Hyperspectral Unmixing Based On Sparse Constraint

With the development of remote sensing technology and imaging spectrometer,hyperspectral imagery has been applied in more and more areas. However, since the spatial resolution of the senor is not enough and the distributions of the materials are very complicated, some materials can jointly occupy a single pixel. This would seriously hinder the practical application of hyperspectral imagery. In recent years, new ideas and approaches for hyperspectral unmixing are springing up. Sparse unmixing now has been a active topic in the field of remote sensing. It is a problem about sparse regression, and its goal is to find the optimal subset of signatures in a big spectral library which can best model the mixed pixel in the scene. But the sparse unmixing focuses on analyzing the hyperspectral imagery without taking advantage of spatial information and so on. We do in-depth research in sparse unmixing on the base of the results of previous research about sparse unmixing. The main contents are as follows:Firstly, it gives a view of linear mixture model and nonlinear mixture model, describes the steps of linear unmixing, including endmember estimated algorithm, endmember extraction algorithm and abundance estimation algorithm.Then,the sparse unmixing is introduced,by assuming that the observed signatures can be expressed in the form of linear combinations of a number of spectral signatures in a library known in advance. Unmixing then amounts to finding the optimal subset of signatures in a spectral library that can best model the mixed pixel. Sparse unmixing is a optimization problem about LO regularization, and always transform into the optimization problem about L1 regularization. In this thesis we utilized the sparse unmixing by variable splitting and augmented lagrangian (SUnSAL). It is a very fast sparse regression algorithm. In order to get better result, we study the relationship between the regular parameter and the punish parameter. We study the influence of the Matrix MC value in unmixing by simulation experiment, and get the following conclusion: the smaller endmember matrix MC value is, the better unmixing result can be achieved.Finally, the research is based on reweighted L1 regularization. Compared to L1 regularized problem,it is closer to the LO regularization problem. It is only mathematical optimization without considering the actual distribution of materials. So a method of sparse unmixing based on corrected reweighted L1 regularization is proposed in this paper. In the process of weights iteration, the spatial information is introduced. The experimental results,conducted with both simulated and real hyperspectral imagery, show that the improved weighted L1 regularization sparse unmixing algorithm can effectively improve the accuracy in low signal-to-noise ratio hyperspectral images.