To solve the problem that the overall and local characteristics of the curves of spectrum are difficultly described by the Gaussian radial basis function and the spectral similarity measurement kernel simultaneously in a hyperspectral imagery detection, a novel kernel anomaly detection method in hyperspectral imagery based on the spectral discrimination method was proposed. A kernel anomaly detection in hyperspectral was conducted by real hyperspectral images collected by airborne visible infra-red imaging spectrometer. The binary graph and receiver operating characteristic curve of the anomaly detection were attained. The results show that, for a lower false alarm rate in a hyperspectral imagery detection, compared with the Gaussian radial basis function and the spectral similarity measurement kernel, the proposed kernel can detect the abnormal targets with a higher accuracy and clarity.