An adaptive-kernel collaborative representation method based on spatial-spectral joint clustering for hyperspectral anomaly detection is proposed, which is well used to solve the abnormal interference in space-spectrum information. The algorithm gives full play to the filtering characteristic of Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for outliers, and is applied to space-spectrum processing. On the basis of removing abnormal spectrum by DBSCAN clustering, random projection for reserved subsets is used to reduce the dimension of the data, so spectral noise and spectral redundancy can be solved properly. Considering the influence of background outliers on collaborative representation detection algorithm, DBSCAN clustering is used to remove the clutter points in the local background pixels. Furthermore, the influence of background dispersion on the selection of kernel parameters is studied. By comparing different kernel estimation methods, an adaptive kernel measure method based on average difference is proposed. The proposed algorithm is used to simulate three sets of AVIRIS and ROSIS data and compared with the international mainstream anomaly detection algorithm, the results show that the proposed algorithm has a good detection performance.