A preliminary study of three training methods for land cover classification by artificial neural networks

Abstract

This paper reports our preliminary study that aims to examine the effectiveness of training methods for land cover classification by artificial neural networks. We consider three training methods, namely, the gradient descent method, the conjugate gradient method, and the Quasi-Newton method. We apply these methods to derive land cover information from a Landsat Enhanced Thematic Mapper Plus (ETM+) scene covering a urban area. Our initial experiment results suggest training methods can affect the overall efficiency of neural networks in terms of land cover classification accuracy.