Modeling and detection in hyperspectral imagery

One aim of using hyperspectral imaging sensors is in discriminating man-made objects from dominant clutter environments. Sensors like Aviris or Hydice simultaneously collect hundreds of contiguous and narrowly spaced spectral band images for the same scene. The challenge lies in processing the corresponding large volume of data that is collected by the sensors. Usual implementations of the maximum-likelihood (ML) detector are precluded because they require the inversion of large data covariance matrices. We apply a Gauss-Markov random field (GMRF) model to derive a computationally efficient ML-detector implementation that avoids inversion of the covariance matrix. The paper details the structure of the GMRF model, presents an estimation algorithm to fit the GMRF to the hyperspectral sensor data, and finally, develops the structure of the ML-detector