A new technique for hyperspectral image analysis with applications to anomaly detection

The paper describes a new approach to hyperspectral image analysis using spectral signature mixture models. In this new approach spectral End-member extraction and spectral unmixing are co-dependent objectives. Previous methods tended to serialize these tasks. Our approach shows that superior hyperspectral modeling can be obtained through a parallel objective approach. The new approach also implements natural constraints on the end-members and mixtures. These constraints allow us to adopt a physical interpretation of the hyperspectral image decomposition. This new modeling technique is useful for the detection of Snown signatures and, more significantly, for the detection of unknown, partially occluded scene anomalies. The anomaly detection algorithm is aided by the newly developed Quad-AR filter which acts as an efficient optimal adaptive clutter rejection filter. Examples are given using a 3-band color image and 210-band HYDICE forest radiance data. The results show these new techniques to be quite effective.