550 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
morphological operators, the obtained binary image was 
vectorized. 
Furthermore, (Zhang and Couloigner 2006) evaluated the 
effectiveness of angular texture signature to discriminate among 
parking lots and roads using high resolution satellite images. In 
their research, spectral and textural information were used 
separately for detection of roads and for eliminating of non 
road pixels respectively. 
In this research, a two stages road extraction methodology is 
presented consisted of road detection and a vectorization 
processes. 
Road detection is performed on high-resolution pan-sharpened 
RGB Quick Bird and IKONOS satellite images, using texture 
parameters in artificial neural network algorithms. 
The vectorization procedure is made up of two steps of road 
key point identification and generating road connections. Road 
key point identification is performed using c-means clustering 
on road raster map. For this reason, at first the possibility of 
genetically guided clustering is evaluated. Then a novel 
methodology for flexible road key point determination, called 
increasing ellipse, is proposed. 
Figure 1. The methodology of road extraction 
In this research, a novel method of road raster map clustering is 
developed to identify road key points, where a fuzzy shell 
clustering provides the required information to generate 
vectorized road networks. 
When road key points as the centre of different adjacent road 
patches are determined, a fuzzy shell clustering provides the 
clues for establishment of road segments. 
In section 2, the proposed methodology for both steps of road 
detection and vectorization are described. Section 3 presents the 
obtained practical results and accuracy assessment parameters. 
It should be mentioned that the vectorization step can be 
implemented independently from the road detection step. Hence, 
it could be applied on any road raster map generated from 
different road detection methodologies. 
In the following, the detailed methodologies for each of these 
two main steps are explained in different sections. 
2. METHODOLOGY 
Road networks in high resolution satellite and aerial images are 
presented as elongated homogeneous areas having a distinct 
brightness differences from the background. 
Therefore, the common practice of automatic road extraction 
from high resolution satellite images, as it is implemented in 
this research, consists of two main steps entitled as “Road 
Detection” and “Road Vectorization”. 
Figure 1 shows the diagram of the implemented methodology of 
road extraction in this research. 
The first step of road detection concentrates on discriminating 
between road and background pixels. It is considered as an 
image segmentation process where a meaningful value is 
assigned to each image pixel that can be used as the criterion to 
distinguish between road and non-road pixels. 
In this research, neural networks are applied for road detection 
where different spectral and texture parameters are uses as their 
input parameters. The result of road detection is a binary image, 
representing all detected road pixels which is called “road rater 
map”. 
The vectorization step aims at extracting the road network 
centreline and its sides from the previously produced road raster 
map. 
2.1 Road detection 
In this research, the most common back propagation neural 
networks are used as the image classifiers for road detection. 
Figure 2 shows the designed neural network structure for this 
reason. 
BP NN 
Figure 2. Road detection using neural networks 
As shown in figure 2, the input layer consists of neurons the 
same number as road feature vector dimension where each input 
neuron is in charge of receiving one normalized input parameter. 
Only one hidden layer is designed in the neural network while 
the number of neurons in this layer can be varied. 
The output layer has only one neuron, expressing the neural 
network’s response in the range of [0, 1] as the road association 
value for the interest pixel. 
After applying the trained neural network on the entire input 
image, the road raster map can be produced assuming a 
threshold on the road association value of input image pixels.