The sparse unmixing model is improved based on the spatial constraint which represents the similarity and difference between the adjacent pixels.Thus, the accuracy of the hyperspectral image unmixing is increased.The unmixing spectrum library is generated by compressing the primary spectrum library to increase the influence of the spatial constraint on the unmixing model, and reduce the sparsity of hyperspectral image in the unmixing spectrum library.According to unmixing spectrum library, the improved unmixing model is constructed by sparse unmixing and manifold regularization, which represent the similarity and difference between the adjacent pixels.The improved unmixing model is solved by means of the convex optimization algorithm such as alternating directions method of multipliers.Experimental results show that the proposed algorithm has high spectral unmixing accuracy and strong performance.