Effect of the learning algorithm on the accuracy of sub-pixel land use classifications with multilayer perceptrons

Abstract

Timely and accurate information on the location and the extent of land use types is high up the agenda of several governmental and scientific organizations. Remote sensing, through image classification at the sub-pixel level, is an attractive source of this type of information. The remote sensing community has recognized the multilayer perceptron (MLP) as a popular machine learning technique for performing land use classifications, both at the pixel and at the sub-pixel level. However, theoretical advances in the machine learning community are not easily adopted by the classification practice. An example is the continued use of the gradient descent algorithm for MLP training. In this paper, the accuracy of this standard first order learning algorithm was compared to that of five alternative, second order learning algorithms for performing a sub-pixel classification of land use in Flanders. The result are clear: all second order algorithms perform markedly better than gradient descent, thereby illustrating the importance of translating theoretical advances in MLP training to the classification practice.