A spatial stochastic model for contextual pattern recognition

A contextual classification algorithm using spatial stochastic model (Markov random field) is proposed. The requirements for the joint probability function on the two-dimensional lattice are discussed. The distinction between the spatial correlation context and the transition probability context is made. The procedures for construction of the model are given with details left out but conceptually clear. Coding technique toward parameter estimation is presented. Extension of the model in the multivariate site variable case is derived to handle the multispectral satellite data. Experiments with remote sensing data are performed and results are compared with simple (no context) rule result. Less frequently occurred classes like highway, commercial areas were found to be classified better using the contextual algorithm with only reasonable computation increase.