Clutter modeling for subsurface detection in hyperspectral imagery using Markov random fields

Hyperspectral imagery provides high spectral and spatial resolution that can be used to discriminate between object and clutter occurring in subsurface remote sensing for applications such as environmental monitoring and biomedical imaging. We look at using a noncausal auto-regressive Gauss-Markov Random Field (GMRF) model to model clutter produced by a scattering media for subsurface estimation, classification, and detection problems. The GMRF model has the advantage that the clutter covariance only depends on 4 parameters regardless of the number of bands used. We review the model and parameter estimation methods using least squares and approximate maximum likelihood. Experimental and simulation model identification results are presented. Experimental data is generated by using a subsurface testbed where an object is placed in the bottom of a fish tank filled with water mixed with TiO₂ to simulate a a mild to high scattering environment. We show that, for the experimental data, least square estimates produce good models for the clutter. When used in a subsurface classification problem, the GMRF model results in better broad classification with loss of some spatial structure details when compared to spectral only classification.