A RFS-SVDD algorithm for hyperspectral global anomaly detection

A RFS-SVDD algorithm was presented in order to decrease false-alarm rate when the SVDD detector is used for global anomaly detection in hyperspectral imagery. The original hyperspectral imagery was divided into some sub-regions by an improved spatial clustering algorithm proposed in this paper. Those sub-regions larger than the target size were defined as background sub-regions, and others were defined as potential anomaly sub-regions. The average spectrum of all the pixels in each background sub-region, which is called Region Feature Spectrum (RFS), was computed and used as training sample for SVDD. Those training samples were used to model the support region of the background spectrum. Because the RFS is the average spectrum of the background sub-region which contains no anomalous pixel, the modeling would not be affected by the anomalous spectrums and random noise in the hyperspectral imagery. The simulated results on the HYMAP and AVIRIS data show that the RFS-SVDD can eliminate the interference induced by the anomalous spectrum and/or random noise, and finally decreases the false-alarm rate when SVDD is used for global anomaly detection in hyperspectral imagery.