Tree Structure Optimization Based on NeRF Reconstruction: Single-Tree Segmentation and Skeleton Extraction

Tree topology reconstruction is vital for precision agriculture applications, including canopy structure analysis and yield prediction. However, Neural Radiance Fields (NeRF)-based three-dimensional reconstruction, which focuses on scene reconstruction, is susceptible to noise and data loss, resulting in incomplete skeleton branches during extraction. To address this challenge, we propose a method for single-tree segmentation and complete skeleton extraction. Our approach begins by extracting the tree shape using normal vectors and clustering techniques, followed by defining a spatial range based on the extracted shape. High-precision segmentation of single-tree point clouds is achieved by combining this spatial range with the original scene point cloud. Subsequently, graph structure clustering and DBSCAN are employed to segment incomplete tree point clouds and extract independent skeleton groups. Potential branch connections are identified based on distance and angular features of the local branches, and erroneous connections are removed through closedloop structure filtering to complete the skeleton. Finally, the skeleton structure is optimized by incorporating the global growth direction. Validation results demonstrate that our method achieves an accuracy of 85.96% on 13 real plum trees and 91.26% on 200 simulated trees, providing reliable support for tree structure reconstruction and analysis.