Background joint sparse representation for hyperspectral image subpixel anomaly detection

A novel sparsity-based sub-pixel anomaly detection framework is proposed for hyperspectral imagery. The proposed approach consists of the following steps. First, a joint sparsity model is utilized to simultaneously represent the surrounding local background pixels and to automatically prune the rough overcomplete dictionary as a reliable, compact base for the following center test pixel representation. An unconstrained linear unmixing approach based on the compact dictionary is then utilized to decompose the abundance of the center test pixel. The unmixing result is finally compared to the former background joint sparse representation step, and the energy disparity is utilized to reflect the anomaly test result. The experimental results confirm that the proposed algorithm outperforms the classical RX-based anomaly detector and the orthogonal subspace projection based detector, and gives a desirable and stable performance.