Hyperspectral Unmixing Theory And Techniques Based On Nonnegative Matrix Factorization

With the advance of remote sensing images in these years, hyperspectral imagery is being applied widely in military and civilian fields. However, due to the limit of spatial resolution, mixed pixels are widespread in hyperspectral imagery. The mixed pixels problem not only influences the precision of object recognition and classification, but also becomes an obstacle to quantification analysis of remote sensing images. Hence, how to effectively interpret the mixed pixels is a critical problem for hyperspectral remote sensing applications.Non-negative matrix factorization (NMF), a new method proposed by Daniel. D. Lee et al, decomposes a positive matrix into a product of two positive matrices and has been used for hyperspectral data unmixing. As will be described later, the NMF finds a set of nonnegative basis vectors that approximates the original data through liner combinations. These basis vectors thus play a similar role as the endmembers. However, the direct application of the standard NMF algorithm to the decomposition of mixed pixels will result in the problem of local minimum and slow convergence. In this paper, by systemly analyzing former studied, the nonnegativity and continuity of both endmember spectral and abundances, and also the sparseness constraint of only abundances, were combined with the NMF algorithm to decompose mixed pixels in hyperspectral remote sensing images.The major works and contribution of this dissertation are as follows:1) In the decomposition of mixed pixels of hyperspectral remote sensing images, the vertex component analysis (VCA) needs that there is at least one pure pixel for every endmember existing in the images, and nonnegative matrix factorization (NMF) easily result in the problem of local minimum, owing to the influence of algorithm initializations. To solve the problems, this paper presents a new scheme based on constrained NMF(MCNMF). We adopt the endmembers extracted by VCA, and abundance obtained by least square as the initial values of NMF iteration, then use the smooth constrained NMF(CNMF) iteration method to achieve the decomposition of mixed pixels. Experimental results obtained from both artificial simulated and real-world hyperspectral remote sensing data demonstrate that the proposed scheme for decomposition of mixed pixels is better than VCA and CNMF .2) The dilemma of local minimum is unlikely to be completely exterminated with the MCNMF algorithm. However, genetic algorithm (GA) is a global search method that can avoid local optima by parallel probing the search space. We presents a new scheme based on MCNMF algorithm and GA to achieve the decomposition of mixed pixels. The endmembers obtained by MCNMF is adopted as the initial individual population values of GA, the optimal solution of GA is in reverse as the new initial endmembers in the next running of MCNMF, repeat this procedure until the global optimal solution is achieved. Experiment results based on simulated data and real hyperspectral imagery demonstrate that the proposed scheme outperforms MCNMF.3) In this paper, we develop a complexity and minimum volume constrained NMF (CMVC-NMF) that incorporates the complexity and volume constrains into NMF. Three important facts are exploited: first, the spectral data are nonnegative; second, the simplex volume determined by the endmembers is the minimum among all possible simplexes that circumscribe the data scatter space; third, the variation of the material spectra and abundance images is smooth in time and space respectively. The experimental results validate the efficiency of the approach.