Spatial–spectral hyperspectral classification using local binary patterns and Markov random fields

Abstract. Local binary patterns (LBPs) have been extensively used to yield spatial features for the classification of general imagery, and a few recent works have applied these patterns to the classification of hyperspectral imagery. Although the conventional LBP formulation employs only the signs of differences between a central pixel and its surrounding neighbors, it has been recently demonstrated that the difference magnitudes also possess discriminative information. Consequently, a sign-and-magnitude LBP is proposed to provide a spatial–spectral class-conditional probability for a Bayesian maximum a posteriori formulation of hyperspectral classification wherein the prior probability is provided by a Markov random field. Experimental results demonstrate that the performance of the proposed approach is superior to that of other state-of-the-art algorithms, tending to result in smoother classification maps with fewer erroneous outliers even in the presence of noise.