as input parameters and 
output layer is made up of just one neuron that shows whether 
the input parameters can represent a road pixel or not. Usually, 
one hidden layer is sufficient, although the number of neurons 
in the hidden layer is often not readily determined (Richard, 
1993). 
In this research, a back propagation neural network with one 
hidden layer is implemented that uses 500 road and 500 
background pixels as its training set. 
An adaptive strategy is used to avoid trail and error learning 
rate and momentum assignment. In this method, both 
parameters are adjusted downwards as half after some training 
intervals if the overall training error has increased and upward 
1.2 times if the overall error has decreased (Heerman and 
Khazeinie, 1992). 
One of the most important factors in employing ANNs is to 
decide what type of information should be extracted from input 
image to be fed through the network as its input parameters. 
The discrimination ability of the network is highly affected by 
chosen input parameters. In continue different input vectors are 
evaluated to find the optimum network design. 
2.2 Network Input Parameters Design 
As a case study, a part of a pan-sharpened true colour (RGB) 
Ikonos image with the size of 550*550 pixels from Kish Island 
in Iran was chosen. Figure 1 shows the original image and its 
Thus, three neurons are designed in input layer in charge of 
receiving spectral values for each pixel in entire image. Figure 
2 shows the network structure and obtained results are 
presented in Table 1. 
Figure 1; a) RGB Ikonos Image from Kish Island in Iran, b) Manually 
produced reference map. 
Hi.N 
Best 
Iteratio 
n 
RCC 
BCC 
RMSE 
Kappa 
Coeff. 
Over 
all 
Acc. 
5 
15000 
77.66 
88.96 
0.2459 
67.92 
93.8 
1 
10 
5000 
73.29 
88.85 
0.2240 
68.34 
94.6 
6 
15 
5000 
73.60 
89.97 
0.2204 
69.46 
94.7 
5 
20 
10000 
74.79 
90.85 
0.2235 
69.30 
94.4 
7 
Table 1: Three spectral values as input parameters 
In this table, RCC and BCC are the road and background 
correctness coefficients that show the percentage of true 
functionality of the network about road and background pixels 
respectively. The RMSE value is computed by comparing of 
network responses in the output neuron and desired value from 
manually produced reference map (1 for road and 0 for 
background pixels). Kappa coefficient and overall accuracy 
parameters are obtained the same way as conventional 
classification methods where the network response about each 
pixel is classified into two classes as road and background 
pixels using an appropriate threshold value. 
In order to improve network’s functionality and considering the 
fact that roads are presented as homogeneous areas in high