Three-dimensional (3D) crop phenotyping supports the discovery of genes linked to plant architecture, yield, and stress responses. Organ segmentation is a first step, because traits such as leaf area, leaf number, stem length, and fruit size are measured from individual organs. Inductive deep networks dominate this task, but they usually require thousands of richly annotated samples and point-level labels. In agriculture, point clouds are often noisy, incomplete, and structurally complex, making high-quality annotation slow and expensive. Given these challenges, there is a need for in-depth research on annotation-efficient learning methods that can segment organs from sparsely labeled 3D plant data.

A team from Donghua University in Shanghai and the Beijing Academy of Agriculture and Forestry Sciences has introduced Dual-branch Graph Convolutional Network (DBGCN), a transductive graph learning framework for organ instance inference on sparsely labeled plant point clouds. The study was published (DOI: 10.1016/j.plaph.2026.100188) in 2026 in Plant Phenomics. DBGCN uses a dual-branch graph convolutional architecture to combine static global structure with dynamic local feature relationships, enabling simultaneous training and inference on a single plant graph. The authors report strong performance across tobacco, tomato, sorghum, and soybean point clouds.

DBGCN is built around four modules. A Graph Representation Module (GRM) converts each point cloud into a graph and enriches each node by concatenating its 3D coordinates with a Gaussian similarity matrix built from K-nearest-neighbor (KNN) neighborhoods. A Dual-branch Graph Convolutional Module (DBGCM) then runs two parallel convolutions. The static-feature-space graph convolutional branch (SGGB) uses a fixed adjacency structure similar to GCNII to capture global plant architecture. The dynamic-feature-space graph convolutional branch (DGGB) searches KNN in the evolving feature space at every layer, allowing it to track local changes and refine boundaries. A Fusion Module (FM) aligns and adds the two feature streams, while a Downstream Task Module (DTM) uses a multilayer perceptron (MLP) and softmax to assign organ labels. On the PlantNet sub-dataset, DBGCN reached 93.00% mean accuracy of node classification (mAcc) with only 1.95% manual annotation. On the Soybean-MVS sub-dataset, it reached 91.05% mAcc with 4.88% annotation. It outperformed several mainstream graph and point-cloud models, including graph convolutional network (GCN), GCNII, graph attention network (GAT), Graph Random Neural Network (Grand), Jumping Knowledge Network (JKnet), Graph Sample and Aggregate (GraphSage), PlantNet, and PSegNet. The model also generalized to Paris-Lille-3D street-view light detection and ranging (LiDAR) data, achieving 92.37% and 95.06% accuracy on two slices.

The authors said the goal was to turn a handful of labeled points into a complete organ map, instead of asking annotators to label every point. They said the static branch captures the overall plant structure, while the dynamic branch adapts to changing local relationships in deeper feature space. This combination, they said, helps preserve leaf edges, small buds, and boundaries between touching organs. They added that the framework could make 3D phenotyping faster and more practical for breeding programs, especially when labeled data are scarce.

DBGCN could reduce the annotation bottleneck in 3D crop phenotyping, where manual point labeling is a major cost. By generating high-quality organ labels from sparse input, it may help researchers build large training sets for inductive deep networks and extract organ-level traits such as leaf area, leaf count, stem length, and fruit size. The authors also tested the method on street-view LiDAR data, suggesting broader use in 3D scene segmentation. Current limits include graph size, with the model tested on fewer than 50,000 nodes. Future work aims to scale DBGCN to larger point clouds and add transformer-like components.

###

References

DOI

10.1016/j.plaphe.2026.100188

Original Source URL

https://doi.org/10.1016/j.plaphe.2026.100188

Funding information

This work was supported in part by the National Natural Science Foundation of China(32472001), and in part by Donghua University 2025 Cultivation Project of Discipline Innovation (xkcx-202505).

About Plant Phenomics

Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cuttingedge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into realworld solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing highthroughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.