基于深度学习融合OBIA的黄土高原小流域淤地坝系提取  

Extraction of check dam system in small watershed of Loess Plateau based on deep learning and OBIA

作　　者：钱伟 王春[2,3] 代文 卢旺达 李敏 陶宇 李梦琪 QIAN Wei;WANG Chun;DAI Wen;LU Wangda;LI Min;TAO Yu;LI Mengqi(School of Remote Sensing and Surveying Engineering,Nanjing University of Information Science and Technology,Nanjing 211800,Jiangsu,China;School of Geographic Information and Tourism,Chuzhou University,Chuzhou 239000,Anhui,China;Anhui Provincial Key Laboratory of Real Scene Simulation and Environmental Display,Chuzhou University,Chuzhou 239000,Anhui,China;School of Geography Science,Nanjing University of Information Science and Technology,Nanjing 211800,Jiangsu,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,江苏南京211800 [2]滁州学院地理信息与旅游学院,安徽滁州239000 [3]实景地理环境安徽省重点实验室,安徽滁州239000 [4]南京信息工程大学地理科学学院,江苏南京211800

出　　处：《干旱区地理》2023年第11期1803-1812,共10页Arid Land Geography

基　　金：安徽高校省级自然科学研究重大项目(KJ2021ZD0130);实景地理环境智能科技滁州市“113”产业创新团队;江苏省高等学校自然科学研究项目(22KJB170016);实景地理环境安徽省重点实验室开放课题(2022PGE013);安徽省高等学校科研计划项目(2022AH010066)资助。

摘　　要：淤地坝对于防治黄土高原水土流失有不可替代的作用,因此精确提取淤地范围和淤地坝点位对研究黄土高原水土有重要意义。现有图像分类方法中缺乏对淤地坝地形特征的考虑,容易被误判为梯田或土堆。除此之外,自动提取研究多集中于淤地范围提取,淤地坝点位仍依赖人工判读。因此,提出一种自动提取淤地坝系的方法:通过深度学习融合面向对象的影像分析(OBIA)方法提取韭园沟流域淤地范围,再利用水文分析方法提取淤地坝点位。结果表明:本方法提取的淤地范围精准率、召回率、F1Score分别为81.97%、90.94%、89.70%,F1Score与仅使用OBIA方法相比提升了21.94%。淤地坝点位的自动识别准确率为81.08%,完整率为88.89%,与前人目视解译的准确度相近,并实现了淤地坝范围和淤地坝点位的全要素提取。研究结果可为黄土高原淤地坝空间布局优化和水土流失评估等分析提供重要基础数据。Check dams play an irreplaceable role in preventing and controlling soil erosion in the Loess Plateau of China.Accurately extracting check dam areas and their locations holds immense importance for soil and water conservation studies in the Loess Plateau of China.However,existing image classification methods lack consideration of the topographic features of check dams,leading to many misidentifications.This study proposes a method that combines object-based image analysis(OBIA)and deep learning to precisely identify check dam areas and their locations.First,a multiscale segmentation method based on an object-oriented approach is used to segment the multisource data.Deep learning is applied to identify check dams.The results from these processes were fused using majority voting to enhance the accuracy of check dam area extraction.Second,the river network within the basin is extracted based on digital elevation model data.Finally,the check dam areas and river network are intersected to determine the check dam locations.The results show that the precision and recall rates of this method are 81.97%and 90.94%,respectively,and an F1Score value of 89.70%,which is 21.94%higher than the methods using only OBIA.Additionally,the accuracy and completeness of automatically identified check dam locations are 81.08%and 88.89%,respectively.This method provides important data for the optimal spatial layout of check dams in the Loess Plateau and for the evaluation of soil and water loss.

关 键 词：淤地范围提取 淤地坝点位提取 面向对象的影像分析(OBIA) U-Net框架 黄土高原 

分 类 号：S157[农业科学—土壤学]