SOY3DSEG: A high-precision universal point cloud segmentation model for soybean full growth period based on improved point transformer.

Abstract

Three-dimensional (3D) reconstruction technologies for crops are of significant importance in the context of smart breeding and precision agriculture, as they enable accurate characterization of crop spatial architecture and developmental dynamics. Such capabilities provide essential phenotypic information for the rapid selection of breeding materials and informed agronomic decision-making. A critical requirement for the practical application of crop 3D models is high-accuracy organ-level segmentation. However, the absence of a stage-universal segmentation framework capable of operating across complete soybean growth cycle remains a major bottleneck hindering progress in this field. To address this issue, we propose SOY3DSEG-a high-precision framework based on an improved Point Transformer, designed to support the full developmental spectrum of soybean (V1-R7). The framework incorporates a novel down sampling strategy termed Dynamic Multi-Stage Sampling Strategy (DMSS), alongside multi-scale feature enhancement and a local geometry-aware attention mechanism, enhancing segmentation accuracy and efficiency. Performance evaluations across 12 consecutive soybean growth stages (V1 to R7) indicate that SOY3DSEG achieved an average mean Intersection-over-Union (mIoU) of 93.34 % for stem-leaf segmentation-surpassing RandLA-Net, BAAF-Net, PointNet++, and PointConv by over 30 %, and outperforming the baseline Point Transformer by 14.18 %. A moderate accuracy decline appears at R6-R7 due to dense canopies and strong occlusion, yet SOY3DSEG retains clear superiority over the baseline Point Transformer, demonstrating robustness under complex morphology. In cross-crop transfer tests limited to early seedling stages of maize and tomato, the model achieves an mIoU of approximately 99 %, indicating strong early-stage transferability while mature-stage generalization across species remains open for future study. SOY3DSEG thus provides a stage-robust and scalable solution for full-cycle soybean phenotyping and growth monitoring, contributing to precision agricultural practice.

Plant phenotyping relevance

大豆の3D点群から器官レベル形態を抽出する分割フレームワークを開発・評価しており、植物表現型取得手法が研究の中心である。

Code and data availability

The authors state that the dataset (Soybean-MVS point clouds) and program code used in this study are publicly available at their GitHub repository, which is an allowed URL.