Band selection method based on target to background separation for anomaly detection

Hyperspectral anomaly detection is to detect abnormal pixels that are different from background pixels when the target spectral characteristics are unknown. Significantly, using full-band hyperspectral data for anomaly detection will produce the so-called Hughes phenomenon and high computational cost. Besides, the redundancy of full-band data will drown out the abnormal target characteristics and reduce the detection performance. Therefore, it is particularly important to select the band in advance. However, the band selection methods currently used for anomaly detection lack relevance to the task, resulting in a large number of selected bands and failing to effectively improve the performance of anomaly detection. To address theses problems, this article proposes a band selection method for anomaly detection based on the separation of target to background. Specifically, we take the RX algorithm as the core, design an indicator function to measure the separation of target to background based on the Mahalanobis distance. Besides, we use a forward update search algorithm to quickly approximate the optimal band combination. The former is associated with anomaly detection tasks and compatible with multiple types of abnormal targets, while the latter fully explores possible band combinations and avoids the huge overhead of the exhaustive method. We have conducted extensive tests on different scene data sets and different anomaly detection algorithms. Experimental results demonstrate that the proposed method achieves superior performance in subsequent anomaly detection compared to its competitors, even with fewer selected bands.