Hyperspectral Unmixing Based on Constrained Nonnegative Matrix Factorization

To obtain an improved optimal solution, a nonnegative matrix factorization method based on abundance and endmember constraints for hyperspectral unmixing is proposed. First, considering the sparseness of the abundance matrix, a weighted sparse regularization is introduced to the Nonnegative Matrix Factorization (NMF) model to ensure the sparseness of the abundance matrix. The weights are updated adaptively according to the abundance matrix. Second, given the priori knowledge of the distribution of adjacent pixels, a total variation regularization is further added to the NMF model to promote the smoothness of the abundance map. Finally , a new constraint given by a potential function from the Markov random field model is introduced to improve the spectral smoothness of the endmembers. Experiments are conducted to evaluate the effectiveness of the proposed method based on three different data sets, including a synthetic data set and two real-life data sets (Jasper Ridge and Cuprite) respectively. From the experimental results, it is found that the proposed method got better performances both on the spectral similarity and the estimation accuracy for abundance.