An endmember dissimilarity based Non-negative Matrix Factorization method for hyperspectral unmixing

Non-negative Matrix Factorization has recently been proposed for application in the field of hyperspectral imagery. And for hyperspectral unmixing, high mixing degree and signature variability always affect the unmixing accuracy. To solve this, this paper proposed a novelmethod based on NMF to unmix hyperspectral data. Using the low similarity between endmember signatures in hyperspectral image, we proposes a constraint named endmember dissimilarity constraint which employs the spectral information divergence between signatures in the basic NMF to search a set of vectors with least similarity. This is consistent with the endmember property of the hyperspectral image. The minimum volume constraint NMF and the Piecewise Smoothness NMF with Sparseness Constraint are used to evaluate the proposed method in different mixing degree and signature variability. The experimental results in synthetic data shows that the proposed method performs best in higher mixing degree and signature variability than the other two approaches and the real AVIRIS with highly mixing degree data results also demonstrate that the proposed method performs well in identifying highly mixed endmembers.