Spectral Contextual Classification of Hyperspectral Imagery With Probabilistic Relaxation Labeling

In this paper, a spectral-spatial classification framework based on probabilistic relaxation labeling using compatibility coefficients is proposed for hyperspectral images. It is a two-stage classifier that uses maximum a posteriori (MAP) estimation to maximize posterior probabilities of classification map obtained in first stage to incorporate spatial information for better classification accuracy. Two different forms of compatibility coefficients based on correlation and mutual information are used for MAP estimation. The initial probability estimates are obtained from probabilistic support vector machine (SVM) classifier. The combination of SVM with MAP estimation is investigated and compared with benchmark Markov random field and extended morphological profile-based approaches and some other recent methods. The experimental results are presented for three airborne hyperspectral images. The results reveal that incorporation of contextual information with both forms of compatibility coefficients statistically significantly improved SVM results. The compatibility coefficients based on correlation produced the best results among the relaxation methods outperforming many existing methods.