Hyperspectral Band Selection for Anomaly Detection: The Role of Data Gaussianity

Abstract

We evaluate the anomaly detection performance of a hyperspectral detection algorithm based on selecting specific bands of the hyperspectral cube. The best bands to be chosen are, as expected, based on the degree that the resulting target and background signatures differ from one another, which is determined by the target energy in the whitened space. In addition, we show that the closer the background distribution resembles a Gaussian model, the better the band performs in the detection algorithm. A comparison is made between choosing bands on the basis of their similarity to Gaussian distribution and one based on minimum variance. Examples are provided both on theoretical simulations and on experimental data.