Nonnegative matrix factorization (NMF) algorithm has powerful information processing and problem solving capabilities, and the nonnegative constraints are required to solve many practical problems. NMF algorithm is widely used in the engineering field [
5]. The objective function of NMF algorithm has obvious non-convexity, possessing local minimum values. To make the NMF algorithm to be applicable in different fields, the corresponding constraint conditions are required to be more complex according to the physical characters of different applications. At present, the conditions of smoothness constraint (SC), minimum volume constraint (MVC) and sparsity constraint, especially the sparsity constraint that based on clustering regularization, have already been proposed [
6–
9]. In this paper, in order to improve the unmixing precision, combining with NMF, an unmixing algorithm, which may be called endmember constraint nonnegative matrix factorization (EC-NMF), was presented. In this algorithm, the correlations and differences between endmembers were used as constraint conditions. The validity of algorithm was verified by the simulated hyperspectral images and real hyperspectral images.