Research On Super-resolution Algorithm For Hyperspectral Remote Sensing Images Based On Tensor And Matrix Decomposition

Due to the rich spectral information,hyperspectral images have very important applications in the fields of ocean exploration,geological survey,precision agriculture and biomedicine.However,the spatial resolution of hyperspectral images is generally not high due to high sensor cost,multiple imaging platforms,and low signal-to-noise ratio.It significantly affects the application of hyperspectral images in these fields.In the same scene,compared with hyperspectral images,multispectral images have high spatial resolution and low spectral resolution.Therefore,in order to obtain hyperspectral images with high spatial resolution,people usually use image super-resolution techniques to process the original hyperspectral images.Specifically,by fusing the high-resolution multispectral image(HR-MSI)and the low-resolution hyperspectral image(LR-HSI)of the same scene,the purpose of improving the spatial resolution of hyperspectral images can be achieved.This paper focuses on the research hotspot of hyperspectral image super-resolution technology,focusing on matrix decomposition and tensor decomposition.The main work and contributions are as follows:(1)To address the problem that the current tensor decomposition-based super-resolution algorithm for hyperspectral images does not fully consider the high-dimensional flow structure of hyperspectral images,and is sensitive to both outliers and noise.In this paper,a new method based on Joint Regularization Tensor Decomposition(JRLTD)is proposed to solve the HSISR problem from the tensor perspective.The model operates on hyperspectral data using the classical Tucker decomposition and introduces graph regularization and the unidirectional total variation(TV)regularization,which effectively preserves the spatial and spectral structures in the fused hyperspectral images while reducing the presence of anomalous noise values in the images.(2)Aiming at the problem that the current hyperspectral image super-resolution algorithm based on matrix decomposition destroys the spatial structure of hyperspectral images and the correlation between each mode,and causes spectral distortion.In this paper,we propose a hyperspectral image super-resolution algorithm based on nonlocal low-rank(NLLR),which introduces gamma function as a low-rank regularization term into the proposed model,extracts the full-band tensor blocks from the images using the non-local similarity of hyperspectral images,clusters them in blocks,decomposes the tensor blocks in each class afterwards,and finally reconstructs the fused target images class by class.The experimental results confirm that the NLLR algorithm can not only preserve the connection between individual modes of hyperspectral images very well,but also effectively avoid image distortion and preserve image features.In the experimental part of the paper,two public datasets: the Pavia University dataset,the Washington DC dataset,and a real dataset: the Ningxia Sand Lake dataset are used to simulate the generation of HR-MSI and LR-HSI to evaluate the two algorithms proposed in this paper.The proposed algorithm is compared with several state-of-the-art super-resolution algorithms for hyperspectral images,including Coupled Nonnegative Matrix Factorization(CNMF),Coupled Sparse Tensor Factorization(CSTF),etc.The experimental results show that the proposed algorithm can effectively improve the spatial resolution of hyperspectral images and has good fusion performance.