Hugging Face's logo

Datasets:
mbsoft31
/
agri-foundation-v1

Tasks:
Image Classification
Size:
100K<n<1M
Tags:
agriculture
plant-disease
phytopathology
leaf-disease
License:
Dataset card Files
xet
 Community
agri-foundation-v1 / README.md
mbsoft31's picture
mbsoft31
Add files using upload-large-folder tool
aacfae8 verified
preview code
|
raw
history blame contribute delete
3.52 kB
metadata 
license: cc-by-sa-4.0
task_categories:
  - image-classification
tags:
  - agriculture
  - plant-disease
  - phytopathology
  - leaf-disease
size_categories:
  - 100K<n<1M

Agri-Foundation-145k: A Unified Agricultural Foundation Dataset

Dataset Summary

Agri-Foundation-145k is a large-scale, taxonomically harmonized dataset designed for pre-training agricultural foundation models. It aggregates, cleans, and normalizes data from eight major open-access repositories, resolving semantic ambiguities and removing over 50,000 duplicate images to create a robust benchmark for Domain Generalized Plant Disease Detection.

Key Features

  • Size: 144,751 Images
  • Classes: 215 Unified Biological Classes (Crop + Disease/Healthy)
  • Sources: PlantVillage, PlantDoc, New Plant Diseases (Kaggle), Tomato Leaf (Mendeley), Cassava Leaf Disease, Wheat Leaf Disease, PlantSeg, PlantWild.
  • Normalization: Mapped disparate labels (e.g., "Tomato_Blight" vs "Tomato_Early_Blight") into a single scientific taxonomy.
  • Quality:
    • Deduplicated: 51,323 duplicate images (cross-dataset contamination) removed via MD5 hashing.
    • Merged: 128 redundant classes (e.g., _google, _bing scrapes) merged into 55 base biological entities, rescuing 54 classes from being underrepresented.
    • Verified: All images validated for file integrity.

Supported Tasks

  • Pre-training: Ideal for self-supervised learning (MAE, DINO) or supervised pre-training of backbones (MobileNet, ResNet, ViT).
  • Domain Adaptation: Contains both lab-controlled (PlantVillage) and "in-the-wild" (PlantDoc) images for the same classes, enabling Sim-to-Real research.
  • Fine-grained Classification: Distinguishes between visually similar pathogens (e.g., Cercospora vs. Septoria).

Dataset Structure

The dataset follows the standard ImageFolder format:

data/
├── apple_black_rot/
│   ├── plantvillage_image_001.jpg
│   ├── new_plant_diseases_image_045.jpg
│   └── ...
├── tomato_early_blight/
│   └── ...
└── ...

A metadata.csv is provided containing:

  • filename: Image filename.
  • label: The unified class label.
  • source: The original dataset source (provenance).
  • crop: The crop type.
  • disease: The specific pathology.
  • release_path: Path relative to the dataset root.

Construction Methodology

  1. Aggregation: Eight public datasets were ingested.
  2. Normalization: A "Fuzzy Alignment" algorithm mapped 200+ raw directory names to canonical biological entities.
  3. Deduplication: MD5 hashing identified and removed duplicate images caused by cross-pollination between source datasets (e.g., New Plant Diseases containing subsets of PlantVillage).
  4. Refinement: Classes with search-engine suffixes (e.g., _google, _baidu) were merged into their base classes to improve class balance.
  5. Verification: All images were verified for file integrity.

Licensing

This dataset is a composite work.

  • Original images retain the licenses of their respective source datasets (mostly CC-BY-SA 3.0 or CC-BY 4.0).
  • This unified compilation and metadata are released under CC-BY-SA 4.0.

Citation

If you use this dataset, please cite:

@dataset{agri_foundation_145k,
  author = {Your Name},
  title = {Agri-Foundation-145k: A Unified Agricultural Foundation Dataset},
  year = {2026},
  publisher = {Hugging Face},
  version = {1.0}
}