Tsunami Prediction Using Neural Networks

A deep learning tensorflow model for binary tsunami classification based on earthquake catalog parameters. This repository contains a complete end-to-end pipeline including preprocessing, scaling, model training, evaluation, and ROC visualization.

---

1. Overview

This project explores whether tsunami occurrence can be predicted directly from earthquake parameters using a deep neural network. The workflow includes:

• Dataset ingestion (tsunami.csv)
• Feature preprocessing and scaling
• Model training with Swish activation
• Saving the trained model and scaler
• Evaluation using classification metrics and ROC–AUC
• Visual output (ROC curve)

The model achieves strong performance (AUC ≈ 0.88), indicating meaningful discriminative power even with minimal catalog-based features.

---

2. Folder Structure

TSUNAMI PREDICTION/
│
├── .venv/                # Python virtual environment
├── train.py              # Training script
├── evaluate.py           # Evaluation script
├── tsunami.csv           # Dataset (input)
├── tsunami_model.h5      # Saved Keras model
├── scaler.pkl            # Saved Scaler
└── roc_curve.png         # ROC curve plot

---

3. Dataset

The dataset tsunami.csv is downloaded from Kaggle

---

4. Model Architecture

The neural network uses:

• Swish activation for all hidden layers
• 150 epochs
• Adamax optimizer (stable for wide-range gradients)
• Binary cross-entropy loss

---

5. Training the Model

Run:

python train.py

This script performs:

1. Loading CSV

2. Feature/label separation

3. Data scaling using MinMaxScaler

4. Model compiling

5. Model training (150 epochs)

6. Saving:

   • tsunami_model.h5
   • scaler.pkl

---

6. Evaluating the Model

Run:

python evaluate.py

This script loads the trained model and scaler, then produces:

• Predictions on the test set
• Accuracy, precision, recall, F1-score
• Confusion matrix
• ROC–AUC score
• ROC curve (roc_curve.png)

---

7. Example Evaluation Results

(Sample results based on current dataset)

• Accuracy: 0.81
• Precision (class 1): 0.72
• Recall (class 1): 0.85
• F1-score (class 1): 0.78
• ROC–AUC: 0.88

These values indicate strong signal detection capability despite noisy geophysical data.

---

8. Future Work

Potential extensions:

• Incorporate waveform-based features (seismograms)
• Time–series modeling (RNN / Transformer)
• Physics-informed constraints (rupture depth thresholds, energy release bounds)
• Regionalization of predictions
• Real-time inference using streaming seismic feeds

---

9. License

This project is released under the MIT License.

---