A Bayesian MRF framework for labeling terrain using hyperspectral imaging

Studies of hyperspectral images point to non-Gaussian statistics of pixels values, and consequently, standard Gaussian models may not perform well in hyperspectral image analysis. This paper presents novel probability models that capture non-Gaussian statistics of hyperspectral images, and uses them in automated classification of terrain sites. After the data are preprocessed using standard dimension-reduction tools, we use: 1) a nonparametric density estimate for capturing spectral variation at each site and 2) two parametric families-generalized Laplacian and Bessel K form-to capture non-Gaussian statistics of difference pixels. Assuming an Ising-type prior on site labels, favoring a smooth classification, we formulate a Markov random field-maximum a posteriori estimation problem and use a Markov chain to estimate site classifications. Results are presented from application of this framework to Washington, DC Mall and Indian Springs rural area datasets.