The broader project includes experiment planning, dataset notes, release documentation, notebook scaffolding, evaluation utilities, and model card drafting. This Hugging Face repository is intended specifically for publishing trained model artifacts and release-ready documentation.

Scope

In scope

Binary classification of healthy vs. unhealthy leaves
Multiclass disease category classification
PyTorch-based transfer learning models
Release-ready model packaging
Usage documentation
Model card documentation

Out of scope

Redistribution of raw PlantVillage image files
Claims of real-world field robustness
Mobile or edge deployment optimization
New dataset collection or annotation

Dataset

Dataset source: PlantVillage
Homepage: https://plantvillage.psu.edu/
Access source used in project planning: Hugging Face Datasets (mohanty/PlantVillage)

Licensing note

According to the original PlantVillage paper, the dataset was released under Creative Commons Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0), with the additional clarification that algorithms trained on the data fall under the same license.

Important dataset note

This repository does not redistribute PlantVillage image files.

Only trained model weights, metadata, and documentation are provided here unless redistribution is later reviewed and explicitly approved by the project release process.

Attribution note

If you use this repository, please also attribute the original PlantVillage dataset and paper.

Classification Tracks

1. Binary Track — Healthy vs. Unhealthy

This framing relabels all images into:

healthy
unhealthy

Motivation: a simpler baseline suitable for quick screening and early triage-style use cases.

2. Multiclass Track — Disease Category

This framing preserves the original disease class labels, such as:

Tomato___Early_blight
Potato___Late_blight

Motivation: more informative predictions for disease-specific classification.

Data Leakage Warning — leaf_id Split Integrity

A critical constraint in this project is leaf-level split integrity.

PlantVillage may contain multiple images of the same physical leaf captured under different angles, crops, or lighting conditions. If train/validation/test splits are created at the image level, the same physical leaf may appear in multiple partitions, causing data leakage and artificially inflated metrics.

For this reason:

all train/validation/test splits must be created at the leaf_id level
all images belonging to the same physical leaf must remain in the same partition
split logic must be enforced programmatically
split manifests should be documented and versioned separately from raw image files

Controlled Background Limitation

PlantVillage images are captured under controlled laboratory conditions with relatively uniform backgrounds.

This matters because:

a model may partially learn background cues instead of disease-relevant leaf features
measured benchmark accuracy may be higher than true field performance
performance reported from this dataset should be interpreted as controlled-condition performance

Real-World Generalization Warning

This project does not claim real-world robustness.

Models trained and evaluated only on PlantVillage may perform substantially worse on:

field images
natural backgrounds
mobile phone images
varied lighting and weather conditions
different regions or growing seasons

Any deployment or real-world use claim would require separate external validation.

Repository Contents

This Hugging Face repository may contain:

trained model checkpoints
configuration files
label mappings
lightweight inference examples
release documentation
model card updates

Intended Use

This repository is intended for:

academic experimentation
benchmarking
student project documentation
reproducible model release

It is not intended for:

direct agronomic decision-making
commercial plant diagnosis pipelines
unsupervised real-world deployment

Usage

Example usage will depend on the final released checkpoint format.

A minimal PyTorch inference example will be added alongside the released model weights.

Limitations

Controlled-background dataset bias
Potential domain shift outside PlantVillage
Class imbalance concerns
Binary and multiclass models should be evaluated separately
Results are only as trustworthy as the enforced split policy

Attribution

This release is part of the Atılım AI Club Plant AI Project.

Please also refer to the main GitHub repository for:

project planning
dataset notes
release checklist
model card drafting
implementation details

Citation

If you use this project or its released artifacts, please cite:

Hughes, D. P. & Salathé, M. (2015). An open access repository of images on plant health to enable the development of mobile disease diagnostics
the PlantVillage dataset
the Atılım AI Club Plant AI Project repository
this Hugging Face model repository

atilimai/plantvillage-leaf-disease-classifier

PlantVillage Leaf Disease Classification

Project Context