A heuristic approach to learning new graph structures for remote sensing images

A probability graph model can effectively model spectral and spatial dependencies within remote sensing images for land cover classification. The most common structure used to unify this probabilistic information is a second order Markov network that encapsulate unary and pairwise potentials. In this paper we explore various heuristics to discover new graph structures that will assist with classifying land cover. Experiments were conducted to compare classification accuracies in two study areas; one homogeneous and one heterogeneous located in the Kwazulu-Natal province, South Africa.