Spectral/spatial modeling of hyperspectral imagery with spectral mixture and multigrid Gibbs-based partition processes

This research develops a new integrated physical-based stochastic model, processing algorithm, and computer code to exploit hyperspectral imagery for the remote identification of material compositions in a scene. The practical goal is to support mapping and targeting applications, with an emphasis on mapping. A more theoretically rigorous goal is to develop a mathematically precise model for underlying spectral and spatial processes to effectively exploit the imagery in an optimal way.; An enhanced method of spectral un-mixing is investigated for hyperspectral imagery of moderate-to-high scene complexity in the context of terrain mapping, where either a large set of fundamental materials may exist throughout, or where some of the fundamental members have spectra that are similar to each other. For a complex scene, the use of one large set of fundamental materials as the set of “endmembers” for performing spectral unmixing can cause unreliable estimates of material compositions at sites within the scene. In such cases, partitioning this large set of endmembers into a number of smaller sets is appropriate, where the smaller sets are associated with certain regions in a scene. A spectral/spatial Gibbs-based approach to partitioning hyperspectral imagery into homogeneous regions is investigated, such that spatial consistency is imposed on the spectral content of sites in each partition. An algorithm is developed that provides an estimator of an underlying and unobserved process called a “partition process” that coexists with other underlying (and unobserved) processes, one of which is called a “spectral mixing process.” The partition algorithm exploits the properties of a Markov Random Field (MRF). A form of Bayesian estimation, Maximum A Posteriori (MAP) estimation, is computed through the localized sampling of a Gibbs distribution defined over a neighborhood system and which is implemented as a multigrid process. Appropriate energy functions and neighborhood graph structures are investigated, which model spectral disparities in an image using spectral angle and/or Euclidean distance, as well as spatial disparities based on the Kolmogorov Smirnov statistic.; Three experiments are performed to investigate the ideas and validate the approach. Suitable parameter ranges are investigated, and the behavior of the partition algorithm is characterized using the individual and combined measures of spectral/spatial disparity. The experiments also validate the combined algorithm, where the spectral mixture process is conditioned on the partition process. Two of the experiments are conducted on scenes from HYDICE 210-band imagery collected over an area adjacent to Fort Hood, Texas, and that is supported with ground truth. The experience gained in these two experiments is then applied in a third experiment, where the values of the parameters developed for the algorithm are applied without change to a new scene. This third experiment is conducted on a GERIS 22-band scene collected over an area of Northern Virginia where the sensor provides far less spectral information and less spatial resolution than the sensor used in Experiments 1 and 2, but where the scene is larger in size and geographic coverage. In aggregate, the scenes help develop and validate the algorithm with respect to a diverse range of terrain features, different spectral resolutions and range, and different spatial resolutions.