Non-negative Matrix Factorization (NMF) consists in factorizing a nonnegative data matrix by the product of two-rank nonnegative matrixes. It has been successfully applied as data analysis technique in hyperspectral unmixing. However, the direct application of the standard NMF algorithm to the decomposition of mixed pixels will result to the problem of local minimum and slow convergence. The basic theory of NMF algorithm is introduced firstly. Then, the endmember matrix was initialized by the automatic morphological endmember extraction method, so the endmembers selected would be close to the real endmembers. The NMF algorithm is extended by incorporating the nonnegativity and sparseness constraining to unmix hyperspectral data and make sure the error is as small as possible. The measurement of sparseness is implemented by non-smooth NMF and NMF with sparseness constraints algorithms respectively. The optimization results is got by continuous iterative. The repeated iterative calculation is included in one iterative more than once. The experimental result proves the efficiency of the approach.