A feature-clustering-based subspace ensemble method for anomaly detection in hyperspectral imagety

Anomaly detection is one of the most important applications for hyperspectral images. In this paper, a new ensemble learning algorithm for anomaly detection in hyperspectral imagery is proposed, which integrates feature grouping and anomalous signal subspace estimation. Main contribution of the proposed algorithm consists in two aspects. First, feature grouping in original hyperspectral images are firstly performed to form feature subsets with more diversity. In the subsets, conventional RX detector can better learn its model parameters. Second, an iterative orthogonal projection processing is given to estimate rare signal subspace for anomalous targets in each feature subset so as to more effectively remove background clutters. Finally, the RX detection is carried out with the estimated signal subspace in the subsets, and the detection results are combined by majority voting. Numerical experiments are conducted on real hyperspectral images and the experimental results show that the proposed algorithm outperforms several existing algorithms.