henhaobio
Sign in

Plant PhenomicsPlant BiologyPeer-reviewed journal article

The Global Wheat Full Semantic Organ Segmentation (GWFSS) dataset

Zi Helen HuangiD, Radek ZenkliD, Latifa GrecheiD, Benoît de SolaniD, Lucas Bernigaud SamataniD, Safaa Ouahid, Andrea VisioniiD, Carlos A. Robles‐ZazuetaiD, Francisco PintoiD, Ivan Perez-Olivera, Matthew Paul ReynoldsiD, Chen ZhuiD, Shouyang LiuiD, Marie-Pia D'argaignoniD, Raúl López‐LozanoiD, Marie WeissiD, Afef MarzouguiiD, Lukas RothiD, Sébastien DandrifosseiD, Alexis CarlieriD, Benjamin DumontiD, Benoît MercatorisiD, Javier A. FernándeziD, Scott Cameron ChapmaniD, Keyhan NajafianiD, , Andreas HundiD

PDFDOIPreprintData
In-depth full-text digest available · Jump to digest ↓

Abstract

Computer vision is increasingly used in farmers fields and agricultural experiments to quantify important traits. Imaging setups with a sub-millimetre ground sampling distance enable the detection and tracking of plant features, including size, shape, and colour. Although todays AI-driven foundation models segment almost any object in an image, they still fail for complex plant canopies. To improve model performance, the global wheat dataset consortium assembled a diverse set of images from experiments around the globe. After the head detection dataset (GWHD), the new dataset targets a full semantic segmentation (GWFSS) of wheat organs (leaves, stems and spikes) covering all developmental stages. Images were collected by 11 institutions using a wide range of imaging setups. Two datasets are provided: i) a set of 1096 diverse images in which all organs were labelled at the pixel level, and (ii) a dataset of 52,078 images without annotations available for additional training. The labelled set was used to train segmentation models based on DeepLabV3Plus and Segformer. Our Segformer model performed slightly better than DeepLabV3Plus with a mIOU for leaves and spikes of ca. 90%. However, the precision for stems with 54% was rather lower. The major advantages over published models are: i) the exclusion of weeds from the wheat canopy, ii) the detection of all wheat features including necrotic and senescent tissues and its separation from crop residues. This facilitates further development in classifying healthy vs. unhealthy tissue to address the increasing need for accurate quantification of senescence and diseases in wheat canopies.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (27 May 2025) · CC BY 4.0

A 1,096-image pixel-labelled wheat organ dataset plus 52,078 unlabelled field images from 11 institutions enables Segformer leaf and spike segmentation near 90% mIoU, while stems lag.

Prior wheat image datasets mostly address one organ or task (head detection, green canopy, disease) with limited geographic, genotypic and growth-stage coverage, and colour-based or foundation segmenters struggle with dense canopies containing senescent tissue, weeds and residues. GWFSS extends the head-focused GWHD effort to joint pixel-level labelling of leaves, stems and heads across all developmental stages, with baseline models for organ phenotyping.

Key results

  1. ShownGWFSS pools a large unlabelled corpus of proximal RGB wheat canopy images from 67 field sites with a pixel-annotated subset drawn with near-equal institutional representation, spanning emergence through maturity.

    52,078 unlabelled images from 11 institutions; ground sampling distance 0.09 to 0.71 mm per pixel; labelled set of 1,096 holds 110 images from each of seven institutions, 109 from ULiege and UTokyo, 108 from CIMMYT, picked near k-means centres of ImageNet-pretrained ResNet features; crops 512 × 512 (UTokyo 1024 × 1024). · 1 Introduction; 2.1 Field Experiments; 2.3 Data Selection for the Annotation Pool; 4.1 The Challenges of Organ Labelling in Complex…

  2. ShownSegformer (B1) outperformed DeepLabV3Plus (R101) under both split regimes, with the margin widening under cross-region distribution shift and concentrated in the head and stem classes.

    mIoU higher by 2.8% in the random 70/10/20 split and by 8.5% in the region split (training on seven institutions, UTokyo for validation, UQ as test); background and leaf scores differed little between models in the region split; both models dropped noticeably from random to region split. · 2.5.1 Impact of Distribution Shift on…; 3.3 Baseline Segmentation Models

  3. ShownAblations on training-set size and model size suggest the current labelled set roughly saturates Segformer for leaves and heads, while stem accuracy plateaus at a low level and the largest model degrades, which authors attribute to too little training data.

    SegFormer-B0 trained on 1% to 100% subsets improved sharply up to 60% (460 images), then levelled off; heads and leaves passed IoU 75% with 10% of data; stems plateaued near IoU 40%. mIoU rose from B0 to B4 but fell for B5, whose parameter space nearly doubles that of B4. · 2.5.2 Impact of Training Data Scale and Model…; 3.3 Baseline Segmentation Models; 4.4 Baseline Model to Guide the Size of the…

  4. ShownStems are the weakest class. The authors attribute this to thin, partly occluded structures, a small pixel share, loss of the cylindrical stem shape during booting, and label noise from confusing stems with soil residues or senescent and rolled leaves.

    Chosen Segformer: stem IoU 44.92% vs IoU 80% for leaves, heads and background; abstract cites ca. 90% for leaves and spikes and 54% stem precision. Stems occur in 847 and heads in 739 images vs 1,080 (leaves) and 1,090 (background), and both occupy few pixels. · Abstract; 3.2 Balance of Developmental Stages and Labelled…; 4.4 Baseline Model to Guide the Size of the…

  5. SuggestedThe model assigned weeds and detached crop residues to background and still segmented chlorotic or necrotic attached wheat tissue, although accuracy declined in senescing canopies.

    Every weed in the five weed-containing random-split test images was predicted as background (visual review only). On 73 early-senescence-to-maturity test images, background IoU was 73.5 vs 84.5 and leaf IoU 73.7 vs 82.3 on the full test set. · 3.4 Visual Inspection of Segmented Images; 4.4 Baseline Model to Guide the Size of the…

Methods, data and reproducibility

Proximal RGB field images of wheat from 11 institutions were pooled; 1,096 chosen by ResNet-feature k-means stratified sampling were centre-cropped and pixel-labelled as leaf, stem, head or background (weeds, residues as background). DeepLabV3Plus and SegFormer baselines were compared under random and cross-region splits with data-size and model-size ablations.

  • GWFSS v1.0 labelled: field RGB images with pixel masks for leaf, stem, head and background, stratified across contributing institutions · wheat · n = 1,096 images · 10.3929/ethz-b-000734546
  • GWFSS v1.0 full: unlabelled proximal RGB wheat canopy images from 67 field sites covering all growth stages, intended for self-supervised pretraining · wheat · n = 52,078 images · 10.3929/ethz-b-000734546
  • MLCAS2025 challenge set: unlabelled 512×512 crops plus labelled images sampled from all institutions except Arvalis · wheat · n = 65,000 crops from 2,000 images; 100 labelled
  • Late-season subset (early senescence to maturity) drawn from the random-split test set · wheat · n = 73 images
  • CIMMYT CENEB trials, Ciudad Obregon, Mexico, 2020-2021: elite, pre-breeding and exotic germplasm under irrigated, drought and heat stress · spring wheat · n = 319 genotypes
  • Rothamsted Field Scanalyzer images from tillering to maturity, including nitrogen-input (NIT) and mapping-population subsets · wheat · n = 391 genotypes
  • ICARDA Merchouche (Morocco) durum trials: 2023 breeding-program entries imaged with PHENOBUGGY, plus 3 trials in 2024 (wild-relative panel, landrace panel, root rot) · durum wheat · n = 960 entries (2023); 60, 288 and 24 lines (2024)
  • NJAU Xinxiang trial 2022-2023 with introduced and domestic cultivars · wheat · n = 565 cultivars in 640 plots
Methods & tools
DeepLabV3Plus (R101 backbone), SegFormer (Mix Transformer MiT-B0 to MiT-B5), ResNet-50 pretrained on ImageNet (feature extraction), k-means clustering (annotation sample selection), UMAP, V7 Darwin annotation tool, BRENDA Tissue Ontology (BTO), EMBL-EBI Ontology Lookup Service, Mean IoU and mean pixel accuracy metrics, Imaging platforms incl. LITERAL, Phenomobile V2, Field Scanalyzer, UFPS, PhenoArm, PHENOBUGGY
Reproducibility gaps
Labelled images withheld at preprint time for the MLCAS25 challenge, with conflicting release dates (2025-06-25 vs July 15, 2025); challenge subsets excluded from the publication; no training code repository cited.
Limitations (6)
  • Stem segmentation is poor (IoU 44.92%); authors cite thin, occluded stems, low pixel share, loss of cylindrical shape during booting and label noise, and suggest class-imbalance losses or many more stem masks. · stated by the authors
  • Developmental stages were estimated visually from images rather than recorded in the field, and stage balance was not achieved: emergence has 8 images versus 282 for early senescence. · stated by the authors
  • Awns were treated as invisible and sheath, peduncle and bare stem were merged into one stem class; authors note reliable awn labels would likely need GSD below 0.2 mm and that these image-defined organs do not map cleanly onto ontology terms. · stated by the authors
  • Accuracy drops under the region split (UQ held out) and in senescing canopies, and authors judge the labelled data insufficient for SegFormer-B5 and for substantial stem gains without far more labels or other training strategies. · stated by the authors
  • Claimed advantages over published wheat or vegetation segmenters (EWS, VegAnn) rest on qualitative inspection; no head-to-head metrics against those models are reported, and weed handling was checked on only five test images. · noted in this summary
  • Figures differ across sections (Results equate 60% of the training subset with 460 images while Discussion cites ca. 600 as 60% of all data; abstract gives ca. 90% for leaves and spikes vs IoU 80% in Discussion), and results come from single splits with no variance across seeds. · noted in this summary

Read this if you build wheat or cereal canopy segmentation models, need diverse multi-site field images for self-supervised pretraining, or quantify organ-level senescence from proximal RGB imagery.

Citations
21
FWCI
11.43
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISNajafian '22Anderegg '23Pound '17Boss '26Miao '20Yasrab '20
68% semantic matchPlant Biology
K. Najafian, A. Ghanbari, M. S. Kish+5 more·bioRxiv·Aug 9, 2022