Researches Of Reconstruction And Feature Extraction Techniques For Hyperspectral Remote Sensing Imagery Based On Sparse And Low-Rank Learning

The continuous development of remote sensing imaging technology provides massive high-resolution remote sensing images for earth observation targets.Through image analysis and interpretation technologies,hyperspectral remote sensing images are widely used in precision agriculture,ocean monitoring,surface inversion,ecological environment,urban planning,sustainable development,and even lunar exploration due to their refined spectral diagnostic capabilities.However,when using hyperspectral remote sensing images to solve practical application problems,they still face a series of quality problems,such as noise and low-spatial resolution.These problems affect the accuracy and credibility of the images in subsequent applications.Therefore,before analyzing and interpreting the images,it is necessary to apply preprocessing methods to improve their quality and increase the their applicability.That is,data preprocessing provides quality assurance for subsequent analysis and interpretation,and lays data foundation for the subsequent applications of hyperspectral remote sensing images.The thesis starts from the imaging principle and imaging mechanism of hyperspectral remote sensing imagery,reviews and analyzes the problems that degrade image quality and affect the subsequent applications of images.The thesis summarizes three problems faced by hyperspectral remote sensing images in practical applications:stripes noise,low-spatial resolution and high data dimensionality.By analyzing and characterizing structure properties of hyperspectral images,based on the sparse/low-rank analysis,the problems of stripes noise and low spatial resolution are discussed and studied from the perspective of reconstruction;the problem of high spectral dimensionality is discussed and explored from the perspective of feature extraction.The main research contents can be summarized as follows:1.In view of the dense stripes noise for hyperspectral imagery,research on improving the quality of striped hyperspectral images is carried out to lay a data foundation for the subsequent interpretation analysis and application research of hyperspectral images.The dissertation proposes a hyperspectral stripes noise removal method(WDLRGS)considering low-rank/group-sparse decomposition in the wavelet domain.For the first time that the low-rank/group-sparse decomposition is deployed in a transformation domain but not the original spatial-spectral domain.In the transformation domain,the low-rank prior is used to characterize the directional structure of the stripes noise,and the image information is captured by the group-sparse prior.Through the experimental analysis on the synthetic data set and the real Tiangong-I data set,the relevant experimental results prove the advanced nature and performance superiority of this method.Compared with the existing methods,the proposed method can remove stripes noise while effectively maintaining the spatial details and texture information.2.The hyperspectral imaging spectrometer has mutual constraints and limitations between spatial resolution and spectral resolution in the imaging process.Hyperspectral images with nanometer-level spectral resolution tend to have low-spatial resolution,which leads to each pixel includes spectral components of different targets,limiting the applications such as refined ground target survey and identification.Combining the imaging mechanism of hyperspectral imagery(i.e.,imaging observation model),by introducing multispectral imagery with high-spatial resolution,the thesis studies the reconstruction of hyperspectral spatial resolution based on multi-source remote sensing image fusion.This research proposes a hyperspectral spatial resolution reconstruction algorithm(LRTA)based on low-rank tensor approximation.Different from traditional matrix representation,tensor approximation can describe the intrinsic high-dimensional structural characteristics of hyperspectral images.Combined with the spatial-spectral imaging mechanism of hyperspectral images,it can enhance the spatial resolution of hyperspectral images while maintaining the spectral curve information.Through experimental analysis on the synthetic data set and the real Zhangye data set,the relevant experimental results prove the performance of this method in enhancing the spatial resolution of hyperspectral images.3.In view of the fact that massive unlabeled hyperspectral imagery is facing the current situation of large and time-consuming and labor-intensive data labeling in practical applications,unsupervised hyperspectral image feature extraction is studied.Based on graph embedding framework,the collaborative representation ability of samples from the same class and the competitive representation relationship between samples from different classes are fully explored.Collaboration-competition preserving graph embedding(CCPGE)method is proposed with the research basis on sparse and low-rank analysis for hyperspectral imagery.The proposed CCPGE is able to characterize the intrinsic manifold structure and global characteristics of hyperspectral imagery more accurately while extracting the most informative and discriminative data features.Classifications are carried out on data sets covering different scenes and regions:Indian Pines,Pavia University,Salinas,Trento,as well as the Xiongan,Hebei Province,China.The classification results demonstrate the feasibility and application potential of collaboration-competition graph embedding under graph signal processing framework in feature extraction of hyperspectral imagery.