Once clutter training samples mix with interfering targets, the clutter suppression performance by space-time adaptive processing will decline. In order to solve the problem, a robust training samples detection method based on target knowledge and partly sparse recovery was proposed. Firstly, the object region in unit to be detected was locked. Then the sparse complete base was obtained by covering the whole angleDoppler plane. After that, the corresponding object region in sparse complete base by transformation matrix was hollowed out to obtain the super resolution clutter space-time spectrum, which helps to estimate the clutter covariance matrix. Finally, the method was combined with the generalized inner product method to realize non-homogeneous training sample detection. Compared with GIP (generalized inner product) method, the proposed method can detect interfering targets in different intensity. Simulation analysis demonstrates that the test statistics of the proposed method have excellent discrimination validity, and can drastically eliminate interfering targets, thus improving the target detection performance of STAP (space-time adaptive processing).