基于U-ConvHDNet模型的戈壁砾幕层提取  

Extraction of gobi desert gravel layer based on U-ConvHDNet

作　　者：马于博 张爱国 王浩宇 刘帅琪 靳镜宇 沈占锋 李均力[1,4] MA Yubo;ZHANG Aiguo;WANG Haoyu;LIU Shuaiqi;JIN Jingyu;SHEN Zhanfeng;LI Junli(Key Laboratory of Ecological Safety and Sustainable Development in Arid Lands,Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,Urumqi 830011,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China;Nanjing Institute of Environmental Sciences,Ministry of Ecology and Environment of the People s Republic of China,Nanjing 210042,China;Key Laboratory of GIS&RS Application Xinjiang Uygur Autonomous Region,Urumqi 830011,China)

机构地区：[1]中国科学院新疆生态与地理研究所干旱区生态安全与可持续发展全国重点实验室,新疆乌鲁木齐830011 [2]中国科学院空天信息创新研究院,北京100101 [3]生态环境部南京环境科学研究所,江苏南京210042 [4]新疆遥感与地理信息系统应用重点实验室,新疆乌鲁木齐830011

出　　处：《测绘通报》2025年第4期68-74,共7页Bulletin of Surveying and Mapping

基　　金：哈密市戈壁生态调查评价与保护修复(202306160842);新疆天山科技创新团队(2022TSYCTD0006);第三次新疆综合科学考察(2021xjkk1403)。

摘　　要：砾幕层是戈壁生态系统的重要组成部分,大尺度的砾幕层遥感监测对戈壁生态系统保护具有重要意义。针对砾幕层结构松散、异质性强的特点,本文提出了一种基于U-ConvHDNet语义分割的砾幕层自动信息制图方法,利用2023年8月的哈密全区域的Sentinel-2影像提取戈壁砾幕层信息。结果表明,U-ConvHDNet模型的F1分数为0.918,优于参与对比的7个主流语义分割模型,消融试验表明骨架网络的改进与上下采样模块的联合使用有效提升了精度。双重感受野滑窗策略优化了拼接线附近不稳定的现象,提取出哈密戈壁砾幕层总面积为1.026×105 km 2,其信息提取精度的F1分数为0.921。本文研究可为戈壁砾幕层的监测和戈壁生态系统治理提供技术支撑。The gravel layer is an essential component of the gobi desert ecosystem.Conducting large-scale remote sensing monitoring of the gravel layer is of great significance for protecting the gobi desert ecology.In response to the loose structure and strong heterogeneity of the gravel layer,this paper proposes an automatic information mapping method for the gravel layer based on the U-ConvHDNet semantic segmentation model.This method utilizes Sentinel-2 imagery from the entire Hami region captured in August 2023 to extract information on gobi gravel layer.The results indicate that the F1 score of the U-ConvHDNet model is 0.918,which is superior to that of the other seven semantic segmentation models.Ablation experiments demonstrate that the combined use of the improved backbone network,upsampling and downsampling modules effectively enhances the accuracy.The dual receptive field sliding window strategy optimizes the instability near stitching lines,enabling the extraction of the total area of the gobi gravel layer in Hami at 1.026×105 km 2,with an information extraction precision of F1 score 0.921.This study provides technical support for monitoring of gobi gravel layer and the management of gobi ecosystems.

关 键 词：遥感影像 语义分割 深度学习 砾幕层 

分 类 号：P23[天文地球—摄影测量与遥感]