A 3D grain-level phenotypic extraction framework for rice using contour-guided NeRF reconstruction and mamba-based point cloud segmentation

Grain-level three-dimensional phenotyping of rice provides quantitative traits for breeding and yield assessment, but remains challenging because dense panicles exhibit severe organ adhesion, boundary ambiguity, and large-scale point-cloud complexity. We developed PRSNeRF, a grain-level phenotyping workflow that combines contour-guided neural radiance field reconstruction with point cloud instance segmentation, and evaluated it on rice grown in controlled indoor environments. PRSNeRF integrates CoNeRF for reconstruction and PRSNet for segmentation. CoNeRF combines image deblurring, contour-guided sampling, and rendering-projection consistency to improve geometric fidelity and suppress floating artifacts. PRSNet uses a Mamba-based encoder, adaptive Gaussian detail enhancement, multi-scale feature fusion, and three-dimensional dynamic anchor boxes to separate densely adjacent grains. Under a fixed 5,000-iteration budget on the Blur Rice benchmark, CoNeRF reduced the scaled NeRF-space Chamfer distance from 10.61 to 7.81. Relative to the 30,000-iteration NeRFacto baseline, it shortened training time from 368 s to 77 s and point cloud export time from 120 s to 10 s. On the RiceNeRF-IS dataset, plant-level five-fold cross-validation showed that PRSNet achieved an overall AP of 87.31%, outperforming OneFormer3D by 22.30 percentage points. The reconstructed point clouds enabled automated estimation of grain count, grain dimensions, and voxel-based occupancy traits. Together, these results connect reconstruction, instance separation, and trait extraction within one workflow. PRSNeRF supports grain-level rice panicle phenotyping from multi-view images under controlled indoor acquisition. It improves reconstruction fidelity, separates densely adhered organs, and supports downstream trait extraction within an end-to-end processing time of about 3.3 min per plant.