测绘科学技术学报2025，Vol.41Issue(6)：597-603,7.DOI:10.3969/j.issn.1673-6338.2025.06.007
一种结合多尺度特征提取与注意力聚合的遥感影像道路提取方法
A Road Extraction Method for Remote Sensing Images Combining Multi-scale Feature Extraction and Attention Aggregation
摘要
Abstract
In response to the problem of poor road extraction performance in remote sensing images due to the com-plex and diversity background features,an improved remote sensing image road extraction method based on U-Net combining multi-scale feature extraction and attention aggregation(MSUNet)was proposed.In this method,the pyramid feature aggregation operation is used to fuse the multi-layer features of the encoder and preserve the spatial and semantic information of roads.Then,multi-scale feature extraction and attention aggregation based on strip convolution are introduced to solve the problem of pixel feature loss in dilated convolution.It strengthens the linear feature extraction and enriches the contextual information of roads.On the Massachusetts Roads dataset,its preci-sion,recall,F1,and IoU have increased by 0.73％,1.44％,1.09％and 1.46％compared to U-Net,respectively.On the autonomously annotated Jilin-1 Optical-A roads dataset,its four indicators have increased by 3.97％,1.82％,2.85％and 3.96％compared to U-Net,respectively.The experimental results show that MSUNet is superi-or to the compared methods,which can effectively improve the accuracy of road extraction.关键词
遥感影像/道路提取/多尺度特征/注意力聚合/条形卷积Key words
remote sensing images/road extraction/multi-scale feature/attention aggregation/strip convolution分类
天文与地球科学引用本文复制引用
范禹静,刘继东,靳国旺,杨鹤..一种结合多尺度特征提取与注意力聚合的遥感影像道路提取方法[J].测绘科学技术学报,2025,41(6):597-603,7.基金项目
国家重点研发计划项目(2023YFB2604001) （2023YFB2604001）
2022年度河南省重大科技专项项目(221100210600) （221100210600）
2022年度河南省交通运输科研资金计划项目(2022-3-2). （2022-3-2）
