A spectral anomaly detector in hyperspectral images based on a non-Gaussian mixture model

Anomaly Detection (AD) in remotely sensed airborne hyperspectral images has been proven valuable in many applications. Within the AD approach that defines the spectral anomalies with respect to a statistical model for the background, reliable background PDF estimation is essential to a successful outcome. This paper proposes a new Bayesian strategy for learning a non-Gaussian mixture model for the background PDF based on elliptically contoured distributions. The resulting estimated background PDF is then used to detect spectral anomalies, characterized by a low probability of occurrence with respect to the global background, through the Generalized Likelihood Ratio Test (GLRT). Real hyperspectral imagery is used for experimental evaluation of the proposed strategy.