Multispectral image classification using a new bayesian approach with weighted markov random fields

Abstract

This paper presents a novel nonparametric supervised spectral-spatial classification method for multispectral image. In multispectral images, if an un- known pixel shows similar digital number (DN) vectors as pixels in the training class, it will obtain higher posterior probability when assuming DN vectors of different classes follow a type of statistical distribution. The proposed method assumes the DN vectors follow a Gaussian mixture distribution in each class. Particularly, we use Bayesian nonparametric method to adaptively estimate the parameters in Gaussian mixture model. Then, we construct an anisotropic multilevel logistic spatial prior to capture the spatial contextual information provided by multispectral image. Finally, simulated annealing optimization algorithm is used to accomplish the maximum a posteriori classification. The proposed approach is compared with recently advanced multispectral image classification methods. The comparison results of classification suggested that the proposed approach outperformed other classifiers in overall accuracy and kappa coefficient.