面向病虫害防控的茶苗分枝夹角三维提取方法

Three-dimensional Extraction Method for Angle between Branches of Tea Seedlings for Pest and Disease Control

摘要:

茶苗枝干三维点云的分枝夹角精确提取，是量化茶树冠层结构并揭示其如何通过改变微气候影响病虫害发生的关键前提。针对茶苗枝条细小、叶片遮挡明显、叶柄残留易干扰分枝节点识别与夹角测量等问题，本研究提出一种融合双视角视频三维重建、PointNet++枝叶语义分离、枝干点云净化增强和Tea-SCAG算法的茶苗枝干分枝夹角自动提取方法。以70株一年生"龙井43"茶苗为研究对象，基于转台式拍摄平台分别从平视和俯视45°方向采集环绕视频，并按固定时间间隔抽帧形成多视角图像序列;利用二维高斯泼溅方法完成茶苗三维重建，结合 PLANesT-3D公开数据集联合训练PointNet++网络，实现茶苗枝干与叶片的语义分离;针对语义分割后仍存在的叶柄残留和外围干扰点，采用离群点剔除方法对枝干点云进行削薄净化，并通过局部球域随机补点恢复枝干点云的连续性和局部厚度。在此基础上，通过主干贴近约束、上方双分枝验证、下方主干支撑判断和相近节点合并等策略，构建适用于茶苗细枝结构的Tea-SCAG算法，提高分枝节点定位与夹角计算稳定性。研究结果表明，PointNet++枝叶分离总体准确率为89.3%，平均交并比为83.7%;Tea-SCAG分枝节点检测召回率、精确率和F1值分别为82.3%、84.8%和83.4%，分枝夹角预测结果与人工测量结果相关系数为0.910。本方法可准确提取茶苗枝干三维结构、分枝节点及分枝夹角，为茶苗株型数字化表型分析、病虫害易发环境判断、精准修剪和抗病抗虫个体筛选提供技术支撑。

Abstract:

Accurate extraction of branch angles from three-dimensional point clouds is important for quantifying tea seedling canopy structure and analyzing pest- and disease-prone microenvironments.To address the problems of slender branches, severe leaf occlusion, and residual petioles interfering with branch node detection, an automatic branch angle extraction method integrating dual-view three-dimensional reconstruction, PointNet++ segmentation, point cloud purification and enhancement, and Tea-SCAG was proposed.Seventy one-year-old "Longjing 43" tea seedlings were used.Surrounding videos were collected from horizontal and 45° top-oblique views using a turntable platform, and multi-view image sequences were obtained by frame extraction.Tea seedlings were reconstructed using 2D Gaussian splatting.PointNet++ was jointly trained on the LJ-43 and PLANesT-3D datasets to separate stems from leaves.Residual petioles and peripheral noise were removed by statistical outlier filtering, while local spherical point supplementation was used to restore stem continuity and local thickness.Tea-SCAG was then developed by introducing main-stem proximity constraint, upper dual-branch validation, lower main-stem support judgment, adjacent node merging, and overlapping candidate merging.The overall accuracy and mean intersection over union of stem-leaf segmentation reached 89.3% and 83.7%, respectively.The recall, precision, and F-score of branch node detection were 82.3%, 84.8%, and 83.4%, respectively, and the correlation coefficient between predicted and manually measured branch angles was 0.910.The proposed method can provide reliable three-dimensional structural traits for tea seedling phenotyping, precision pruning, identification of pest- and disease-prone canopy structures, and early screening of potentially resistant seedlings.