Individual tree segmentation using multi-temporal unmanned aerial vehicle (UAV) data

Remote‑sensing enabled phenotyping is driven by individual tree detection (ITD) and segmentation (ITS), which provide the essential basis for extracting individual tree structural and functional traits. Changing climate regimes require scalable and cost-effective methods, positioning Unmanned Aerial Vehicles (UAV) equipped with LiDAR as an effective tool to obtain repeated high-resolution measurements. Despite widespread ITS research, its accuracy and consistency across the multi-temporal nature of UAVs remain largely unexplored. This study evaluated four ITS algorithms, namely Dalponte2016, Silva2016, Li2012, and Watershed, applied to four years of UAV-LiDAR data (2022–2025) collected over a coastal Douglas-fir genetic trial site established in 2003 near Jordan River, Vancouver Island, British Columbia. Normalized point clouds with a density of approximately 100 points/m² were processed using the lidR package in R. Algorithm performance was assessed against a field-referenced dataset of 1,526 live trees using precision, recall, and F-score metrics. Silva2016 and Dalponte2016 achieved the highest overall performance, with peak F‑scores of 0.78 and 0.77, respectively, showing balanced and consistent tree detection across all four years. The Watershed algorithm exhibited the highest precision (0.91–0.92) but very low recall (0.30–0.36), consistently missing over 1,000 trees. Algorithm performance was influenced by more than two decades of unmanaged stand development, which reduced the structural regularity of the gridded trial plot and created canopy conditions more typical of complex natural forests. High tree mortality (~42%) further degraded performance by introducing dead trees and canopy gaps, particularly affecting the Watershed and Li2012 algorithms. These results suggest that Dalponte2016 and Silva2016 are more suitable for repeated surveys in structurally dense trial plots, highlighting the need to consider stand structure, development history, and mortality when selecting ITS methods for forest phenotyping.

基于遥感的表型分析依赖于单木检测（Individual Tree Detection, ITD）与单木分割（Individual Tree Segmentation, ITS）技术，二者是提取单木结构与功能性状的核心基础。不断变化的气候格局亟需可扩展且具成本效益的研究方法，而搭载激光雷达（LiDAR）的无人机（Unmanned Aerial Vehicles, UAV）正是获取重复高分辨率测量数据的高效工具。尽管单木分割（ITS）的相关研究已较为广泛，但针对无人机多时序场景下的算法精度与一致性，目前仍鲜有探索。本研究针对四种单木分割算法（即Dalponte2016、Silva2016、Li2012与Watershed）展开评估，所用数据为2022至2025年的无人机激光雷达（UAV-LiDAR）时序数据，采集自不列颠哥伦比亚省温哥华岛乔丹河附近一处建于2003年的海岸花旗松遗传试验林。本研究使用R语言中的lidR工具包对密度约为100点/平方米的归一化点云数据进行处理。以包含1526株活立木的野外参考数据集为基准，采用精准率、召回率与F1分数三种指标评估各算法的性能表现。Silva2016与Dalponte2016的综合性能最优，其最高F1分数分别达到0.78与0.77，在四年的观测周期中均展现出均衡且稳定的单木检测效果。Watershed算法的精准率最高（0.91~0.92），但召回率极低（0.30~0.36），始终漏检超过1000株树木。算法性能受到二十余年未经营林分发育的影响：该因素降低了网格化试验林的结构规整性，使得林冠条件更接近复杂天然林的状态。高达42%的树木死亡率进一步加剧了性能退化：死亡木与林冠空隙的出现干扰了算法运行，其中Watershed与Li2012算法受影响最为显著。本研究结果表明，Dalponte2016与Silva2016更适用于结构致密的试验林重复调查工作，同时也凸显了在选择森林表型分析所用的单木分割算法时，需综合考量林分结构、发育历史与树木死亡率等因素。