A Bayesian approach to identification of gaseous effluents in passive LWIR imagery

Typically a regression approach is applied in order to identify the gaseous constituents present in a hyperspectral image, and the task of species identification amounts to choosing the best regression model. Common model selection approaches (stepwise and criterion based methods) have well known multiple comparisons problems, and they do not allow the user to control the experiment-wise error rate, or allow the user to include scene-specific knowledge in the inference process. A Bayesian model selection technique called Gibbs Variable Selection (GVS) that better handles these issues is presented and implemented via Markov chain monte carlo (MCMC). GVS can be used to simultaneously conduct inference on the optical path depth and probability of inclusion in a pixel for a each species in a library. This method flexibly accommodates an analyst’s prior knowledge of the species present in a scene, as well as mixtures of species of any arbitrary complexity. A modified version of GVS with fast convergence properties that is tailored to unsupervised use in hyperspectral image analysis will be presented. Additionally a series of automated diagnostic measures have been developed to monitor convergence of the MCMC with minimal operator intervention. Finally, the applicability of aggregating inference from adjacent pixels will be discussed. This method is compared against stepwise regression for model selection and results from LWIR data from the Airborne Hyperspectral Imager (AHI) are presented. Finally, the applicability of this method to operational scenarios and various sensors will be discussed.