Statistical modeling of hyperspectral data using Gauss-Markov random fields and its application to classification

Hyperspectral imagery provides high spectral and spatial resolution that can be used to discriminate between object and natural clutter in environmental monitoring applications such as coastal environment and coral reef monitoring. High dimensionality of the data set makes it difficult to apply statistical models to the full image. This thesis presents a 3D noncausal Gauss-Markov Random Field (GMRF) model that, under the assumption of near spatial stationarity and using the nearest spectral and spatial neighbors, results in a statistical model that is parameterized only by four parameters. We study the model parameter estimation using least squares (LS) and approximate maximum likelihood (AML) techniques. We study the parameter estimation and model validation using synthetic and measured data. We also apply the model to classification. Our results show that including the spatial information in hyperspectral data improves the classification performance. The performance of the classification is studied using synthetic and hyperspectral data. The MRF-based classifier is better than spectral-only classifiers.