Without considering background interferences, the classical algorithm of support vector data description (SVDD) has relatively low detection probability in hyperspectral anomaly detection. To solve the problem, this paper presented a new algorithm based SVDD, which includes hyperspectral unmixing to separate
target information from complicated background clutter. After spectral unmixing, the error datum includes abundant target information while effectively suppresses the background interference. The error datum is then mapped into a high-dimensional feature space with nonlinear SVDD. By exploiting nonlinear information between the spectral bands of hyperspectral imagery, anomaly targets are detected. The results show that the proposed algorithm can improve detection performance and decrease false alarm probability.