COMPARING INFERENCE METHODS FOR CONDITIONAL RANDOM FIELDS FOR HYPERSPECTRAL IMAGE CLASSIFICATION

Classification of hyperspectral images is an important method for various object-based-analysis applications in remote sensing. We propose a two-level learning algorithm combining Support Vector Machines (SVMs) and Conditional Random Fields (CRFs) to achieve accurate classification of hyperspectral images. The hyperspectral data is initially processed by SVMs into a local, pixel based classification which serves as the observations in the CRFs model for generating unary and pairwise potentials. Three inference algorithms: mean field, tree-reweighted belief propagation, and loopy belief propagation are compared in the CRF inference procedure. This two-step algorithm is tested with the publicly available AVIRIS Indian Pines data set, and results from the three listed inference methods are discussed.