| With the development of computer technology, digital signal processing technology, and image processing technology, the hyperspectral remote sensing image processing technology has become an attractive research field in the real application. Electromagnetic energy scattered reflected by the surface object can be measured by image spectrometers in their instantaneous field view in hundreds or thousands of spectral channels with higher spectral resolution, which contributes to material identification via spectral signatures. However, due to the low spatial resolution of image spectrometers and topographic change etc in the real situation, mixed pixels that are the combination of different materials and contained in hyperspectral images make it difficult to analyze the nature of the hyperspecral image. For solving the mentioned problem effectively, spectral unmixing technology that realizes endmember extraction and abundance estimation is an available and effective choice.Basing on the basic linear mixture model in this thesis, under the consideration of the spatial structure information of the hyperspectral data and the high mutual coherence of spectral signatures in the spectral library, we mainly study group sparse unmixing to further improve the accuracy of endmember extraction and abundance estimation. The main contributions of this thesis are as follows:1. Spatial-spectral information based abundance constrained for group sparse unmixing is proposed. Under the consideration of the spatial structure information of the hyperspectral data, the adaptive segmentation can be realized on the data by the mean shift algorithm and sparse unmixing based on group sparse regression is conducted on each group of the hyperspectral pixels. A dictionary pruning strategy guided by abundance is used to further weaken the influence of the high mutual coherence of the spectral library on the basis of the group sparse unmixing. In conclusion, under the consideration of the characteristics of spectral library combined with the spatial information of hyperspectral images, a new framework based on group sparse unmixing is proposed to improve the sparse unmixing.2. A method of based on the hierarchical row sparse for hyperspectral image unmixing is proposed. This method is on the base of the assumption that the estimated abundance matrix is hierarchical sparse, row sparse and sparse in each row, which considers the spatial structure information of the hyperspectral data and the group information of the spectral library. With the above mentioned characteristics of the estimated abundance matrix, we can formulate a new mathematical expression solved by a bi-group sparse unmixing algorithm to solve the spectral unmixing. As a result, the proposed method that considers for spatial information of the hyperspectral data and group sturcture information of spectral library through a novel model solves spectral unmixing effectively and receives better sparse unmixing.3. Graph regularized non-negative group sparse coding is proposed and used for endmember extraction. The method aims at retaining the low mixed pixels regarded as the candidate set for the following endmember extraction and ripping off the high mixed pixels from the hyperspectral image. The pure pixels almost exist in the spatial homogeneous areas in which contain the low mixed pixels. In this thesis, a new optimization object function based on the above ideas is established and a graph regularized group sparse coding algorithm is presented. Using the obtained coefficient matrix, we can get the endmember candidate set. In the end, we extraction endmembers chosen from the endmember candidate set with traditional method. Through the experimental results executed on the simulated data and real data, its effectiveness and practicability is proved.This work was supported by the National Natural Science Foundation of China(No. 61272282), and the Program for New Century Excellent Talents in University(NCET-13-0948)... |
PDF Full Text Request