Cabbage is a major leafy vegetable crop whose growth status directly affects productivity and supply stability. In open-field cultivation, complex and dynamic conditions make non-destructive, high-precision, continuous monitoring essential. Remote sensing platforms such as satellites and unmanned aerial vehicles (UAVs) enable dynamic crop monitoring, but they struggle with high-precision trait extraction at the individual plant scale due to spatial resolution and occlusion. Two-dimensional (2D) vision methods based on RGB or multispectral imagery suit fine-scale analysis but lack three-dimensional (3D) structural information, limiting accurate extraction of plant height and head volume. Three-dimensional reconstruction offers promise, yet Structure-from-Motion (SfM) and Multi-View Stereo (MVS) methods remain susceptible to occlusion and illumination, with high computational costs. Due to these challenges, in-depth research on full-cycle, high-precision 3D phenotyping of field-grown cabbage is needed.
The study was conducted by researchers at China Agricultural University, the National Engineering Research Center for Information Technology in Agriculture, the Beijing Academy of Agriculture and Forestry Sciences, and the Key Laboratory of Digital Village Technology of the Ministry of Agriculture and Rural Affairs. It was published (DOI: 10.1016/j.plaph.2026.100219) in Plant Phenomics on 2026. The team developed a 3D point cloud dataset and an adaptive segmentation network for organ-level segmentation of cabbage across developmental stages under complex open-field conditions, enabling quantitative characterization of growth dynamics throughout the full growth cycle.
The network improves on OneFormer3D, using a 3D U-Net backbone for multi-scale feature extraction and a transformer decoder for joint instance, semantic, and panoptic segmentation. Three domain-specific modules were introduced. The Head Refinement Module (HRM) enhances head detection by addressing morphological similarity between heads and leaves. The Leaf Instance Segmentation Module (LISM) handles overlapping leaf structures in rosette and heading stages through geometric feature enhancement, boundary detection, instance embedding, and adaptive clustering. The Cross Module Interaction (CMI) uses a point-wise gated mechanism conditioned on head probability to coordinate features between branches and suppress cross-category interference. A growth-stage classification module enables discrimination among developmental phases. For semantic segmentation, mean Intersection over Union (mIoU) reached 0.767, with 94.8% point classification accuracy. The model has 54.25 million parameters and a 0.76 s response time. For instance segmentation, Average Precision (AP) improved by 2.3% for heads and 3.8% for leaves, while Average Recall (AR) rose 6.9%. Ablation studies confirmed that HRM mainly enhanced head detection, with AP at 50% IoU (AP50%) rising from 0.4699 to 0.6985, while LISM improved leaf performance. Introducing CMI mitigated feature competition between branches.
The authors said the work addresses a critical gap in extending 3D phenotyping from controlled indoor environments to open-field applications. They explained that the three submodules contribute differently to full-cycle monitoring: the HRM improves discriminative capability for the head class and reduces feature confusion with leaves; the LISM increases leaf recall and improves segmentation quality; and the CMI integrates multi-dimensional features to balance contributions, enhancing overall detection for both head and leaf classes. They added that segmentation accuracy takes precedence over marginal gains in inference speed for this application context, and that the method maintained robust performance across all growth stages despite decreasing accuracy with increasing morphological complexity.
The proposed framework enables non-destructive, full-cycle monitoring of cabbage phenotypes under open-field conditions, offering a practical tool for precision agricultural management. By extracting multidimensional growth parameters such as plant height, canopy spread, leaf area, and head diameter, the method supports accurate growth assessment and yield-related trait evaluation. Its ability to maintain robust segmentation performance across developmental stages and variable weather conditions makes it adaptable to real-world field scenarios. The approach provides a novel technical pathway for extending 3D phenotyping from controlled environments to open-field applications, supporting precise crop monitoring and the development of digital twin agriculture. Future work will extend the method to field-scale analysis by integrating georeferenced robotic scouting data with the individual plant phenotyping pipeline.
###
References
DOI
10.1016/j.plaphe.2026.100219
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100219
Funding information
This work was supported by National Key Research and Development Program of China Project (Grant No.2024YFD2000805), the earmarked fund for CARS Project (CARS-23-D07) and Beijing Smart Agriculture Innovation Consortium Project (BAIC10-2026).
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.