Realistic matched filter performance prediction for hyperspectral target detection

In hyperspectral imaging applications, the detection performance of the matched filter is typically evaluated or predicted based on Gaussian (or normal) probability density models. However, it is well known that hyperspectral imaging data exhibit nonGaussian behavior. In this paper, we propose modeling the distribution of hyperspectral backgrounds using a more accurate model based on the elliptically contoured multivariate t-distributions. Then, using a Gaussian distribution model for the target class and an elliptical t-distribution for the background class, we show that detection performance predictions based on solely Gaussian models can be inaccurate and overly optimistic. Therefore, conclusions based on such models should be used with extreme care.