Hvperspectral Endmember Extraction And Its Fast Implementation Based On Simplex Growing Theory

Hyperspectral remote sensing has been the advanced technology and plays a more and more important role in various fields for its multi-band, high spectral resolution, large data amount, etc. However, the mixed pixels exist extensively due to the limitation of imaging system’s spatial resolution and the complex diversity of the ground targets, thus, spectral unmixing is important in remote sensing. As the key step in linear spectral unmixing, endmember extraction (EE) research is very important in hyperspectral image analysis. The new simplex growing algorithm (NSGA), which is developed as a linear alternative to the SGA algorithm, still suffers from several issues such as its inapplicability for nonlinear mixing model and the computational complexity. This paper aims at solving the two above issues and the major work and contribution of this dissertation are as follows:(1) On account of inapplicability of NSGA for hyperspectral imagery where linear mixing model isn’t appropriate due to the multiple scattering effects, this paper extends NSGA to its kernel version by utilizing kernel method and puts forward the kernel-based algorithm named kernel new simplex growing algorithm (KNSGA).(2) In order to solve the excessive computation involved in the two algorithms NSGA and KNSGA which is caused by the iterations of simplex volume calculation, a novel approach is proposed by utilizing partitioned matrix determinant formula to simplify the volume calculation and thus two fast implementation algorithms named FNSGA and FKNSGA are present in the paper.(3) In order to solve the excessive computation involved in the two algorithms NSGA and KNSGA, another new method is put forward by taking advantage of the modified Cholesky factorization to decompose the volume matrix into triangular matrices beforehand, which helps avoid directly computing the determinant tautologically in the simplex volume formula. And FNSGACF and FKNSGACF are the corresponding fast algorithms.Experiments on both simulated and real spectral data demonstrate that the proposed NSGA and the kernel-based algorithm KNSGA can accurately extract endmembers and the four fast implementation algorithms for linear and nonlinear models can significantly reduce computational complexity while their performance remains invariant.