An Application of Combined Neural Networks to Remotely Sensed Images

Abstract

Studies in the area of pattern recognition have indicated that in most cases a classifier performs differently from one pattern class to another. This observation gave birth to the idea of combining the individual results from different classifiers to derive a consensus decision. This work investigates the potential of combining neural networks to remotely sensed images. A classifier system is built by integrating the results of a plurarity of feed-forward neural networks, each of them designed to have the best performance for one class. Fuzzy Integrals are used as the combining strategy. Experiments carried out to evaluate the system, using a satellite image of an area undergoing a rapid degradation process, have shown that the combination may yield a better performance than that of a single neural network.