摘要
遥感影像超分辨率重建有助于丰富地物细节,从而更全面地反映地物目标信息。为了解决目前基于深度学习的超分辨率重建方法难以同时兼顾影像高、低频信息的问题,提出了一种并联式遥感影像超分辨率重建方法。该方法并联了密集深层反投影网络和浅层多尺度网络,利用密集深层反投影网络精确预测遥感影像的高频内容;同时利用浅层多尺度网络来增加目标可分辨能力,并保留影像的低频部分来提升影像的质量。这个方法在GF-1和GF-2数据集上进行了实验,并在Landsat 8和ASTER异源遥感影像数据集上进行了泛化验证。研究结果表明,相较于增强深度残差网络(enhanced deep residual networks for single image super-resolution,EDSR)、深层和浅层端到端卷积网络(end-to-end image super resolution via deep and shallow convolutional network,EEDS)和密集深层反投影网络(deep back-projection networks for super-resolution,DBPN),峰值信噪比(peak signal to noise ratio,PSNR)分别提升了2.30、2.23、0.25 dB,结构相似度(structural similarity,SSIM)性能指标分别提升了0.1316、0.1085、0.0096。本文方法有助于从数据端改善遥感影像目标识别、地物分类等应用的精度,进一步提高遥感数据在资源调查、环境监测、灾害预报等领域的应用效能。
The super-resolution reconstruction of remote sensing images helps to enrich the details of ground objects,and reflect the information of ground targets more comprehensively.In order to solve the problem that the current super-resolution reconstruction method based on deep learning is difficult to take into account the high-frequency and low-frequency information of the image,a new super-resolution reconstruction method based on parallel convolutional neural network was proposed.The method connected dense deep back projection network and shallow multi-scale network in parallel.The dense deep back projection network was used to accurately predict the high-frequency content of remote sensing images.On the other side,the shallow multi-scale network was used to increase target resolution and preserve low-frequency parts of the image to improve the quality of the reconstructed images.This method was tested on the GF-1 and GF-2 datasets and generalized validated on the Landsat 8 and ASTER datasets.Comparison of EDSR(enhanced deep residual networks for single image super-resolution),EEDS(end-to-end image super resolution via deep and shallow convolutional network)and DBPN(deep back-projection networks for super-resolution)three common deep learning super resolution reconstruction networks,the method proposed achieved the best results,with PSNR improvements of 2.30 dB,2.23 dB,and 0.25 dB and SSIM improvements of 0.1316,0.1085,and 0.0096,respectively.The proposed method is helpful to improve the accuracy of remote sensing image target recognition and land classification applications from a data perspective.Additionally,it can further enhance the application efficiency of remote sensing data in fields such as resource investigation,environmental monitoring,and disaste.
作者
李薇
杜东升
邓剑波
陈良宇
LI Wei;DU Dong-sheng;DENG Jian-bo;CHEN Liang-yu(Hunan Institute of Meteorological Sciences,Changsha 410118,China;Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction,Changsha 410118,China;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China)
出处
《科学技术与工程》
北大核心
2023年第27期11513-11521,共9页
Science Technology and Engineering
基金
基于湖南省气象局2022年重点科研课题(XQKJ22A001)。
关键词
遥感影像
超分辨率重建
卷积神经网络
多尺度
并联式
super-resolution reconstruction
convolution neural network
multi-scale
parallel
remote sensing images
