Authors

Abstract

This study aims to construct digital fruit trees with high-precision geolocation and high-quality canopy phenotypic details, supporting the development of digital fruit tree technology and the establishment of smart orchards. The Neural Radiance Fields (NeRF) theory was integrated with georeferencing technology. Firstly, multiple ground control points were placed around the tree, and their WGS-84 coordinates were recorded using an RTK surveying instrument. Next, a drone captured multi-view images of the fruit tree, recording the camera poses during the image acquisition. The multi-view fruit tree images undergo ray casting, hierarchical sampling, and high-frequency position encoding before being input into a Multilayer Perceptron (MLP). The MLP was then supervised through volume rendering to obtain a convergent radiance field that reflects the true form of the fruit tree, resulting in the generation of a fruit tree point cloud. Finally, by establishing correspondences between the points in the fruit tree point cloud and the ground control points in the real world, a rigid transformation matrix was computed to convert the point cloud from a local coordinate system to WGS-84 coordinates, yielding a geographically informed digital fruit tree. The experiments demonstrate that the constructed digital fruit tree exhibits excellent phenotypic details and accurately represents multi-scale characteristics. The accuracy of tree morphology indicators, such as tree height, crown length, and width, reached 99.12%, 99.34%, and 99.22%, respectively. Compared to point clouds generated by traditional Structure from Motion-Multi View Stereo (SFM-MVS) methods, the root mean square errors were reduced by 61.24%, 73.48%, and 62.32%, respectively. Additionally, the georeferencing accuracy achieved millimeter-level precision, with registration errors generally below 2 mm. The proposed method can construct digital fruit trees with high geolocation accuracy, detailed phenotypic information, and scale consistency, overcoming key barriers in the development of digital fruit tree technology. It can provide comprehensive data for various production operations in smart orchards.

Abstract in Chinese

本研究旨在构建具有高水平地理定位精度与高品质叶冠表型细节的数字果树，支持数字果树技术体系与智慧果园建设。将神经辐射场（Neural Radiance Fields，NeRF）理论与地理坐标配准（Georeferencing）技术相结合，以初果期的桃树作为研究对象。首先，在果树周围地面上布设多个地面控制点并通过RTK测量仪记录地面控制点中心位置的WGS-84坐标；其次，使用无人机环绕拍摄果树多视角图像并记录拍摄时相机位姿；然后，将多视角果树图像进行光线投射法分层采样和高频位置编码后输入多层感知机（Multilayer Perceptron，MLP），通过体积渲染（Volume Rendering）监督训练过程以获取收敛且能反映果树真实形态的辐射场并导出果树点云；最后，通过果树点云中与现实世界中地面控制点的对应关系，计算刚性变换矩阵，将果树点云从局部坐标系转换至WGS-84坐标系，得到具有地理信息的数字果树。试验表明，本研究构建的数字果树具有良好的表型细节，可准确表征果树多尺度表型细节。该方法构建的果树点云在树高、冠层长度与宽度等树形指标方面的精度分别达到99.12%、99.34%、99.22%，相较于传统的运动恢复结构-多视图立体匹配（Structure from motion-Multi View Stereo，SFM-MVS）方法构建的果树点云，均方根误差分别减小61.24%、73.48 %、62.32%。同时，其地理坐标配准精度达到毫米级，配准误差普遍小于2mm。该研究提出的方法能构建具有高地理坐标定位精度、高表型细节与高尺度一致性的数字果树，突破了制约数字果树技术体系发展的关键瓶颈，能够为智慧果园的数字化果树表型组学研究、数字化树形管理、数字化生长监测等领域提供关键技术支撑。