A Study On Spatial Super-resolution Reconstruction Of Hyperspectral Remote Sensing Images Based On Deep Learning

Hyperspectral imaging contains a large number of narrow spectral bands to achieve continuous spectral coverage.However,due to the limitations of imaging mechanisms and optical devices,collection equipment often sacrifices spatial resolution for higher spectral resolution.This can lead to limitations in identifying and interpreting small features,which limits the application of hyperspectral imaging in various fields.Therefore,improving the spatial resolution of hyperspectral imaging is crucial.Hyperspectral image super-resolution reconstruction is a post-processing technique that improves the spatial resolution of input hyperspectral images without changing the optical sensor.Currently,common solutions involve using deep learning to obtain images with high resolution in both spatial and spectral domains.However,there are issues with low processing efficiency and incomplete extraction of complex details for images with different spatial complexity.Using the same convolution method with equal weight values to fuse spatial information across different bands inevitably results in noise mixing in each band,leading to spectrum mismatching.This paper proposes two methods for hyperspectral image super-resolution reconstruction: a single-channel regionalized approach and a full-channel spatial-spectral dual-path residual method guided by spectral response functions.Firstly,to improve the spatial resolution of hyperspectral images and simplify computation,a single-channel regionalized hyperspectral image super-resolution reconstruction method is designed.To address uneven spatial information distribution in hyperspectral images,a sub-map shunts module is proposed to use inter-image texture information to effectively use local spatial features and improve image reconstruction efficiency.In addition,to fully retain image texture details and eliminate unnecessary noise,high-and low-frequency information is separated and enhanced using different strategies to improve image quality.A large number of experimental results show that this method is superior to state-of-the-art CNN-based methods in terms of quantitative metrics and visual quality,saving time while retaining spatial details.Secondly,to improve spatial resolution in hyperspectral imaging while maintaining spectral correlation,a full channel Spatial spectral dual-path network based on spectral response functions is designed.To address the problem of spectral disorder caused by modeling all spectral bands simultaneously,a spectral response function is used to guide the grouping of bands with spectral correlation and the spectral dimension attention mechanism to alleviate spectral confusion.To improve the fusion ability of same-level spatial-spectral information,a spatial-spectral dual-path parallel structure is used,allowing the exploration of spatial-spectral features simultaneously.The spatial path is used to extract spatial-spectral features and reconstruct sharp edges and realistic textures,while the spectral path is used to model spectral correlation to refine spectral features and maintain spectral correlation.Experiments on the QUST-1 satellite show that this method significantly improves quantitative evaluation metrics and visual quality in various scenarios while effectively improving spatial resolution and maintaining spectral correlation.