Research On Hyperspectral Image Resolution Enhancement Based On Spatial-Spectral Correlation

Hyperspectral image(HSI)includes abundant of spatial information and spectral features,which enable discrimination of different materials.HSI has achieved wide applications in several of military and civil fields.However,due to the factors of imaging optics,it is challenging for acquiring HSI with wide swath and high spatial resolution,which defects the performance of subsequent applications.Modifying imaging equipment will result in long period and high costs.Improving the swath and spatial resolution of HSI through image processing are highly efficient and cost saving.As the spatial and spectral features in HIS are always redundant,hyperspectral image processing based on sparse representation enables extracting spatial information and spectral features,training over-completed dictionary and obtaining mapping schemes between different resolution images to achieve high resolution HSI.In this thesis,we focus on the topic of HSI resolution enhancement based on spatial-spectral correlation including researches on spatial-spectral features and joint processing of multi-task,which are summarized as follows:1.To address issues of how to deal with the previous tasks by referring to requirement of subsequent tasks and reduce transferred errors,we proposed a joint processing model of multi-task for HSI.Feasibility of the proposed model is firstly demonstrated by analyzing the mutual promotion relationship between HSI image enhancement and spectral interpretation.Then,HSI spatial resolution enhancement and spectral unmixing are combined into a joint processing framework,which makes the two procedures alternatively regularize each other and simultaneously improves the performance of them.Experiments on different types of landscapes(simulated,urban,vegetation and mineral)show the effectiveness of the proposed method.Results of the proposed method are superior to the state-of-the-art spatial resolution enhancement methods and unmixing methods.2.To address issues of how to deal with sub-space mapping schemes between different resolution images and refer to the performance of subsequent applications,we proposed a HSI fusion method based on spatial and spectral correlations.The proposed method includes spatial feature fusion and spectral feature fusion,where spatial-and spectral correlation among multiple resolution images are projected into the sub-spaces of spatial-and spectral features.Additionally,HSI fusion and spectral unmixing are iteratively solved by unifying into a joint processing framework to achieve better fusion performance.Experiments on USGS dataset,AVIRIS dataset and Sentinel 2/Hyperion real datasets show the superiority of the proposed method.3.To address issues of how to deal with spatial inconsistency and insufficient spatial constraints,we designed a spectral resolution enhancement method for MSI based on spectral dictionary learning.The proposed method is implemented by spectral improvement strategy and spectral preservation strategy.For spectral improvement strategy,spectral resolution enhancement result is achieved by estimating spectral response function of MSI/HSI and learning band-adaptive spectral dictionaries.In spatial preservation strategy,the abundances consistency of MSI and the desired HSI is used as spatial constraint to ensure spatial consistency and reduce spatial errors.Spectral improvement strategy and spectral preservation strategy are solved iteratively in a joint processing framework to improve the performance of the proposed method.Experimental results of CAVE dataset,AVIRIS dataset and ALI/Hyperion real datasets demonstrate the superiority of the proposed method than other state-of-the-art spectral resolution enhancement methods.4.To address issue of how to simultaneously achieve both spatial resolution improvement and spectral resolution enhancement,we proposed a joint spatial-spectral resolution enhancement method for MSI based on spectral matrix factorization for the first time.The proposed method recovers a high spatial resolution HSI from the input panchromatic(PAN)image and MSI.The recovered HSI should have the same spatial resolution as PAN image.Firstly,two virtual intermediate variables are used to formulate spectral observation model and spatial observation model.The spectral observation model and spatial observation model are then unified into a joint multi-task processing framework,where spectral dictionary and high resolution abundances are alternately solved to reconstruct the desired HSI.Experimental results on AVIRIS dataset and ALI dataset show that the proposed method outperforms the state-of-the-art spatial-and spectral-resolution enhancement methods.5.To prove the reliability of the proposed resolution enhancement methods,we evaluate the performance of them using different types of landscapes(urban,vegetation and mineral).Spectral unmixing is firstly implemented on the reconstruction results of three proposed methods to acquire spectral reflectance and abundances of different objects,whose accuracy is then evaluated by several evaluating indicators to verify the object identification ability and spatial distribution conditions of proposed methods.Ultimately,the applicability of three methods are given,which demonstrate applying conditions for the proposed HSI spatial resolution enhancement method,MSI spectral resolution method and joint spatial-spectral resolution enhancement method,respectively.