Contextual bayesian classification of remotely sensed data

Abstract

We consider the problem of Bayesian contextual classification with spatial correlation structure. We assume there is training data from pixels which are physically far from the pixel with observation z0 to be classified and its immediate neighbors, so that they can be considered independent. We adopt the assumption of local spatial continuity. We classify an unknown observation z0 into one of K normal populations by choosing that population which maximizes the predictive classification probability, which is shown to be proportional to a matrix T-density in z0 and its neighbors, and centered at a matrix of generalized means of the training data set. Results are studied using simulated data.