EDGE CONSTRAINED MRF METHOD FOR CLASSIFICATION OF HYPERSPECTRAL IMAGERY

This paper proposes an edge constrained Markov random field (MRF) method for accurate classification of hyperspectral imagery. The characteristic of the proposed method is using an alterable spatial weighting coefficient, which is acquired based on the edge information, instead of a traditional fixed coefficient. In this way, the spatial contributions in MRF model are different for pixels inside objects or on the boundaries. Therefore, the "salt and pepper" inside object can be removed and the "overcorrection" can be resolved. Experimental results of the synthetic hyperspectral data and the real hyperspectral image demonstrate that the proposed method works better on both homogeneous regions and class boundaries with improved classification accuracies.