Sparsity-based approaches have been considered useful for target detection in hyperspectral imagery. Based on the sparse reconstruction theory, the vectors representing the spectral signature of hyperspectral pixels can be a linear combination of linearly dependent training vectors. The training vectors constitute an overcomplete dictionary, which allow for sparse representations for test pixel vectors as only a few of training vectors are used. Such sparsity can be applied in hyperspectral target detection. However, since the sparse decomposition has the potential instability, similar data often have different estimates. In this letter, we propose a nonlocal similarity regularized sparsity model to deal with the problem. Nonlocal similarity enhances classical sparsity model as it preserves the manifold structure of original data and makes more stable estimations for similar data. In addition, the nonlocal sparsity model is effectively solved with a developed greedy algorithm. Experimental results suggest an advantage of the nonlocal sparsity model over conventional sparsity models and a better performance of the proposed algorithm compared with conventional sparsity-based algorithms.