Unmixing Of Hyperspectral Imagery Based On Grouping Fisher Discriminant Analysis

Hyperspectral imagery (HSI) gets the sacrificial signatures and is obtained by imagingspectrometers, which can capture the spatial information and near-continuous spectra at thesame time. With the development of remote sensing techniques, HSI has been applied in moreand more fields, and the applications call for advanced techniques for hyperspectral dataprocessing. HSI is generated by imaging spectrometer simultaneously to the same surfacescenery at dozens even hundreds bands. Mixed pixels are widely existent in HSI for its lowspatial resolution. This problem must be solved in order to make better use of the HSI. Whenall classes included are known, the problem is the proportion occupied by each class in mixedpixel. Spectral unmixing is forward aiming at this problem.It is required to establish the spectral mixing model to unmix the mixed spectral. Thereare liner (LMM) mixing model and nonlinear mixing model. Linear mixing model is easier tounderstand compared with nonlinear mixing model, and its physical meaning is clear. LMMassumed that there are several pure materials spectral in the mixed pixel, and the mixed pixelspectral is linear combination of the pure materials spectral. This dissertation researchedLMM and HSI data, and the major contribution as follows.First, unmixing of hyperspectral imagery based on grouping Fisher discriminant analysis(FDA). Spectral variability always exists in practical situations，which reduces the accuracy ofmixed pixel decomposition. In order to solve the problem, the spectral data were translated bygrouping Fisher discriminant. This translation is linear translation. Fisher discriminantanalysis searches a linear combination of the spectrum, of which the endmember spectra havethe largest separation degree, indicating that it has small variability inside one endmembergroup but a large difference among endmember groups. The synthetic data and realhyperspectral data experiments show that the proposed method gives higher unmixingaccuracy than traditional LMM method.Second, unmixing of hyperspectral imagery combined with Fisher discriminant analysisand nonnegative matrix factorization (NMF). Nonnegative matrix factorization was proposedat the end of the twentieth Century. It is suitable to unmixing of hyperspectral imagery due tothe hyperspectral data is nonnegative. It requires a large number of iterations in thecalculation process, so the efficiency of the method is low. The dissertation successfully combined the FDA and NMF in the unmixing of hyperspectral imagery; this can reduce thedimension of the original data and reduce the spectral variability at the same time. Theexperiments show that the efficiency of the new method had greatly improved, and theunmixing accuracy is higher than the original NMF method, too.Third, LMM is one of the most classic and effective model for unmixing hyperspectralimagery till now. There are many scholars are still studying unmixing hyperspectral imagebased on LMM, and many new idea emerge，liner spectral unmixing using a hierarchicalBayesian model for hyperspectral imagery is one of them. Dissertation researched the methodand introduces the groping FDA idea into the method. Though the result of this method is notas well as we expect, scholars may get some enlightenments by the result analysis.