rick003/plant-disease-clean-v1

Clean Plant Disease Detection Dataset v1

This repository contains the cleaned dataset used by the repaired plant disease detection project.

Dataset summary

• 1,896 images
• 17 disease classes
• 1,328 train images
• 284 validation images
• 284 test images
• 0 invalid YOLO labels
• 0 near-duplicate images across train, validation and test splits under the audit threshold

Healthy-leaf-only images are retained as negative samples. Duplicate semantic classes and classes with only one or two usable images were removed. Empty source labels, conflicting near-duplicates and excess negative samples were excluded.

Files

• plant_disease_clean_v1.zip: complete YOLO dataset
• metadata/data.yaml: class names and split configuration
• metadata/audit_summary.json: split and class statistics
• metadata/manifest.csv: cleaned-to-source mapping
• metadata/excluded.csv: excluded files and reasons
• colab/plant_disease_yolo11s_t4.ipynb: validated Colab T4 notebook
• tools/sync_colab_results.py: download uploaded runs for local Web use

Colab training

1. Download colab/plant_disease_yolo11s_t4.ipynb.
2. Open Google Colab and upload the notebook.
3. Select Runtime -> Change runtime type -> T4 GPU.
4. Run cells from top to bottom.
5. The notebook trains YOLO11s, selects best.pt using the validation set, evaluates the final model on the independent test set, exports ONNX and optionally uploads all results back to this repository.
6. After uploading results, run the local sync_colab_results.py script to install the new weights and reports into the repaired Web project.

Source and license

The source images originate from the public Roboflow Universe Plant Diseases dataset by Muhammad Salah, licensed under CC BY 4.0:

https://universe.roboflow.com/muhammad-salah/plant-diseases-xy8x3

The cleanup, class filtering, deduplication and split reconstruction were performed for this project.