| Recently, Hyper-spectral imaging is becoming one of hot research area as a brand new earth observation technology. By recording tens to hundreds of spectral band information and rich spatial information on all types of surface features, Imaging spectrometer makes accurate classification, identification, recognition and sub-pixels location possible. However, with the improvement of spectral resolution, spatial resolution has been limited. Since that, one pixel of hyper-spectral image is always containing several different types of surface materials, so-called "mixed pixel." What have to bring much effect on the accuracy of the identification and classification of surface features. Spectral unmixing is meaning that spectral materials vector is a combination of one or several basic types of components(i.e end-member) and their corresponding mixing proportions(i.e abundance). By use of accurate abundance estimation results will improve on the effect of the subsequent data application. Hence, abundance estimation is a major problem for the quantitative remote sensing technology development, being badly in need of solving.NMF(Non-negative Matrix Factorization, NMF) will decompose a high-dimensional non-negative matrix factorization into two lower-dimensional non-negative matrices. Similar to linear mixed models, NMF models always have been solved the spectral unmixing problem. However, NMF is an ill and under-determined problem. Hence, when NMF is applied directly to the hyper-spectral image, it is always resulting in local minimum problem in case of pixel decomposition. To offset this problem and disadvantage, this thesis researches on digging spatial structural information, build graph Laplacian manifold and spatial structural regularization to design a variety of non-negative matrix decomposition for abundance estimation and unmixing techniques. Details are as follows:1. Because of traditional NMF abundance estimation methods always obscure spectral-spatial information in scene, a kind of based on geometric structural information NMF(GNMF) abundance estimation algorithm is designed to solve this disadvantage. The method firstly use of local window to dig the deep spatial geometric information, adding to spectral distance graph regularizer to NMF model for abundance estimation. The method experiment on two simulated hyper-spectral data set and two real remote images(AVIRIS and HYDICE).In this section, researching on parameters chose and astringency of this method, paper also analyses the unmixing accuracy and robustness comparing with other advanced algorithms, such as SSACEE, Graphed 2/1l NMF and so on.2. Advanced NMF abundance estimation methods always cost more computing time and more unknown factors. so a kind of spatial-spectral sparse abundance estimation via couple compressing measurement(CC3SAE) is designed to solve these problem. Under the condition of known completed spectral dictionary, the abundance of mixed pixels are sparse, additionally, by research on the spatial distribution structure of abundance and couple sensing to optimize observed matrix, constructing a new graph manifold. Introducing couple non-correlate measurement into the sparse representation, the sparse model can extend to compressing sensing frame. In four simulated data sets and a HYDICE real hyper-spectral image data, this method have been verified advantage than other algorithms, nether in abundance estimation nor in robustness.3. For complex nonlinear mixed models, few NMF algorithm is used to solve this problem. Inspired of above method, NMF method is applied in bi-linear mixed model. By building more concision spatial-spectral graph regularizer, introducing Laplacian graph regularization, a new abundance estimation method has been proposed, which is based on local neighborhood weighted NMF for bi-linear unmixing. Experiments on a simulated hyper-spectral data set and a AVIRIVS real image demonstrate the better performance in bi-linear abundance estimation. |
PDF Full Text Request