33. Istanbul 2004 
nages. Next, the 
'h and low areas, 
>monstrated in a 
f extraction trees 
m that algorithm 
d points in aerial 
istrate that the 
observed points 
future, with this 
forest affect the 
  
€ 
(96) 
Mock 2 
41 
2,33,42 
9,23,16 
10,9,11 
759 
9,28,24 
10,8,15 
6,49,39 
7,4,1 
8-22 
4.13727 
7722 
> forest-covered 
ying Perceptual 
ion and Stereo 
y. IntArchPhRS, 
on of Range and 
| Terrain Model 
of 2^4 Airborne 
10. 
Semi-Automatic 
p.457-461. 
nition from 
omatic Aerial 
ROAD DETECTION FROM HIGH RESOLUTION SATELLITE IMAGES USING 
ARTIFICIAL NEURAL NETWORKS 
M.J Valadan Zoej ^, M. Mokhtarzade 
a,b 
* ValadanZouj@kntu.ac.ir , * 
Faculty of Geodesy and Geomatics Eng., K.N.Toosi University of Technology, Tehran, Iran 
m mokhtarzade(gyahoo.com 
Commission WG III/4 
KEY WORDS: Remote Sensing, Extraction, Neural, Networks, Learning, High resolution, IKONOS, Quickbird 
ABSTRACT: 
In this article, the possibility of using artificial neural networks for road detection from high resolution satellite images is tested on a 
part of RGB Ikonos and Quick-Bird images from Kish Island and Bushehr Harbour respectively. Then, the effects of different input 
parameters on network's ability are verified to find out optimum input vector for this problem. A variety of network structures with 
different iteration times are used to determine the best network structure and termination condition in training stage. 
It was discovered when the input parameters are made up of spectral information and distances of pixels to road mean vector in a 
A 
3*3 window, network's ability in both road and background detection can be improved in comparison with simple networks that just 
use spectral information of a single pixel in their input vector. 
1. INTRODUCTION 
Roads are one of the most important manmade objects in map 
production from satellite images that must be extracted 
accurately. Nowadays when satellite images have improved 
highly in both spatial and spectral resolutions and they are 
available in short time intervals, automatic road extraction 1s 
one of the challenges in remote sensing and photogrammetry. 
Road detection can be considered as the first step in road 
extraction. It is defined as the process of assigning a value to 
each pixel that can be used as a criterion to extinguish road and 
non-road pixels. 
Artificial Neural Networks (ANN) are computational systems, 
inspired from biological neural networks, in which a set of input 
parameters is related to an output set by a transformation 
encoded in the network weights (Yang,1995). In this respect 
they can be used as a function to receive input data from 
satellite images as their input parameters and calculate the road 
detection criterion mentioned above. 
In this article the possibility of using ANNS as road detectors is 
tested on an RGB Ikonos and Quick-Bird images and the effects 
of input parameters on network's functionality are verified. 
2. ARTIFICIAL NEURAL NETWORKS (ANNS) 
Neural Networks are made up of simple processing units called 
nodes or neurodes. The main task associated with a neurode is 
to receive input from its neighbours (the output of other 
neurodes), compute an output and send that output to its 
neighbours (Yang, 1995). 
Neurods are usually organized into layers with full or random 
connections between successive layers. There are three types of 
layers: input, hidden and output layers in charge of receiving, 
processing and presenting the final results respectively (Figure 
1). 
  
  
  
INPUT OUTPUT 
DATA DATA 
  
  
(input layer) hidden layer 1 hsdden layer 2 oatpat layer 
  
Figure (1): A typical neural network 
There are two main stages in the operation of an ANN 
classifier: learning and recalling. 
Learning (training) is the process of adapting or modifying the 
connection weights so that the network can fulfil a specific task 
and is usually done in an iterative way. This process is mainly 
carried out using a training set which comprises some known 
input-output samples. This kind of training is called training 
with a teacher or supervised learning. 
Back-propagation is the most common learning algorithm that 
was discovered by Rumelhart. and Parker independently in the 
early 1980s. It is an iterative gradient algorithm designed to 
minimize the Error function. The error function is shown in 
equation 1. 
2 
L 
Ye) cun 
Je 
| 
E 
SF GIN 
Where ^ and  / represent desired output and current 
response of the neurode “j” in the output layer respectively and 
“L” is the number of neurodes in the output layer. In an