基于深度学习的高分四号卫星图像超分辨率重建  

Deep learning-based super-resolution for GF-4 satellite imagery

作　　者：贺智 贺丹 HE Zhi;HE Dan(School of Geography and Planning,Center of Integrated Geographic Information Analysis,Sun Yat-sen University,Guangzhou 510275,China;Guangdong Provincial Key Laboratory of Urbanization and Geo-simulation,Guangzhou 510275,China;City College of Dongguan University of Technology,Dongguan 511700,China)

机构地区：[1]中山大学地理科学与规划学院综合地理信息研究中心,广州510275 [2]广东省城市化与地理环境空间模拟重点实验室,广州510275 [3]东莞理工学院城市学院,东莞511700

出　　处：《遥感学报》2020年第12期1500-1510,共11页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家自然科学基金(编号:41501368);中央高校基本科研业务费用专项资金项目(编号:16lgpy04);2018年东莞理工学院城市学院青年教师发展基金项目(编号:2018QJY001Z);2019年广东大学生科技培育专项资金(“攀登计划”专项资金)(编号:pdjh2019b0623)。

摘　　要：针对高分四号卫星图像中波红外谱段空间分辨率远低于相应的可见光近红外谱段,提出一种基于深度学习的高分四号卫星图像中波红外谱段超分辨率重建方法。首先,设计卷积回归网络进行初步重建,利用8倍下采样的可见光近红外谱段和原始中波红外谱段训练回归网络。然后,设计卷积重建网络进行进一步重建,利用低分辨率初步重建结果和原始中波红外谱段训练重建网络。最后,将原始可见光近红外谱段依次输入到训练好的回归网络和重建网络,得到最终的中波红外谱段超分辨率重建结果。以湖北和江西部分地区真实数据进行试验,结果表明该方法能有效提高中波红外谱段空间分辨率,与其他方法对比均方根误差至少下降了7.54,且目视效果更清晰自然,有利于扩展高分四号卫星的应用范围。GF-4 is a geostationary orbit satellite launched under the support of the major special project of China’s high-resolution earth observation system.Notably,the spatial resolution of the medium-wave infrared channel in GF-4 satellite imagery is much lower than the corresponding visible near-infrared channels of the same scene.The resolution of the medium-wave infrared channel needs to be improved.In this study,a deep learning method termed as GF-4 Super-Resolution Network(GF-4-SRN)is proposed for super-resolution GF-4 satellite imagery.First,a convolutional regression network is designed for preliminary reconstruction of the original image.The convolutional regression network is trained using eight times down-sampled visible near-infrared channels and the original low-resolution medium-wave infrared channel.Thus,the correspondence between the visible near-infrared channels and the medium-wave infrared channel is developed.Second,a convolutional reconstruction network is designed to further reconstruct the preliminary results.This network is trained using the low-resolution preliminary reconstruction results and the original medium-wave infrared image.Notably,the convolutional reconstruction network can build the relationship between the preliminary reconstruction results and the medium-wave infrared image.Finally,the original visible near-infrared channels in GaoFen-4 satellite imagery are fed into the trained GF-4-SRN.The final super-resolution reconstruction results can be obtained by the relationship between visible near-infrared images and the medium-wave infrared image.Experiments are performed on two real-world GF-4 satellite images acquired from Hubei and Jiangxi regions.Experimental results demonstrate that the proposed GF-4-SRN method can effectively enhance the spatial resolution of the medium-wave infrared image.Compared with the root mean square error of state-of-the-art methods,that of the GF-4-SRN is reduced by at least 7.54.Moreover,the visual effects are much clearer and more natural.Therefore,th

关 键 词：高分四号 超分辨率 影像重建 深度学习 卷积神经网络 湖北 江西 

分 类 号：P237[天文地球—摄影测量与遥感] TP391.41[天文地球—测绘科学与技术] TP18[自动化与计算机技术—计算机应用技术]