1.The School of Technology,Beijing Forestry University,Beijing 100083,China
2.College of Forestry,Northwest A&F University,Yangling 712100,China
3.School of Landscape Architecture,Beijing Forestry University,Beijing 100083,China
4.College of Forestry,Beijing Forestry University,Beijing 100083,China
5.Key Laboratory of Forestry Equipment and Automation of the National Forestry and Grassland Administration,Beijing 100083,China
Show less
文章历史+
Received
Published
2026-01-05
2026-07-20
Issue Date
2026-09-24
PDF (6687K)
摘要
针对毛白杨人工林点云语义分割中枝叶遮挡严重、点云重叠复杂、传统基于PointNet++的分割方法存在训练时间长、分割精度不高的问题,提出一种结合超点聚类(superpoint cluster,SC)分割网络与PointNet++网络的改进方法(SC-PointNet++),以提高分割效率与准确性。选取山东省聊城市清平镇国家生态公园的毛白杨人工林样地,使用华测导航RS10手持激光雷达在落叶前期完成扫描。针对原始数据中的噪声和离群点,依次进行重采样、去噪、地面点分离与坐标归一化等预处理,并划分训练集、验证集和测试集。借助超点聚类分割网络计算点云的局部几何特性(线性度、平面度、散射度、垂直度)与辐射特性红绿蓝三原色(RGB)强度,再通过分割追踪算法(Cut-Pursuit)求解“能量最小化”问题将几何、辐射特征相似的点划分为一个超点。用一个多层感知机(multilayer perceptron,MLP)分析其聚合后的特征(比如超点整体的线性度、RGB均值),输出该超点属于“枝干、树叶、地面、其他”各部分的概率,相邻的超点用另一个MLP分析其特征相似度,输出“一致性分数”,高则合并,低则切断。在此基础上,将分割后形成的超点簇作为PointNet++的输入,利用其分层集体抽象模块(Set Abstraction)与特征传播模块(Feature Propagation)进行多尺度特征提取与语义分割。最后将单木分割后建立树木定量结构模型(tree quantitative structure model,TreeQSM)进行林业参数提取。结果表明,SC-PointNet++在测试集上实现了总体准确率(overall accuracy,OA)为0.94,比原始PointNet++提升11.9%;总体分割精度达到0.887,较PointNet++提高5.3%;树叶平均交并比(intersection over union,IoU)达0.874,地面IoU达0.987,枝干IoU也由0.753提升至0.773;模型训练时间较PointNet++缩短约10%。归一化混淆矩阵分析显示,除少数枝干被误分割为树叶外,各部分分割结果与人工标注结果高度一致。将TreeQSM模型提取出的单木参数与人工测量参数建立散点回归图,并计算相关系数(R²)和均方根误差(root mean square deviation,RMSE),研究表明所提SC-PointNet++方法通过超点聚类分割网络有效减少输入规模、增强局部语义一致性,结合PointNet++网络深度特征学习,实现毛白杨人工林点云的高效、精准分割,满足后续胸径、树高等林业参数的提取需要。
Abstract
In view of the severe branch and leaf occlusion and complex point cloud overlap in the semantic segmentation of Populus tomentosa plantations, and the problems of long training time and low segmentation accuracy in traditional PointNet++-based segmentation methods, this paper proposed an improved method (SC-PointNet++) that combines the SuperpointCluster segmentation network with the PointNet++ network to enhance the efficiency and accuracy of segmentation. Firstly, a sample plot of Populus tomentosa plantation in Qingping Town, Liaocheng City, Shandong Province was selected, and the handheld laser scanner RS10 of Huace Navigation was used to complete the scanning before leaf fall. For the noise and outliers in the original data, preprocessing steps such as resampling, denoising, ground point separation, and coordinate normalization were carried out successively, and the training set, validation set, and test set were divided. The local geometric characteristics (linearity, planarity, scattering, and verticality) extracted by the PCA operator and radiometric characteristics red, green and blue primary colors (RGB) intensity of the point cloud were calculated using the SuperpointCluster segmentation network, and the Cut-Pursuit algorithm was used to solve the ‘energy minimization’ problem to group points with similar geometric and radiometric features into a superpoint. A multilayer perceptron (MLP) was used to analyze the aggregated features of the superpoint (such as the overall linearity and RGB mean of the superpoint), and output the probability that the superpoint belonged to ‘trunk, leaf, ground, other’ parts. The feature similarity of adjacent superpoints was analyzed by another MLP, and the ‘consistency score’ was output. If the score was high, the superpoints were merged; if it was low, they were separated. On this basis, the superpoint clusters formed after segmentation were used as the input of PointNet++, and the multi-scale feature extraction and semantic segmentation were carried out by using its hierarchical Set Abstraction and Feature Propagation modules. Finally, the TreeQSM model was established for each tree after single-tree segmentation to extract forestry parameters. SC-PointNet++ achieved an overall accuracy (OA) of 0.94 on the test set, an increase of 11.9% compared to the original PointNet++; the overall segmentation accuracy reached 0.887, an improvement of 5.3% compared to PointNet++; the average intersection over union (IoU) of leaves was 0.874, the IoU of the ground was 0.987, and the IoU of the trunk increased from 0.753 to 0.773; the model training time was shortened by about 10% compared to PointNet++. The normalized confusion matrix analysis showed that except for a few trunks being misclassified as leaves, the segmentation results of each part were highly consistent with the manual annotations. Scatter regression plots were established between the single-tree parameters extracted by the TreeQSM model and the manually measured parameters, and the correlation coefficient (R²) and root mean square error (RMSE) were calculated. The SC-PointNet++ method proposed in this paper effectively reduced the input scale and enhanced the local semantic consistency through the SuperpointCluster segmentation network, and combined the PointNet++ network for deep feature learning, achieving efficient and accurate semantic segmentation of Populus tomentosa plantation point clouds, meeting the needs of subsequent extraction of forestry parameters such as diameter at breast height and tree height.
XUM B, ZHONGX T, ZHONGR F.A method for automatic extraction and individual segmentation of urban street trees from laser point clouds[J].Optics & Laser Technology,2025,180:111431.
[2]
WANGP C, TANGY, LIAOZ F.Road-side individual tree segmentation from urban MLS point clouds using metric learning[J].Remote Sensing,2023,15(8):1992.
GUANT S, WANGC B, WANGW,et al.Path planning method of mobile LiDAR in plantation sample plot survey[J].Journal of Beijing Forestry University,2024,46(5):154-162.
CAOW, CHEND, SHIY F,et al.Progress and prospect of LiDAR point clouds to 3D tree models[J].Geomatics and Information Science of Wuhan University,2021,46(2):203-220.
[7]
NINGX J, MAY S, HOUY Y,et al.Semantic segmentation guided coarse-to-fine detection of individual trees from MLS point clouds based on treetop points extraction and radius expansion[J].Remote Sensing,2022,14(19):4926.
[8]
SUNJ Q, WANGP, GAOZ Y,et al.Wood–leaf classification of tree point cloud based on intensity and geometric information[J].Remote Sensing,2021,13(20):4050.
[9]
YANGJ T, KANGZ Z, CHENGS,et al.An individual tree segmentation method based on watershed algorithm and three-dimensional spatial distribution analysis from airborne LiDAR point clouds[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,2020,13:1055-1067.
SHIJ C.Study on optimization method of large diameter grade Poplar clones in Guanzhong industry based on airborne LiDAR data[D].Xianyang:Northwest A&F University,2023.
[12]
CHARLESR Q, SUH, KAICHUNM,et al.PointNet:Deep learning on point sets for 3D classification and segmentation[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),July 21—26,2017,Honolulu,HI,USA,IEEE:77-85.
[13]
CHARLESR Q, LIY, SUH,et al.PointNet++:Deep hierarchical feature learning on point sets in a metric space[J].arXiv:1706.02413 [cs.CV].
[14]
SHENX Y, HUANGQ Q, WANGX,et al.A deep learning-based method for extracting standing wood feature parameters from terrestrial laser scanning point clouds of artificially planted forest[J].Remote Sensing.2022,14(15):3842.
[15]
RUOPPAL, OINONENO, TAHERJ,et al.Unsupervised deep learning for semantic segmentation of multispectral LiDAR forest point clouds[J].ISPRS Journal of Photogrammetry and Remote Sensing,2025,228:694-722.
[16]
DAMIENR, HUGOR, LOICL.Scalable 3D panoptic segmentation as superpoint graph clustering[J].arXiv:2401.06704 [cs.CV].
[17]
LIJ, LIUJ H, HUANGQ Q.PointDMM:A deep-learning-based semantic segmentation method for point clouds in complex forest environments[J].Forests2023,14(12):2276.
XIAJ Z, SUNH M, HUS H,et al.3D laser point cloud clustering method based on image information constraints[J].Opto-Electronic Engineering,2023,50(2):220148.
[20]
ZHAOX Q, GUOQ H, SUY J,et al.Improved progressive TIN densification filtering algorithm for airborne LiDAR data in forested areas[J].ISPRS Journal of Photogrammetry and Remote Sensing,2016,117:79-91.
[21]
LANDRIEUL, OBOZINSKIG.Cut pursuit:Fast algorithms to learn piecewise constant functions on general weighted graphs[J].SIAM Journal on Imaging Sciences,2017,10(4):1724-1766.
[22]
JONATHONS.A tutorial on principal component analysis[J].International Journal of Remote Sensing,2014,51(2):1684.
[23]
HUGOR, LOÏCL.Cut-pursuit algorithm for regularizing nonsmooth functionals with graph total variation[J].Proceedings of the 35th International Conference on Machine Learning,2018,80:4247-4256.
XIED B, LIUQ W, LEIY K,et al.Suitability analysis of single tree segmentation algorithm in the core area of Winter Olympic Games based on airborne LiDAR data[J].Scientia Silvae Sinicae,2022,58(10):121-130.
[26]
WANGD, MOMOT S, CASELLAE.LeWoS:A universal leaf-wood classification method to facilitate the 3D modelling of large tropical trees using terrestrial LiDAR[J].Methods in Ecology and Evolution,2020,11(3):376-389.
[27]
SHENX Y, HUANGQ Q, WANGX,et al.A method for extracting the tree feature parameters of Populus tomentosa in the leafy stage[J].Forests,2023,14(9):1757.
HUOL N, ZHANGX L.Individual tree information extraction and accuracy evaluation based on airborne LiDAR point cloud by multilayer clustering method[J].Scientia Silvae Sinicae,2021,57(1):85-94.
WUY H, WANGP, LIY X,et al.Performance analysis of tree parameter estimation based on TreeQSM modeling of actual and model trees[J].Chinese Journal of Lasers,2023,50(22):165-174.