Existing localization methods have mismatch problem when applied to the real uncertain ocean, and this will lead to performance degradation. In normal mode models, some modal eigenfunctions remain to be more correlated than others in the presence of environmental uncertainties. Based on this, we have proposed a mode subspace reconstruction robust localization method, which uses stable modes to reconstruct the replica vector to grantee the localization performance. The data from simulation and experiment are used to verify the effectiveness of the proposed method. Performances of the matched field processor (MFP) and the robust ML (maximum localization) estimator are also given here for comparison. Results show that: (1) the generally used MFP method has a low localization performance even at high SNR values; (2) the proposed method outperforms the robust ML estimator and the generally used MFP.