摘要
树木的几何建模在林木性状评价、森林动态经营管理与可视化研究中具有重要意义。现今,从激光雷达(Light Detection And Ranging,LiDAR)数据中重建树体三维模型并精准获取林木空间枝干结构参数是数字林业发展的必然趋势。该研究提出了一种深度学习与计算机图形学相融合的树木骨架重建与参数反演方法。该方法以PR107、CATAS 7-20-59、CATAS 8-79三个品种的橡胶树为实验对象,首先,采用背包移动激光雷达获取三个橡胶树品种的样地数据,并通过体素剖分和数据增广策略来构建橡胶树训练样本集。其次,构造由四层特征编码层和特征解码层所组成的点云分类深度学习网络,并包含优化的PointConv模块与不同尺度的特征插值模块,以实现在多尺度条件下,全面考虑点云的全局和局部优化特征,引导网络实现枝叶点云的精确分类。最后,面向分类后的枝干点云,运用计算机图形学的空间连通性算法与圆柱拟合策略,重建树木骨架模型,并自动解决叶子点云与对应的一级枝干归属问题,进而在叶团簇尺度下开展对单株树的精细描述与参数反演。通过对三块橡胶树测试样地的验证和与实测值的比对表明,该研究提出的深度学习网络枝叶分类总体准确率在90.32%以上。骨架重建与叶团簇分析结果显示,PR107品种橡胶树具有较为发散的树冠、最大的分枝夹角和叶团簇体积;CATAS 7-20-59品种橡胶树冠呈花瓶型,分枝夹角和叶团簇体积较小;而CATAS 8-79品种橡胶树尽管胸径最粗,但不耐寒害处于落叶期导致冠积最小。同时,反演得到的橡胶树一级枝干直径与实测值比对为:决定系数R^(2)不低于0.94,均方根误差(Root Mean Square Error,RMSE)小于3.01 cm;主枝干与一级枝干的分枝角为:决定系数R^(2)不低于0.91,均方根误差RMSE不高于4.94°。同时发现橡胶树一级枝干的直径与对应的叶团簇体积呈正相关分布。该研究�
Fine-scale geometric modeling of actual trees can be an essential prerequisite for forest information and phenotypic characteristics at present. It is necessary to effectively map the tree architecture, and accurately retrieve the plant growth properties from light detection and ranging(LiDAR) data in a wide range of biophysical and ecological processes. Here, a synergistic approach was proposed for the skeleton modeling and property retrieval of the rubber tree using deep learning and computer graphics. The backpack laser scanning data was also collected for three typical rubber tree clones, i.e., PR107,CATAS 7-20-59, and CATAS 8-79. Firstly, the labeled wood and leaf point samples were collected using traditional machine learning assisted by manual segmentation. The acquirement strategy was coupled with the voxelization and data augmentation to achieve a suitable and larger training data set than before. Secondly, the deep learning network of reformulation architecture was developed to comprise four features encoding and decoding layers. The improved PointConv modules were embedded to calculate the local translation-invariant point features and the interpolation modules, further propagating the features from sub-sampled point clouds to a scale-up resolution. As such, both the global and local features of the forest points were comprehensively extracted to greatly reinforce the leaf-wood point cloud classification from the forest scene geometry. Thirdly,the tree skeletons were reconstructed for the spatial connectivity and cylinder fitting scheme using the extracted branch points after the classification of forest point clouds. Finally, the affiliation relationship between foliage points and the first-order branch points was determined to simplify each tree crown with as many foliage clumps and different branch compositions. The overall accuracy of the branch and leaf classification was 90.32% using the deep learning network during the field measurements of the test samples, including three rubber tree plots, wh
作者
丁竹娴
周立军
樊江川
安锋
陈帮乾
王铭慧
蒋玲
薛联凤
云挺
Ding Zhuxian;Zhou Lijun;Fan Jiangchuan;An Feng;Chen Bangqian;Wang Minghui;Jiang Ling;Xue Lianfeng;Yun Ting(College of Information Science and Technology,Nanjing Forestry University,Nanjing 210037,China;Rubber Research Institute,Chinese Academy of Tropical Agricultural Sciences,Hainan Danzhou Agro-ecosystem National Observation and Research Station,Haikou 571737,China;National Engineering Research Center for Information Technology in Agriculture,Beijing Key Lab of Digital Plant,Beijing 100097,China;College of Forestry,Nanjing Forestry University,Nanjing 210037,China;College of Engineering,University of Waterloo,Waterloo N2L 3G1,Canada)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2022年第8期187-199,共13页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家重点研发计划项目(2020YFD1000600)
国家自然科学基金资助项目(31770591,41701510)
南京林业大学虚拟教研室项目。
