Research On Super-Resolution For High-Spatial-Resolution Hyperspectral Image

With the development of aerospace,optics and computer science,hyperspectral remote sensing imaging technology has developed rapidly.Hyperspectral images(HSIs)include spatial distributions and record spectral signatures of observed scene in continuous narrow bands,which can help to accurately detect and identify different land-covers in the observed scene.Since the spectral signatures can reflect the material information of the land-covers,HSIs have wide applications,such as intelligent agriculture,environmental monitoring,and forestry disaster analysis.However,the hardware limitation of existing imaging systems usually leads the acquired HSIs to have low spatial resolution,thereby limiting the application scope of HSIs.In order to obtain the high-spatial-resolution(HR)HSI,this dissertation has carried out two researches on the spatial dimension and spectral dimension of remote sensing images: one is improving the spatial resolution of HSI to reconstruct HR HSI,which is called HSI spatial super-resolution method;the other is improving the spectral resolution of the multispectral image(MSI)by utilizing the HR characteristic of MSI to reconstruct HR HSI,which is called MSI spectral super-resolution method.Although the existing methods have achieved superior performance,there still exists the following problems: 1)The precondition of correct registration of HSI and MSI is needed;2)The potential mapping between observed MSI and HSI in the spectral domain is ignored;3)The spectral signatures of the MSI are not fully utilized,and the neighboring band correlation prior of the HSI is ignored;4)Existing supervised learning-based methods rely on abundant paired MSIs and HSIs.For these problems,this dissertation carries out research on super-resolution for HR HSI from four parts,and main research contents and contributions are listed:(1)Synchronous nonnegative matrix factorization for unregistered HSI and MSI fusion.This method independently estimates HSI abundance and MSI abundance,so as to avoid relying on the premise of correct registration of HSI and MSI.The unregistered HSI and MSI fusion problem is transformed into a bound-constrained optimization problem.Considering the non-negativity of endmembers and abundances,a synchronous projected gradient method is proposed to solve this problem.Experimental results on different datasets prove that this method has superior performance in fusing both registered and unregistered HSI and MSI.The spectral angle mapper of the HSI reconstructed by the proposed method under the unregistered condition is 18.6% better than existing methods on the Washington DC Mall dataset.(2)Self-supervised spectral-spatial residual network for HSI spatial super-resolution.This method utilizes the potential spectral mapping between MSI and HSI,and regards the HSI spatial super-resolution problem as the pixel-wise spectral mapping problem.By learning the spectral mapping between MSI and HSI,HR HSI can be directly reconstructed from HR MSI,which can effectively alleviate the spatial structure distortion of the reconstructed HR HSI.A self-supervised fine-tuning strategy is proposed to promote the network to learn the optimal spectral mapping.The spectral angle mapper of the HSI reconstructed by the proposed method is 16.7% better than existing methods on the Pavia University dataset.(3)MSI spectral super-resolution using spectral-spatial residual attention network.This method fully utilizes the spatial information and spectral signatures of MSIs.A spectral-spatial residual attention network is utilized to directly learn the spectral mapping from MSI to HSI,and the network is lightweight with high computational efficiency.A spatial-spectral residual block is proposed to extract spatial features and spectral features of MSI in parallel.The parallel feature extraction helps to transmit the spatial features and spectral features to deep layers.A neighboring spectral band attention module is proposed to explicitly constrain the reconstructed HSI to maintain the correlation between neighboring spectral bands,which can further improve the quality of the reconstructed HSI.The spectral angle mapper of the HSI reconstructed by the proposed method is 11.3% better than existing methods on the Washington DC Mall dataset.(4)Semi-supervised spectral degradation constrained network for MSI spectral super-resolution.This method can alleviate the problem that existing supervised learning-based methods rely on a large number of paired MSIs and HSIs.This method integrates the learning of spectral mapping from MSI to HSI and the estimation of spectral degradation into a unified framework,and utilizes spectral degradation constraints to promote the learning of the optimal spectral mapping from MSI to HSI.A semi-supervised training strategy is proposed to optimize the weight parameters of the network,which can fully exploit both MSI and HSI pairs and the MSIs without ground truth HSIs for semi-supervised training.The spectral angle mapper of the HSI reconstructed by the proposed method is 4.2% better than existing methods on the Paris dataset.