🐛 Tea Leaves Pest Classifier | Transfer Learning

This project is a custom-trained image classification model designed to detect and classify 10 types of insect pests commonly found on tea leaves. The entire pipeline—from dataset creation to prediction—is built and documented by Birendra, a Computer Science student exploring practical applications of deep learning in agriculture.

🔍 Project Highlights

• ✅ Custom Image Dataset (manually created, cleaned, and structured)
• 📊 10-Class Insect Pest Detection (tea pest species)
• 🔁 Transfer Learning with pretrained VGG-16
• 🌐 Online Image Prediction via URL support
• 📈 Performance Metrics (confusion matrix, accuracy, AUC, classification report)
• 🧪 Test on Custom Split with non-augmented test images

🗃️ Dataset Details

All images are resized to 150×150 pixels and organized into separate train/ and test/ folders.

pest/
├── train/
│   ├── borer/               # 60 images × 5 copies = 300
│   ├── caterpillar/
│   ├── ...
├── test/
│   ├── borer/               # 60 original images
│   ├── caterpillar/
│   ├── ...

🐞 Classes

No.	Class Name	Directory Name
1	Borer	borer
2	Caterpillar	caterpillar
3	Grub	grub
4	Leafworm	leafworm
5	Mite	mite
6	Nematode	nematode
7	Tea Aphid	tea-aphid
8	Tea Mosquito Bug	tea-mosquito-bug
9	Termite	termite
10	Thrip	thrip

🧠 Model Summary

Component	Details
Base Model	VGG-16 (pretrained on ImageNet)
Input Shape	150×150×3
Head Architecture	Flatten → Dense → Dropout → Dense
Optimizer	Adam (lr=1e-4)
Loss	Categorical Crossentropy
Output	10 Softmax units (multi-class)
Metrics Tracked	Accuracy, AUC

🏋️ Training Metrics Visualization

• Tracks accuracy and AUC per epoch
• Shows convergence trend and training stability

📊 Confusion Matrix

• Clear breakdown of model predictions
• Helps identify class-level confusion or imbalance

📋 Classification Report

Classification Report:
                   precision    recall  f1-score   support

        nematode       1.00      1.00      1.00        79
tea-mosquito-bug       1.00      1.00      1.00        78
     caterpillar       1.00      1.00      1.00        78
        leafworm       1.00      1.00      1.00        73
            grub       1.00      1.00      1.00        75
           borer       1.00      1.00      1.00        65
            mite       1.00      1.00      1.00        78
           thrip       1.00      1.00      1.00        54
       tea-aphid       1.00      1.00      1.00        65
         termite       1.00      1.00      1.00        75

        accuracy                           1.00       720
       macro avg       1.00      1.00      1.00       720
    weighted avg       1.00      1.00      1.00       720

• Includes precision, recall, and F1-score
• Offers class-wise insights into performance

🎯 ROC Curve (One-vs-Rest)

• Multi-class ROC curves (One-vs-Rest strategy)
• Evaluates discriminative power of each class boundary

🎲 Random Prediction predict_random()

This function selects a random image from the test set, makes a prediction, and visualizes the result with a probability bar chart.

predict_random()

• Displays: predicted class, confidence scores, image preview

📸 Output Preview

📂 Predict from Test Directory predict(...)

You can directly predict any image from the dataset using a path:

predict("/kaggle/input/tea-leaves-pest-dataset/TeaLeavesPest-Dataset/pest/test/grub/grub_07.jpg")

• Accepts any image path
• Outputs class prediction + confidence visualization

📸 Output Preview

🖼️ Visual Prediction Grid (12 Random Test Images)

This section demonstrates how the trained model performs on 12 randomly selected images from the test directory. Each image is shown alongside its predicted and true class. If the prediction is correct, the label is shown in green; otherwise, in red.

✅ Correct Prediction ❌ Incorrect Prediction

# Automatically samples 12 test images and compares predictions
# Displays each image with predicted and true label

# (See full implementation inside the notebook)

📸 Output Preview

🌐 Predict from Online Image

You can predict using any internet image URL:

url = "https://www.holderspestsolutions.com/wp-content/uploads/2015/11/termites.jpg"
predict_from_url(url, model, class_names)

Result:

• Top prediction
• Full probability graph (bar chart)
• Image preview

📸 Output Preview

📚 How to Use

This repo is designed to be run directly in Kaggle or Colab.

▶️ Steps

1. Open tea-leaves-pest-classifier-transfer-learning.ipynb in Kaggle or Google Colab.
2. Make sure the Tea Leaves Pest - Dataset dataset folder is available via Kaggle dataset or uploaded to Colab.
3. Run the notebook cells in order.
4. Optionally, use the final prediction cell to test any web image.

📜 License

This project is licensed under Apache 2.0 License. You’re welcome to reuse the notebook, dataset, and model structure with credit.

🤝 Acknowledgments

• TensorFlow/Keras for VGG-16
• Matplotlib & scikit-learn for visualization
• PIL & Requests for web image processing

---

Author: Birendra Nath Nandi
Date: July 2025
License: Apache 2.0