IDENTIFY ANOMALY COMPONENT BY SPARSITY AND LOW RANK

Traditional hyperspectral anomaly detection methods either model the global background or the local neighborhood, that bring some apparent drawbacks, such as the unreasonable assumption of uni-modular background in global detectors, or the high false alarms by sliding windows in local detectors. In this paper, a source component-based anomaly detection approach is proposed. It first extracts the source components in the spectral image data cube by using the blind source component separation and then identifies the components that are anomaly (or salient) to other components. We interpret the anomaly detection as a matrix decomposition problem with the minimum volume constraint for the multi-modular background and sparsity constraint for the anomaly image pixels. Experimental results show that the approach is promising for anomaly detection in spectral data cube.