Research On Deep Learning-based Remote Sensing Image Super-resolution

High-resolution remote sensing image present rich spatial information,which is an important information carrier for earth observation and is high-value for applying in areas such as urban,agriculture,and military and defense.However,often remote sensing images are inherently low spatial resolution due to the limitation from physical factors like the imaging distance and revisit cycle of remote sensing system and the influence from the mutual restriction between imaging spatial resolution and spectral and temporal resolutions.For overcoming the inherently low spatial resolution of remote sensing images and promoting further the application of remote sensing images in areas like urban planning and investigation,fine classification of crops,and military and defense,where usually involve complex scene analysis and small object detection,this thesis,based on deep learning theory,studies in depth the super-resolution of complex remote sensing images and the large-scale super-resolution of remote sensing images.Specifically,(1)A Multi-Perception Attention Network for complex remote sensing image super-resolution is developed.Considering the problem that reconstructing complex remote sensing image requires recovering much spatial details but the inherently low spatial resolution of remote sensing images makes poor the available reference clues in image domain,the proposed network,based on Multi-Perception Learning,builds a parallel two-branch structure of which each branch consists of several Residual Channel Attention Groups and each group further contains Enhanced Residual Block and Channel Attention Module,and overcomes the previously mentioned problem by extracting and adaptively fusing multi-scale feature information in feature domain as clues for reconstructing image.The effectiveness of Multi-Perception Learning is quantitatively supported by validation experiments.Experimental results on remote sensing dataset,natural benchmarks,and real-world remote sensing data show further that the proposed Multi-Perception Attention Super-Resolution Network is strong in reconstructing complex scenes.(2)A Dense-Sampling Super-Resolution Network for remote sensing image large-scale reconstruction is presented.The proposed network,which introduces a Dense Upsampling Mechanism,upsamples the features learned at different network depths to obtain high-resolution multi-scale feature information and meet the demand of large-scale reconstruction for high-resolution details.Besides,a Chain Training Strategy is proposed to benefit the network’s large-scale reconstruction ability by transferring the learned mapping parameters from the networks trained for small-scale reconstruction.And a Wide Residual Block is constructed and used as basic building block,to enhance the network’s nonlinear representation capability.Experiments suggest that Dense Upsampling Mechanism and Chain Training Strategy optimize effectively the network performance for large-scale reconstruction,and show that Dense-Sampling Super-Resolution Network is a competitive player in reconstruction tasks of remote sensing dataset,natural benchmarks,and real-world remote sensing data.(3)An Enhanced Back-Projection Super-Resolution Network employing Transfer Training Strategy is studied,considering the lacking benchmark training data of deep learning-based remote sensing image super-resolution.For breaking the limit of lacking remote sensing data for training,the proposed network is first trained using benchmark natural images to learn shareable low-level feature knowledge and then trained using remote sensing images to make its parameters toward the statistic characteristics of remote sensing data.Experiments show that the used Transfer Training Strategy enables the network outperforming those are trained using directly remote sensing images or benchmark natural images and provides a solution for overcoming the limit caused by lacking benchmark training data on study of remote sensing image super-resolution.