Owing to the variability of mine surfaces, it is difficult to obtain the deformation monitoring data of the observation stations by traditional leveling technique. GPS RTK （Real-Time Kinematic） technique was employed to the subsidence observation in this paper, and its main sources of errors including rover pole deflection of the vertical, un-modeled systematic errors （gross error, multipath delay etc.） and the height transformation error, were analyzed systematically. Based on the fundamental theories of spherical fit- ting and Empirical Mode Decomposition （EMD）, the error reduction models were studied exhaustively. And two experiments were done in different environment to test the proposed models. The results show that the proposed methods can achieve a fourth-grade leveling accuracy, with （Root-Mean-Square） RMS in three orthogonal directions （N, E and H） of 4.1, 3.3 and 3.1 ram, respectively, by 3-5 rain continuous shaking of the observation GPS antenna, fully satisfying for mine surface subsidence deformation monitoring.
