tanbul 2004 
Generalized Cooccurrence Matrix to Classify IRS-1D Images using Neural Network 
see ; d = 2 d os A b 
E. Hosseini Aria , M.R.Saradjian , J. Amini , and C. Lucas 
a x i t. : = - . : 
Remote Sensing Division, Surveying and Geomatics Engineering Department, 
Faculty of Engineering, University of Tehran, Tehran, Iran 
aria@engineer.com , sarajian(@ut.ac.ir, jamini(@ut.ac.ir 
  
b : mn d 
Electrical and Computer Engineering Department, 
Faculty of Engineering, University of Tehran, Tehran, Iran 
KEY WORDS: Neural Network, Classification, IRS-image, Feature, Multispectral, Segmentation 
ABSTRACT: 
This paper presents multispectral texture analysis for classification based on a generalized cooccurrence matrix. Statistical and 
texture features have been obtained from the first order probability distribution and generalized cooccurrence matrix. The features 
along with the gray value of the selected pixels are fed into the neural network. Frist, Self Organizing Map (SOM) that is an 
unsupervised network, has been used for segmentation of IRS-1D images. Then a generalized cooccurrence matrix and first order 
probability distribution have been extracted from each kind of segments. Texture features have been obtained from generalized 
cooccurrence matrix. The matrices describe relevant “texture” properties of classes. Next, feature vectors are generated from the 
extracted features. Then the image is classified by Multilayer Perceptron (MLP) network which has been trained separately using the 
selected pixels. The method used in this paper has been tested on the IRS-1D satellite image of Iran. The Experimental result is 
compared to the Maximum Likelihood Classification (MLC) result and it has been shown the MLP method is more accurate than 
MLC method and also is more sensitive to training sites. 
1. INTRODUCTION 
Artificial neural networks can be seen as highly parallel 
dynamical systems consisting of multiple simple units that 
can perform transformation by means of their state response 
to their input information. How the transformation is carried 
out depends on the Neural Network (NN) model and its way 
of learning the transformation. Neural network learns by 
example. In a typical scenario, a neural network is presented 
iteratively with a set of sample, known as the training set, 
from which the network can learn the values of its internal 
parameters. 
During the last few years the number of reported applications 
about the use of neural network in remote sensing, have been 
steadily increasing. The majority of applications have used 
the multilayer perceptron neural network trained with back- 
propagation algorithm although applications employing the 
self-organizing feature maps have also been reported. 
MLP networks are general-purpose, flexible, and nonlinear 
models consisting of a number of units organized into 
multiple layers. The complexity of the MLP networks can be 
changed by varying the number of layers and the number of 
units in each layer. Given enough hidden units and enough 
data, it has been shown that MLPs can approximate virtually 
any function to any desired accuracy. MLPs are valuable 
tools in problems when one has little or no knowledge about 
the form of the relationship between input vectors and their 
corresponding outputs. 
In order to approach higher classification accuracies it is 
necessary to consider texture information and neighborhood 
information around each pixel. A cooccurrence matrix gives 
some texture information. Augusteijn et al. (Augusteijn 
,1995) compared the performance of several texture measures 
for classifying land cover classes in satellite images. One of 
these texture measures was cooccurrence matrices. Their 
experiments showed that neural networks can give excellent 
results with texture features. 
117 
Since the feature extracting is time consuming process for the 
whole image, the image segmentation has been made first. 
Image segmentation is the process of division of the image 
into regions with similar attributes (Pratt, 2001). The self- 
organizing map has been used successfully for image 
segmentation (Ohta, 1980). The algorithms have been 
implemented on the IRS-ID images from which three 
spectral bands have been selected and fused with the PAN 
band to create an image with 5.8 m spatial resolution. 
2. FEATURE EXTRACTION 
Remotely sensed data and the land cover/land use 
classification of urban areas set their own requirement for 
feature extraction. Features should be easily computed, 
robust, insensitive to various distortions and variations in the 
images, and they should support the discrimination of the 
land cover/land use classes. 
In this paper the following two basic feature groups are used: 
-Statistical features showing the intensities and intensity 
variations of pixels. 
-Texture features based on gray level cooccurrence matrix. 
2.1 Statistical features 
The most basic of all image features is some measure of 
image amplitude in terms of luminance, spectral value, or 
other units. One of the simple ways to extract statistical 
features in an image is to use the first-order probability 
distribution of the amplitude of the quantized image. They 
are generally easy to compute and largely heuristic. The first 
order histogram estimate of P(b) is simply 
N(b) (1) 
  
P(A) = 
where bis a gray level in an image,