| Hyperspectral remote sensing technology has widely been used in many important fields such as environmental research,geological research and ocean research.With the emergence and progress of imaging spectrometer technology,the resolution of hyperspectral images（HSI）has been greatly improved.However,The problems of difficult sample labeling,noise interference and curse of dimensionality in HSI pose great challenges for data processing.Therefore,how to effectively extract features from HSI data so that the dimensionality of the data can be reduced without affecting its subsequent image processing is the focus of this thesis.1.An unsupervised spatial-spectral neighbor hypergraph embedding algorithm is proposed for the HSI feature extraction.To improve the problem that traditional graph embedding methods cannot reveal complex structure in HIS,and that only spectral information is considered in the construction of the graph,while ignoring spatial information.The proposed algorithm makes full use of the spatial and spectral information of HSI to construct effective unsupervised spatial-spectral nearest neighbor relationships and reduce the impact of different objects with the same spectrum.The hypergraph embedding constructed by unsupervised spatial-spectral nearest neighbor information can reveal complex multivariate relationships among high-dimensional data and achieve effective discriminative feature extraction.Comparative analysis on the Indian Pines and Salinas datasets show that the proposed feature extraction algorithm can improve the feature extraction and classification performance.2.An unsupervised spatial-spectral collaboration-competition preserving graph embedding algorithm is proposed for the HSI feature extraction.To improve the problem that only spectral information is considered in the construction of the graph,and that the superficial use of spatial-spectral information to construct the graph cannot reveal deep features of HSI.In an unsupervised environment,the proposed algorithm preserves the global geometric structure using a cooperative representation,and preserves the local manifold information using local constraint properties based on spatial and spectral nearest-neighbor information.The spatial-spectral collaboration-competition representation coefficients are used to construct the spatial-spectral collaboration-competition preserving graph embedding which can extract the deep features of HSI.The HSI data are projected into the low-dimensional space by graph embedding and low-dimensional data are obtained.Experiments results on the publicly available datasets Indian Pines,Salinas and PaviaU show that the proposed algorithm in this thesis can preserve both global geometric structure and local spatial-spectral manifold information for data in low-dimensional space,thus improving the feature extraction and classification performance of the algorithm. |
