An improved Markov random field classification approach for hyperspectral data based on efficient belief propagation

Aiming at the problems of imprecise class conditional probability (CCP) estimation and heavy computational cost for the global energy minimum in Markov random field (MRF) based classification algorithm, an improved MRF approach based on efficient belief propagation (EBP) is developed for land-cover classification of hyperspectral data. The estimation accuracy of the CCP is improved by the probabilistic support vector machine (PSVM) algorithm using spectral information of pixels, then the spatial correlation information is introduced by the MRF classification algorithm, thus the spectral information and spatial information is combined effectively. Moreover, an EBP optimization algorithm is developed, by which the computational cost is reduced and the classification accuracy is improved. The experimental results show that the proposed approach is effective. The average classification accuracy is up to 95.78%, Kappa coefficient is 93.34%, and the computational time of EBP is about 25% of that by belief propagation algorithm. Therefore, the proposed approach is valuable in land-cover classification application for hyperspectral data with low computational cost and high classification accuracy.