基于深度学习的无人机造林成活率自动检测方法  

Automatic detection method for survival rate of UAV forestation based on deep learning

作　　者：李亚东[1] 曹明兰[1] 金红玉 姚树军 李亚昆 贾树华 LI Yadong;CAO Minglan;JIN Hongyu;YAO Shujun;LI Yakun;JIA Shuhua(Beijing Key Laboratory of Urban Spatial Information Engineering,Beijing Polytechnic College,Beijing 100042,China;Forest Resources Management Department of National Forestry and Grassland Administration,Beijing 100013,China;Administrative Bureau of Inner Mongolia Ulanba National Nature Reserve,Chifeng 025450,Inner Mongolia,China;Inner Mongolia Dayangshu Forestry Bureau,Hulun Buir 165456,Inner Mongolia,China)

机构地区：[1]北京工业职业技术学院北京市城市空间信息工程重点实验室,北京100042 [2]国家林业和草原局森林资源管理司,北京100013 [3]内蒙古乌兰坝国家级自然保护区管理局,内蒙古赤峰025450 [4]内蒙古大杨树林业局,内蒙古呼伦贝尔165456

出　　处：《中南林业科技大学学报》2023年第10期20-27,79,共9页Journal of Central South University of Forestry & Technology

基　　金：国家自然科学基金项目(U1710123);北京工业职业技术学院重点课题(BGY2022KY-12Z,BGY2022KY-13Z,BGY2021KY-10);环境模拟与污染控制国家重点联合实验室开放基金资助项目(23K01ESPCW)。

摘　　要：【目的】造林是实现林业资源可持续利用的重要手段。造林地块分散、分布偏远、交通条件差,人工造林成活率检查耗费大量人力、物力和时间成本,且效率低、无法及时全面精准掌握造林成活率、采用的抽样检查方法受主观因素和偶然因素影响较大。为解决以上问题,本研究探索了一种基于目标检测模型和多目标追踪的无人机造林成活率自动检测方法。【方法】利用大疆Mavic2多旋翼无人机采集不同时期云杉人工造林地的影像,利用YOLOv5深度学习网络模型训练了苗木对象检测模型,并与DeepSort多目标追踪算法相结合,在安装Mavic2无人机地面站的安卓平板电脑上进行二次开发。通过整合无人机4G图传,将无人机拍摄的视频实时回传至无人机地面站安卓平板电脑上。利用YOLOv5s+DeepSort实现了实时苗木检测与动态多目标追踪,通过对跟踪到的苗木进行类别判定和计数汇总,最终自动计算目标小班的造林成活率。【结果】在内蒙古大兴安岭大杨树林业局乌鲁布铁林场2022年春季造林地块中随机抽取了7个造林小班进行试验,结果表明本研究与人工检查相比成活率检测平均相对精度能达到99.62%,可平均节省50.94%人力工时。【结论】文章采用基于目标检测模型和多目标追踪的无人机造林成活率自动检测方法,不仅效率高、精度可靠、容易操作、可以极大地减少人力工时。与传统人工检查方法相比具有以下优势:1)有效提高了检测效率,减少了人力、时间、交通成本,能够实现对所有造林地块的全面检测,减少了抽样检查方法导致的随机误差;2)实现利用基于视频检测的算法自动完成成活率统计,统一了检测标准,避免了人工检测带来的主观误差;3)在检测视频中标注每株苗木的判定类别,可作为过程资料保存输出,为复核提供了有效依据。【Objective】Forestation is a crucial method for realizing the sustainable development of forestry resources.However,forestation blocks are scattered and distributed in remote areas,and traffic conditions are poor.In general,the examination of the survival rate of forestation costs lots of manpower,material resources,and time,with low efficiency.Besides,the survival rate of forestation cannot be grasped comprehensively,accurately,and timely.The sampling inspection method adopted is significantly affected by subjective and contingent factors.In order to solve the above problems,this paper probed into an automatic detection method of the survival rate of UAV forestation based on the target detection model and multi-target tracking.【Method】The DJI Mavic2 multi-rotor UAV was used to collect images of spruce artificial forestation land at different periods,trained the seedling object detection model using the YOLOv5 deep learning network model,combined it with DeepSort multi-target tracking algorithm,and conducted secondary development on Android tablet computer of Mavic2 UAV ground station.By integrating UAV 4G image transmission,this study sent video captured by the UAV back to Android tablet computer of the UAV ground station in real-time,adopted YOLOv5s+DeepSort to realize real-time seedling detection and dynamic multi-target tracking,and automatically calculated the forestation survival rate of the target based on category judgment and counting summary of tracked seedlings.【Result】Seven small forestation classes were randomly selected from the 2022 spring forestation block in Wulu Butie Forestry Station of Dayangshu Forestry Bureau in Inner Mongolia.Compared with manual inspection,the mean relative accuracy of the survival rate detection by the method in this paper could reach 99.62%,saving an average of 50.94%of manpower and time.【Conclusion】The automatic detection method of the survival rate of UAV forestation based on the target detection model and multi-target tracking is not only efficient,reli

关 键 词：深度学习 无人机 目标检测 多目标追踪 造林成活率 

分 类 号：S758.51[农业科学—森林经理学] S771.51[农业科学—林学]