空洞卷积U-Net的遥感影像道路提取方法  

Road extraction from remote sensing images based on dilated convolutions U-Net

作　　者：林娜[1,2] 张小青 王岚 冯丽蓉 王伟 LIN Na;ZHANG Xiaoqing;WANG Lan;FENG Lirong;WANG Wei(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing Geomatics and Remote Sensing Center,Chongqing 401147,China)

机构地区：[1]重庆交通大学土木工程学院,重庆400074 [2]重庆市地理信息和遥感应用中心,重庆401147

出　　处：《测绘科学》2021年第9期109-114,156,共7页Science of Surveying and Mapping

基　　金：重庆市教委科技项目(KJQN201800747);重庆交通大学研究生教育创新基金资助项目(2020S0001);重庆交通大学教育教学改革项目(1903015)。

摘　　要：针对从遥感影像上提取道路出现的细节特征丢失、提取结果模糊的问题,该文提出了一种基于空洞卷积U-Net的遥感影像道路提取算法:(1)以U-Net为基础网络,将低层细节特征与高层语义特征进行多特征融合,更好地还原道路目标细节;(2)为了进一步提高网络对道路细节特征的识别能力,在U-Net中引入空洞卷积模块,学习更多的语义信息来改善提取结果出现的模糊问题。在Massachusetts roads和高分辨率城市道路影像Cheng roads dataset数据集下的实验结果表明,在召回率、精度和F1-score指标分别达到了82.5%、86.7%、84.5%;93.2%、92.1%、92.6%。与基础的U-Net相比,该算法在解决细节特征丢失和提取结果模糊问题方面更具有应用价值。Aiming at the problems of loss of detail features and blurry extraction results of road extraction from remote sensing images,a road extraction algorithm for remote sensing images based on dilated convolution U-Net was proposed.Firstly,based on the U-Net,the low-level detail features and the high-level semantic features were multi-featured to better restore the road target details.Secondly,in order to further improve the network’s ability to recognize detailed features of roads,a dilated convolution module was introduced in the U-Net to learn more semantic information to improve the blurring of the extraction results.The experimental results using the Massachusetts roads and high-resolution urban road image Cheng roads dataset showed that the recall rate,accuracy and F1-score index reached 82.5%,86.7%,84.5%;93.2%,92.1%,92.6%,respectively.Compared with the basic U-Net,the algorithm had more application value in solving the problem of loss of detailed features and blurring of extraction results.

关 键 词：遥感影像 道路提取 空洞卷积 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]