The study, led by Kunhong Li, Xiaoyang Zhu, and colleagues and published in the journal Smart Agricultural Technology, addresses a problem that has grown urgent as high-throughput phenotyping and intelligent breeding move from concept to practice. Plant organs such as stems, leaves, and pods carry vital phenotypic information reflecting structural development, biomass accumulation, and yield potential. Accurately identifying and separating them is a prerequisite for measuring morphological parameters and conducting quantitative analysis. Traditional assessment, however, relies on manual observation and measurement, which is labor-intensive, inefficient, and vulnerable to subjective bias, making it unsuited to the scale demanded by modern crop breeding programs.

Two-dimensional image segmentation has made impressive strides over the past decade in crops like tomato, maize, sorghum, and wheat, but flat images have an inherent weakness: they fail to capture the three-dimensional topology of a plant, producing incomplete results when organs occlude or overlap one another. Three-dimensional imaging technologies, including depth cameras, Time-of-Flight sensors, LiDAR, structured light scanners, and multi-view stereo reconstruction, solve part of that problem by capturing rich spatial structure, surface texture, and even normal vectors. Yet translating raw 3D scans into organ labels has remained difficult. Earlier computer vision techniques based on geometry, thresholds, or 3D skeletons depended on prior knowledge and hand-tuned rules, and they generalized poorly when confronted with the complex spatial topology of mature plants.

Deep learning offered a way forward. Point cloud networks such as PointNet and PointNet++, which process unordered 3D coordinates directly, have been widely applied in plant phenotyping with strong results. But existing methods struggle with two persistent weaknesses when facing complex structures. First, their feature representation capacity is limited, making it hard to extract key geometric cues. Second, and more fundamentally, the severe spatial entanglement of plant organs produces highly confused feature distributions in the network’s latent space, limiting its ability to discriminate between different organs of the same plant. A pod pressed against a stem can simply look, in feature terms, like more stem.

PVSCNet tackles both weaknesses by integrating two established components into the PointNet++ backbone in a task-specific way. The first is the PVConv module, inserted after the Feature Propagation stage, which performs what the authors call dual-grained feature fusion. A low-resolution voxel branch applies 3D convolutions on a regularized voxel grid to capture coarse-grained local geometric topology, while a high-resolution point branch preserves fine-grained, high-frequency geometric details that voxelization inevitably blurs. After 3D convolution, devoxelization maps the voxel features back to the point domain, where they are fused with point-branch features through element-wise addition. The result is a multi-scale representation that is simultaneously detail-rich and structurally aware.

The second component is a Geometric-Aware Discriminative Loss, or GADL, adapted from the discriminative embedding loss used in the JSIS3D framework. Applied to the semantic feature space of plant organs, it combines pull, push, and regularization terms: the pull term draws features of the same organ type toward their class prototype, enforcing intra-class compactness; the push term activates only when the distance between different class centers falls below a preset threshold, forcing them apart to maximize inter-class separability; and a regularization term constrains feature norms. Together with standard cross-entropy loss, this metric-learning strategy explicitly reshapes the geometry of the feature manifold rather than merely fitting local patterns, and the researchers show that it meaningfully untangles the overlapping feature clusters that plague entangled plants.

To train and test the system, the team built Soybean3D, a dataset of 40 fully annotated soybean point clouds acquired with an EinScan Pro 2X structured light scanner at a research base of Sichuan Agricultural University in Chongzhou City. Each plant was scanned over a full 360 degrees in sessions lasting 15 to 40 minutes, then partitioned with a sliding-window strategy into 10-centimeter cubic blocks and manually annotated in CloudCompare with pod and stem labels. Because whole-plant point clouds are too large for direct network input, the partitioning approach preserves local semantic content while keeping memory demands manageable. Generalization was then tested on the public Pheno4D dataset, using its 77 annotated tomato point clouds, split so that no plant appeared in both training and validation sets.

The results are striking. On Soybean3D, PVSCNet reached a mean Intersection over Union of 91.70 percent and a mean precision of 95.95 percent, improvements of 4.76 and 3.73 percentage points over the PointNet++ baseline. Ablation experiments showed that PVConv alone lifted mIoU to 90.38 percent, while GADL alone raised it to 88.04 percent, and the combination proved complementary. Principal component analysis of the learned features, quantified with KL, symmetric KL, and Jensen-Shannon divergences, confirmed that GADL reshapes the feature space into compact, well-separated clusters; in one example, the stem-versus-pod KL divergence rose from 6.55 to 19.70. In head-to-head comparisons against PointNet, PAConv, PointMLP, SPoTr, and PointMamba, PVSCNet led on every metric, and on Pheno4D it achieved an mIoU of 89.43 percent, exceeding the average of competing methods by 5.47 points, with particularly clean separation of leaves and stems in dense near-ground regions.

Notably, the added accuracy did not cost speed. Although PVSCNet increases the parameter count to 6.26 million and FLOPs to 13.58 billion relative to PointNet++, measured training time per iteration fell about 12 percent to 177 milliseconds and per-plant inference time dropped roughly 10 percent to 72 milliseconds, a practical efficiency the authors attribute to the point-voxel processing design. The pipeline also proved useful beyond segmentation: after isolating stems, HDBSCAN clustering extracted individual pod regions, and oriented bounding boxes yielded pod length and width estimates validated against manual measurements of 50 pods, achieving R-squared values of 0.9517 for length and 0.9728 for width with root mean square errors of 0.0857 and 0.0393 centimeters, respectively.

For a field racing toward automated breeding decisions, the implications are considerable. A framework that can reliably parse the tangled geometry of a mature soybean plant, generalize across species to tomato, and run fast enough for high-throughput pipelines removes one of the stubborn obstacles between 3D scanning hardware and actionable phenotypic data. The researchers position PVSCNet as an efficient and robust segmentation solution for high-throughput plant phenotyping, with significance for intelligent breeding and precision agriculture. As structured light scanners, LiDAR units, and multi-view stereo rigs become cheaper and more common in greenhouses and field stations, tools like this may soon decide how quickly a promising soybean line advances from the scanner to the seed catalog.

Subject of Research: Deep learning-based 3D plant organ segmentation in point clouds for high-throughput crop phenotyping

Article Title: PVSCNet: Point-voxel dual-grained feature fusion and spatial constraints for 3D plant organ segmentation

Article References: Li, K., Zhu, X., Zhang, J., Yang, F., Wang, M., & Yang, C. (2026). PVSCNet: Point-voxel dual-grained feature fusion and spatial constraints for 3D plant organ segmentation. Smart Agricultural Technology, 15, Article 102553. https://doi.org/10.1016/j.atech.2026.102553

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102553

Keywords: 3D plant organ segmentation, point cloud deep learning, soybean phenotyping, point-voxel fusion, PVSCNet, high-throughput phenotyping, discriminative loss, Soybean3D dataset, Pheno4D, precision agriculture, intelligent breeding, Smart Agricultural Technology