Subspace based non-parametric approach for hyperspectral anomaly detection in complex scenarios

Recent studies on global anomaly detection (AD) in hyperspectral images have focused on non-parametric approaches that seem particularly suitable to detect anomalies in complex backgrounds without the need of assuming any specific model for the background distribution. Among these, AD algorithms based on the kernel density estimator (KDE) benefit from the flexibility provided by KDE, which attempts to estimate the background probability density function (PDF) regardless of its specific form. The high computational burden associated with KDE requires KDE-based AD algorithms be preceded by a suitable dimensionality reduction (DR) procedure aimed at identifying the subspace where most of the useful signal lies. In most cases, this may lead to a degradation of the detection performance due to the leakage of some anomalous target components to the subspace orthogonal to the one identified by the DR procedure. This work presents a novel subspace-based AD strategy that combines the use of KDE with a simple parametric detector performed on the orthogonal complement of the signal subspace, in order to benefit of the non-parametric nature of KDE and, at the same time, avoid the performance loss that may occur due to the DR procedure. Experimental results indicate that the proposed AD strategy is promising and deserves further investigation.