AgriField-40K

AgriField-40K is a field-centric agricultural dataset curated from 17 publicly available sources, containing 39,963 RGB images. It is designed for visual representation learning, parameter-efficient continual pretraining, and self-supervised learning in real-world agricultural field settings.

Unlike leaf-centric or controlled-environment plant datasets, AgriField-40K focuses exclusively on field-centric imagery captured under real-world agricultural conditions.

Key Features

• Scale & Diversity: 39,963 images covering over 26 crop species, dozens of weed types, mixed vegetation, pastures, and soil clutter.
• Acquisition Platforms: Captured across multiple sensors — handheld cameras, ground robots, UAV/drones, and shrouded field platforms.
• Environmental Variation: Diverse growth stages, seasonal changes, lighting conditions, and geographic regions.
• Preprocessed for Self-Supervised Learning: Standardized aspect-ratio scaling to 512×512 resolution, temporal de-duplication, and quality filtering.

Dataset Structure & Splits

agrifield40k/
├── train/          # ~80% split (32,136 images)
│   ├── acw_rgb-2022-10-06-17-16-49.jpg
│   ├── acw_rgb-2022-10-06-17-16-51.jpg
│   └── ...
└── val/            # ~20% split (7,827 images)
    ├── acw_rgb-2022-10-06-17-39-39.jpg
    └── ...

Source Datasets

AgriField-40K aggregates and curates images from the following 17 public resources:

Dataset	Year	License	Size	Retained	Domain	Acquisition	Task
MuST-C	2026	CC BY 4.0	7,242	7,242	Sugar Beet, Soybean, Potato, Maize, Wheat, Intercrop	Robot	--
VCD	2022	CC BY 4.0	2,258	2,258	Maize, Bean (Early Stage)	Leek Shrouded Platform	Detection
PalmerAmaranth	2023	CC BY 4.0	614	516	Palmer Amaranth (8 Stages)	H. Cameras	Detection
ACRECropWeed	2023	CC BY 4.0	1,000	791	Maize, Beans, 4 Weeds	Robot	Multi-Task
SorghumWeed	2023	CC BY 4.0	252	172	Sorghum, Grasses, Weeds	H. Cameras	Multi-Task
GrassClover	2019	CC BY-SA 4.0	435	435	Grass, Clover, Weeds	H. Cameras	Segmentation
PhenoBench	2026	CC BY-SA 4.0	29,312	9,606	Sugar Beet, 6 Weeds	Drone	Segmentation
VegAnn	2022	CC BY 1.0	3,775	1,607	26+ Crops	Multiple	Segmentation
Ronin	2021	CC BY 4.0	1,176	135	6 Crops, 8 Weeds	H. Cameras	Detection
LUCASVision	2023	CC BY 4.0	15,876	11,195	12 Crops	H. Cameras	Classification
WE3DS	2023	CC BY 4.0	2,568	1,553	7 Crops, 10 Weeds	Stereo RGB-D	Segmentation
Maize-Weed	2022	CC BY 4.0	843	255	Maize, Weeds	H. Cameras	Detection
RadishWheat	2022	CC BY 4.0	552	534	Wild Radish in Wheat	O. Cameras	Detection
RumexLeaves	2024	CC BY 4.0	809	809	Rumex Obtusifolius	Robot	Detection
SesameWeed	2020	CC0	1,300	1,300	Sesame, Weeds	H. Cameras	Detection
PerennialPlants	2021	MIT	392	240	Weeds in Perennials	H. Cameras	Multi-Task
iNatWeeds	2026	CC BY 4.0	1,315	1,315	Mixed Species	H. Cameras	--
AgriField-40K	2026	CC BY-SA	--	39,963	Field-Centric	Multiple	Pretraining

For the iNatWeeds split, an iNatWeeds_metadata.json file provides the attribution information required under CC BY 4.0.

Processing & Modifications

The source datasets were processed and modified as follows to form AgriField-40K:

• Unsupervised formulation: original labels, bounding boxes, masks, and class annotations were removed to prepare the data for self-supervised learning.
• De-duplication & frame sampling: sequence and video-based datasets were downsampled using fixed frame intervals to remove visual redundancy and near-duplicate frames.
• Quality & relevance filtering: out-of-focus, heavily blurred, corrupt, non-field, or artifact-heavy images were excluded.
• Resizing & center cropping: images were resized using Lanczos interpolation so their shorter edge measures 512px (preserving aspect ratio), followed by a centered 512×512 crop.
• Standardized filenaming: images were renamed using a consistent [dataset_source]_[id] prefix format to guarantee full source tracking back to the original authors.

License & Compliance

The aggregated compilation AgriField-40K is released under CC BY-SA 4.0, covering the rights in the compilation itself and the modifications/contributions made in producing it. Individual source materials remain subject to their respective upstream licenses. Identify the applicable upstream source before reusing or redistributing individual images or subsets.

Upstream licenses by source:

License	Datasets
CC BY-SA 4.0	PhenoBench, GrassClover
CC BY 4.0	MuST-C, LUCASVision, WE3DS, iNat Weeds, VCD, Rumex Leaves, ACRECropWeed, RadishWheat, Palmer Amaranth, Maize-Weed, SorghumWeed, Ronin
CC BY 1.0	VegAnn
MIT	PerennialPlants
CC0 1.0 (Public Domain)	Sesame&Weed

Share-Alike sources (PhenoBench, GrassClover): derived images were incorporated and modified through dataset curation/filtering, (frame) sampling, resizing and center cropping to 512×512, and filename standardization, and remain subject to the original CC BY-SA 4.0 license and its attribution requirements.

VegAnn (CC BY 1.0): derived images were modified as above; original authors retain copyright to their respective contributions.

Original authors retain copyright to their respective contributions across all sources. This dataset is provided "as is", without warranty of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, or non-infringement.

Citation

This dataset is associated with the following paper: arXiv:2608.07984

@article{tzouras2026agrifield,
  title   = {AgriField-40K: Adapting Vision Models to Agriculture With Efficient Continual Pretraining},
  author  = {Tzouras, Vasileios and Pegios, Paraskevas and Nalpantidis, Lazaros},
  journal = {arXiv preprint arXiv:2608.07984},
  year    = {2026}
}