文献类型：期刊文献

中文题名：基于UAV-LiDAR的落叶松子代测定林单木分割方法研究

英文题名：Individual Tree Segmentation Methods for Larch Progeny Test Forest Based on UAV-LiDAR

作者：蔡天润[1] 孙晓梅[1] 陈东升[1] 谢允慧[1]

第一作者：蔡天润

机构：[1]林木遗传育种全国重点实验室,国家林业和草原局林木培育重点实验室,中国林业科学研究院林业研究所,北京100091

年份：2025

卷号：38

期号：6

起止页码：33-47

中文期刊名：林业科学研究

外文期刊名：Forest Research

收录：;北大核心:【北大核心2023】；

基金：中央级公益性科研院所基本科研业务费专项资金(LYSZX202002,CAFYBB2022ZC001);“十四五”国家重点研发计划课题“林木优异种质维持的遗传基础”(2022YFD2200103)。

语种：中文

中文关键词：无人机;激光雷达;单木分割;育种;子代测定林;日本落叶松;脉冲重复频率

外文关键词：UAV;LiDAR;individual tree segmentation;breeding;progeny test forest;Larix kaempferi(Lamb.)Carrière;pulse repetition frequency

分类号：S722.3;S771.8

摘要：[目的]探究适用于子代测定林的激光脉冲重复频率及单木分割算法,为高郁闭度且地形复杂的育种试验林开展基于无人机激光雷达的长期表型监测提供技术参考。[方法]以辽宁大孤家林场37年生日本落叶松家系子代测定林为研究对象,基于380 kHz顺逆融合(380)、550 kHz顺逆融合(550),及2种激光脉冲重复频率融合(380550),获得三组融合点云数据。利用来自真实定位点的种子点结合基于点云距离判别聚类算法(PCS),以及来自3种冠层高度模型的种子点结合标记控制分水岭分割算法(MWS)、区域增长算法(SRG)和PCS,形成了30种单木分割组合。根据单木分割精度及树高提取精度,选择合适的激光脉冲重复频率及单木分割算法,并结合坡度信息分析未正确匹配单木特征。[结果]基于380、380550和550数据分别进行单木分割:利用真实定位点种子点的PCS算法F值分别为0.96、0.96、0.93;利用9种来自冠层高度模型种子点的算法,F均值分别为0.80、0.79、0.74;380 kHz为最适合的激光脉冲重复频率。以380数据进行单木分割的结果显示,利用真实定位点的PCS算法单木分割效果最佳,其F值为0.96,树高提取精度R^(2)为0.84;没有真实定位点的情况下,利用克里金插值和反距离权重方法生成的冠层高度模型探测的种子点,结合MWS和PCS算法的单木分割效果良好,其F值在0.82~0.83,R^(2)在0.83~0.84。未正确匹配的单木主要分布在陡坡和险坡区域,多为其半径4 m内的被压木。[结论]本文获得了5种适用于高郁闭度子代测定林且分割效果较好的无人机激光雷达单木分割方法,最佳分割方法是基于真实定位点的PCS算法,可提高表型性状调查效率,满足育种计划长期表型监测的需求。
[Objective]To identify optimal laser pulse repetition frequencies and individual tree segmentation algorithms for progeny test forests,providing technical references for long-term phenotypic monitoring of breeding trial forests characterized by high canopy density and complex terrain based on UAV LiDAR.[Method]Based on a 37-year-old Japanese larch progeny test forest in Dagujia Forest Farm,Liaoning,three sets of fused point cloud data were obtained through 380 kHz forward and reverse fusion(380),550 kHz forward and reverse fusion(550),and the fusion of two laser pulse repetition frequencies(380550).By combining seed points derived from real locations with the Point Cloud-based Cluster Segmentation(PCS),and seed points from three Canopy Height Models(CHMs)with the Marker-Controlled Watershed Segmentation(MWS),the Seeded Region Growing(SRG),and PCS,a total of 30 single tree segmentation combinations were generated.The appropriate laser pulse repetition frequency and individual tree segmentation algorithm were selected based on the accuracy of individual tree segmentation and tree height extraction,and the characteristics of unmatched individual trees were analyzed in conjunction with slope information.[Results]The PCS algorithm using ground-truth seeds achieved superior performance across all datasets(F-scores:380-PRF=0.96,380550-PRF=0.96,550-PRF=0.93),outperforming CHMbased approaches(mean F-scores:0.80,0.79,0.74 respectively);380-FR was identified as the optimal PRF configuration,yielding 0.96 segmentation F-score and 0.84 R^(2)height accuracy with PCS;MWS and PCS using Kriging/IDW CHM seeds maintained robust performance(F=0.82-0.83,R^(2)=0.83-0.84)without ground truth.Slope analysis revealed 67%of mismatches occurred on slopes>25°,predominantly suppressed trees within 4m radius zones.[Conclusion]This study identifies five UAV-LiDAR individual tree segmentation methods that are effective for high canopy density progeny test forests.The most effective segmentation method utilizes the PCS algorithm with real location points,enhancing the efficiency of phenotypic trait surveys and supporting the long-term monitoring needs of tree breeding.