ImAg4Wheat: Wheat Image Corpus for Pretraining

• ImAg4Wheat is a wheat-specific image corpus with 2.5M high-resolution RGB images that capture diverse genotypes, environmental conditions, and crop stages.
• The dataset supports self-supervised pretraining by addressing the unique visual challenges of wheat canopies, enabling improved segmentation and classification.
• Data acquisition spans multiple countries and devices with hybrid-scale cropping to preserve fine canopy structures and ensure robust model transfer.

ImAg4Wheat is the wheat-specific image corpus introduced as the data foundation for the vision model family "FoMo4Wheat." It is described as the largest and most diverse wheat image dataset to date, comprising 2.5 million high-resolution RGB images collected over 2010–2024, spanning more than 2,000 wheat genotypes, more than 500 environmental conditions, and the full crop cycle from emergence to maturity. Its primary role is not that of a conventional single-task benchmark; rather, it is used as a large unlabeled corpus for self-supervised pretraining of wheat-domain visual representations, with the stated goal of reducing the mismatch between general-domain vision pretraining and in-field wheat imagery (Han et al., 8 Sep 2025).

1. Definition, rationale, and problem setting

ImAg4Wheat was created in response to a specific limitation of mainstream visual pretraining: models initialized on general-domain corpora such as ImageNet or DINOv2 do not reliably capture the visual statistics of wheat canopies and organs under field conditions. The FoMo4Wheat paper characterizes wheat imagery as dominated by dense, repetitive structures, subtle organ boundaries, heavy occlusion, strong genotype-by-environment variation, and appearance shifts induced by illumination, weather, and phenological stage. In that formulation, the difficulty is not only object recognition, but the joint modeling of fine canopy architecture and unstable field context (Han et al., 8 Sep 2025).

The dataset is therefore positioned as a crop-domain pretraining corpus rather than a narrowly annotated benchmark. The paper explicitly treats ImAg4Wheat as the substrate for self-supervised learning on wheat morphology, canopy organization, and field variability. A central claim is that wheat-specific pretraining yields representations that are more robust on wheat tasks and still transferable to other agricultural tasks, including rice and crop–weed segmentation (Han et al., 8 Sep 2025).

This framing places ImAg4Wheat within a broader shift in agricultural computer vision from task-specific pipelines toward domain-adapted foundation models. A plausible implication is that the dataset is intended not merely to support one downstream application, but to define a reusable visual prior for in-field wheat perception.

2. Data composition, acquisition, and curation

The corpus is reported as containing 2.5 million high-resolution RGB images acquired from breeding and experimental wheat fields. The main text states that the data come from 30 sites, while the abstract and results mention 10 countries and the supplementary material lists 29 regions spanning 9 countries. The methods text explicitly names Australia, China, Ethiopia, France, India, Japan, Switzerland, United Kingdom, and United States. The manuscript thus presents ImAg4Wheat as a globally distributed wheat corpus, although site, region, and country counts are reported with different conventions (Han et al., 8 Sep 2025).

Spatial resolution is unusually fine for a field-scale corpus: the reported ground sampling distance ranges from 0.1 mm to 0.4 mm. Native image resolutions vary by device and site, with examples including 4800 × 3200, 5184 × 3456, 4032 × 3024, 1200 × 800, 1024 × 1024, and 1083 × 719. The image sources are heterogeneous and include field plot-scale imaging, plant-scale monitoring of individual growth dynamics, multi-angle canopy observations, and organ-scale imaging of leaves, stems, and spikes. Most images come from non-public institutional sources; the paper identifies FIP 1.0 as the principal public component integrated into the corpus (Han et al., 8 Sep 2025).

The acquisition ecosystem is also multi-device. Devices listed in the supplementary material include Sony RX0, Canon EOS X5, Hunting camera 800M, PhenoArm, Sigma SD14, GECKOCAM, LITERAL, iPad Pro, EOS 5D Mark III, smartphones, and custom Android application systems. This multi-platform construction introduces natural variation in optics, resolution, viewpoint, and color rendering. The paper treats that heterogeneity as an asset for robustness rather than a nuisance to be eliminated (Han et al., 8 Sep 2025).

Standardization is handled mainly through preprocessing for pretraining. The paper reports a hybrid-scale random cropping strategy with crop sizes from 512 × 512 to 1024 × 1024 pixels, used to preserve fine spatial structure while balancing training samples across regions, years, and spatial scales. The dataset is described primarily as an unlabeled RGB corpus for self-supervision. Although the discussion refers to captured and annotated images, the paper does not define a unified annotation schema for ImAg4Wheat itself; detailed labels are concentrated in separate downstream benchmark datasets (Han et al., 8 Sep 2025).

3. Role in FoMo4Wheat pretraining

Within FoMo4Wheat, ImAg4Wheat is used to pretrain a plain ViT backbone in a DINOv2-style teacher–student framework combining distillation and masked image modeling. The paper defines patch tokens as

$$X_p = [x_1, x_2, \ldots, x_N]^T \in \mathbb{R}^{N \times d},$$

with

$N = (H/14)\times(W/14).$

It then appends learnable tokens and positional embeddings before standard Pre-LN transformer blocks. Self-attention is written as

$$\mathrm{Attn}(X)=\mathrm{Softmax}\!\left(\frac{XW_Q (XW_K)^T}{\sqrt{d_k}}\right) XW_V,$$

and multi-head attention as

$$\mathrm{MHA}(X)=\mathrm{Concat}(\mathrm{Attn}_1(X),\mathrm{Attn}_2(X),\ldots,\mathrm{Attn}_H(X))W_O.$$

The block update is given as

$$X_{\text{out}} = \tilde{X} + \mathrm{FFN}(\mathrm{LayerNorm}(\tilde{X})), \qquad \tilde{X} = X + \mathrm{MHA}(\mathrm{LayerNorm}(X)).$$

The self-supervised objectives are expressed through a patch-level reconstruction term and a class-token distillation term:

$$\mathrm{LOSS}_{\text{rec}} = \mathrm{CE}(\mathrm{Teacher}(X_u)_{\text{patch}}, \mathrm{Student}(X_v)_{\text{patch}})$$

and

$$\mathrm{Loss}_{\text{dis}} = \mathrm{CE}(\mathrm{Teacher}(X_u)_{\text{cls}}, \mathrm{Student}(X_v)_{\text{cls}}).$$

Here, $X_u$ denotes an unmasked view and $X_v$ a masked view. The paper interprets this combination as coupling contrastive/distillation-style invariance learning with masked image modeling for local detail recovery (Han et al., 8 Sep 2025).

A domain-specific modification concerns the Sinkhorn–Knopp clustering loss used in original DINOv2 recipes. The paper reports that repetitive foliage textures and weak foreground–background separation in agricultural images make this clustering term less reliable, and therefore its weight is reduced or the term is removed entirely in some settings. This is one of the clearest statements that the visual statistics of wheat imagery require adaptation of generic self-supervised objectives (Han et al., 8 Sep 2025).

Training scale is correspondingly large. Pretraining was conducted on a 6-node HPC cluster with 48 A100-80GB GPUs under distributed mixed precision. The reported schedule uses 224 × 224 global crops for an initial stage of about five days, followed by transfer to 518 × 518 patches for approximately three additional days. The largest model, FoMo4Wheat-Giant, is initialized from DINOv2 weights; the smaller Large and Base variants are then obtained by distillation from the Giant model rather than by direct training from scratch (Han et al., 8 Sep 2025).

4. Downstream benchmark ecosystem built around the corpus

Although ImAg4Wheat itself functions mainly as a pretraining corpus, the FoMo4Wheat study evaluates it through 10 downstream tasks spanning canopy-level wheat classification, organ-level wheat perception, rice transfer, and mixed crop–weed segmentation. The benchmark suite is as follows (Han et al., 8 Sep 2025):

Level	Task	Dataset and reported split
Canopy-level wheat	Growth stage classification	WGSP, 3,353 images, 671 train / 2,682 test
Canopy-level wheat	Disease classification	Combined CerealConv, WFD, AWDD, MSWDD, LWDCD, 4,000 images, 320 train / 3,680 test
Organ-level wheat	Ground-based head detection	GWHD, 4,700 RGB images, 190,000 labeled heads, 3,607 train / 1,448 val / 1,382 test
Organ-level wheat	UAV head detection	Self-collected, 100 images at GSD 0.6 mm and 120 images at GSD 1.2 mm
Organ-level wheat	Wheat leaf counting	Self-collected, 1,887 images, 1,508 train / 379 test
Organ-level wheat	Wheat organ segmentation	GWFSS, 1,096 images, 876 train / 220 test
Rice transfer	Rice leaf counting	Self-collected, 1,700 images, 1,360 train / 340 test
Rice transfer	Rice organ segmentation	RiceSEG, 3,078 images, 2,462 train / 616 test
Multi-crop transfer	Multi-crop segmentation	VegAnn, 3,775 images, 3,020 train / 755 test
Crop–weed transfer	Crop and weed segmentation	CropAndWeed Fine24, 7,705 images, ~112,000 instances, 6,164 train / 1,541 test

This benchmark composition is significant because it makes explicit how ImAg4Wheat differs from earlier wheat datasets. Datasets such as GWHD and GWHD 2021 were designed primarily for wheat-head localization and cross-domain detection benchmarking, with box annotations and domain-structured splits, whereas ImAg4Wheat is used upstream as a domain-scale pretraining corpus feeding a multi-task evaluation stack (David et al., 2020, David et al., 2021).

5. Empirical performance and transfer behavior

The reported gains are consistent across all 10 tasks, but they are not uniform in magnitude. The largest improvements appear in segmentation, especially on fine plant structures. On wheat organ segmentation, the FoMo4Wheat-Giant model reaches 76.49 mIoU against a 70.57 mIoU DINOv2-based baseline, with stem IoU improving from 37.48 to 52.73. On rice organ segmentation, the corresponding Giant result is 74.54 mIoU versus 69.97. On multi-crop segmentation, the Giant model reaches 90.71 mIoU versus 78.36, and on crop-and-weed segmentation it reaches 69.99 mIoU versus 64.45 (Han et al., 8 Sep 2025).

Classification gains are smaller in absolute terms but remain systematic. For growth stage classification at full training data, the paper reports 96.8 mAP for FoMo4Wheat-Giant versus 95.6 for the SOTA baseline; for wheat disease classification, it reports 96.5 versus 92.9. The low-label regime is more revealing. At 25% of training data, the FoMo4Wheat-Large model reaches 95.1 mAP on growth stage classification versus 91.4, and 88.9 on disease classification versus 78.8. The paper states that FoMo4Wheat can match state-of-the-art classification performance using only about 30% of training data on these wheat classification tasks (Han et al., 8 Sep 2025).

Detection results show a different pattern. On ground-based GWHD wheat spike detection at GSD 0.4 mm, the Giant model improves from 36.0 AP to 36.6 AP. The relative advantage is larger under cross-platform transfer: on UAV imagery at GSD 0.6 mm, Giant improves from 34.2 AP to 35.4 AP, and at GSD 1.2 mm from 27.4 AP to 28.3 AP. The paper interprets this as evidence that wheat-specific pretraining improves robustness under unseen acquisition conditions rather than only in-distribution accuracy (Han et al., 8 Sep 2025).

Counting shows positive but more modest gains. On wheat leaf counting, the Giant model improves from 14.26 MSE, 0.88 $R^2$ to 14.11 MSE, 0.89 $N = (H/14)\times(W/14).$0; on rice leaf counting, it improves from 15.17 MSE, 0.90 $N = (H/14)\times(W/14).$1 to 14.72 MSE, 0.91 $N = (H/14)\times(W/14).$2. The Base model exhibits the clearest relative improvement, which the paper takes as evidence that domain-specific pretraining can compensate for reduced model scale (Han et al., 8 Sep 2025).

Taken together, these results support two conclusions emphasized by the authors. First, the main benefit of ImAg4Wheat pretraining is strongest in structurally demanding dense-prediction tasks. Second, the learned representation is not limited to wheat-only transfer; the same wheat-domain pretraining improves rice and mixed crop–weed tasks as well.

6. Research position, limitations, and significance

ImAg4Wheat occupies a distinct position within the wheat-vision literature. Earlier datasets typically targeted specific supervised tasks: GWHD and GWHD 2021 concentrated on wheat-head localization from high-resolution RGB field images, while WisWheat introduced a three-tier wheat vision–language resource for captioning, quantitative VQA, and management-oriented instruction tuning. ImAg4Wheat differs from both by prioritizing large-scale wheat-domain self-supervision over a unified annotation taxonomy, and by treating dataset scale and diversity as prerequisites for a wheat-specific foundation model (David et al., 2020, David et al., 2021, Yuan et al., 6 Jun 2025).

The paper nevertheless identifies several open issues. Geographic breadth is substantial but not yet equivalent to a universal crop corpus. The manuscript itself contains discrepancies in whether the corpus spans 30 sites, 29 regions, 9 countries, or 10 countries. The model side also reveals a trade-off between domain specialization and broader universality: the authors explicitly present FoMo4Crop as a future direction requiring still larger and more diverse multi-crop data. On the optimization side, the instability of the Sinkhorn–Knopp clustering term under repetitive crop textures is a reminder that agricultural imagery does not always conform to assumptions inherited from natural-image pretraining. Finally, deployment remains nontrivial: even the Base model is described as heavy enough that compression, quantization, and further distillation are still relevant (Han et al., 8 Sep 2025).

Despite those limitations, ImAg4Wheat is important for two reasons. Methodologically, it operationalizes the idea that crop vision can benefit from domain-scale self-supervised pretraining rather than repeated dependence on general-domain backbones. In dataset terms, it reframes wheat imaging from a collection of isolated task datasets into a broad morphological and environmental corpus spanning genotype, site, year, device, and phenological stage. The dataset and models are reported as publicly available through the PheniX-Lab GitHub repository and Hugging Face release, which further positions ImAg4Wheat as infrastructure rather than a one-off benchmark (Han et al., 8 Sep 2025).