一种改进的融合不同尺度特征的遥感影像道路提取新方法  

A new and improved method for road extraction from remote sensing images by fusing different scale features

作　　者：闫志恒 任超 李毅 徐宁辉 张胜国 YAN Zhiheng;REN Chao;LI Yi;XU Ninghui;ZHANG Shengguo(College Geomatics and Geoinformation,Guilin University of Technology,Guilin 541004,China;Guangxi Key Laboratory of Spatial Information and Geomatics,Guilin 541004,China;Guangxi Zhuang Autonomous Region Natural Resources and Real Estate Registration Centre,Nanning 530000,China;Nanning Survey and Mapping Geographic Information Institute,Nanning 530000,China)

机构地区：[1]桂林理工大学测绘地理信息学院,广西桂林541004 [2]广西空间信息与测绘重点实验室,广西桂林541004 [3]广西壮族自治区自然资源和不动产登记中心,广西南宁530000 [4]南宁勘察测绘地理信息院,广西南宁530000

出　　处：《测绘通报》2022年第9期58-62,共5页Bulletin of Surveying and Mapping

基　　金：国家自然科学基金(42064003)。

摘　　要：针对高分辨率遥感影像中细小道路纹理特征不明显、信息提取困难的问题,本文提出并实现了一种融合不同尺度特征的深度学习道路提取新方法。首先引入CoT模块构建残差网络,以充分利用局部与全局上下文信息提取不同尺度道路特征;然后构建特征金字塔注意力模块,融合不同层级道路特征信息;最后使用全局注意力上采样模块结合全局背景对道路细节进行恢复。试验结果表明,该方法的召回率、交并比均优于已有方法,能够较完整准确地提取遥感影像中的道路信息,提升道路提取效率。Aiming at the problem of inconspicuous fine road texture features and difficult information extraction in high-resolution remote sensing imagery,a new method of deep learning road extraction fusing different scale features is proposed and implemented.The method firstly introduces a CoT module to build a residual network to extract road features at different scales by making full use of local and global contextual information.Secondly,a feature pyramid attention module is built to fuse different levels of road feature information.Finally,a global attention upsampling module is used to recover road details in conjunction with the global context.The experimental results show that the proposed method is better than the existing methods in terms of recall and intersection ratio,and can extract the road information in remote sensing images more completely and accurately,which improves the efficiency of road extraction.

关 键 词：遥感影像 道路提取 CoT模块 特征金字塔 全局注意力 

分 类 号：P237[天文地球—摄影测量与遥感]