水稻是全球主要粮食作物，准确测量其分蘖期表型性状对于育种和产量评估至关重要。传统测量方法耗时费力且易受主观误差影响，为此，本研究提出一种基于神经辐射场（NeRF）的三维重建方法，用于实现分蘖期水稻表型参数的高精度、无损提取。该方法首先利用智能手机采集水稻植株的环绕视频，并采用自适应抽帧算法获取高质量图像序列；随后，基于运动恢复结构（SfM）技术估计相机位姿，并利用改进的 Instant-NGP 算法进行高效三维重建。与原始 NeRF 相比，本方法在平均峰值信噪比上提升 17.3%，显存消耗降低 54.3%，重建时间缩短 99.4%。进一步对重建得到的点云进行下采样、降噪、坐标校正和分割等预处理，最终提取到株高、茎粗、分蘖数、分蘖角度、投影面积、最小包围盒体积及叶片数等表型参数。实验结果表明，系统自动测量结果与人工测量结果高度一致，株高、茎粗、分蘖数、分蘖角度和叶片数的决定系数（R²）分别为 0.98、0.94、1.00、0.95 和 0.97，平均绝对百分比误差分别为 2.38%、5.16%、0%、7.15% 和 2.20%。本研究为水稻品种选育和精准栽培提供了有效的技术支撑。

Rice is a major staple food crop worldwide, and accurate measurement of phenotypic traits during the tillering stage is essential for breeding programs and yield assessment. Conventional measurement methods are often time-consuming, labor-intensive, and susceptible to subjective errors. To overcome these limitations, this study introduces a 3D reconstruction approach based on Neural Radiance Fields(NeRF) for high-precision, non-destructive extraction of phenotypic parameters of rice at the tillering stage. The method begins by capturing multi-view videos of rice plants using a consumer-grade smartphone, followed by an adaptive frame extraction algorithm to obtain high-quality image sequences. Camera poses are then estimated using Structure-from-Motion (SfM), and an improved Instant-NGP algorithm is applied for efficient 3D reconstruction. Compared to the original NeRF, the proposed method achieves a 17.3% improvement in peak signal-to-noise ratio, a 54.3% reduction in GPU memory usage, and a 99.4% decrease in reconstruction time. The resulting point clouds undergo preprocessing—including downsampling, denoising, coordinate correction, and segmentation—to extract key phenotypic traits such as plant height, stem diameter, tiller number, tiller angle, projected area, bounding box volume, and leaf count. Experimental results show strong agreement between automated and manual measurements, with coefficients of determination (R²) of 0.98, 0.94, 1.00, 0.95, and 0.97 for plant height, stem diameter, tiller number, tiller angle, and leaf number, respectively. The corresponding mean absolute percentage errors were 2.38%, 5.16%, 0%, 7.15%, and 2.20%. This research offers reliable technical support for rice breeding and precision cultivation.