A Hybrid Statistics and Representation-Based Anomaly Detector for Hyperspectral Images

摘要

Anomaly detection (AD) is an important research topic in the hyperspectral remote sensing field. However, owing to the complex background distributions and the interference of clutter noise in practical situations, the AD problem is far from being addressed satisfactorily. In this paper, a novel AD method by joining the statistical model and the representation theory for a predominant detection is proposed. It mainly consists of two parts. A Mahalanobis distance-based anomaly characterization criterion is first designed to acquire the initial detection result. In order to model the background accurately and efficiently, a fast matrix decomposition algorithm is utilized to eliminate the anomalyoise information from the raw data. Moreover, both the decomposed components are taken into account in the measuring formulation for enhancing the distinguishability between anomalies and the background. Second, a local representation process is performed on some selected pixels according to an improved image segmentation method. By using an improved outlier determination criterion, some latent false alarms in these pixels can be located, and consequently their response values are suppressed effectively through an adaptive weight function. Experimental results on one synthetic and two real hyperspectral datasets validate the outstanding performance of our proposed method compared with some state-of-the-art anomaly detectors.