3D Point Cloud Preprocessing for High-Quality Crop 3D Modeling

Efficient and accurate crop phenotyping is essential to address the growing global food demand due to rapid population growth. While conventional two-dimensional (2D) imaging techniques are widely used, they do not fully capture complex three-dimensional (3D) plant structures, limiting the accuracy of morphological measurements such as biomass estimation and leaf-stem angle measurements. In this study, we use neural radiance fields (NeRF) to reconstruct high-fidelity 3D models of crop specimens from a 360° RGB image sequence taken with a Galaxy S20 Ultra smartphone. Approximately 230-240 frames per specimen were sampled and registered using COLMAP, which was then integrated with the Nerfacto algorithm to generate a detailed point cloud representation. Scale correction was performed using checkerboard markers and CloudCompare software. To increase model fidelity, we compare two denoising strategies: an AI-driven cropping function within NeRF Studio and a geometry-based pipeline that incorporates density clustering (DBSCAN with different eps multipliers), bounding box extraction, statistical outlier removal (SOR), and voxel-based downsampling. The geometry-based approach effectively removes fine residual points, especially those between neighboring stems and leaves, producing cleaner, more uniform point clusters. However, the manual selection of clustering parameters remains a limitation. Future work will focus on automating parameter extraction within a geometry-based workflow and integrating the preprocessing directly into the NeRF reconstruction pipeline. These advances will aim to streamline 3D phenotyping workflows and support more accurate quantitative trait extraction, ultimately contributing to accelerating crop breeding and optimizing agricultural production.