Abstract
Poplars are essential to China's forestry, contributing to timber production, ecological restoration, and shelterbelt construction. Branch architecture critically influences tree growth, demanding scalable solutions beyond manual methods to assess phenotypic variation in large-scale poplar breeding programs. Unmanned aerial vehicle light detection and ranging (UAV LiDAR) provides an efficient alternative; however, existing methods focus on conifers, leaving a gap in approaches for the more complex morphology of poplar branches. This study proposes a poplar branch reconstruction algorithm utilizing material transport flux and object-level geometric features from low-cost UAV LiDAR data. First, a voxel-based near-centroid method is used to extract skeleton points from tree point clouds. Next, a material transport flux model identifies individual branches, and geometric features of transport paths, including path length and curvature, are used to reconstruct each branch. Finally, branch parameters are estimated based on reconstructed branches. Data from a 5-ha plot were collected using the DJI Zenmuse L1 UAV LiDAR at the Shishou National Poplar Breeding Station, Hubei Province, China. Results demonstrate the proposed algorithm achieves high accuracy in first-order branch identification (F1-score = 1), with second-order branches having an average F1-score of 0.69. Branch length estimation demonstrates an RMSE of 0.47 m, while branch angles show an RMSE of 7.06°. The study also reveals structural variability in branch traits, with the highest variability observed in the second-order branch length (coefficient of variation = 29.68%), and a moderate positive correlation between first- and second-order branch lengths (correlation coefficient = 0.34), providing insights into tree growth patterns. This approach offers a framework for high-throughput phenotyping, which provides an efficient solution towrads advanced tree breeding using UAV LiDAR.