Abstract: Field phenotyping is fundamental to crop variety evaluation. However， traditional data acquisition methods are often destructive， costly， and subjective， and conventional 3D reconstruction techniques struggle to capture fine plant structures. To address these challenges， this study proposed a high-throughput phenotyping method for individual oat plants based on smartphone video acquisition and 3D Gaussian Splatting （3DGS） technology. The experiment utilized 187 oat plants at the milky stage as tested materials. Multi-view video data captured by smartphones were used to generate 3DGS point cloud models， from which canopy structural parameters—including projected area， maximum canopy width， and average canopy width-were extracted. Canopy volume was calculated using both Convex Hull and Voxelization （voxel size 0.01 m and 0.005 m） methods. Furthermore， biomass （fresh and dry weight） prediction models based on canopy volume were constructed using Random Forest （RF） and Support Vector Machine （SVM） algorithms. The results showed that： （1） 3DGS technology achieved high-fidelity reconstruction of the complex morphology of oats， effectively resolving issues related to slender leaves， thin stems， and severe self-occlusion， thereby demonstrating significant advantages over traditional reconstruction methods in restoring intricate structural forms； （2） Biomass estimation models based on 3DGS point cloud volume were reliable， with dry weight prediction accuracy （Accuracy≈0.88） generally superior to that of fresh weight； notably， high-resolution voxelization （voxel size 0.005 m） demonstrated a slight advantage in accurately describing the physical canopy structure. The workflow established in this study-“Smartphone Video Acquisition-3DGS Modeling-Phenotypic Parameter Extraction”-offered a low-cost， high-precision， and non-destructive solution， providing effective technical support for high-throughput phenotypic analysis of fields.

Key words: Gaussian splatting, Oat, Biomass prediction, Machine learning, Canopy volume, High-throughput phenotyping

摘要： 作物田间表型测量是品种选育的基础。传统作物表型采集存在破坏性大、成本高、主观性强等问题，难以满足当前品种选育工作的要求。本研究提出一种基于智能手机视频采集与3D高斯泼溅（3D Gaussian Splatting，3DGS）技术的饲用燕麦单株表型分析方法。试验以187株乳熟期饲用燕麦为研究对象，利用智能手机获取多视角视频数据并生成3DGS点云模型，提取投影面积、最大冠幅及平均冠幅等冠层结构参数，分别采用凸包法和体素化法（体素大小0.01 m，0.005 m）计算冠层体积。结合随机森林与支持向量机算法，构建生物量（鲜重、干重）预测模型。结果表明：（1）3DGS技术能够高保真地还原燕麦叶片细长、茎秆纤细及自遮挡严重的复杂形态，在复杂形态还原上展现出相对传统重建方法的显著优势；（2）基于3DGS点云体积的生物量估测模型精度可靠，干重预测精度（Acc≈0.88）普遍优于鲜重，且高分辨率（体素大小0.005 m）的体素化方法在描述冠层实体结构方面略占优势。本研究构建的“手机视频采集—3DGS建模—表型参数提取”工作流具有低成本、高精度且非破坏性的特点，为田间作物高通量表型分析提供了有效的技术支撑。

关键词: 高斯泼溅, 饲用燕麦, 生物量预测, 机器学习, 冠层体积, 高通量表型