🌿 Medicinal Leaf Disease Detection with AI-MedLeafX + XAI

Deep learning for medicinal plant leaf disease classification with ResNet, EfficientNetV2-S, ConvNeXt-Tiny, and explainability using Grad-CAM & t-SNE.

---

📌 Overview

This project focuses on medicinal plant leaf disease classification using deep learning on the AI-MedLeafX dataset. In addition to classification performance, the project also emphasizes model interpretability through Explainable AI (XAI) techniques such as Grad-CAM and t-SNE.

Main objectives

• Classify diseases on medicinal plant leaves from RGB images.
• Compare multiple CNN architectures on the same dataset.
• Analyze model decisions using XAI.
• Build a solid baseline for future real-world agricultural applications.

---

✨ Highlights

• ✅ Fine-tuning pretrained CNN models on AI-MedLeafX
• ✅ Comparison of ResNet50, EfficientNetV2-S, and ConvNeXt-Tiny
• ✅ Explainability with Grad-CAM and t-SNE
• ✅ Strong performance with 98.85% accuracy from EfficientNetV2-S
• ✅ Organized notebooks for training, evaluation, and XAI experiments

---

🖼️ Project Preview

Sample leaf images

Example result visualizations

🎯 Problem Statement

Early detection of plant diseases is important for reducing crop damage and improving decision-making in smart agriculture. Unlike many previous works that focus on common agricultural crops, this project targets medicinal plants, which are less studied and have more limited public datasets.

The task is to classify leaf images into disease/healthy categories using deep learning models and provide visual explanations for model predictions.

---

🧪 Dataset

The project uses AI-MedLeafX, a medicinal plant disease dataset containing 4 plant species and 13 classes.

Plant species

• Camphor
• HariTaki
• Neem
• Sojina

Disease categories

• Bacterial Spot
• Shot Hole
• Powdery Mildew
• Yellow Leaf
• Healthy Leaf

Input setting

• Image size: 224 × 224
• Data format: RGB images
• Preprocessing based on ImageNet normalization

Split used in notebook

The notebook uses the augmented image directory and splits the dataset as follows:

• Train: 70%
• Validation: 20%
• Test: 10%

Observed dataset size in notebook

• 65,178 images
• 13 classes
• Train: 45,624
• Validation: 13,035
• Test: 6,519

Note: The original report mentions the base AI-MedLeafX dataset at around 10,858 images. The notebook appears to use an augmented version, resulting in a much larger number of samples.

---

🧠 Models

Three representative CNN architectures were studied:

1. ResNet50

A strong baseline CNN with residual connections that helps stabilize deep training.

2. EfficientNetV2-S

A more compute-efficient architecture balancing model size and classification performance.

3. ConvNeXt-Tiny

A modern ConvNet inspired by design ideas from Vision Transformers, achieving the best performance in this project.

---

⚙️ Methodology

The overall pipeline follows these main steps:

Input Images
   ↓
Preprocessing & Augmentation
   ↓
Train / Validation / Test Split
   ↓
Fine-tune Pretrained CNN Models
   ↓
Evaluation Metrics
   ↓
XAI Analysis (Grad-CAM, t-SNE)

Preprocessing

Implemented using torchvision.transforms:

• Resize to 224x224
• Random horizontal flip
• Random rotation
• Convert to tensor
• Normalize with ImageNet mean/std

Training strategy

Based on the report:

• Pretrained ImageNet weights
• AdamW optimizer
• Cross-Entropy Loss
• Early stopping
• ReduceLROnPlateau
• Two-stage training strategy for stable convergence

Evaluation metrics

• Accuracy
• Precision
• Recall
• F1-score
• ROC-AUC
• Confusion Matrix

---

📊 Results

Quantitative comparison

Model	Accuracy (%)	Precision	Recall	F1-score	Params
ResNet50	95.03	0.95	0.95	0.95	~24M
EfficientNetV2-S	98.85	0.99	0.99	0.99	~21M
ConvNeXt-Tiny	98.16	0.98	0.98	0.98	~28M

Key observations

• ConvNeXt-Tiny achieved the best overall performance.
• EfficientNetV2-S provided a strong balance between accuracy and efficiency.
• ResNet50 remained a reliable baseline for comparison.

---

🔍 Explainable AI

This project does not stop at prediction performance. It also investigates why the model makes a prediction.

Grad-CAM

Grad-CAM helps identify the image regions that most influence the final prediction.

Findings from the report:

• Good localization on diseases such as Bacterial Spot and Shot Hole.
• Distributed attention over the leaf surface for Healthy Leaf samples.
• More difficulty with Powdery Mildew, where symptoms are diffuse and spread across the leaf.

t-SNE

t-SNE is used to visualize learned feature embeddings in 2D.

Observations:

• ResNet50 forms reasonably separated clusters but still shows overlap.
• EfficientNetV2-S produces the clearest feature separation.
• ConvNeXt-Tiny forms denser clusters yet still achieves the highest classification accuracy.

---

🚀 Installation

Recommended environment: Python 3.10+

Install dependencies

pip install torch torchvision scikit-learn matplotlib seaborn pillow opencv-python tqdm jupyter lime grad-cam kaggle

Kaggle API setup

Place your Kaggle API file at:

~/.kaggle/kaggle.json

Set permissions:

chmod 600 ~/.kaggle/kaggle.json

---

⚡ Quick Start

1. Launch Jupyter

jupyter notebook

2. Open the main training notebook

medleaf-disease-detection (2).ipynb

3. Download dataset from Kaggle

Inside the notebook, a command similar to this is used:

kaggle datasets download -d mrlocbap/ai-medleafx

4. Extract data

The notebook extracts the archive into:

content/ai-medleafx

5. Train and evaluate

Main notebook tasks include:

• loading images by class directory
• creating train/val/test splits
• fine-tuning pretrained models
• saving checkpoints
• evaluating classification results
• visualizing XAI outputs

---

🖼 Inference on New Images

The test.ipynb notebook includes helper functions for:

• single-image prediction
• Grad-CAM visualization
• LIME explanation
• loading trained .pth weights for inference

Typical inference flow:

1. Load model checkpoint
2. Apply the same test transform
3. Run forward pass
4. Return predicted class and confidence score

This can be extended into:

• a simple desktop app
• a Streamlit/Gradio web app
• a mobile app
• a field-support diagnostic tool

---

⚠️ Limitations

Despite strong results, several limitations remain:

• The training data is mostly collected in controlled conditions, not fully reflecting field environments.
• Diffuse diseases such as powdery mildew remain harder to localize and explain.
• The system has not yet been deployed as a real-time production application.

---

🔮 Future Work

Possible next steps include:

• training on real-world field images for better generalization;
• experimenting with ViT, Swin Transformer, or hybrid CNN-Transformer models;
• using additional XAI tools such as SHAP and deeper LIME analysis;
• building a more user-friendly diagnosis interface;
• deploying the model on mobile devices or agricultural drones.

---

🙏 Acknowledgement

• Supervisor: TS. Lê Thị Vĩnh Thanh
• Course: Thị giác máy tính và ứng dụng
• Institution: Trường Đại học Công nghiệp TP.HCM

---

📄 Internal Sources

This README was consolidated from:

• Nhom_01_NhanDienBenhTrenLaCay.docx
• medleaf-disease-detection (2).ipynb
• test.ipynb

---

⭐ Suggested Next Improvements

If you want to make this repository even more professional, the next best additions would be:

• a clean architecture diagram;
• confusion matrix images with labels;
• Grad-CAM comparison figure per model;
• a requirements.txt file;
• a demo.ipynb or app.py for quick demonstration.