Description of component model for automated generation of scene statistics and comparison of algorithm performance applied to both natural and hypothetical spectral scenes

There is a need to assess hyperspectral image processing algorithms in a way that does not require applying the algorithm to a large set of spectral scenes. The statistical nature of hyperspectral scenes can be modeled as a set of means and covariances. In this model, each mean-covariance pair describes some physical component of the scene (for example grass, trees, soil, road, etc.). Modeling the scene in this fashion allows non-gaussian nature of scene to be explored, with the assumption that the scene statistics are linear sums of gaussians. Once this component model of a scene is constructed, filter performance can be assessed quickly by applying the filter to the ensemble of means and covariances. Furthermore, filter performance can be predicted for scenes not yet collected, as scene models may be artificially generated from statistics of physical components. As a validation of the model we generate plots of target probability of detection versus probability of false alarm for natural scenes and models based on those scenes.