LeafNet

LeafNet is a large image and text dataset of plant leaf diseases, built to train vision-language foundation models for plant pathology. The full dataset has about 186,000 leaf images from 22 crop species, sorted into 97 classes: 62 diseases caused by fungi, bacteria, oomycetes, viruses and mites, plus healthy leaves for each species. Each image is paired with a caption that names the crop and the disease and describes its symptoms. The symptom descriptions were taken from trusted sources such as NIH, NIFA and published studies, and experts reviewed samples of the data.

The images come from public datasets and from the authors' own collection in 7 countries, and most of them are field photos. The authors also built LeafBench, a test set of 2,570 images and 13,950 multiple-choice questions over 6 tasks, and used it to compare vision-language models. Their own model, SCOLD, reached 96.28% on disease identification with no fine-tuning and 99.15% after fine-tuning.

This release is the public training part of LeafNet, about 70% of the full dataset, with 121,337 image and caption pairs.

• Paper: LeafNet: A Large-Scale Dataset and Comprehensive Benchmark for Foundational Vision-Language Understanding of Plant Diseases
• Related paper: A Vision-Language Foundation Model for Leaf Disease Identification (SCOLD)
• Original dataset: enalis/LeafNet
• License: the original dataset card lists cc-by-4.0, but access is gated and users must agree to use the data for non-commercial purposes only (academic research or education).

Structure

train-0000-of-0001.parquet
images/
  image-train-000-of-002.zip ... image-train-001-of-002.zip
  path_to_shard.parquet

Each row has:

• images: list with one {"bytes": null, "path": <sha256>}. The path is the SHA256 of the image bytes and points into a zip shard through images/path_to_shard.parquet.
• id: leafnet_train_<counter>, for example leafnet_train_0.
• messages: one user turn (the image only) and one assistant turn (the caption).
• raw_metadata: JSON string with file_name and raw_image_paths.

Example:

{
  "images": [{"bytes": null, "path": "79bd8128155e97c0f3bd240960b997432e4a6faaa6d15ccc6f207fc1eafcd7d0"}],
  "id": "leafnet_train_0",
  "messages": [
    {"role": "user", "content": [{"type": "image", "text": null}]},
    {"role": "assistant", "content": [{"type": "text", "text": "a image of Tea healthy leaves with leaves appearing normal and healthy"}]}
  ],
  "raw_metadata": "{\"file_name\": \"D:/AI/Dataset/dataset/dataset/000000_Healthy_747.jpg\", \"raw_image_paths\": [\"000000_Healthy_747.jpg\"]}"
}

Changes from the original data

• Built from the 7 parquet files in data/. All 121,337 rows are kept in the same order, in one train split. The paper's 6:2:2 split is not reproduced.
• The source has no question, only an image and a caption. No question was added: the user turn holds just the image, and the assistant turn is the caption.
• Captions are kept exactly as written, including their spacing. There are 89 different captions across all rows.
• raw_metadata keeps file_name as given in the source, which is a path on the authors' machine. The image's own path from the source parquet is kept in raw_image_paths. The caption is already in messages.
• The images were stored inside the source parquet files. They are moved into zip shards and stored once, deduplicated by SHA256, which leaves 80,371 unique images. Every row's image is packed.

Usage

from agml import loadImageTextToTextDataset

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

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

Citation

@article{quoc2026leafnet,
  title={LeafNet: A Large-Scale Dataset and Comprehensive Benchmark for Foundational Vision-Language Understanding of Plant Diseases},
  author={Quoc, Khang Nguyen and Dao, Phuong D and Quach, Luyl-Da},
  journal={arXiv preprint arXiv:2602.13662},
  year={2026}
}

@article{NGUYENQUOC2025130084,
  title = {A Vision-Language Foundation Model for Leaf Disease Identification},
  journal = {Expert Systems with Applications},
  pages = {130084},
  year = {2025},
  issn = {0957-4174},
  doi = {https://doi.org/10.1016/j.eswa.2025.130084},
  url = {https://www.sciencedirect.com/science/article/pii/S0957417425037005},
  author = {Khang {Nguyen Quoc} and Lan Le {Thi Thu} and Luyl-Da Quach}
}

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.