Bayesian Gaussian mixture model for spatial-spectral classification of hyperspectral images

We propose a Bayesian Gaussian mixture model for hyper-spectral image classification. The model provides a robust estimation framework for small size training samples. Defining prior distributions for the mean vector and the covariance matrix, we are able to regularize the parameter estimation problem. Especially, we can obtain invertible positive definite covariance matrices. The mixture model also takes into account the spatial alignments of the pixels by using non-stationary mixture proportions. Based on the classification results obtained on Indian Pine data set, the proposed method yields better classification performance especially for small size training samples compared to state-of-the-art linear and quadratic classifiers.