Research On Key Issues Of Processing And Analysis For High-resolution Hyperspectral Images

Hyperspectral image(HSI) is a revolutionary combination of the spectrum related to target material and the spatial geometric relationship,which greatly improves the ability of object interpretation and identification.Hyperspectral image with high resolution can provide efficient processing and analysis,and thus,it has become an important supporting technology in national great strategic demand such as resource investigation,space reconnaissance,hazard monitoring,etc.However,there are still some key problems in the high-resolution HSI data,which consists primarily of as:(i)The spectral-spatial structures of high-resolution HSI are very complex.(ii)It is difficult to annotate a large number of high-quality training samples.(iii)The hyperspectral image has high spectral dimension.(iv)The quality of the captured HSI data is susceptible to external environment.All of which lead to the performance bottleneck on traditional HSI processing and analyzing.How to accurately characterize,efficiently integrate,and reconstruct the complex spectrum and spectral-spatial structure of high-resolution hyperspectral images,and to achieve high-precision recognition of hyperspectral images in the case of sample scarcity,are the frontier challenges and problems in the field of information processing and the commanding heights of international competition.Therefore,it is urgent to develop new researches on the key technologies of high-resolution HSI processing and analyzing.On the basis of analyzing and summarizing current researches on hyperspectral image processing,this thesis focuses on three main problems mentioned above including quality degradation,fusion enhancement,and accurate identification of land covers.By fully utilizing the structure characteristics and prior information of different ground objects,we construct sun glint and fog degradation models for hyperspectral image with high resolution.Furthermore,we propose spectral-spatial multi-channel fusion enhancement and multi-scale intrinsic structure extraction methods.These approaches can achieve higher accuracy in complicated scenes.Moreover,The superiority and effectiveness of the proposed methods has been verified in multiple real scenarios.The main contributions of this thesis are as follows:(1)Aiming at the problem of image degradation caused by external environment such as sun glint and fog,this thesis reveals the spectral response mechanism of sun glint and fog in HSIs,and constructs degradation models of sun glint removal and dehazing for HSIs.Based on these models,this thesis proposes models-guided sun glint and fog removal approaches for high-resolution HSI,which achieves high fidelity restoration of HSI spectral-spatial structure information.Compared to other state-of-the-art sun glint removal method,the proposed method can improve the spectral fidelity by 78.94% on average.(2)Focusing on the high dimensionality of HSI that makes it hard to directly reflect differences of different ground objects,based on the mechanism of human vision perception(HVP),this thesis constructs a spectral-spatial multi-channel fusion enhancement framework and proposes a HVP inspired multi-band fusion enhancement method.Experiments demonstrate that the proposed method not only decreases the spectral dimension but also significantly boost the contrast between the background and object,which is beneficial for visual interpenetration and subsequent analyses.More importantly,the objective indexes obtained by the proposed method is increased by 46.15% on average compared to other image enhancement methods.(3)In order to solve the problem that it is difficult to represent the complicated spectralspatial structures of HSI,this thesis elaborates uniform representation mechanism of multi-scale spectral-spatial structural features in a nonlinear space and proposes nonlinear edge-preserving filtering-based feature extraction method.Moreover,a multi-scale intrinsic spectral-spatial feature fusion framework is developed,which can accurately represent the complex spectral-spatial structures of HSI.Experimental results show that the classification accuracy obtained by the proposed method is increased by 21.23% on average compared to other feature extraction techniques.(4)Aiming at the problem of limited classification performance at the situation of scarce training samples because of sample labeling,high cost,rare quantity,etc.,this thesis constructs isolation forest-based hyperspectral object detection framework in the case of very limited samples.Moreover,an extended random walker-based ensemble learning method is proposed for mapping different land covers.Experimental results demonstrate that the classification accuracy obtained by the proposed method is increased by 9.86% on average with respect to other state-of-the-art object recognition approaches.(5)Finally,several representative case studies are used to verify the effectiveness of multiple feature extraction and classification approaches in this dissertation,including: 1)The proposed approaches are utilized to identify different types of wetland vegetation in the Yellow River Delta to analyze the spatial distribution of invasive species.2)The proposed feature extraction and classification techniques are employed to detect the spilled oil in Penglai 19-3c platform,which is helpful for obtaining the spatial distribution of oil spill.3)The proposed approaches are applied to identify different land covers in Matiwan cillage,Xiong’an new area.4)The proposed methods are applied for mapping crops in Honghu area,Wuhan city.5)The proposed approaches are performed on the fused dataset of hyperspectral and visible images to identify minerals in silinyervi,Finland.