By analyzing the limitation of the traditional sparse representation based classification, a novel classification framework called two dimensional Local-constrained Coding and Sparse Representation (2D-LSRC) is proposed for Synthetic Aperture Radar (SAR) images recognition. Different from other recent popular vector-based representation, 2D-LSRC preserves the global linear coding coefficients between the input matrix and these elementary matrices, as well as the local data structure. Extensive experimental results of MSTAR datasets show the effectiveness of the proposed algorithms and its robustness for the number of the training dataset.