Classification of remote sensing data using fast learning neural networks and topology selection algorithms

Abstract

The classification of remote sensing data has been a topic of interest since the early 70s. Recent studies have demonstrated that neural networks can be effective when used as a classifier of remote sensing data. One problem encountered using neural networks, though, is the time required to train the networks. This amount of time is directly related to the size of the network as well as to the amount of data used during the training. Earlier work demonstrated that a fast learning (FL) training algorithm provided a means to efficiently train neural networks in a fraction of the time required using conventional back-propagation learning algorithms. Additional decreases in the training time can be achieved by minimizing the network topology and analyses of the information content of the data. In this study, a fast learning (FL) neural network (NN) with optimized topology is applied to the problem of classification. Recently, new methods have been devised to determine the optimal size of the network required for a given set of data. The classification of remote sensing data using NNs trained with the FL and topology selection methods is examined. Several different SAR scenes are used to illustrate the technique. It is shown that the methods used are effective at classifying different land types while, at the same time, minimizing the size of the required networks and the time required for training the networks. The minimization of topology size, in addition to reducing training times, also reduces the amount of time for processing additional data. >