Hierarchical Sub-Pixel Anomaly Detection Framework for Hyperspectral Imagery.

Anomaly detection is an important task in hyperspectral processing. Some previous works, based on statistical information, focus on Reed-Xiaoli (RX), as it is one of the most classical and commonly used methods. However, its performance tends to be affected when anomaly target size is smaller than spatial resolution. Those sub-pixel anomaly target spectra are usually much similar with background spectra, and may results in false alarm for traditional RX method. To address this issue, this paper proposes a hierarchical RX (H-RX) anomaly detection framework to enhance the performance. The proposed H-RX method consists of several different layers of original RX anomaly detector. In each layer, the RX's output of each pixel is restrained by a nonlinear function and then imposed as a coefficient on its spectrum for the next iteration. Furthermore, we design a spatial regularization layer to enhance the sub-pixel anomaly detection performance. To better illustrate the hierarchical framework, we provide a theoretical explanation of the hierarchical background spectra restraint and regularization process. Extensive experiments on three hyperspectral images illustrate that the proposed anomaly detection algorithm outperforms the original RX algorithm and some other classical methods.