Local covariance equalization of hyperspectral imagery: advantages and limitations for target detection

The operational implementation of many conventional hyperspectral detection algorithms can be greatly simplified by adaptively preprocessing all test pixels and target signatures to equalize first- and second-order background statistics. The process is equivalent to expressing spectral radiance in a locally Euclidean coordinate system. Removing hyperspectral curvature in this way greatly simplifies both the data archiving function and the mathematical forms of standard detectors. Here we show why the equalization procedure does not compromise performance for conventional detection methods. More advanced algorithms cannot, however, be implemented with equalized data alone. We show how this limitation can nonetheless be overcome by temporarily storing a few parameters, with no archiving penalty