Pruning projection pursuit models for improved cloud detection in AVIRIS imagery

A projection pursuit (PP) method is used to find structure and reduce the complexity of high-dimensional remote sensing data. Individual projection pursuit networks extract features from gray-level difference vector distributions, sum and difference histograms, or simple normalizations of raw pixel intensity from one of four spectral bands used in the study. A PP pruning technique based on an online perturbation analysis similar to that of LeCun, Denker, and Solla (1990), is used to remove parameters of low significance. The four AVIRIS spectral channels studied here were chosen because of their similarity to those which will be available from the multi-angle imaging spectro-radiometer, an instrument which will be on EOS satellites. Ensemble models, which combine features extracted from AVIRIS imagery by multiple projection pursuit networks, use backward error propagation with a cross-entropy objective function to obtain pixel classifications. Predicted cloud masks are compared against human interpretation masks