Bushfire-prediction

This repository provides two PyTorch-based models to predict daily Fire Radiative Power (FRP) from MODIS satellite data over Australia:

• DeepFRPNet (MLP): a deep feed-forward regression network
• ImprovedFireModel (LSTM + CNN + Self-Attention): a sequence model with 30-day sliding windows

It also includes scripts for training, inference and summarising performance.

---

📁 Repository Structure

Bushfire-prediction/
├── datasets/                      # MODIS CSV files: modis_YYYY_Australia.csv
├── LSTM.py                        # Train ImprovedFireModel (CNN + biLSTM + attention) 
├── MLP.py                         # Train DeepFRPNet MLP model 
├── apply_lstm_modis1.py          # Inference script for LSTM model 
├── apply_mlp_modis1.py           # Inference script for MLP model 
├── apply_summary1.py             # Compute & save combined metrics 
├── best_model.pth                # Checkpoint of best LSTM model
├── best_frp_model.pt             # Checkpoint of best MLP model
├── LICENSE                        # MIT License
└── proj.code-workspace           # VS Code workspace settings

---

🛠 Installation

1. Clone the repo

   git clone https://github.com/ErrorChen/Bushfire-prediction.git
   cd Bushfire-prediction

2. (Optional) Create a virtual environment

   python3 -m venv venv
   source venv/bin/activate   # Windows: venv\Scripts\activate

3. Install dependencies

   pip install numpy pandas scikit-learn matplotlib
   pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

---

📦 Data Preparation

Place your daily MODIS data CSVs in the datasets/ folder, named exactly as:

datasets/modis_YYYY_Australia.csv

where YYYY is a four-digit year.

• LSTM.py aggregates these files via glob('datasets/modis_????_Australia.csv')
• MLP.py uses glob('datasets/modis_*.csv')

Each CSV must include at least:

acq_date, acq_time, latitude, longitude,
brightness, bright_t31, scan, track,
confidence, frp, satellite, instrument,
daynight, type, version

---

⚙️ Model Training

1. Train the MLP (DeepFRPNet)

python MLP.py

• Loads & preprocesses all datasets/modis_*.csv
• Splits into train/val/test
• Trains with mixed precision & early stopping
• Saves best weights to best_frp_model.pt

2. Train the LSTM-based model (ImprovedFireModel)

python LSTM.py

• Aggregates daily summaries with a 30-day lookback window
• Trains CNN→biLSTM→Self-Attention network
• Saves best weights to best_model.pth

---

🚀 Inference

MLP Inference

python apply_mlp_modis1.py

• Loads best_frp_model.pt and all datasets/modis_*.csv
• Outputs modis_mlp_frp_results.csv with true vs predicted FRP

LSTM Inference

python apply_lstm_modis1.py

• Loads best_model.pth and all datasets/modis_YYYY_Australia.csv
• Outputs modis_lstm_frp_results.csv with true vs predicted FRP

---

🤝 Contributing

1. Fork this repo
2. Create a feature branch:

   git checkout -b feature/your-feature

3. Commit your changes:

   git commit -m "Add awesome feature"

4. Push and open a Pull Request

Please include tests and update this README as needed.

---

📜 License

This project is licensed under the MIT License. See LICENSE for details.