Support vector data description for fast anomaly detection in hyperspectral imagery based on sample segmentation

Support vector data description (SVDD) is particularly good at anomaly detection in hyperspectral imagery but has extremely high computational complexity which limits the real-time implementation. To reduce the computational cost of support vector extraction, a fast SVDD method was proposed by dividing training samples into subsets. The mathematic models of the support vector computation complexity and the sample segmentation subset amount were established, the optimum choosing method of the subset amount was given. The optimal segmentation method of target and background windows was proposed by segmenting background widow into subsets whose shape and size are the same as that of target window, with which only a half of training sample is updated when target window moves one step to reduce the computation amount of extracting support vector from training samples in whole image. The simulation results on HYMAP data demonstrate that the computation time used by the proposed method is less than 10% of that used by SVDD with different sizes of images.