基于深度学习的遥感图像超分辨率重建方法综述  

Overview of Methods for Remote Sensing Image Super-resolution Reconstruction Based on Deep Learning

作　　者：成科扬[1,2,3] 荣兰 蒋森林 詹永照 CHENG Keyang;RONG Lan;JIANG Senlin;ZHAN Yongzhao(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212013,China;Zhenjiang Zhaoyuan Intelligent Technology Co.,Ltd.,Zhenjiang 212013,China;Jiangsu Province Big Data Ubiquitous Perception and Intelligent Agricultural Application Engineering Research Center,Zhenjiang 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013 [2]镇江昭远智能科技有限公司,江苏镇江212013 [3]江苏省大数据泛在感知与智能农业应用工程研究中心,江苏镇江212013

出　　处：《郑州大学学报（工学版）》2022年第5期8-16,共9页Journal of Zhengzhou University（Engineering Science）

基　　金：国家自然科学基金资助项目(61972183);镇江“金山英才”高层次领军人才培养计划培养对象科研项目。

摘　　要：基于深度学习的遥感图像超分辨率重建方法是计算机视觉中的重要方法之一。传统的遥感图像超分辨率重建方法已无法满足地物目标识别和土地检测等应用的需求,如何利用深度学习来重建遥感图像的分辨率是目前要解决的问题。结合国内外最新研究现状,将基于深度学习的遥感图像超分辨率重建方法分成3大类:单幅遥感图像超分辨率重建方法、多幅遥感图像超分辨率重建方法和多/高光谱遥感图像超分辨率重建方法。系统梳理了基于深度学习的单幅遥感图像超分辨率重建方法,包括基于多尺度特征提取的方法、结合小波变换的方法、沙漏状生成网络的方法、边缘增强网络的方法以及可跨传感器的方法。总结了基于深度学习的多幅遥感图像和多/高光谱遥感图像超分辨率重建方法中目前主流的方法。通过实验结果分析了遥感图像超分辨率重建方法目前效果最好的单幅图像超分辨率重建方法是基于GAN的方法,但是多幅遥感图像和多/高光谱遥感图像超分辨率重建效果仍然不佳,存在配准融合、多源信息融合等问题。最后,对基于深度学习的遥感图像超分辨率重建方法未来可能的发展趋势进行了展望,指出构建针对遥感图像特点的神经网络结构,无监督学习的遥感图像超分辨率重建方法,以及多源遥感图像的超分辨率重建方法是未来的研究趋势。Remote sensing image super-resolution reconstruction based on deep learning is one of the most important methods in computer vision.The traditional super-resolution reconstruction method of remote sensing image could not meet the needs of ground object recognition,detailed land detection and other applications,This study aimed to solve the problem by using deep learning to reconstruct the resolution of remote sensing image.After reviewing the latest research status at home and abroad,this paper divides deep learn-based remote sensing image super-resolution reconstruction methods were classified into three categories,includeing single remote sensing image,multi-remote sensing image and multi-hyperspectral remote sensing image super-resolution reconstruction methods.The methods of super-resolution reconstruction of single remote sensing image based on deep learning were systematically examined,including multi-scale feature extraction method,combined with wavelet transform method,hourglass generation network method,edge enhancement network method and cross-sensor method.The current mainstream methods of multi-remote sensing image and multi-hyperspectral remote sensing image super-resolution reconstruction were also examined based on deep learning.Through the analysis of the experimental results,the best single image reconstruction method is based on GAN,but the effect of multi-remote sensing image and multi-hyperspectral remote sensing image reconstruction was still not good enough,there were several prablems,such as registration fusion,multi-source information fusion and other soon.Finally,the future development trend of remote sensing image super-resolution reconstruction method based on deep learning was explored,The future research trend could be building neural network structure according to the characteristics of remote sensing image,unsupervised learning remote sensing image super-resolution reconstruction method,and multi-source remote sensing image super-resolution reconstruction method.

关 键 词：遥感图像 超分辨率重建 深度学习 卷积神经网络 生成对抗网络 

分 类 号：P237[天文地球—摄影测量与遥感] TP391[天文地球—测绘科学与技术]