Analysis and optimization of neural networks for remote sensing

Abstract A technique for improving the topology of a trained neural network, used for an inversion or classification problem, is presented. The technique models the multilayer perceptron as a power series, which allows us to (1) remove units from the network which are well‐approximated by zero‐degree or first‐degree polynomials, (2) measure the effect of removing a hidden layer, and (3) determine the degree of the overall polynomial discriminant which approximates the network. The smaller, pruned networks can process data faster than can the larger original networks. The network degree is a direct measure of the nonlinearity inherent in the particular inversion or classification problem of interest. Neural networks for inversion of surface scattering parameters and classification of sea ice are analyzed to illustrate the technique.