An AI-powered plant disease detection system using deep learning to identify 38 different plant diseases with 98.55% accuracy.
This project implements a deep learning model based on MobileNetV2 architecture to detect and classify plant diseases from leaf images. The system can identify diseases across 14 different plant species with high accuracy.
- ✅ 98.55% Classification Accuracy
- ✅ 99.89% Top-3 Accuracy
- ✅ 38 Disease Classes across 14 plant species
- ✅ Real-time Detection (<2 seconds per image)
- ✅ Web-based UI using Streamlit
- ✅ REST API for integration
The model can detect diseases in the following crops:
| Crop | Number of Classes | Diseases Detected |
|---|---|---|
| 🍅 Tomato | 10 | Bacterial Spot, Early Blight, Late Blight, Leaf Mold, Septoria Leaf Spot, Spider Mites, Target Spot, Mosaic Virus, Yellow Leaf Curl Virus, Healthy |
| 🥔 Potato | 3 | Early Blight, Late Blight, Healthy |
| 🍎 Apple | 4 | Black Rot, Apple Scab, Cedar Rust, Healthy |
| 🍇 Grape | 4 | Black Rot, Esca (Black Measles), Leaf Blight, Healthy |
| 🌽 Corn/Maize | 4 | Cercospora Leaf Spot, Common Rust, Northern Leaf Blight, Healthy |
| 🌶️ Pepper (Bell) | 2 | Bacterial Spot, Healthy |
| 🍒 Cherry | 2 | Powdery Mildew, Healthy |
| 🍑 Peach | 2 | Bacterial Spot, Healthy |
| 🍓 Strawberry | 2 | Leaf Scorch, Healthy |
| 🍊 Orange | 1 | Huanglongbing (Citrus Greening) |
| 🎃 Squash | 1 | Powdery Mildew |
| 🫐 Blueberry | 1 | Healthy |
| 🫛 Raspberry | 1 | Healthy |
| 🫘 Soybean | 1 | Healthy |
Total: 38 Classes
- Total Images: 54,305
- Image Resolution: 224×224 pixels (resized)
- Format: JPG
- Classes: 38 disease categories
| Split | Images | Percentage |
|---|---|---|
| Training | 37,985 | 70% |
| Validation | 8,158 | 15% |
| Test | 8,162 | 15% |
data/
├── sample_test_images/ # 76 sample images (2 per class) - included in repo
├── processed/ # Full dataset (not in repo - download separately)
│ ├── train/ # 37,985 images
│ ├── val/ # 8,158 images
│ └── test/ # 8,162 images
└── raw/ # Original PlantVillage dataset
📦 Sample Data in Repository
This repository includesdata/sample_test_images/with 76 sample images (2 per class) for quick testing and demonstration purposes.📥 Full Dataset Download
The complete PlantVillage dataset (1.8GB, 54,305 images) is not included in this repository due to GitHub's file size limitations. To download the full dataset, run:python download_dataset.pyThis will download and organize the complete training, validation, and test datasets.
- Architecture: MobileNetV2
- Technique: Transfer Learning + Fine-tuning
- Input Size: 224×224×3
- Output: 38 classes (softmax)
- Parameters: ~3.5M
- Model Size: 25 MB
| Parameter | Value |
|---|---|
| Base Learning Rate | 0.0001 |
| Optimizer | Adam |
| Loss Function | Categorical Crossentropy |
| Batch Size | 32 |
| Initial Epochs | 20 |
| Fine-tuning Epochs | 10 |
| Total Training Time | ~7 hours |
- Random rotation (±20°)
- Width/Height shift (±0.1)
- Horizontal flip
- Zoom (±0.1)
- Fill mode: nearest
| Metric | Score |
|---|---|
| Validation Accuracy | 98.55% |
| Top-3 Accuracy | 99.89% |
| Training Accuracy | 99.01% |
| Test Accuracy | 94.88% |
| Tier | Accuracy Range | Number of Classes |
|---|---|---|
| Perfect | 100% | 6 classes |
| Excellent | 95-99% | 23 classes |
| Good | 85-95% | 7 classes |
| Fair | <85% | 2 classes |
| Rank | Class | Accuracy | Correct | Total |
|---|---|---|---|---|
| 1 | Apple - Black Rot | 100.00% | 93 | 93 |
| 2 | Peach - Healthy | 100.00% | 54 | 54 |
| 3 | Orange - Huanglongbing | 100.00% | 826 | 826 |
| 4 | Grape - Healthy | 100.00% | 64 | 64 |
| 5 | Grape - Leaf Blight | 100.00% | 162 | 162 |
| 6 | Squash - Powdery Mildew | 100.00% | 276 | 276 |
| 7 | Soybean - Healthy | 99.48% | 760 | 764 |
| 8 | Corn - Common Rust | 99.44% | 178 | 179 |
| 9 | Corn - Healthy | 99.43% | 174 | 175 |
| 10 | Cherry - Healthy | 99.22% | 127 | 128 |
| 11 | Peach - Bacterial Spot | 99.13% | 342 | 345 |
| 12 | Strawberry - Leaf Scorch | 98.80% | 165 | 167 |
| 13 | Raspberry - Healthy | 98.21% | 55 | 56 |
| 14 | Cherry - Powdery Mildew | 98.10% | 155 | 158 |
| 15 | Grape - Esca (Black Measles) | 98.08% | 204 | 208 |
| 16 | Apple - Healthy | 97.98% | 242 | 247 |
| 17 | Tomato - Bacterial Spot | 97.81% | 312 | 319 |
| 18 | Tomato - Yellow Leaf Curl Virus | 97.26% | 782 | 804 |
| 19 | Strawberry - Healthy | 97.10% | 67 | 69 |
| 20 | Pepper Bell - Bacterial Spot | 96.00% | 144 | 150 |
| 21 | Pepper Bell - Healthy | 95.95% | 213 | 222 |
| 22 | Tomato - Healthy | 95.82% | 229 | 239 |
| 23 | Apple - Cedar Apple Rust | 95.24% | 40 | 42 |
| 24 | Grape - Black Rot | 94.92% | 168 | 177 |
| 25 | Apple - Apple Scab | 94.74% | 90 | 95 |
| 26 | Blueberry - Healthy | 93.36% | 211 | 226 |
| 27 | Potato - Early Blight | 92.67% | 139 | 150 |
| 28 | Tomato - Late Blight | 91.64% | 263 | 287 |
| 29 | Corn - Cercospora Leaf Spot | 89.61% | 69 | 77 |
| 30 | Potato - Late Blight | 89.33% | 134 | 150 |
| 31 | Tomato - Mosaic Virus | 89.29% | 50 | 56 |
| 32 | Tomato - Leaf Mold | 88.81% | 127 | 143 |
| 33 | Corn - Northern Leaf Blight | 85.14% | 126 | 148 |
| 34 | Tomato - Septoria Leaf Spot | 84.59% | 225 | 266 |
| 35 | Tomato - Spider Mites | 78.57% | 198 | 252 |
| 36 | Tomato - Target Spot | 78.20% | 165 | 211 |
| 37 | Tomato - Early Blight | 64.67% | 97 | 150 |
| 38 | Potato - Healthy | 60.87% | 14 | 23 |
Best Performing: Apple - Black Rot (100%)
Needs Improvement: Potato - Healthy (60.87% - limited samples)
Understanding why certain classes have lower accuracy helps improve future iterations:
- Primary Issue: Severe class imbalance
- Sample Size: Only 23 test images (smallest in dataset)
- Confusion: Often misclassified as Potato Early Blight or Late Blight
- Reason: Healthy potato leaves can show early signs of stress that resemble disease onset
- Recommendation: Needs more training samples for robust performance
- Visual Similarity: Early Blight symptoms (concentric rings) overlap with Late Blight and Septoria Leaf Spot
- Disease Progression: Different stages of Early Blight look significantly different
- Multiple Lesion Types: Can present as small spots or large necrotic areas
- Confusion Matrix: Often confused with Tomato Late Blight and Septoria Leaf Spot
- Challenge: Requires fine-grained feature discrimination
- Symptom Similarity: Circular lesions similar to Early Blight and Bacterial Spot
- Variable Appearance: Lesions vary greatly in size and color development
- Limited Dataset: Fewer distinct visual patterns compared to other classes
- Background Noise: Often occurs with other diseases simultaneously
- Subtle Symptoms: Early damage appears as tiny white/yellow stippling
- Not Directly Visible: Actual mites not visible in leaf images
- Progressive Damage: Early, mid, and late-stage damage look completely different
- Confusion: Resembles nutrient deficiency or environmental stress
- Special Case: Pest damage rather than disease, requires different visual cues
- Small Lesions: Numerous tiny circular spots can blend together
- Similar Patterns: Overlaps with Bacterial Spot and Leaf Mold symptoms
- Variable Stages: Early small spots vs. late coalescent lesions differ greatly
- Elongated Lesions: Long cigar-shaped lesions can vary in size (1-6 inches)
- Color Variation: Gray-green to tan coloring overlaps with other corn diseases
- Confusion: Often mistaken for Cercospora Leaf Spot due to similar lesion shapes
-
Class Imbalance
- Classes with <50 test samples show reduced performance
- Potato Healthy (23), Tomato Mosaic Virus (56) have limited data
-
Visual Similarity
- Tomato diseases share many overlapping symptoms
- Spot-like lesions across multiple classes create confusion
-
Disease Progression
- Early vs. late stage symptoms look dramatically different
- Model trained on mixed stages may struggle with edge cases
-
Dataset Characteristics
- PlantVillage images are lab-controlled
- Real-world field conditions may introduce more variation
-
Multi-Disease Presence
- Some images may show early signs of multiple diseases
- Healthy classifications most challenging due to subtle stress indicators
To address these limitations:
- Collect more samples for underrepresented classes (especially Potato Healthy)
- Implement focal loss to handle class imbalance
- Add disease progression-aware augmentation
- Use ensemble methods for similar-looking diseases
- Implement hierarchical classification (plant → health status → specific disease)
Python 3.8+
TensorFlow 2.x
Streamlit- Clone the repository
git clone https://github.com/Faheem8585/plant_disease_detection.git
cd plant-disease-detection- Install dependencies
pip install -r requirements.txt- Download the dataset (optional)
python download_dataset.pystreamlit run ui/app.pyThen open http://localhost:8501 in your browser.
uvicorn api.main:app --reloadAPI will be available at http://localhost:8000
python quick_test.pyplant-disease-detection/
├── api/ # FastAPI REST API
│ ├── main.py
│ └── routes.py
├── configs/ # Configuration files
│ ├── model_config.yaml
│ ├── preprocessing_config.yaml
│ └── training_config.yaml
├── data/ # Dataset directory
│ ├── processed/ # Processed train/val/test splits
│ └── raw/ # Raw PlantVillage dataset
├── logs/ # Training logs
├── models/ # Saved models
│ └── pretrained_replacement.h5 # Best model (98.55%)
├── notebooks/ # Jupyter notebooks for experiments
├── results/ # Training results and visualizations
├── src/ # Source code
│ ├── data/ # Data loading and preprocessing
│ ├── inference/ # Model inference
│ ├── models/ # Model architectures
│ ├── training/ # Training scripts
│ └── utils/ # Utility functions
├── tests/ # Unit tests
├── ui/ # Streamlit web interface
│ └── app.py
├── analyze_per_class.py # Per-class accuracy analysis
├── download_dataset.py # Dataset downloader
├── quick_start.py # Quick start training script
├── quick_test.py # Quick inference test
├── train_better_model.py # Main training script
├── visualize_predictions.py # Prediction visualizer
├── requirements.txt # Python dependencies
└── README.md # This file
python train_better_model.pyTraining parameters can be adjusted in configs/training_config.yaml.
View training history and predictions:
python visualize_predictions.pyAnalyze per-class performance:
python analyze_per_class.py- Image Upload: Drag & drop or browse for plant leaf images
- Real-time Prediction: Get results in <2 seconds
- Confidence Scores: See prediction confidence and top-3 alternatives
- Treatment Recommendations: Basic guidance for detected diseases
- Responsive Design: Works on desktop and mobile devices
Upload an image and get disease prediction.
Request:
{
"image": "<base64_encoded_image>"
}Response:
{
"crop": "Tomato",
"disease": "Early Blight",
"confidence": 0.97,
"top_predictions": [...]
}- Best performance on clear, well-lit leaf images
- Trained on PlantVillage dataset (controlled conditions)
- May have reduced accuracy on field images with:
- Multiple diseases present
- Environmental damage
- Poor lighting or unusual angles
- Not a replacement for professional diagnosis
- Use as initial screening tool only
If you use this project, please cite the PlantVillage dataset:
Hughes, D.P. and Salathe, M. (2015).
"An open access repository of images on plant health to enable the
development of mobile disease diagnostics."
arXiv preprint arXiv:1511.08060.
⚠️ Permission Required
If you wish to use, modify, or redistribute this code or trained model, you must:
- Request Permission: Contact the author for explicit permission before use
- Provide Attribution: Cite this work in your project, paper, or application
- Acknowledge Source: Include a clear reference to this repository
If you use this code or model in your research or project, please cite as:
@software{plant_disease_detection_2024,
author = {Faheem},
title = {Plant Disease Detection System using Deep Learning},
year = {2024},
url = {https://github.com/Faheem8585/plant_disease_detection},
note = {MobileNetV2-based transfer learning model achieving 98.55% accuracy on 38 plant disease classes}
}Or in text format:
Faheem (2024). Plant Disease Detection System using Deep Learning.
GitHub repository: https://github.com/Faheem8585/plant_disease_detection
For permission requests or collaborations, please open an issue on this repository or contact via GitHub.
This project is available for:
- ✅ Learning and education (with proper citation)
- ✅ Academic research (with permission and citation)
- ✅ Personal projects (with attribution)
- ❌ Commercial use (requires explicit written permission)
This project is licensed under a custom license requiring permission and attribution. While the code is publicly available, you must:
- Obtain permission before use
- Provide proper citation and acknowledgment
- Not use commercially without explicit permission
See the repository for full terms.
Contributions are welcome! Please:
- Fork the repository
- Create a feature branch
- Make your changes
- Submit a pull request
For major changes, please open an issue first to discuss proposed changes.
For questions, permission requests, or feedback:
- GitHub Issues: https://github.com/Faheem8585/plant_disease_detection/issues
- Repository: https://github.com/Faheem8585/plant_disease_detection
Note: This system is designed for educational and research purposes. Always consult with agricultural professionals for treatment decisions.
© 2024 Faheem. All rights reserved.