Research On Spectral Unmixing Technique For Hyperspectral Imagery

With the development of digital signal processing technology, computer technology and communication technology, remote sensing imagery processing takes increasingly important effects in the fields of military and civil applications. Mixed pixels are widely existent in hyperspectral imagery (HSI) for its low spatial resolution. The analysis and processing of mixed pixels are of more importance and significance. As a major mixed-pixel processing technology, spectral unmixing aiming to work out the proportion of mixed composition within mixed pixels is a more precise classification. For mixed pixels, when all classes included are known, there is a kind of main information to be interested. That is the proportion occupied by each class in mixed pixel. Spectral unmixing is put forward aiming at the problems.First, fixed number and fixed shape of endmembers are used in traditional linear spectral mixing modeling (LSMM), leading to a discount unmixing accuracy. Aiming at this problem and to constructing a new spectral unmixing way based on flexibly determined endmembers, the method of spectral shape modifying regionally and the method of endmember subset determining locally are proposed. Experiments show that the proposed method gives higher unmixing accuracy than traditional LSMM method.Second, traditional linear spectral mixing modeling (LSMM) based spectral unmixing method is solved in terms of iteration manner, suffering a heavy computational burden. In this case, a geometric solving method is proposed for LSMM to take replace of the iteration manner. Experiments show that the proposed method gives higher unmixing efficiency than traditional LSMM method. A new implementation method of spectral unmixing based on the SVM modeling is proposed at the same time, which is free of dimensional reduction and makes use of distance measure instead of volume one. Experiments show that the computational burden is decreased greatly by using the SVM modeling.Third, traditional spectral unmixing method based on linear spectral mixing modeling (LSMM) with non-negative and sum-to-one constraints is solved in terms of iteration manner, suffering higher complexity. In this case, parameter substitution is introduced to remove the non-negative and sum-to-one constraints. And so, the process of spectral unmixing is resorted to an optimization problem for finding extreme value of minimum mean square error based fitness function. Then Taguchi optimization algorithm is used to solve the optimization problem iteratively. At the same time, the initialization method is researched. Experiments implemented on synthesized data and truth hyperspectral data show that the proposed method gives higher unmixing efficiency and unmixing accuracy than traditional LSMM method.