AI sees through soybean canopies from seedling to seed
image: Overview of the SOY3DSEG workflow. (A) Dynamic Multi-Stage Sampling Strategy (DMSS): For each soybean point cloud spanning growth stages V1–R7, the algorithm computes bounding-box volume and point-cloud density, derives the coefficient of variation, and adaptively sets a two-stage down-sampling ratio to generate a structurally faithful yet lightweight point set.(B) SOY3DSEG point-cloud segmentation network: An encoder–decoder architecture first extracts multi-scale geometric features with a geometry-aware Transformer and Transition Down layers, then restores resolution through Transition Up and feature-fusion blocks, finally classifying individual points into stem and leaf categories.(C) Cross-crop transfer validation: The trained model is applied—without additional fine-tuning—to maize and tomato datasets, producing accurate organ-level segmentations and demonstrating strong cross-species generalization.
Credit: Plant Phenomics
Three dimensional point cloud technology can capture plant architecture in vivid detail, but no existing model has been able to reliably separate stems from leaves across an entire crop’s life cycle—from the first true leaf to full reproductive maturity. Researchers now report a new segmentation framework, SOY3DSEG, that achieves over 93% average accuracy across 12 consecutive growth stages of soybean (V1–R7). The system uses an adaptive sampling strategy that allocates more computational attention to complex, densely packed regions, enabling it to maintain clear organ boundaries even when canopies become heavily overlapping and occluded.
Conventional imaging methods falter when crops mature. Two dimensional approaches lose spatial fidelity under occlusion and variable lighting, while existing 3D models rely on fixed sampling ratios that either discard critical details in sparse zones or waste computation on redundant points in dense foliage. Most networks trained on early stage plants fail catastrophically when confronted with the structural chaos of reproductive stage canopies, where leaves stack and stems twist into near impenetrable thickets. Based on these challenges, the team set out to build a segmentation system that could adapt its focus as the plant itself transformed—a model that would not simply recognise a leaf, but would know how a leaf looks at every stage of its life.
The work, published (DOI: 10.1016/j.plaph.2026.100167) online on 3 September 2026 in the journal Plant Phenomics, was conducted by researchers from the College of Engineering, College of Arts and Sciences, School of Public Administration & Law, and College of Agriculture at Northeast Agricultural University, together with the National Key Laboratory of Smart Farm Technologies and Systems in Harbin, China. The team trained SOY3DSEG on the Soybean MVS dataset—high density point clouds with sub millimetre resolution, spanning five cultivars and 12 growth stages, manually annotated point by point as stem or leaf—and then tested its cross crop transferability on maize and tomato seedlings using the Pheno4D dataset, achieving near perfect accuracy without fine tuning.
At the core of SOY3DSEG is a Dynamic Multi Stage Sampling Strategy (DMSS) that replaces fixed ratio down sampling. The algorithm first computes two complementary variability measures: the coefficient of variation (CV) of point occupancy across voxels (capturing global structural spread) and the CV of local point density via k nearest neighbours (KNN) (capturing local crowding from leaf overlap). These feed into a dynamic weight that determines how aggressively to sample—preserving more points in geometrically intricate regions like stem leaf junctions while efficiently culling redundant points in simpler areas. The model then passes the adaptively sampled cloud through an enhanced Point Transformer architecture featuring multi scale neighbourhood aggregation with ball query radii of 0.05, 0.10 and 0.15 (normalised by plant height), alongside a geometry aware attention mechanism that explicitly weights nearby points by their spatial proximity. This dual strategy enables SOY3DSEG to maintain segmentation accuracy above 95% even at the R6 and R7 stages, where conventional models such as PointNet++ and RandLA Net plummet to as low as 25% mean Intersection over Union (mIoU). Ablation studies confirm that removing DMSS drops mIoU from 93.34% to 87.95%, while stripping out geometry aware attention reduces stem accuracy from 95.57% to 91.85%, proving that each component pulls its weight.
“The challenge with plant point clouds is that a soybean at V1 and a soybean at R7 are practically different objects—different density, different occlusion patterns, different geometric complexity,” the authors said. “What we’ve done is build a system that doesn’t treat every plant the same way. It looks at the structure first, then decides where to look more carefully. That’s why it can keep stems and leaves straight even when you can barely see through the canopy.” They emphasised that the framework is backbone agnostic, meaning the DMSS and multi scale modules could be paired with other point cloud architectures beyond the Point Transformer used in this study.
The implications extend well beyond soybean. In cross crop tests on maize and tomato seedlings, SOY3DSEG achieved mIoU values of 99.8% and 99.2%, respectively, demonstrating strong early stage transferability without retraining. This suggests that a single model, trained once, could serve as a universal phenotyping tool across multiple crop species—at least during early development. For breeders and agronomists, automatic organ level segmentation throughout the full growth cycle unlocks high throughput extraction of traits such as stem diameter, leaf area index and branching patterns, all critical for selecting resilient varieties and optimising field management. The Soybean MVS dataset and code are publicly available on GitHub, providing a benchmark for future research in 3D crop phenotyping. The team plans to extend the framework to instance level segmentation and to test its performance on a wider range of species and stress affected morphologies.
###
References
DOI
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100167
Funding information
This work was supported by the Research and Application of Key Technologies for Intelligent Farming Decision Platform (an Open Competition Project of Heilongjiang Province, China) under Grant No. 2021ZXJ05A03; the National Key R&D Program Project funded by the Ministry of Agriculture and Rural Affairs of the People's Republic of China under Grant No. 2025YFD2300401; the Programs for Science and Technology Development of Heilongjiang Province under Grant No. 2024ZX01A07; and the Heilongjiang Province Key Research and Development Plan (Innovation Bases), China under Grant No. JD2023GJ01.
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 cutting edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.