In this paper, we modify the classic manifold separation technique (MST), aiming to reduce its dependence on high signal-to-noise ratio (SNR) measuring environment. According to the analysis of the array response, it is demonstrated that to maintain a correct phase relationship between the received data at different calibration angles is indispensable for the application of MST. Thus, we slightly change the structure of the traditional calibration system, so that a phase reference for the measurements can be obtained. Besides, unlike the classic MST, where only a single snapshot measurement is utilized for calibration, multi-snapshot information is exploited in the novel method by using the subspace decomposition technique. Simulation results verify the superiorities of the proposed subspace-based calibration method in 1-D and 2-D scenarios.