AI-Powered Rockfall Prediction System

An intelligent system developed for the Smart India Hackathon (SIH) to proactively predict rockfall events in open-pit mines. By leveraging time-series analysis with a deep learning model, we shift from reactive monitoring to a predictive safety framework, safeguarding personnel and equipment.

---

Core Idea

Traditional safety protocols often rely on manual inspection or post-event analysis. Our system transforms this paradigm by analyzing temporal data patterns leading up to a potential rockfall. We process historical weather and topographical data into 10-day sequences, feeding them into a Bidirectional LSTM with an Attention mechanism. This model learns to identify subtle, critical precursors to instability, providing a probability-based forecast of rockfall risk.

---

System Architecture

The project follows a modular architecture, from raw data ingestion to a user-facing web application.

1. Data Acquisition: Collects historical landslide data from the NASA Global Landslide Catalog and corresponding weather data.
2. Preprocessing & Feature Engineering: Cleans, integrates, and transforms the data into time-series sequences (X_lstm_ready.npy, y_lstm_ready.npy) suitable for the LSTM model.
3. Deep Learning Model: A trained Bi-LSTM with Attention model (rockfall_bilstm_attention.pth) serves as the predictive core, analyzing sequences to forecast risk.
4. Backend API: A Flask server (app.py) wraps the model, exposing prediction endpoints.
5. Frontend Dashboard: An intuitive web interface for users to input data, view predictions, and visualize risk zones.

---

Project Structure

The repository is organized to separate data, notebooks, and application source code for clarity and maintainability.

.
├── 📄 .gitignore
├── 📄 README.md
├── 📁 data/
│   ├── 📄 enhanced_landslide_dataset.csv
│   ├── 📄 featured_weather_data.csv
│   ├── 📄 X_lstm_ready.npy      # Feature data (sequences) for the model
│   └── 📄 y_lstm_ready.npy      # Target data for the model
├── 📁 notebook/
│   └── 📄 rockfall_lstm.ipynb   # Jupyter Notebook for experimentation & model dev
├── 📁 src/
│   └── 📁 backend/
│       ├── 📁 static/         # CSS, JS files
│       ├── 📁 templates/      # HTML templates for Flask
│       ├── 🐍 app.py          # Main Flask application
│       ├── 🐍 combine.py      # Utility scripts
│       └── ...
├── 🧠 rockfall_bilstm_attention.pth  # The trained PyTorch model
├── 🔧 scaler.gz                     # The fitted data scaler
└── ...

---

Tech Stack

• Backend: Flask
• Deep Learning: PyTorch
• Data Manipulation: Pandas, NumPy
• Geospatial: Rasterio, earthaccess
• Machine Learning: Scikit-learn
• Development: Jupyter Notebook

---

Getting Started

To get a local copy up and running, follow these simple steps.

Prerequisites

• Python 3.8 or higher
• A virtual environment tool like venv or conda

Installation & Execution

1. Clone the repository:

   git clone https://github.com/mohitsharmas97/rockfall.git
   cd rockfall

2. Create and activate a virtual environment:

   # For venv
   python -m venv venv
   source venv/bin/activate  # On Windows, use `venv\Scripts\activate`
   
   # For conda
   conda create --name rockfall_env python=3.8
   conda activate rockfall_env

3. Install the required dependencies:

   pip install flask torch pandas numpy scikit-learn

4. Navigate to the backend directory and run the Flask application:

   cd src/backend
   python app.py

5. Open your browser and go to http://127.0.0.1:5000 to see the application in action!

---

Project Workflow

Our project was developed in four key phases, resulting in a fully integrated system.

Phase 1: Data Acquisition & Preprocessing (✅ Completed)

• Input Data: Utilized the NASA Global Landslide Catalog.
• Geographic Filtering: Isolated events in high-risk zones to create a relevant dataset.
• Data Enrichment: Fetched and integrated elevation and historical weather data for each event.
• Output: A clean, feature-rich dataset ready for sequence creation.

Phase 2: Time-Series Feature Engineering (✅ Completed)

• Sequence Generation: Transformed the static data into 10-day sequences, capturing the temporal dynamics leading up to each landslide event.
• Data Scaling: Normalized the features using a standard scaler to ensure model stability.
• Output: The final X_lstm_ready.npy and y_lstm_ready.npy files used for training.

Phase 3: Model Development & Training (✅ Completed)

• Model Selection: Chose a Bidirectional LSTM with an Attention mechanism to effectively capture long-range dependencies and focus on the most influential time steps.
• Training: Trained the model on the prepared time-series data until convergence.
• Evaluation: Assessed the model's performance using metrics like Accuracy, Precision, Recall, and F1-Score to confirm its predictive power.

Phase 4: Deployment & Visualization (✅ Completed)

• API Development: Encapsulated the trained model and scaler into a lightweight Flask API (app.py).
• Dashboard Creation: Developed a user-friendly web dashboard using HTML/CSS and Flask's template engine to visualize predictions.
• Alert System: Implemented a basic alert mechanism within the frontend to notify users of high-risk predictions.

---

Future Scope

• Real-time Data Integration: Integrate live weather data and geotechnical sensor feeds (displacement, strain) for on-the-fly predictions.
• Model Enhancement: Experiment with more advanced architectures like Transformers for time-series forecasting.
• Cloud Deployment: Deploy the application on a cloud platform (AWS, GCP, Azure) for scalability and accessibility.
• Advanced Visualization: Create more detailed risk maps with temporal heatmaps and interactive charts.

---

Contributors

• Pushkar Sharma
• Ved Thorat
• []

---

Acknowledgments

• Smart India Hackathon for providing the platform and opportunity.
• NASA for making the Global Landslide Catalog and NASADEM datasets publicly available.