Bayesian classification by data augmentation

A typical remote sensing data clustering is the maximum likelihood supervised procedure. It consists of the estimation of a suitable mixture of distributions, based on training samples only, and in the subsequent pixel-by-pixel classification, performed by maximizing the likelihood ratio. In this way all the information on the parameters of the distributions, contained in the unsurveyed samples, is lost. In the paper it is proposed to apply a suitable Bayesian method, known as a data augmentation algorithm, to fully exploit the information contained in the data. The method is presented in detail and applied to an elementary simulated example proving its capability of achieving almost the theoretical limit for the classification error. Comparisons with current classification methods as well as an application to a real dataset are reported.