AgroOmni

AgroOmni is a large-scale multi-view agricultural dataset introduced in "AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning" (arXiv:2603.14342). It spans ground, UAV, and satellite imagery to address the ground-level bias of existing agricultural multimodal models, with 288K visual question answering pairs covering 56 specialized task categories across 14 task types. The paper's proposed model, AgroNVILA, reports 62.32% on the AgroMind benchmark (+15.03% over GPT-5.2).

This repo: 288,831 QA rows over 107,488 unique images (35.9GB total). Rows are deduplicated by image path — many rows reference the same underlying image (e.g. multi-turn or multi-region QA over one image), so images are stored once and referenced by path.

Not auto-loadable via load_dataset(). Requires the companion loader (load_agml_dataset()), which reads images lazily and directly out of the zip shards below — no extraction step. The configs: block above documents the dataset's config/split layout for the Hub dataset card and viewer; it does not make load_dataset() itself work for this repo.

Layout

train-0000-of-0001.parquet   # 288,831 rows: id, messages, raw_metadata, images
images/                      # 7 shards, ZIP_STORED (uncompressed), ~5GB target each
  image-train-000-of-007.zip
  ...
  image-train-006-of-007.zip
  path_to_shard.parquet              # 107,488 rows: path -> shard_file

images holds {"bytes": None, "path": ...} per referenced image (original relative path, e.g. CropLand/time1/00000.png) instead of embedded bytes. path_to_shard.parquet maps each unique path directly to the shard filename it lives in.

Configs

config	split	rows	file pattern
train	train	288,831	train-*.parquet

Only one config exists today. A dataset with multiple configs (e.g. one per category) would repeat this whole layout — its own metadata parquet(s) + its own images/ folder of shards — nested under a folder named for that config, with a matching entry added to configs: above.

Source datasets

AgroOmni is assembled from 14 existing agricultural datasets/challenges, spanning ground photography, UAV imagery, and satellite/remote sensing:

Source dataset	Contribution
ACFR Orchard Fruit Dataset (2016)	Almond, apple, mango instance detection with fruit radius annotations
2018 AI Challenger Dataset	Crop leaf disease images, 3-level severity labels across 26 diseases and 9 crops
IP102 Dataset (2019)	Large-scale pest collection, VOC-format boxes for 102 insect species
Agriculture-Vision Challenge (2020)	Multi-spectral aerial imagery with pixel-level masks for six field anomalies
OilPalmUAV Dataset (2021)	High-resolution UAV imagery with bounding boxes for oil palm growth stages
CropHarvest Dataset (2021)	Global remote sensing time-series with satellite imagery and climate data for crop classification
GWHD Dataset (2021)	Global wheat head images with bounding box annotations and growth stage metadata
CLCD Dataset (2022)	Bi-temporal satellite image pairs for cropland change detection
PhenoBench Dataset (2023)	UAV-captured agricultural field images with crop/weed semantic segmentation
Tomato Dataset (2023)	Multi-illumination tomato images with maturity labels and bounding boxes
EarthVQA Dataset (2024)	High-resolution remote sensing imagery with 8-category semantic masks
OAM-TCD Dataset (2024)	Global aerial imagery with MS-COCO format annotations for tree cover mapping
Maize Tassel Identification Dataset (2025)	Field imagery with bounding boxes for yield estimation
Private Parcel Dataset (2025)	Internal multi-spectral GeoTIFFs and shapefiles for farmland analysis

(Per the source paper, §Data; see the paper for full construction and annotation details.)

Usage

from agml import loadImageTextToTextDataset

ds = loadImageTextToTextDataset("Project-AgML/AgroOmni", token=HF_TOKEN)
print(ds)              # DatasetDict({'train': ...})
ds["train"][0]         # images decoded lazily on access

Citation

If you use this dataset, please cite the original AgroOmni paper:

@misc{agroomni2026,
  title  = {AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning},
  author = {Zhang, Jiarui and Hu, Junqi and Mai, Zurong and Liu, Yang and Chen, Yuhang and Lou, Shuohong and Huang, Henglian and Cheng, Hong and Zhao, Lingyuan and Huang, Jianxi and Lu, Yutong and Fu, Haohuan and Zheng, Juepeng},
  year   = {2026},
  eprint = {2603.14342},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url    = {https://arxiv.org/abs/2603.14342}
}

---

This dataset is indexed and structured on https://project-agml.github.io/ as part of the AgML python library. This dataset was reformatted from its original format to match HuggingFace's Imagefolder standards but requires an external module (agml) that processes and returns a HF Dataset object faster than HF module functions.