The Research On Hyperspectral Imagery Unmixing Technology Based On Kernel Methods

Hyperspectral remote sensing images can provide enough information for spectral unmixing, owing to its high spectral resolution and hundreds of spectral channels ranging from 0.4 to 2.5 micrometers. However, the modern spectrometer could not bring us to the same high spatial resolution, so the mixed pixels are widespread in hyperspectral imagery. It can improve the precision of imagery classification and target detection by use of the decomposition of the subpixels in the hundreds of imagery data. It is a good research direction for the detection and analysis of the micro-scale and anomaly objection. There are two main hyperspectral unmixing methods that include linear-based and nonlinear-based algorithms. Recently, many scholars develop kinds of new methods in the area of hyperspectral decomposition. The dissertation mainly studies the kernel-based hyperspectral imagery nonlinear umixing techniques and applications. By systematically analysing the traditional spectral decomposition theory, the thesis focuses on extending the algorithms based on linear spectral model to nonlinear feature space in terms of the kernel functions for resolving the difficulties of hyperspectral nonlinear unmixing problem.The major works and contribution of this dissertation are as follows:(1) The orthogonal subspcace projection approach is an traditional supervised unmixing method. It can suppress undesired or interfering spectral signatures, and detect the presence of a spectral signature of interest. This operation is an optimal interference suppression process in the least squares sense. In our thesis, we extend it to nonlinear space to exploit the high order features between the spectral data, and get better robust performance.(2) We extend the traditional nonnegtive matrix factorization method to nonlinear space, and then get a kernel-based nonnegative matrix factorization method that includes pure pixels kernel NMF and null pure pixels kernel NMF. It is a nonlinear version of NMF by using the kernel function to transform the dot product without any knowledge of the nonlinear mapping function. The simulated data and the real hyperspectral imagery experiments show that the methods can provide good performance.(3) According to the spatial continuity of the spectral abundances, we develop a kernel spatial complexity-based nonnegative matrix factorization by incorporating the spatial complexity into a nonlinear version of NMF method. We realize the nonlinear decomposition of hyperspectral imagery via kernel methods. It can improve the performance of the spectral umixing by exploiting the high order features in the high-dimensional feature space.