Research On Super Resolution For Remote Sensing Images Via Deep Learning With Attention Mechanism

Multispectral remote sensing images and hyperspectral remote sensing images with high spatial resolution,which have rich texture,shape,spectrum and other detailed information,have been widely used in urban and rural planning,environmental monitoring,precision agriculture,military reconnaissance and other application fields.The spatial resolution of remote sensing images not only determines the ability to capture small details of target,but also can greatly impact the performance of practical applications.However,there is always a tradeoff between the spatial,spectral,and temporal resolutions in remote sensing,and it often happens that there are no available multispectral and hyperspectral remote sensing images with high spatial resolution at a specified time and place.Therefore,it is still necessary to improve the spatial resolution of multispectral remote sensing images and hyperspectral remote sensing images.Single image super resolution reconstruction method for remote sensing images aims to obtain remote sensing images with high spatial resolution that cannot be directly observed by remote sensing sensors based on the limited information of only one remote sensing image with low spatial resolution.Although good progress has been made in the research of deep-learning-based super-resolution algorithms for remote sensing image in recent years,these algorithms still have many challenging problems in terms of reconstruction performance,generalization performance and application capability when applied to real-world application scenarios,such as how to better integrate the multi-scale and multi-level features,how to better balance spectral and spatial features,how to better balance the pixel error and visual effect,how to better decompose super-resolution complex tasks,and how to adaptively deal with various degradation factors.In this paper,the super-resolution method for multispectral and hyperspectral remote sensing images is proposed to improve the reconstruction performance,and a multi-step generalization method of blind super-resolution for remote sensing images based on the pre-trained models is proposed to improve the generalization performance.Finally,the application capability of the proposed superresolution method is verified in classification applications.The research is summarized as follows:(1)We propose a super-resolution method for multispectral remote sensing images using attention mechanism and adversarial training strategy.In terms of model structure design,a super-resolution network is designed by combining local and global attention mechanism,dense connection,residual learning and other strategies,which can fully integrate and utilize the rich features of shape,scale,and texture in remote sensing images.In terms of model training strategy,various loss functions are used to balance the pixel error and visual effect of the super-resolution reconstruction,and an effective adversarial learning algorithm with gradient penalty and relativistic adversarial loss is used to promote the model to generate more realistic high-resolution details.(2)We propose a super-resolution method for hyperspectral remote sensing images using grouped attention and gradient reconstruction.In terms of model structure design,a grouped attention module is used to increase the discrimination ability among hundreds of bands and features,and the gradient information is fully fuesd in the reconstruction process to promote the generation of sharp edges and realistic textures.In terms of model training strategy,the progressive super-resolution framework with decomposition and fusion strategy is proposed to reduce the reconstruction difficulty from task level,spectral level and feature level,and various loss functions are used to balance the reconstruction effects in both space and spectrum.(3)We propose a multi-step generalization method of blind super-resolution for remote sensing images to adaptively handle the degradation with unknown blur kernel.In terms of reconstruction theory,the generalization feasibility of pre-trained models in blind super-resolution tasks is deduced,and a multi-step generalization framework for blind super-resolution of remote sensing images based on pre-trained models is designed.In terms of model structure design,a blur kernel estimation network and an image correction network based on blur kernel are designed,and a multi-step loss function considering regular terms of blur kernel prior and contrastive learning is adopted to promote more accurate blur kernels and clearer images.(4)In order to verify the effectiveness for practical application of the superresolution method proposed in this paper,the application experiments of scene classification for multispectral remote sensing images and land cover classification for hyperspectral remote sensing images are carried out.Experiments show that our proposed generalized super-resolution method can effectively improve the spatial resolution of remote sensing images and better preserve the spectral characteristics of hyperspectral images,which can provide a wider range of data sources for practical applications and effectively improve the application level of scene classification and land cover classification for remote sensing images.In summary,this thesis expects to improve the reconstruction performance and generalization performance of super-resolution methods for remote sensing images,obtain high-resolution remote sensing images with better spatial integrity,more accurate spectrum and richer details in real application,and further improve the practical application level of remote sensing images.