JK-TK/PlantDiseaseDetection

Multi-Domain Plant Disease Detection: A YOLOv11x Benchmark on 116 Classes

---

Overview

This repository contains a complete implementation of a state-of-the-art plant disease detection system using YOLOv11x trained on 23,689 images across 116 disease classes covering major food crops. The model combines two complementary datasets — controlled laboratory images and real-world field images — to achieve strong generalization across domains.

• plantwild — 18,533 lab images, 89 classes (controlled environment, uniform backgrounds)
• FieldPlant — 5,156 field images, 27 classes (real field conditions, pathologist-annotated)

The model achieves 69.8% mAP50, substantially outperforming existing benchmarks on field data (~38–48%) while handling 4× more classes.

---

Model Performance

Metric	Value
mAP50	69.8%
mAP50-95	68.4%
Recall	69.5%
Precision	62.2%
Total Classes	116
Total Images	23,689
Architecture	YOLOv11x (57M parameters)

Top Performing Classes

Rank	Class	mAP50
1	Corn Stripe	99.5%
2	Corn Leaf Blight	98.8%
3	Tomato Brown Spots	97.5%
4	Raspberry Leaf	96.1%
5	Cassava Mosaic	95.5%
6	Celery Leaf	95.4%
7	Peach Leaf Curl	95.3%
8	Corn Smut	93.9%
9	Cassava Healthy	92.6%
10	Basil Leaf	92.4%

Comparison with Published Benchmarks

Study	Dataset	Classes	Architecture	mAP50
Moupojou et al. (2023)	FieldPlant	27	Faster R-CNN	~38%
PlantDoc Benchmark	PlantDoc	27	YOLOv5	43–48%
PlantVillage Typical	PlantVillage	38	YOLOv5/v8	85–90%
Our Model	Combined	116	YOLOv11x	69.8%

---

Dataset

Split	Images	Percentage
Training	16,582	70%
Validation	4,737	20%
Test	2,370	10%

Source Composition:

Source	Type	Images	Classes	Share
plantwild	Laboratory	18,533	89	78.2%
FieldPlant	Field	5,156	27	21.8%

---

Training Configuration

Parameter	Value
Architecture	YOLOv11x (57M parameters)
Input Size	640 × 640 pixels
Batch Size	64 (A100) / 8 (T4)
Optimizer	AdamW
Initial Learning Rate	0.001
Final Learning Rate	0.00001 (cosine decay)
Weight Decay	0.0005
Momentum	0.937
Warmup Epochs	3
Total Epochs	100 (early stopping at epoch 49)
Mixed Precision	Enabled
Augmentations	Mosaic (100%), MixUp (20%), Copy-Paste (20%), Rotation (±10°), Translation (±10°), Scaling (±50%), Flip (50%), HSV adjust

---

Export & Inference

Model Export Sizes

Format	Size	mAP50 (Validation)
PyTorch (FP32)	457 MB	69.8%
TFLite (FP16)	114 MB	69.5%
TFLite (INT8)	62 MB	68.2%
ONNX	113 MB	69.8%

Inference Speed

Hardware	Precision	Time per Image (ms)	FPS
NVIDIA A100	FP16	4.4	227
NVIDIA T4	FP16	15.2	66
iPhone 14	FP16 (Core ML)	~120	8
Android Flagship	INT8 (TFLite)	~180	5

---

Quick Start

Option 1: Google Colab

1. Click the badge above.
2. Select Runtime → Change runtime type → A100 GPU (or T4 if unavailable).
3. Run cells sequentially from top to bottom.

Option 2: Local Installation

# Clone repository
git clone https://github.com/SIV-TK/A-YOLOv11x-Benchmark-on-116-Classes-Using-Combined-Laboratory-and-Field-Images.git
cd A-YOLOv11x-Benchmark-on-116-Classes-Using-Combined-Laboratory-and-Field-Images

# Install dependencies
pip install ultralytics opencv-python matplotlib seaborn scikit-learn pandas pyyaml tqdm kagglehub

# Run the training notebook
jupyter notebook "YOLO(2).ipynb"

Option 3: Load Pre-trained Model

from ultralytics import YOLO

# Load model
model = YOLO('best.pt')  # or load from https://huggingface.co/JK-TK/PlantDiseaseDetection

# Run inference
results = model('path/to/leaf_image.jpg', conf=0.25)

# Display results
results[0].show()

# Print detections
for box in results[0].boxes:
    class_id = int(box.cls[0])
    confidence = float(box.conf[0])
    class_name = model.names[class_id]
    print(f"Detected: {class_name} ({confidence:.2%})")

---

Resume Training

The notebook automatically saves checkpoints. To resume an interrupted training run:

from ultralytics import YOLO

model = YOLO('/content/drive/MyDrive/PlantDisease_TrainingBackup_*/training_run/weights/last.pt')
model.train(data='dataset/data.yaml', epochs=100, resume=True)

---

Test on Custom Images

Using Python:

from google.colab import files
from ultralytics import YOLO

model = YOLO('best.pt')
uploaded = files.upload()

for img_name in uploaded:
    results = model(img_name, conf=0.25)
    results[0].show()

Using the command line:

yolo predict model=best.pt source='path/to/image.jpg' conf=0.25

---

Repository Structure

├── YOLO(2).ipynb                  # Complete training notebook
├── README.md                      # This file
├── LICENSE                        # MIT License
├── best.pt                        # Pre-trained weights (69.8% mAP50)
├── dataset/
│   ├── train/images/              # 16,582 training images
│   ├── train/labels/
│   ├── val/images/                # 4,737 validation images
│   ├── val/labels/
│   ├── test/images/               # 2,370 test images
│   ├── test/labels/
│   └── data.yaml                  # Dataset configuration
└── publication/
    ├── comparison_bar_chart.png
    ├── training_curves.png
    ├── confusion_matrix.png
    ├── dataset_source_pie.png
    ├── dataset_split_bar.png
    ├── overall_performance.csv
    └── comparison_table.csv

---

Key Features

• 116 disease classes across major food crops (corn, tomato, cassava, apple, grape, potato, banana, citrus, and more)
• Combined-domain training (laboratory + field) for real-world generalization
• Complete pipeline from dataset download to model export
• Per-class mAP50, confusion matrix, and comprehensive test evaluation
• TensorFlow Lite export for mobile deployment (FP16 and INT8)
• Publication-ready visualizations generated automatically
• Resume training with checkpoint saving
• Fully reproducible with fixed random seeds

---

Citation

If you use this model or dataset in your research, please cite:

@software{kariuki_2026_plant_disease,
  author = {Kariuki, James Kariuki},
  title  = {Multi-Domain Plant Disease Detection: A YOLOv11x Benchmark on 116 Classes},
  year   = {2026},
  url    = {https://github.com/SIV-TK/A-YOLOv11x-Benchmark-on-116-Classes-Using-Combined-Laboratory-and-Field-Images/tree/main/YOLO.ipynb},
  note   = {mAP50: 69.8\%, 116 disease classes, 23,689 images}
}

---

Acknowledgments

• plantwild — high-quality laboratory images
• FieldPlant — real-world field annotations
• Ultralytics — YOLOv11x implementation and training framework
• Google Colab — free A100 GPU access
• KaggleHub — easy dataset downloading

---

License

This project is licensed under the MIT License.

---

Contact

Author: Kariuki James Kariuki
GitHub: SIV-TK
Hugging Face: JK-TK/PlantDiseaseDetection
Paper / Notebook: YOLO.ipynb
Email: ST62550392025@students.ouk.ac.ke
Phone: +254 718 845 849

For questions, issues, or collaboration requests, please open an issue on GitHub.

---

If you find this work useful, please consider starring the repository and citing it in your publications.