Hyperspectral Anomaly Detection With Otsu-Based Isolation Forest

Hyperspectral anomaly detection involves in many practical applications. Traditional anomaly detection methods are mainly proposed based on statistical models and geometrical models. This article proposes an Otsu-based isolation forest method, which applies the assumption that anomaly pixels are more sensitive to be isolated from the alternative pixels. The proposed article trains an isolation forest by assembling multiple binary trees. To construct a more discriminative binary tree, Otsu-based splitting criterion is applied to split subsamples into two groups at each division. Then, it feeds each tested pixel into isolation forest and obtains its anomaly score via the average path length throughout isolation forest. Considering the pixels with anomaly attribute values, path length refinement strategy based on distance weight is applied to better distinguish anomaly scores of tested pixels. Experimental results on three datasets reveal that the proposed method can effectively separate anomalies from backgrounds compared with other anomaly detection methods.