Several Key Technologies Of Spectral Unmixing In Hyperspectral Imagery

Hyperspectral remote sensing integrates both detained spectral and geometric structure information about ground features, which provide human being unprecedented knowledge. In this thesis, we are going to discussion the spectral unmixing problem. Spectral unmixing is the procedure by which the measured spectrum of a mixed pixel is decomposed into a collection of constituent spectra, or endmembers, and a set of corresponding fractions, or abundances, that indicate the proportion of each endmember present in the pixel. Collecting data in hundreds of spectral bands, hyperspectral sensors have demonstrated the capability of performing spectral unmixing. The main research works and contributions are centered on the two aspects, they are described as following:(1) Fraction abundances estimation by linear mixing model (LMM) and its extension. A key aspect of inversion is the incorporation of the dual physical constraints that abundances must obey, full additivity and nonnegativity. Quadratic programming is adapted to give the solution. To address endmember variability, the LMM is extended by two aspects, one is using weighted linear spectral analysis; another is that endmembers are supposed to be Gaussian vectors, which finally turned into a normal compositional model (NCM).In NCM model the abundances are estimated by a hierarchical Bayesian algorithm. Considering the nonlinear nature of the mixing process, a simple but effective method is adopted, where the multiplication of each pair of endmembers results in virtual endmember representing multiple scattering effects. Then it is followed by linear unmixing for abundance estimation. After the experiments analysis, we get the following conclusions:1)When the participated endmembers for mixed pixel unmixing are the same with the actual endmembers in the mixed pixel or are more larger than those, constrained unmixing results are better than unconstrained unmixing, and FCLS is the best;2) If the variance in each bands is the same, NCM unmixing result is better than FCLS;3)Weighted LMM has enhanced the contrasted between classes;4) Polynomial based nonlinear spectral unmixing can indicate the interaction between neighboring objects;5)The purity of an endmember may degrade the performance of spectral unmixing, because the less purity endmember weaken the independence between endmembers;6)When the participated endmembers is much more than the actual endmembers in the mixed pixel, FCLS became over fitting; on the contrary, when the participated endmembers is less than that, FCLS and SCLS cause large errors.(2) Automatic endmembers extraction. To implement an automatic endmember extraction algorithm, two problems need to resolved. They are the estimation of the number of endmembers and the extraction of endmember signatures. Determination the number of endmember is a very challenging problem, we use a Neyman-Pearson detection theory based eigen threshold method, referred to as HFC method. The difficulty of endmember extraction algorithm is its robustness; and it is reflected that when there are none pure pixels corresponding to the endmember, the algorithm can also approach an "endmember" qualified for spectral unmixing error constraint. So, we design our algorithms according to geometrical endmember theory, which selects endmembers from the vertices of a simplex, a polyhedron or a convex cone that minimally encloses or is maximally contained in the data in a scene. Firstly, we a minimally enclosing convex cone based endmember approaching method is deduced. The method directly used a simplex volume to formulate an objective function with convex hull constraint. But nonconvexity of the objective function makes the solution difficult. So a cyclic minimization algorithm for approximating the problem is developed using linear programs, this is the MVE method. Secondly, we combine the minimization of squared error and simplex criterion to formulate objective function. One type of the function is using a determinant simplex volume to form the regular terms, one type is using a vertices distance of the simplex to form the regular terms. They are NMVE and DVE respectively. The two objective function are solved in an alternative framework, the difference is that NMVE is based on the nonnegative matrix factorization, DVE is based alternative least square. The experiments demonstrate that when not every endmember has its corresponding pure pixels, the traditional geometric method are unable to get satisfied endmember estimation, but new developed methods (MVE, NMVE and DVE) approach the ideal endmembers very well. It is also show that MVE is sensitive to noise, NMVE and DVE are relative more robust.(3) Sparse constraint based spectral unmixing. It is mainly based on a fact and two questions. The fact is that the endmember abundance vector is sparse. Because the mixed pixel is a combination of a limited components, and the available endmembers may be larger than the number of the components even be much larger. The first question is each pixel in a scene may utilize a different subset of endmembers, but the tradition methods rarely do like this. The second question is that when the number of endmembers is larger than the size of bands, it may lead to an under-determined system. The newest theory about sparse regression revealed that the under-determined problem can be solved if the evaluated vector is sparse. The first question can also be achieved by the sparse theory, because the sparse constraint restricts endmembers participating in the mixed pixel. In this thesis we utilized the split augmented lagrangian shrinkage algorithm-SALSA. It is a fast sparse regression method, including two important parameter—regular parameter and punish parameter. In the experiment, we use a sample error analysis method to estimate them. After the experiments analysis, we get the following conclusions:1) The smaller endmember matrix MC value is, the higher accuracy results can be achieved by sparse regression;2)When use the same endmember matrix is used for sparse unmixing, the less components is in a mixed pixel, the higher accuracy can be achieved;3) The larger endmember matrix is, the lower accuracy results can be get by sparse regression;4) SALSA sparse unmixing results are better than FCLS,NNLS,ISMA, and it is more faster than the traditional methods;5) The under-determined spectral unmixing problem is resolved by SALSA method.(4)Mixed spectral model based fluorescence hyperspectral image analysis. The purpose of fluorescence hyperspectral image interpretation is to separate the different fluorescence signal and estimate their relative intensity. But because of the cross talk problem between fluorescence signals and the significant autofluorescence signal, the interpretation becomes difficult. Spectral unmixing is a useful technique. We first used the geometric based endmember approaching method NMVE to extract the endmember information, and at the same time we have utilized the auxiliary information (available autofluorescence endmember and the range of excitation spectrum). The simulated mixed spectral samples constituted by some representative fluorescence are used for testing our method. The result indicated NMVE approaching the fluorescence endmember signal successfully. It was also convinced by the result of a multiple fluorescence signals in vivio hyperspetcral image analysis. Then, we used NMVE as an auto-unmixing tool to location the tumor signal in an in vivio fluorescence hyperspetcral image, and the similar procedure was performed on a tissue fluorescence hyperspetcral image. The results indicated that our method is succeeded in interpretation the fluorescence hyperspectral image.The study in this paper is not only important to hyperspectral image analysis, but also contributes to the fluorescence image analysis.