A novel algorithm for endmember extraction is presented. The approach follows the linear mixture model for hyperspectral data. Endmembers are identified based on sparsity consideration. Theoretical and experimental results suggest the potential of the method.