一种深度学习的无人机影像道路自动提取方法  

An automatic road extraction method from UAV imagery based on deep learning

作　　者：王晓霏 叶虎平 廖小罕 岳焕印 施冬[2] WANG Xiaofei;YE Huping;LIAO Xiaohan;YUE Huanyin;SHI Dong(State Key Laboratory of Resource and Environmental Information Systems,Institute of Geographic Sciences and Natural Resources Research,CAS,Beijing 100101,China;School of Geosciences,Yangtze University,Wuhan 430100,China;Institute of UAV Application Research,Tianjin and CAS,Tianjin 301800,China;The Research Center for UAV Applications and Regulation,CAS,Beijing 100101,China)

机构地区：[1]中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室,北京100101 [2]长江大学地球科学学院,武汉430100 [3]天津中科无人机应用研究院,天津301800 [4]中国科学院无人机应用与管控研究中心,北京100101

出　　处：《测绘科学》2021年第11期106-113,共8页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41771388,41971359);中国科学院重点部署项目(ZDRW-KT-2020-2-1);天津科技计划项目智能制造专项(Tianjin-IMP-2018-2)。

摘　　要：针对无人机影像道路提取自动化程度低、道路信息不完整及道路交叉口不连通等问题,该文提出了一种结合拓扑结构和全局上下文感知的无人机影像道路提取方法,通过构建一种编码/解码模式的深度学习方法实现自动化提取。在网络模型中,设计了聚合特征模块及增强型扩张卷积模块以获取更多的道路信息,并引入拓扑感知损失函数以保证道路的连通性,实现道路拓扑结构特性的反演。实验结果表明,基于改进后的网络模型对道路信息的提取效果较好,在无人机影像测试集上的准确率、召回率、F1得分和交并比(IoU)分别达到了89.07%、84.74%、86.86%和72.45%;在马萨诸塞州道路公共影像集通用性测试中,提取原始遥感图像的道路信息也表现了出色的提取性能。For the problem of the low degree of automation of road extraction from UAV images,incomplete road information and disconnection of road intersections,this paper designs a road extraction method from UAV images that combines topology and global a UAV image road extraction method combining topological structure and global context-aware.This method constructs a deep learning method of encoding/decoding mode to realize automatic extraction.In the network model,we designed an aggregate feature module and an enhanced expanded convolution module to obtain more road information.Considering the topological structure of the road,a topology-aware loss function is introduced to ensure the connectivity of the road.The experimental results show that the extraction effect of road information based on the improved network model in this paper is better.Respectively,the accuracy,recall,F1-score and IoU on the UAV image set reached 89.07%、84.74%、86.86% and 72.45%.In addition,in the general test of the public road image set in Massachusetts,the road information extracted from the original remote sensing image also showed excellent extraction performance.

关 键 词：拓扑结构 全局上下文 无人机 道路提取 聚合特征 

分 类 号：P237[天文地球—摄影测量与遥感]