Hyperspectral Unmixing Based On Nonnegative Matrix Factorization

Due to the resolution limitation of the sensors and the variability of the ground surface, the observation of one pixel in a hyperspectral remote sensing image may contain several disparate substances, causing it to be a "mixed pixel". In order to utilize the hyperspectral data, these mixed pixels have to be decomposed into a set of constituent spectra, called endmember signatures, and their corresponding proportions, called abundances. Mixed pixels exist in almost all hyperspectral images, thus how to unmix these pixels has become an important problem of hyperspectral imagery, for whose application in the identification and detection of ground targets. In recent years, new ideas and approaches for hyperspectral unmixing are springing up. Using unsupervised method to decompose mixed-pixels more efficiently and accurately has become a hotspot in the research area of remote sensing. As an advanced statistical approach, Nonnegative Matrix Factorization (NMF) has been frequently researched recently for hyperspectral unmixing, because it can ensure the nonnegativity of results. However, it has some disadvantages, such as large amount of local minima and high computational complexity. Focusing on the application of nonnegative matrix factorization in hyperspectral unmixing, this thesis has made a lot of research, and the main works and innovations are as follows:1. A constrained NMF approach has been proposed. Because of the local minima in the objective function, the traditional NMF algorithm is sensitive to the initial value when being applied to hyperspectral unmixing. In order to solve the problem, a new approach based on constrained NMF is proposed for decomposition of mixed pixels by introducing constraints of abundance separation and smoothness into the objective function of NMF. The algorithm can also satisfy the abundance nonnegative and sum-to-one constraints, which are necessary for hyperspectral unmixing. Experimental results on simulated and real hyperspectral data demonstrate that the proposed approach can overcome the shortcoming of local minima, and obtain better results. Meanwhile, the algorithm performs well for noisy data, and can also be used for the unmixing of hyperspectral data in which pure pixels do not exist.2. On account of the low computational speed of NMF, a multi-core parallel processing method is proposed to increase its efficiency. The method decomposes the constrained NMF algorithm to several blocks, and choose appropriate parallel strategy according to the features of hyperspectral imagery. Experimental results show that, the multi-core CPU is utilized sufficiently after the execution of parallel processing, and the computational speed of unmixing has got an obvious improvement.3. The correlation between neighboring bands of hyperspectral imagery is very high, while NMF is a statistical approach, thus a large amount of redundant information will be processed if the original data is used directly. In order to solve this problem, a maximum-information-based band selection approach for NMF is proposed. It aims at preserving the maximum amount of information, and removes bands one by one iteratively. Experimental results show that, the proposed approach can obviously reduce the number of dimensions of hyperspectral imagery and increase the processing efficiency, and obtain accurate results at the same time.4. Some general free datasets are usually used for the experimental research of hyperspectral unmixing. These datasets are mostly captured over a decade ago by U.S.-made devices. In this thesis, datasets lately obtained by China-made devices are also used for experiments, in order to evaluate the performance of algorithms for practical application.