Researches On Hyperspectral Image Unmixing Based On Nonnegative Matrix Factorization

Compared with traditional color images, hyperspectral images have dozones, perhaps hundreds, of spectral bands which provides rich spectral information of the observed scenes. However, due to the low spatial resolution of hypecspectral imager, in the real hyperspectral images, one pixel may contain some kinds of materials, that is to say make up the "mixed pixels". Hyperspectral unmixing aims decomsed the "mixed pixels" into spectral vectors (endmember) of various materials and their corresponding proportions (abundance). This technique is able to obtain subpixels information of hyperspectral images, which benefits the subsequent task of hypersptral image processing and recogniton.Nonnegative Matrix Factorization (NMF) aims decompsed a nonnegative matrix into the multiply of two nonnegative matrix. The decomposition model of NMF has been similarly with hyperspcetral unmixing, so NMF can be used to solve the problem of hyperspectral unmixing. We were consider the geology and physical properties of hyperspetral image ground covers in this paper, added effective regular constraint term in the orignal NMF model, and then the solving methods were optimized. The main works of this paper are as follows:1. Random initialization for endmember and abundance matrices may suffers from low speed of iteration, and making results trapped in local minimum. In order to solve those problems, we were used VCA-FCLS algorithm for initialize endmember and abundance matrices in this paper. VCA algorithm can successful extracted endmembers when pure pixels existed, but using this algorithm as our initialition methods can speed up convergence and make results avoid trapped in local minimum.2. The algorithm of hyperspectral unmixing base on sparse and minimum volume constrained nonnegative matrix factorization (SMVCNMF). Traditional NMF methods highly rely on the iterative initial value and suffer from poor robustness. The proposed method takes advantage of the sparseness of abundance maps and the volume properties of simplex formed by endmembers. The sparse constraints determined by abundance and minimum volume contraint determided by endmenbers are combined to regularize objective function of NMF. The experiment result has showed the proposed method can be avoided trap in local minimum value, and the method has good convergence and robustness.3. The algorithm of hyperspectral unmixing base on advanced spectral spatial information contrained Nonnegative Matrix Factorization (ASSNMF). In view of the problem that regular constrained non-negative matrix factorization algorithm do not use spectral information effectively. We fully exploit the information of the spectrum, and add the spectral spatial information to the non-negative matrix factorization in this paper. In addition, the sparse constraint term in the objective function were improved. The experimental results show that the proposed method can effectively improve the accuracy of hyperspectral unmixing, and also speed up the convergence of the algorithm.