Improved hyperspectral anomaly detection in heavy-tailed backgrounds

A new metric for anomaly detection in hyperspectral imagery is developed to account for anisotropic heavy tails in covariance-whitened data. The anisotropy, consisting of a variation in tail heaviness with principal component number, commonly occurs when the number of linearly independent components representing the data to within the noise level is less than the number of data dimensions. The detection metric is generated by representing the probability density function of the data with an empirical anisotropic super-Gaussian model for the probability density function. Its performance exceeds that of the RX and Subspace RX methods in examples from CAP ARCHER and HyMap imagery.