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

摘要

Abstract: There is a need to assess hyperspectral imageprocessing algorithms in a way that does not requireapplying the algorithm to a large set of spectralscenes. The statistical nature of hyperspectral scenescan be modeled as a set of means and covariances. Inthis model, each mean-covariance pair describes somephysical component of the scene. Modeling the scene inthis fashion allows non-gaussian nature of scene to beexplored, with the assumption that the scene statisticsare linear sums of gaussians. Once this component modelof a scene is constructed, filter performance can beassessed quickly by applying the filter to the ensembleof means of covariances. Furthermore, filterperformance can be predicted for scenes not yetcollected, as scene models may be artificiallygenerated from statistics of physical components. As avalidation of the model we generate plots of targetprobability of detection versus probability of falsealarm for natural scenes and models based on thosescenes. !6