This project implements a Aago F-APM: Fire Area Prediction Model using a Transformer-based deep learning architecture. The model is trained on synthetic fire spread data generated using a probabilistic BFS approach, which considers environmental factors such as wind speed, wind direction, and humidity. The goal is to predict future fire spread patterns based on past fire data and weather conditions.
The simulations for fire spread are inspired by real-world satellite images of wildfires. These satellite images, such as the examples below, are translated into heatmaps representing fire intensity. These heatmaps are then used alongside weather data (e.g., wind speed, wind direction, humidity) to train the model for predicting fire progression.
Source: OroraTech's Satellite Image
Fire spread data is generated by simulating fire propagation on a grid. The dataset includes:
- Grid-based fire spread simulation
- Weather conditions per timestep (wind speed, wind direction, humidity)
- Dynamic environmental updates (wind variation, humidity variation, wind-driven jumps)
- 1,000 training simulations and 100 test simulations stored as
.pklfiles
The dataset is preprocessed using MinMaxScaler for fire grids and StandardScaler for weather conditions. The FireDataset class handles:
- Sequence creation with a specified window size
- Normalization of fire grids and weather data
- Sliding window batching for training
The fire prediction model is based on a Transformer encoder with the following components:
- Positional Encoding for sequence-aware learning
- Multi-Head Self-Attention for capturing spatial dependencies
- Feed-Forward Network (FFN) for feature transformation
- Layer Normalization & Dropout for stability and regularization
- Sigmoid Activation to predict fire spread probabilities (scaled back to [0,255])
The training pipeline includes:
- Data Loading: Splitting dataset into 80% training and 20% validation
- Model Initialization: Using Transformer-based architecture
- Loss Function: Mean Squared Error (MSE)
- Optimizer: Adam optimizer
- Training Loop:
- Forward pass
- Backpropagation and optimization
- Validation after each epoch
- Checkpointing: Saving the best model based on validation loss
To run this model, install the required dependencies:
pip install -r requirements.txtTo generate the fire spread dataset:
python dataset_maker.pyRun the following command to start training:
python trainer.pyTo use the trained model for fire spread prediction:
python tester.pyfire_prediction/
│── dataset/
│ ├── fire_dataset.pkl
│ ├── fire_dataset_test.pkl
|── test_results/
| │── first_prediction-0-0.png
| │── first_prediction-0-1.png
| │── ...
│── model.py # FireTransformer model implementation
│── dataset.py # FireDataset class
│── dataset_maker.py # Fire spread simulation
│── trainer.py # Model training script
│── tester.py # Inference script
│── config.py # Configuration settings
│── checkpoints/ # Saved model weights
|── docs/ # Supporting image files
│── requirements.txt # Requirements file
│── README.md # Project documentation
The model generates images in batches , where each image shows the original fire spread vs. the predicted fire spread. This visual comparison helps analyze the model's accuracy and improvements over time.
Example images from test_results directory:
This project is licensed under the MIT License.