A class of detection filters for targets and anomalies in multispectral / hyperspectral imagery

In this paper we present a class of detection filters based on variations of the spectral screening. Spectral screening is a technique used in reducing the multispectral / hyperspectral data to a representative subset of spectra. The subset is formed such that any two spectra in it are dissimilar and, for any spectrum in the original image cube, there is a similar spectrum in the subset. Spectral screening is performed in a sequential manner, at each step, the subset being increased with a spectrum dissimilar from all the spectra already selected. The procedure, using the spectral angle as similarity measure is employed in a variety of algorithms for linear unmixing and data compression. We modified the algorithm such that at the selection step the spectrum with the largest distance from the set is selected. While not introducing additional computational complexity, the Maximum Spectral Screening (MSS) algorithm ensures that the overlap among the representatives is minimized. MSS is further improved by choosing as first spectra the one with the largest distance to the diagonal, thus transforming the screening into a deterministic process. Second, we investigated the alternative approach where at each step the spectra with the smallest distance (but larger than a threshold value) is selected (Minimum Spectral Screening - mSS). The detection filters were obtained as the classification projector matrices based on the spectral subset. The developed algorithms were tested on HYDICE hyperspectral data using the spectral angle and the spectral information divergence. The results indicate that MSS outperforms regular spectral screening detection and mSS.