CLASSIFICATION OF REMOTE SENSING IMAGERY USING AN UNSUPERVISED NEURAL NETWORK 
C.F.Chen?*, S.W.Chen*, andS.D.Shyn* 
& Center for Space and Remote Sensing Research 
National Central University 
Chung-Li, TAIWAN, R.O.C. 
# Dept. of Information and Computer Education 
National Taiwan Normal University 
Taipei, TAIWAN, R.O.C. 
Commission VII, Working Group 1 
KEYWORDS: Land Use, Classification, Neural Network, Spot 
ABSTRACT 
This paper presents an unsupervised method for classification of remotely sensed imagery. The main body of the 
method is an unsupervised artificial neural network model called the adaptive resonance theory 2 (ART2). The primary 
use of the model is to produce a fine classification from multispectral data. Since ART2 is able to learn and classify by 
self-organization without the help of the training data, it is therefore used to generate numerous spectral classes. Then 
the spectral classes are regrouped using a hierarchical clustering technique. The main objective of the clustering is 
simply to regroup all the relevant spectral classes to the appropriate information classes. The proposed method is 
tested using an artificial image and a Spot image. The results indicate that the method can provide a successful and 
fine discrimination between different spectral classes and the final regrouping accuracy can reach a satisfied level. It 
appears that the proposed method is feasible and useful for classifying remotely sensed data. 
1. INTRODUCTION 1988), to perform an initial multispectral classification. 
The main objective of this stage is to obtain spectral 
Information of  land-cover/land-use extracted from classes as fine as possible. The resultant classes is 
remotely sensed imagery has represented a very then re-grouped by a hierarchical clustering algorithm 
important data for natural resource planning. Recently, (Jain, 1988) to attain a small set of information classes at 
the use of neural network for image classification has second stage. In this paper, the fundamental idea of 
received a great deal of interests. For instance, the ART2 and hierarchical clustering are addressed in 
neural network has been applied to Landsat TM (Bischof ^ section 2. The following section 3 and 4 present the test 
and et al., 1992), Spot (Dreyer 1993), and radar image (Hara, data used in this study and their results and discussion. 
1994). All these studies have demonstrated the Finally, conclusion remarks are given in section 5. 
usefulness of neural networks in various remote sensing 
data. However, most effort of these studies has put in 
the use of supervised neural network. The supervised 2. METHODS 
approach normally requires training set to classify the 
image into useful categories. The selection of training 2.1 ART2 Neural Classifier 
set ordinarily needs the aids from the user, such as the 
prior knowledge about the region to be classified, the Many neural network models tend to forget old 
manual identification of appropriate training region, and information if they attempt to learn new information. The 
intensive analysis of the training data. As a result, the adaptive resonance theory (ART) neural network can 
supervised classification is a highly user-dependent learn many things without necessarily forgetting things 
process (Lillesand, 1995). It is obvious that the learned in the past. In addition, the network is able to 
accuracy of the unsupervised classification is highly self-adapt configurations in real time for retaining codes 
dependent on the training data selected. Unsupervised of categories in response to input pattern, that can be 
neural network has been successfully applied to presented in any order. They are several available 
recognize the characters, speech, and patterns (Kosko, types of the ART neural networks. In this study, the 
1990). but very few has been applied to remotely sensed ART2 network is employed for the classification because 
image. of its potential in dealing with gray scale images. Figure 
This study proposes an unsupervised neural network 1 depicts the ART2 architecture which consists of two 
classification for remote sensing image. Two stages are major modules: the attentional and orienting modules. 
included in the proposed algorithm. At first stage, we The attentional module is further divided into two fields: 
use an unsupervised neural network, called Adaptive an input representation field F1 and a category 
Resonance Theory 2 (ART2) (Carpenter and Grossberg, representation field F2. There are six layers, w, x, v, u, 
112 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996