改进Deeplabv3+的高分辨率遥感影像道路提取模型  

An information extraction model of roads from high-resolution remote sensing images based on improved Deeplabv3+

作　　者：赵凌虎 袁希平 甘淑[1,2] 胡琳 丘鸣语 ZHAO Linghu;YUAN Xiping;GAN Shu;HU Lin;QIU Mingyu(School of Land and Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China;Surveying and Mapping Geo-Informatics Technology Research Center on Plateau Mountains of Yunnan Higher Education,Kunming 650093,China;School of Earth Science and Engineering,West Yunnan University of Applied Sciences,Dali 671000,China)

机构地区：[1]昆明理工大学国土资源工程学院,昆明650093 [2]云南省高校高原山区空间信息测绘技术应用工程研究中心,昆明650093 [3]滇西应用技术大学地球科学与工程学院,大理671000

出　　处：《自然资源遥感》2023年第1期107-114,共8页Remote Sensing for Natural Resources

基　　金：国家自然科学基金项目“滇中星云湖高原湖泊流域聚落空间格局演化研究”(编号:41561083);“东川小江泥石流迹地的多尺度遥感探测试验分析研究”(编号:41861054)共同资助。

摘　　要：针对传统的道路提取方法在高分辨率遥感影像中存在提取效果差和提取速度慢的问题,提出了改进Deeplabv3+的高分辨率遥感影像道路提取模型。采用MobileNetv2主干特征提取网络与Dice Loss函数相结合,较好地平衡了高分辨率遥感影像道路提取精度与速度的矛盾,实现较高提取精度的同时减少了模型参数,满足了时效性的要求。基于开源道路提取数据集的实验结果表明:①该文提出的道路提取模型在高分辨率遥感影像上具有可行性,提取道路的整体精度达到98.71%,具有较高的提取精度;②在提取道路的速度方面该方法平均帧数达到120.05,模型参数量仅为5.81 M,总体上比原模型更加轻量化,表明该方法满足了时效性的要求。该方法在大幅减少参数量、满足时效性的同时保证了提取的精确度,为提高基于高分辨率影像的道路提取精度和速度提供了一种新的改进思路和方法。Aiming at the problems of poor extraction effect and slow extraction speed of traditional road extraction methods in the information extraction of roads from high-resolution remote sensing images,this study proposed a new information extraction model based on improved Deeplabv3+.In the new model,the combination of the MobileNetv2 backbone feature extraction network with the Dice Loss function effectively balanced the contradiction between the precision and speed of road information extraction from high-resolution remote sensing images.As a result,high extraction precision was achieved while meeting timeliness requirements by reducing model parameters.The experimental results based on the open-source road information extraction dataset show that:①The road information extraction model proposed in this study was feasible for high-resolution remote sensing images,with overall accuracy of up to 98.71%;②In terms of the information extraction speed,the new model had an average frame number of 120.05 and parameter amount of only 5.81 M.Therefore,the new model was more lightweight lighter than original models,meeting the timeliness requirements.Therefore,the model proposed in this study meets the timeliness requirements by greatly reducing the parameter amount while ensuring high extraction accuracy.This study provides a new philosophy and method for improving the accuracy and speed of road information extraction from high-resolution images.

关 键 词：遥感影像 道路提取 深度学习 语义分割 DeepLabv3+模型 

分 类 号：P2[天文地球—测绘科学与技术]