基于深度学习的遥感图像道路分割  

Road segmentation of remote sensing image based on deep learning

作　　者：李旭涛 杨寒玉 卢业飞 张玮[1,3] LI Xutao;YANG Hanyu;LU Yefei;ZHANG Wei(Faculty of Computer Science and Technology,Qilu University of Technology(Shandong Academy of Sciences),Jinan 250353,Shandong,China;School of Mathematics and Statistics,Hubei University,Wuhan 430062,Hubei,China;Shandong Provincial Key Laboratory of Computer Networks,Shandong Computer Science Center(National Supercomputer Center in Jinan),Jinan 250000,Shandong,China)

机构地区：[1]齐鲁工业大学(山东省科学院)计算机科学与技术学部,山东济南250353 [2]湖北大学数学与统计学学院,湖北武汉430062 [3]山东省计算中心(国家超级计算济南中心)山东省计算机网络重点实验室,山东济南250000

出　　处：《山东大学学报（工学版）》2022年第6期139-145,共7页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(61802233);山东省自然科学基金资助项目(ZR2021LZH001);山东省自然科学基金资助项目(ZR2020LZH010)。

摘　　要：为在遥感图像中提取出来道路信息,利用深度学习技术,引入U^(2)-Net模型进行遥感图像道路分割。相比于传统的道路提取方法,基于U^(2)-Net方法可以实现道路的自动化提取。为验证U^(2)-Net模型分割效果,选取U-Net、DeepLabV3+等近几年较为流行的语义分割方法进行对比试验,并进一步分析U^(2)-Net显著图融合模块中卷积核对道路提取效果的影响。试验结果表明,U^(2)-Net模型能较有效地提取出道路信息,模型在测试集上的平均交并比达到了76.49%,Kappa达到了0.701 2,分割精度优于U-Net、DeepLabV3+等语义分割方法。基于U^(2)-Net模型的深度学习方法可以用于解决遥感图像中的道路分割问题,并具有较好的分割效果。Deep learning technology was used and the U^(2)-Net model was introduced for road segmentation of remote sensing images.Compared with the traditional road extraction method,the method based on U^(2)-Net could realize automatic road extraction.Popular semantic segmentation methods in recent years such as U-Net,DeepLabV3+for comparative experiments were selected to verify the segmentation effect of U^(2)-Net model,and the influence of the convolution kernel in the U^(2)-Net saliency map fusion module on the road extraction effect was further analyzed.The experimental results showed that road information could be extracted by the U^(2)-Net model more effectively.The mean intersection over union of the model on the test set reached 76.49%,and the Kappa coefficient reached 0.701 2.The segmentation accuracy of the model was better than that of U-Net,DeepLabV3+ and other semantic segmentation methods.The deep learning method based on U^(2)-Net model could be used to solve the problem of road segmentation in remote sensing images,and had good segmentation effect.

关 键 词：语义分割 深度学习 遥感图像 U^(2)-Net模型 高分辨率 

分 类 号：TP363[自动化与计算机技术—计算机系统结构]