AgroOmni

AgroOmni is a large agricultural training set for vision-language models that mixes three views: ground photos, UAV (drone) images and satellite images. Most models only learn from ground-level photos and get confused at other scales, for example mistaking farmland for a wall or floor. AgroOmni has 288,831 question and answer pairs over 107,488 images, in 56 task categories across 14 task types.

By pair count, 68.0% of the data is ground view, 24.1% satellite and 7.9% UAV. The pairs were made from the source labels with rule-based templates and with logic built from the raw labels, then filtered and checked by hand. The data was also cleaned against the AgroMind benchmark so the two do not overlap.

The authors trained AgroNVILA on it. It scored 62.32% on AgroMind, 15.03 points above GPT-5.2.

• Paper: AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning
• Original dataset: anonymous7777/AgroOmni
• License: the original dataset card marks it as other and does not name a license.

Tasks

Dimension	Task types
Spatial Perception	Spatial Localization, Spatial Relationship, Boundary Detection
Object Understanding	Object Classification, Pest/Disease Diagnostics, Growth Status Recognition
Scene Understanding	Counting, Area Statistics, Scene Comparison, Temporal Variation
Scene Reasoning	Visual Prompt Reasoning, Anomaly Reasoning, Climate Type Reasoning, Planning

Each row's task_file in raw_metadata names the source file it came from, one per task category. The numbers in the file name match the task codes in the original dataset card, for example train_json_1_1_1.json holds task 1.1.1.

View	Rows
GROUND	196,293
SATELLITE	69,713
UAV	22,825

255,396 rows have one image, 14,870 have two, 12,050 have four and 6,515 have five.

Structure

train-0000-of-0001.parquet
images/
  image-train-000-of-008.zip ... image-train-007-of-008.zip
  path_to_shard.parquet

Each row has:

• images: list of {"bytes": null, "path": <sha256>}, in the order the images appear in the question. The path is the SHA256 of the image bytes and points into a zip shard through images/path_to_shard.parquet.
• id: agroomni_train_<counter>, for example agroomni_train_0.
• messages: one user turn (question with images) and one assistant turn (answer).
• raw_metadata: JSON string with view_type (GROUND, UAV or SATELLITE), task_file and raw_image_paths.

Example:

{
  "images": [
    {"bytes": null, "path": "8df98fdbcb2483eaba2d5be169859db0f60c106b461f32fbaf8622440089bc9f"},
    {"bytes": null, "path": "fe73274fdc1728d4607631dcc39744aff34679743ebf39a394a0b0b38aa878e1"}
  ],
  "id": "agroomni_train_0",
  "messages": [
    {"role": "user", "content": [
      {"type": "text", "text": "image1: "},
      {"type": "image", "text": null},
      {"type": "text", "text": ", image2: "},
      {"type": "image", "text": null},
      {"type": "text", "text": "Compare the pictures of two time periods to determine whether there is a mutual transformation between farmland and non-farmland at the specified coordinate pixel (351, 444)? (The upper left corner is (0,0); x is horizontal, y is vertical; The image is 512x512 in size.)"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "Yes"}]}
  ],
  "raw_metadata": "{\"view_type\": \"SATELLITE\", \"task_file\": \"train_json_3_4_2.json\", \"raw_image_paths\": [\"./CropLand/time1/00000.png\", \"./CropLand/time2/00000.png\"]}"
}

Changes from the original data

• Built from train.json. All 288,831 rows are kept, matching the count in the paper, including 3 exact duplicate rows in the source.
• task_file comes from the train_json/ folder, which holds the same rows split into one file per task.
• Turns are converted from from/value to role/content: human becomes user and gpt becomes assistant. All other text is kept as in the source.
• Each <image> token in the question becomes an image item at the same place. Single-image questions start with <image> and a line break, so the image comes first and that line break is dropped. In multi-image questions the images stay between the text, so labels such as A. or image1: stay next to their image.
• raw_metadata keeps view_type from the source. The image paths as written are kept in raw_image_paths.
• All 117,158 source image files are packed, including 9,670 that no row uses, such as tree masks, extra tomato images, sample images and cropland labels. Images are stored once, deduplicated by SHA256, which leaves 115,035 unique images. The rows use 106,548 of them.

Usage

from agml import loadImageTextToTextDataset

dataset = loadImageTextToTextDataset("Project-AgML/AgroOmni")

sample = dataset["train"][0]
sample["images"]    # list of PIL images
sample["messages"]

Citation

@misc{zhang2026agroomni,
      title={AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning},
      author={Jiarui Zhang and Junqi Hu and Zurong Mai and Yang Liu and Yuhang Chen and Shuohong Lou and Henglian Huang and Hong Cheng and Lingyuan Zhao and Jianxi Huang and Yutong Lu and Haohuan Fu and Juepeng Zheng},
      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 Project AgML @ UC Davis. 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.