HYPERSPECTRAL ANOMALY DETECTION BASED ON ISOLATION FOREST WITH BAND CLUSTERING

In this paper, we propose a hyperspectral anomaly detection approach based iForest (Isolation Forest) with band clustering. Instead of random selecting a feature from all bands in iForest algorithm, the proposed approach designed the following three main steps. Firstly, all bands were divided into several groups by use of the correlation among bands and optimal clustering. Then, one of the groups was randomly selected as a candidate. Finally, a band was randomly selected from the candidate group as an attribute for tree node splitting. Compared with other anomaly detection methods, our approach for attributes selection can not only handle high dimensional problem, but also reduce the probability that important information was ignored. The experiments demonstrate its robustness and competitive performance.