The use of backpropagating artificial neural networks in land cover classification

Abstract

Artificial neural networks (ANNs) are used for land cover classification using remotely sensed data. Training of a neural network requires that the user specifies the network structure and sets the learning parameters. In this study, the optimum design of ANNs for classification problems is investigated. Heuristics proposed by a number of researchers to determine the optimum values of network parameters are compared using two datasets. Those heuristics that produce the highest classification accuracies are tested using two independent datasets. Comparisons are also made among the ANNs designed using optimum settings, the ANNs based on the worst performing heuristics, and the maximum likelihood classifier. Results show that the use of ANNs with the settings recommended in this study can produce higher classification accuracies than either alternative. A number of guidelines are constructed from the experiences of this study for the effective design and use of artificial neural networks in the classification of remotely sensed image data.

Acknowledgments

Dr Kavzoglu was supported by a grant from the Turkish Government for postdoctoral research at the University of Nottingham. While the SIR-C SAR data were kindly made available by the NASA Jet Propulsion Laboratory, Pasadena, California, USA, the SPOT HRV data was provided by Logica UK Ltd.