Research On Theory And Algorithms Of Mixed Pixel Decomposition For Hyperspectral Remote Sensing Image

Hyperspectral Remote Sensing technology is the frontiers of Remote Sensing development.Thanks to the high spectral resolution and images integrated with the spectral information of thehyperspectral imagery, it plays a more and more important role in the civilian and military fields.But, the mixed pixels which are widespread in the hyperspectral image(HSI) become an mainobstacle to the in depth development for quantification analysis of HSI. Mixed pixeldecomposition(MPD) which is the most effective method to solve the mixed pixel problem canbreak through the limitation of spatial resolution, obtain the property information of the mixedpixels on the subpixel precision and improve the classification precision. A general approach forMPD is first to build a mathematical model for the spectral mixture, and then apply this model toimplement the MPD. Because of the simpleness and clear physical meanings of the linearspectral mixture model(LSMM), several theory and algorithms of MPD were studied based onLMM in this dissertation. The main research works and corresponding creations are listed asfollows:1. The theoretical foundations of MPD were summarized and analyzed, which include themixed pixel decomposition model, spectral data's dimension reduction algorithms, the estimationmethods of endmember number, the endmember extraction algorithms, abundance estimationalgorithms, the evaluation methods for the MPD accuracy. The four endmember identificationalgorithms(N FINDR, VCA, SGA and OSP) and two endmember generation algorithms (ICEand MVC NMF) were compared with the simulated and real hyperspectral imagery, the meritsand drawbacks of endmember identification algorithm and endmember generation algorithmwere concluded.2. A maximum volume simplex endmember extraction algorithm(MVSEEA) based on thevolume calculation method of the simplex in the original spectral space was proposed. In view ofthe relationship of the LSMM and the convex simplex theory, it used the volume calculationmethod of the convex simplex that doesn't need the dimention reduction procedure into theN FINDR and SGA algorithms, this avoids the information loss caused by the dimensionreduction. The experimental results showed that the endmember extraction results of theMVSEEA based on the N FINDR principle are similar to N FINDR, but it is faster thanN FINDR, the MVSEEA based on the SGA principle is better than N FINDR, SGA and VCA onprecision.3. After analyzing the influence on spectral unmixing error when the endmembers are notincluded properly, the optimal endmember subset selection algorithm based on the endmembervariable theory was giver. To improve the unmixing accuracy, it sorts the endmembers according to the abundance estimation result using the initial endmember set, then estimates the abundanceusing the unconstrained least squares method and the sorted endmember results, at last, theoptimal endmember subset can be fixed on according to the abundance estimation results.Experiments showed that this algorithm is right and effective.4. Based on the similarity of the LSMM and nonnegative matrix factorization(NMF) , theMPD algorithms based on the constrained NMF were brought forward. When the NMF is usedfor MPD directly, the result is not unique and doesn't fulfill the abundance sum to one constraint,the sum of squared distance constraint NMF can solve this problem. According to thecorresponding relations of the endmember variable theory and the sparseness of the abundancematrix, the smoothed L0 norm constraint for the abundance matrix was introduced into the sumof squared distance constraint NMF algorithm, this further improves the MPD accuracy.Experiments showed that proposed MPD algorithms based on the constrained NMF outperformsthe MVC NMF and MVSEEA.