AgroOmni

AgroOmni is a multi-view multimodal agricultural QA dataset covering remote sensing, plots, orchards, pest and disease recognition, crop growth status, and temporal change understanding.

Links

• Code (SFT training): https://github.com/Jrookie2005/AgroOmni
• Paper: AgroOmni: A Large-Scale Multi-View Agricultural Dataset for Cross-Scale Multimodal Reasoning, NeurIPS 2026 (Evaluations & Datasets Track)
• Evaluation benchmark: AgroMind

Dataset Overview

• Training set: 268,119 QA pairs
• View distribution:
  • GROUND: 192,800
  • SATELLITE: 56,732
  • UAV: 18,587
• Image-count distribution:
  • Single-image: 234,786
  • Two-image: 14,870
  • Four-image: 11,948
  • Five-image: 6,515
• Unique images: 97,926
• Main source prefixes:
  • Pest, Leaf_diseases, Agriculture, Wheat, Oil_palm_trees, Fruit, Trees, Rural, CropHarvest, EarthVQA, crop, Tomato, CropLand

Task Taxonomy

The taxonomy spans 4 dimensions, 14 task types, and 56 fine-grained task categories.

1 Spatial Perception

1.1 Spatial Localization

• 1.1.1 Determine the main aggregation area of trees
• 1.1.2 Locate the type of abnormal regions
• 1.1.3 Judge the distribution area of fruits
• 1.1.6 Tomato boundary localization

1.2 Relative Spatial Relationship Reasoning

• 1.2.1 Judge the spatial relationship between two objects (8 directional relations)
• 1.2.2 Select images that satisfy the given spatial relation

1.3 Field Boundary Localization

• 1.3.1 Field bounding box delineation
• 1.3.2 Field bounding box selection
• 1.3.3 Four-option image selection

2 Object Understanding

2.1 Object Category Recognition

• 2.1.1 Plant recognition
• 2.1.2 Fruit recognition
• 2.1.3 Pest insect recognition
• 2.1.4 Pest discrimination
• 2.1.5 Fruit discrimination
• 2.1.7 Farmland existence discrimination

2.2 Pest and Disease Diagnosis

• 2.2.1 Leaf disease type identification
• 2.2.2 Disease severity assessment
• 2.2.3 Leaf disease type identification (four-image selection)

2.3 Crop Growth Status Recognition

• 2.3.1 Oil palm growth status assessment
• 2.3.2 Count fruits within a specified radius
• 2.3.3 Tomato ripeness estimation in images
• 2.3.4 Wheat growth status identification
• 2.3.5 Select images showing specific wheat growth stages

3 Scene Understanding

3.1 Scene Comparison

• 3.1.1 Select the image with the highest coverage rate
• 3.1.2 Select the image with the largest number of trees
• 3.1.6 Select tomatoes with higher ripeness
• 3.1.7 Compare quantitative differences in wheat abundance

3.2 Counting

• 3.2.1 Pest quantity counting (rice leaf diseases and pests)
• 3.2.2 Fruit counting
• 3.2.4 Oil palm counting (text answer)
• 3.2.5 Farmland plot counting
• 3.2.6 Independent plant counting
• 3.2.7 Oil palm counting (multiple-choice)
• 3.2.8 Count the number of farmland plots

3.3 Coverage Statistics and Area Estimation

• 3.3.1 Forest coverage calculation
• 3.3.2 Agricultural land coverage calculation
• 3.3.3 Coverage estimation of specific abnormal areas
• 3.3.4 Cultivated land coverage calculation
• 3.3.5 Forest coverage area estimation
• 3.3.6 Cultivated land area estimation

3.4 Temporal Change Comparison

• 3.4.1 Calculate the cultivated land change rate between two images
• 3.4.2 Judge whether the cultivated land at a given location has changed
• 3.4.3 Compare the consistency of wheat growth status in dual images

4 Scene Reasoning

4.1 Visual Cued Reasoning

• 4.1.1 Canopy diameter estimation in red-boxed regions (multiple-choice)
• 4.1.2 Canopy diameter estimation in red-boxed regions (text answer)

4.2 Anomaly Reasoning (Anomaly Detection & Interpretation)

• 4.2.1 Abnormal region detection (double plants, weed areas, etc.)
• 4.2.2 Identify pest species affecting crops
• 4.2.3 Determine whether specific pests cause crop damage

4.3 Climate Type Reasoning

• 4.3.1 Climate zone classification (multiple-choice)
• 4.3.2 Select images matching specific climate zone characteristics
• 4.3.3 Ecoregion type recognition
• 4.3.4 True/False judgment for climate zone classification

4.4 Planning & Decision Reasoning

• 4.4.1 Describe cultivated land conditions
• 4.4.3 Assess plant occlusion degree
• 4.4.4 Select weeding strategies based on weed coverage
• 4.4.5 Identify farmland types in remote sensing imagery

Directory Structure

AgroOmni/
├── Agriculture.zip
├── CropHarvest.zip
├── CropLand.zip
├── EarthVQA.zip
├── Fruit.zip
├── Leaf_diseases.zip
├── Oil_palm_trees.zip
├── Pest.zip
├── Rural.zip
├── Tomato.zip
├── Trees.zip
├── Wheat.zip
├── crop.zip
├── train.json
├── train_json/
│   └── train_json_<level1>_<level2>_<level3>.json   (56 files)
└── README.md

Each entry in train.json contains:

• image_path — one image, for example ./Trees/images/xxx.jpg
• image1, image2 — two images, for the two-image tasks
• options — for the four-image and five-image tasks the option values are image paths
• question, answer
• type_id, item_id, level1_id, level2_id, level3_id — taxonomy indices

Every image path is relative to the repository root and starts with the source prefix.

The shard names such as train_json_1_1_1.json are aligned with the taxonomy numbering above; for example, 1_1_1 corresponds to 1.1.1. One shard is provided per task category, so the train_json/ folder contains 56 files.

Notes

This dataset is designed for agricultural visual understanding and reasoning. Please keep the original directory structure unchanged so that relative paths continue to work correctly.

AgroOmni aggregates imagery and annotations derived from publicly available agricultural datasets as well as a proprietary remote sensing parcel collection. The upstream source datasets keep their own licenses and terms of use, so please consult each source before redistribution or commercial use; per-source attribution is listed in the paper's appendix.

License and Attribution

Our contribution — the question–answer annotations and the instruction templates built on top of the source annotations — is released under CC BY-SA 4.0.

Images are aggregated from the datasets listed below and keep their original licenses. The copyright of the third-party imagery belongs to its original owners; this project does not hold and cannot grant any rights over it.

Exception — portions that are not for commercial use. Two sources carry stricter terms, and the content derived from them inherits those terms:

• EarthVQA — CC BY-NC-SA 4.0 (the imagery originates from Google Earth and is additionally subject to Google's terms of use)
• Pest — IP102, declared by its authors as free for academic use only

Please check the original license of a source before using the corresponding portion commercially. When you use a portion of the data coming from a specific source, please also cite that source's original paper.

Source prefix	Original dataset	License	Link
Pest	IP102 (2019)	Academic research use only	https://github.com/xpwu95/IP102
Leaf_diseases	2018 AI Challenge (2018)	No license stated	https://aistudio.baidu.com/datasetdetail/76075
Wheat	GWHD 2021 (2021)	CC BY 4.0	https://zenodo.org/records/5092309
Agriculture	Agriculture-Vision (2020)	Custom terms and conditions	https://github.com/SHI-Labs/Agriculture-Vision
Oil_palm_trees	OilPalmUAV / MOPAD (2021)	No license stated	https://github.com/rs-dl/MOPAD
Fruit	ACFR Orchard Fruit (2016)	No license stated	https://data.acfr.usyd.edu.au/ag/treecrops/2016-multifruit/
Trees	OAM-TCD (2024)	CC BY 4.0	https://zenodo.org/records/11617167
Rural	Self-collected parcel imagery (2025)	Released with this project	—
CropHarvest	CropHarvest (2021)	CC BY-SA 4.0	https://zenodo.org/records/5828893
EarthVQA	EarthVQA (2024)	CC BY-NC-SA 4.0	https://github.com/Junjue-Wang/EarthVQA
CropLand	CLCD (2022)	No license stated	https://github.com/liumency/CropLand-CD
crop	PhenoBench (2023)	MIT	https://github.com/PRBonn/phenobench
Tomato	Tomato Dataset (2023)	No license stated	https://www.sciengine.com/CSD/doi/10.11922/11-6035.csd.2023.0154.zh

Citation

@inproceedings{zhang2026agroomni,
  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},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}