The current anomaly detection algorithms are shortage to solving anomaly detection problem because hyperspectral imagery has a higher order features and complicated background information distribution characteristics. By analyzing the spectral features and the spatial features, and exploiting spectral unmixing and subspace divided methods, based on statistical learning theory, an algorithm of anomaly detection of hyperspectral imagery selectivity band subsets is proposed based on spectral unmixing(UNBS-KRX). At first, hyperspectral imagery reduce the background interference and prominent anomaly target information by using spectral unmixing methods, which extract endmembers spectral, that is great influence on background information distribution of hyperspectral imagery. Then, the algorithm divides the whole bands space to a few subspaces. The size of the subspace is different, and non-Gaussian measurement criterion is used to extract the characteristic bands in each subspace. The bands are rich in anomaly target information. At last, as an anomaly detection operator, the kernel RX completes anomaly target detection. The real hyperspectral data sets are used in the experiments, and the result shows the UNBS-KRX is effective and reasonable, and has an execllent detection performance.