As one of the hot issues of remote sensing imaging, the traditional method of remote sensing image compression has problems that widespread a long reconstruction time, and the quality of the reconstructed image needs to be improved. According to the remote sensing images of different typical surface feature, the K-SVD dictionary learning method is utilized in the paper. In the process of reconstruction, through multiple iterations on the part of the image blocks, the original image can be solved by a linear representation of the atoms from the corresponding surface feature of an overcomplete dictionary. Then the atoms are given preferentially as the initial value to calculate the residual of the image blocks in the neighborhood, to reduce the number of iterations. The remote sensing image information content on typical surface and the similarity between image blocks are fully exploited. Compared with the general dictionary structured by non-redundant orthogonal base or non-classified learning dictionary, the proposed method outperforms in the reconstructed image quality and reconstruction speed.