Robust detection using M-estimators for hyperspectral imaging

Hyperspectral data have been proved not to be multivariate normal but long tailed distributed. In order to take into account these features, the family of elliptical contoured distributions is proposed to describe noise statistical behavior. Although non-Gaussian models are assumed for background modeling and detectors design, the parameters estimation is still performed using classical Gaussian based estimators; as for the covariance matrix, generally determined according to the SCM approach. We discuss here the class of M-estimators as a robust alternative for background statistical characterization and highlight their outcome when used in an adaptive GLRT-LQ detector.