基于神经辐射场和路径分析的油茶树表型参数提取  

Extracting the phenotypic parameters of Camellia oleifera using neural radiation field and path analysis

作　　者：朱幸辉[1] 张杨聪 谭泗桥[1,2] 李斌 ZHU Xinghui;ZHANG Yangcong;TAN Siqiao;LI Bin(College of Information and Intelligence,Hunan Agricultural University,Changsha 410126,China;Hunan Engineering Technology Research Center of Agricultural&Rural Informatization,Changsha 410126,China)

机构地区：[1]湖南农业大学信息与智能科学技术学院,长沙410126 [2]湖南省农村农业信息化工程技术研究中心,长沙410126

出　　处：《农业工程学报》2025年第4期201-210,共10页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(62202163);湖南省自然科学基金(2022JJ40190);湖南省教育厅科学研究项目(22A0145);湖南省学位与研究生教学改革研究项目(2023JGYB140)。

摘　　要：为提取油茶树的表型参数,解决复杂冠层结构植物和田间环境下表型提取速度慢且精度低的问题,该研究将传统的重建算法和聚类分割算法进行改进,提出一种基于神经辐射场和路径分析的表型参数提取方法。通过多视角相机获取油茶树图像,训练神经辐射场生成三维点云模型,然后采用路径分析方法分割树干和叶片点云,提取油茶树的表型参数。试验结果表明:相比基于运动结构恢复的多视立体几何方法,采用神经辐射场重建的时间平均减少约90%,自由视角渲染图像峰值信噪比提升约10%。茎叶分割结果在召回率、精确率和分割时间等指标上优于几何特征法和区域生长法。计算的树高、冠幅、冠层高度和树干与人工测量结果的误差分别为0.519%、0.325%、0.364%、4.491%,叶长、叶宽、叶面积和叶形指数的决定系数分别为0.98、0.94、0.97、0.93。该方法不仅能快速构建真实形态的油茶树点云模型,而且能精准获取油茶树各器官的表型参数,可为复杂冠层植物的田间表型研究提供参考。Camellia oleifera has been widely cultivated as a vital economic tree species for its oil-rich seeds.Accurate phenotypic analysis is essential to the conservation,breeding,and optimization.Precise measurement of phenotypic parameters can also advance the breeding programs for the high crop yields and optimal genetic traits.However,the complex canopy structure of Camellia oleifera has presented the significant challenges to extract the phenotypic data under field conditions.Manual measurements and multi-view stereo(MVS)reconstruction can often limited to the accuracy,efficiency,and scalability,particularly when applied to the dense and intricate tree canopies.In this study,an innovative approach was introduced to combine the neural radiance fields(NeRF)and path analysis.The high-precision extraction of phenotypic parameters was also performed on the camellia oleifera in fields.Multi-view images were captured from the fixed positions around the tree.A NeRF model was trained to reconstruct an accurate three-dimensional point cloud of the tree.Both structural and high-frequency were preserved after reconstruction.Path analysis was subsequently employed to segment the tree stem and leaf clusters within the point cloud,thus enabling the precise extraction of key morphological and phenotypic parameters.Experimental results show that the NeRF significantly outperformed the traditional MVS,in terms of both efficiency and reconstruction quality.Specifically,the NeRF approach was reduced the point cloud reconstruction time by approximately 90%.While the reconstruction quality was improved with the peak signal-to-noise ratio(PSNR)by 10%.Furthermore,the path analysis-based segmentation demonstrated that the superior performance was achieved to separate the stems from leaves.There was also the increase in the recall,precision,F1 score,and overall accuracy of segmentation,compared with the conventional geometric feature-based and region-growing approaches.Respectively,compared with the geometric feature and the region growing a

关 键 词：油茶树 三维重建 神经辐射场 路径分析 表型参数 

分 类 号：S126[农业科学—农业基础科学]