Research On Unmixing Algorithms For Hyperspectral Image

Because of the low spatial resolution, there will be a lot of mixed pixels. To improve the results of classification and target detection, it is necessary to study a subpixel hyperspectral image processing technology called hyperspectral unmixing algorithms to separate endmembers and estimate relative aboudances of all the pixels. Researchers around the world have made great efforts in studying such kind of algorithms.Algorithms studied in this paper are based on the well-known linear mixed model. After summarized the mathematical basis of the linearly unmixing problem selectively, the author have proposed three different unmixing algorithms.By replaceing the non- negativity constraint with a negative pena lty term and relaxing the pure pixel assumption in convex geometry based approaches, we proposed a new algorithm named Robust Minimum- Volume Enclosing Simplex(RMVES). A cyclic minimization algorithm for approximating the RMVES problem is developed using the technique named ADMM, which is able to solve large scale of convex problems. In this algorithm, we further provide a method to estimate the regularization parameters automatically. Some simulations and real data experiments are presented to illustrate the efficacy of the proposed RMVES algorithm over some existing hyperspectral unmixing methods.To improve the circumstances of trapping into local minimum in None-negative matrix factorization(NMF) based unmxing algorithms, a substance dependance constraint is considered. To further describe the substance dependance correlation in the image, the well-known 1l-Graph is added and we proposed a new unmxing algorithm based on NMF called 1l-Graph Substance Dependance Constrained NMF(1l SDSNMF). Experiments on both simulation data and real data prove the efficacy of the proposed 1l SDSNMF algorithm. We also prove the convergence of the proposed 1l SDSNMF.The famous sparse regression can also be applied in unmixing framework. To conquer the difficulty brought by high coherence of spectral dictionary and to further utilize spatial information of the hyperspectral image, hypergraph theory is applied to describe the spatial and spectral relationships between pixels. To construct the spatial structure of the image, a hypergraph regular term is added in the sparse regression unmxing model and we proposed a new unmixing method named Hyper Graph Based Sparse Unmixing(HGSU). After prove the target function is a convex, we proposed the detailed solutions of HGSU under ADMM framework. Experiments on both simulation data and real data show that the performance of the proposed algorithm is better than other state-of-the-art sparse regression algorithms.