Abstract
Neural network architecture enables us to attain flexibility for pattern recognition in remote sensing. In this paper, we propose a pattern recognition system of remote sensing data using two kinds of neural networks. It uses both self-organizing and back-propagation (BP) methods. Although the BP method can easily train the network, it is difficult to find exact training data. The self-organization method can calculate connection weights autonomously and select the training data easily. But we must prepare a large size network for input data.
In order to improve ability of pattern recognition, we divide the pattern recognition system into three subsystems. Each of them has been processed in a sequential way. The first one clusters the remote sensing data into some regions based on Kohonen's feature map. At the second step using the clustered images and geographic knowledge, training data sets are classified into many categories based on the BP method. The third step corrects miss-classified pixels by using IF THEN rules. The rules are described on the basis of human criterion for pattern recognition. Effectiveness of the proposed method is illustrated by simulation results of remote sensing TM data for checking data sets observed around Tokushima City in Japan.
