基于深度学习语义分割模型的喀斯特山区乔木林地识别  

Arbor Forest Identification of Karst Mountainous Areas Based on Deep Learning Semantic Segmentation Model

作　　者：乔宝 黄兆志 田秋东 向娟 邓帮健 常新辉 张本国 安旭琴 QIAO Bao;HUANG Zhaozhi;TIAN Qiudong;XIANG Juan;DENG Bangjian;CHANG Xinhui;ZHANG Benguo;An Xuqin(The Second Surveying and Mapping Institute of Guizhou Province,Guiyang 550004,China;School of Geography and Environmental Science,Guizhou Normal University,Guiyang 550025,China)

机构地区：[1]贵州省第二测绘院,贵阳550004 [2]贵州师范大学地理与环境科学学院,贵阳550025

出　　处：《河南科学》2025年第2期200-206,共7页Henan Science

基　　金：贵州省科技支撑项目(黔科合支撑[2023]一般175)。

摘　　要：深度学习技术的发展为地物识别提供了全新的思路和方法,准确地识别林地类型,能为喀斯特山区不同植被利用类型环境下,提供更好的管理和监测区域乔木林地条件,对生态监测、植被管理和环境保护具有重要意义。本研究基于ArcGIS Pro平台,以贵州省铜仁市思南县数据为基础,采用DeepLabV3和U-Net模型分别与ResNet-50和ResNet-34骨干模型结合对思南县乔木林地进行自动化识别,开发一种高效,准确的贵州喀斯特山区乔木林识别与提取方法。研究结果表明:高密度乔木林地在影像中呈现出密集的树冠,整体郁闭度较高;中密度乔木林地树冠间有一定的间隔,整体树木分布不集中;低密度乔木林地则表现为树冠之间有较大的空隙,树木分布较为稀疏。DeepLabV3相对于U-Net在这组数据中表现更优秀,特别是结合ResNet-50的DeepLabV3模型的精度、F1值和召回率分别达到了89.8%,90.0%和90.0%。本研究所提出的方法具有较好的泛化能力,可推广应用于其他类似生态环境的植被识别任务中,为区域资源管理和环境保护提供技术支持。The development of deep learning technology has provided a new approach and method for land cover recognition,which can accurately identify forest types.This can provide better management and monitoring of the conditions of the arbor forest in karst mountainous areas under different vegetation utilization types,and has significant implications for ecological monitoring,vegetation management,and environmental protection.Based on the ArcGIS Pro platform and the data of Si’nan County,Tongren City,Guizhou Provence,this study uses the DeepLabV3 and U-Net models combined with the ResNet-50 and ResNet-34 backbone models respectively to automatically identify arbor forests in Si’nan County.A highly efficient and accurate method for identifying and extracting arbor forests in the karst mountainous areas of Guizhou Province is developed.The research results show that high-density arbor forests in the image appear with dense canopies,with a high overall canopy closure;the tree canopies of medium density arbor forests have some gaps,and the trees are not concentrated in a specific area;low density arbor forests have larger gaps between tree canopies,and the trees distribute sparsely.DeepLabV3 performs better than U-Net in this data set,especially the DeepLabV3 model combined with ResNet-50 in its precision,F1 value and recall,reaching 89.8%,90.0%and 90.0%,respectively.The method proposed in this study has good generalization ability and can be applied to other similar ecological environment vegetation recognition tasks,providing technical support for regional resource management and environmental protection.

关 键 词：语义分割 喀斯特山区 深度学习 乔木林地 遥感识别 

分 类 号：X173[环境科学与工程—环境科学] TP751[自动化与计算机技术—检测技术与自动化装置]