A Bayesian framework for abundance estimation in hyperspectral data using Markov random fields

A model is proposed which uses neighborhoods of pixels as priors in a Bayesian setting to extract abundance information from a hyperspectral image. It is assumed that elements of the abundance vector for a pixel are independent, but that corresponding elements of abundance vectors for neighboring pixels are correlated. A posterior density encourages estimated abundances in neighboring pixels to be similar. Minimum mean-square error estimates are obtained by averaging samples from this density, where the samples are obtained by Gibbs sampling.