In order to solve problem of the limited training samples and keep consistency in one region, a new two-level classification scheme is proposed, which combines sparse auto-encoder (SAE) and Boundary-preserved Wishart-markov random fields (BWMRF). In the first layer, an SAE classifier is applied to obtain an initial classification and more accurate regional boundaries. In the second layer, Boundary-preserved Wishart-markov random fields have been used to correct the previous classification results. Meanwhile, the boundaries classified by sparse auto-encoder are preserved, and a new error correction strategy is applied to ensure the classification accuracy. Therefore, accurate region boundaries supplied by SAE are explored to divide different regions, and the coherent in each region will be realized during the BWMRF process. Compared with other classification methods, this method obtains higher classification accuracy and proves the validity of the new scheme.