Research On Fusion Methods Of Hyperspectral Remote Sensing Images Based On Block Term Tensor Decomposition

Due to the limitation of the satellite optical sensor.Multispectral images have high spatial resolution and low spectral resolution.Hyperspectral images have high spectral resolution and low spatial resolution.These images do not meet the needs in many application areas such as feature classification and target identification.Therefore,in order to obtain high spatial reso-lution hyperspectral images(HRHS),researchers have performed hyperspectral image fusion by combining high spatial resolution multispectral images(HR-MSI)and low spatial resolution hyperspectral images(LR-HSI)from the same scene in order to enhance the spatial resolution of hyperspectral images.As for the existing fusion algorithms,there are always some problems,such as:the fusion process tends to lose part of the spatial structure and spectral information,the fused images suffer from spectral distortion,and the physical interpretation of potential factors cannot be introduced into the framework.Therefore,this thesis will focus on multispectral and hyperspectral image fusion techniques,with block term tensor decomposition as the research focus,to address the above problems.The main work and contributions are as follows:(1)To address the problem that typical Polyadic Decomposition(CPD)and Tucker models cannot introduce the physical interpretation of potential factors into the framework,and it is dif-ficult to generate high-quality fused images using known attributes.In this thesis,we propose a Multispectral and Hyperspectral Image Fusion Based On Regularized Coupled Non-negative Block Term Tensor Tensor Decomposition(2D-CNBTD)model to estimate ideal high spatial resolution hyperspectral images,firstly,characterize the sparsity by imposing L1 parametriza-tion and introduce Total Variation(TV)to describe the segmental smoothness;secondly,define and introduce different operators in two directions to characterize their segmental;finally,Prox-imal Alternating Optimization(PAO)algorithm and Alternating Directional Multiplier Method(ADMM)are used to solve the model iteratively.Experiments on two standard datasets and two local datasets show that the algorithm proposed in this thesis can remove noise well,reduce spectral distortion and retain spatial information compared with existing fusion algorithms.(2)In order to address the problems of noise/artifact interference experiments caused by un-known tensor rank,high complexity of the algorithm and the existence of scaling/anti-scaling effects in the model,this thesis proposes an image fusion algorithm based on joint structured sparse tensor decomposition of block terms.Firstly,the method forms two abundance matri-ces into a block matrix and imposes L2,1 parametric to promote structured sparse and eliminate the scaling effect in the model;secondly,the inverse scaling effect is achieved by imposing L2parametric on the end element matrix;finally,the focus is on solving the noise/artifacts aris-ing from the inability to accurately estimate the tensor rank and overestimation of the rank in the block term tensor decomposition model.The chunking matrix and the end element matrix are coupled together to reorganize the matrix,to which the L2,1 parametric boost is applied to eliminate the chunks,and the problem is solved using an extended Iteratively Reweighted Least Squares(IRLS)method.Experiments were performed on two standard datasets and one local dataset,and then the fusion results of each algorithm on the University of Pavia dataset were classified using a spectral-spatial convolutional network(LS~2CM).Compared with existing fu-sion algorithms,the algorithm proposed in this paper retains spatial information better,reduces noise and blurred blocks,and fuses images with little spectral distortion.