Anomaly detection in hyperspectral imagery using stable distribution.

摘要

In hyperspectral imaging applications, the background generally exhibits a clearly non-Gaussian impulsive behavior, where valuable information stays in the tail. In this thesis, we propose a new technique, where the background is modeled using the stable distribution for robust detection of outliers. The outliers of the distribution can be considered as potential anomalies or regions of interests (ROIs). In this thesis, we effectively utilized the stable model for detecting targets in impulsive hyperspectral data. To decrease the false alarm rate, it is necessary to compare the ROI with the known reference using a suitable technique, such as the Euclidian distance. This representation compensates a drawback of the Gaussian model, which is not well suited for describing signals with impulsive behavior. In addition, thresholding is considered to avoid misclassification of targets. Test results using real life hyperspectral image datasets are presented to verify the effectiveness of the proposed technique.