Maize (Zea mays L.) is a global staple for food, feed, and industry, and breeders increasingly need large-scale trait data. Manual ear phenotyping is slow, labour-intensive, and often destructive: counting kernels can require marking each seed, while ear volume is measured by water displacement. Many digital methods rely on one or a few two-dimensional (2D) images, missing the whole three-dimensional (3D) ear surface and failing to orient traits in space. Some multi-view systems rotate ears but do not reconstruct 3D structure. Consequently, key traits such as kernel row number, ear volume, and kernel number are rarely measured together across the entire ear. Given these challenges, or because of these issues, in-depth research is needed on accurate, non-destructive, high-throughput 3D maize ear phenotyping.
Researchers at the University of Hohenheim in Germany, with collaborators at China Agricultural University, published (DOI: 10.1016/j.plaphe.2026.100274) the study in Plant Phenomics online 26 August 2026. They developed an ear scanner that rotates a maize ear through 100 viewpoints, captures RGB images, and converts them into a 3D voxel grid. Deep learning models detect and segment kernels, then map 2D masks onto the 3D model. The pipeline measures ear and kernel traits in about 30 seconds of capture and 47 seconds of computation per ear.
The workflow combines object detection and instance segmentation. YOLOv8m, a You Only Look Once (YOLO) model, proposes kernel bounding boxes; the Segment Anything Model (SAM) refines them into masks, achieving a Dice score (DS) of 0.94 and a mean average precision (mAP) at intersection-over-union (IoU) thresholds from 0.50 to 0.95 of 0.720. Each mask is linked to a voxel in the carved 3D grid, allowing pixel measurements to be scaled to millimetres and enabling tracking of individual kernels across views. Against manual measurements on 105 morphologically diverse ears, the pipeline predicted kernel number (KN) with R² = 0.98, ear volume (EV) with R² = 0.97, ear length (EL) with R² = 0.96, maximum kernel row length (KRL) with R² = 0.87, ear width (EW) with R² = 0.82, and kernel row number (KRN) with R² = 0.77. The lowest mean absolute percentage error was 2.1% for KN; no trait differed by more than 5.7%. In a phosphorus (P) fertilisation trial with 44 unfertilised control ears and 45 ears receiving diammonium phosphate (DAP), the pipeline detected significant increases in EW, EL, EV, kernel size (KS), mean box width (BW), and mean box height (BH), while KN, KRL, KRN, and kernel shape ratio remained unchanged.
The authors said the scanner was designed to remove a long-standing bottleneck in maize research. They said that capturing 100 views in 30 seconds and computing all traits in 47 seconds makes the pipeline practical for breeding programmes and nutrition studies, where hundreds or thousands of ears must be compared. They added that mapping 2D kernel masks to a 3D voxel grid was key: it allowed reliable tracking, real-world scaling, and measurement of traits such as EV that are difficult to obtain from single images. They said the approach is non-destructive, leaving ears available for further analysis.
The pipeline could accelerate maize breeding by linking genomic data to detailed ear and kernel traits at scale. In plant nutrition, it can reveal how nutrient supply reshapes ear development, as shown by the P trial, where yield-related size traits responded while kernel arrangement remained stable. Because the method uses RGB cameras, a rotating stage, and open-source deep learning, it may be adapted to other crops with compact, repeatable structures. Future work could compare predicted KS with thousand-kernel weight (TKW), a key breeding trait, and test the pipeline across more genotypes, locations, and years. The authors have released source code and result data, supporting wider adoption.
###
References
DOI
10.1016/j.plaphe.2026.100274
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100274
Funding information
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 328017493/GRK 2366 (Sino-German International Research Training Group AMAIZE-P).
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.