Anomaly discrimination and classification for hyperspectral imagery

Anomaly detection finds data samples whose signatures are spectrally distinct from their surrounding data samples. Unfortunately, it generally cannot discriminate its detected anomalies one from another. One common approach is to measure closeness of spectral characteristics among detected anomalies to determine if the detected anomalies are actually targets of different types. However, this also leads to a challenging issue of how to find an appropriate criterion to threshold their spectral similarities. Interestingly, this issue has not received much attention in the past. This paper investigates the issue of anomaly discrimination without using any spectral measure. The idea is to take advantage of an unsupervised detection algorithm, automatic target generation process (ATGP) coupled with an anomaly detector to discriminate detected anomalies. Experimental results show that the proposed methods are indeed very effective in anomaly discrimination.