Markov chain random fields, spatial Bayesian networks, and optimal neighborhoods for simulation of categorical fields

Abstract

The Markov chain random field (MCRF) model/theory provides a non-linear spatial Bayesian updating solution at the neighborhood nearest data level for simulating categorical spatial variables. In the MCRF solution, the spatial dependencies among nearest data and the central random variable is a probabilistic directed acyclic graph that conforms to a neighborhood-based Bayesian network on spatial data. By selecting different neighborhood sizes and structures, applying the spatial conditional independence assumption to nearest neighbors, or incorporating ancillary information, one may construct specific MCRF models based on the MCRF general solution for various application purposes. Simplified MCRF models based on assuming the spatial conditional independence of nearest data involve only spatial transition probabilities, and one can implement them easily in sequential simulations. In this article, we prove the spatial Bayesian network characteristic of MCRFs, and test the optimal neighborhoods under the spatial conditional independence assumption. The testing results indicate that the quadrantal (i.e., one nearest datum per quadrant) neighborhood is generally the best choice for the simplified MCRF solution, performing better than other sectored neighborhoods and non-sectored neighborhoods with regard to simulation accuracy and pattern rationality.