Fine-grained 3D phenotypic analysis of rice plays a vital role in rice breeding and yield estimation. However, a comprehensive rice data acquisition and segmentation pipeline is still lacking. While Neural Radiance Fields (NeRF) have shown impressive results in crop-level 3D reconstruction, their high sensitivity to data volume and camera viewpoints often leads to reconstruction failures for rice. In addition, the large-scale rice point clouds, coupled with heavy occlusion and visual similarity among grains, pose significant challenges for fine-grained trait extraction. To address the challenge of reconstructing rice point clouds under low-quality data conditions, we propose a novel method named Multi-Scale NeRF(MSNeRF). This method incorporates a structure-detail collaborative reconstruction mechanism and a dynamic initialization density scheduling strategy. Furthermore, we introduce a multimodal and multitask rice dataset (MMR) as a benchmark resource for future research. For rice point cloud segmentation, we develop Vision Rice Knowledge Graph Network(VRKGNet), which comprises an image segmentation module, a projection module, and a point cloud segmentation module enhanced with a Transformer to enlarge the receptive field. VRKGNet performs standalone point cloud segmentation and integrates image segmentation results from multiple viewpoints as prior knowledge to enhance semantic and instance-level segmentation. Extensive experiments demonstrate that MSNeRF achieves high-fidelity point cloud reconstruction with as few as 10 viewpoints. VRKGNet achieves superior rice plant segmentation with a semantic segmentation mIoU of 88.79% and an instance segmentation AP₂₅ of 84.55%, outperforming mainstream algorithms.

Advanced plant phenotyping technologies are vital for trait improvement and accelerating intelligent breeding. Due to the species diversity of plants, existing methods heavily rely on large-scale high-precision manually annotated data. For self-occluded objects at the grain level, unsupervised methods often prove ineffective. This study proposes IPENS, an interactive unsupervised multi-target point cloud extraction method. It utilizes radiance field information to lift 2D masks, segmented by SAM2 (Segment Anything Model 2), into 3D space for target point cloud extraction. A multi-target collaborative optimization strategy addresses the challenge of segmenting multiple targets from a single interaction. On a rice dataset, IPENS achieves a grain-level segmentation mean Intersection over Union (mIoU) of 63.72 %. For phenotypic trait estimation, it achieves a grain voxel volume coefficient of determination R² = 0.7697 (Root Mean Square Error, RMSE = 0.0025), leaf surface area R² = 0.84 (RMSE = 18.93), and leaf length and width prediction accuracies of R² = 0.97 and R² = 0.87 (RMSE = 1.49 and 0.21). On a wheat dataset, IPENS further improves segmentation performance to a mIoU of 89.68 %, with exceptional phenotypic estimation results: panicle voxel volume R² = 0.9956 (RMSE = 0.0055), leaf surface area R² = 1.00 (RMSE = 0.67), and leaf length and width predictions reaching R² = 0.99 and R² = 0.92 (RMSE = 0.23 and 0.15). Without requiring annotated data, IPENS rapidly extracts grain-level point clouds for multiple targets within 3 min using single-round image interactions. These features make IPENS a high-quality, non-invasive phenotypic extraction solution for rice and wheat, offering significant potential to enhance intelligent breeding.