Fast training of neural networks for remote sensing

Abstract

Abstract Fast training algorithms are presented for the training of neural networks used for inversion and classification problems. The algorithms build upon a previously described technique in which linear equations are solved for the network's output weights. First, the method is motivated by an analysis of the multilayer perceptron, based on polynomial basis functions. A conjugate gradient solution to the output weight equations is introduced, which works even when the equations are ill‐conditioned. Techniques are described which can be used to improve hidden unit weights. The output weight optimization technique is extended to classification networks, which have nonlinear output unit activations. Neural networks for inversion of surface scatter parameters and classification of sea ice are designed to illustrate the techniques. It is seen that the techniques are significantly faster than back‐propagation.