Uncertainty-Aware Deep Learning for Wildfire Danger Forecasting

This repository contains the code and pretrained models used in the paper "Uncertainty-Aware Deep Learning for Wildfire Danger Forecasting".

The project implements both deterministic and uncertainty-aware deep learning models to predict wildfire danger.

Table of Contents

• Installation
• Usage
• Configuration
• Pretrained Models
• Citation

---

Installation

Clone this repository and install the required dependencies:

git clone <repository_url>
cd uncertainty-wildfires
pip install -r requirements.txt

---

Dataset

You can download the dataset from: Zenodo

Usage

The repository includes configuration files for all models in the configs and configs_test directories.

Training

To train a model, specify the corresponding configuration file. For example, to train a Variational Inference based Bayesian Neural Network:

python train.py --config configs/config_bnn.json

You can modify the configuration files to change hyperparameters, model type, or enable aleatoric uncertainty (via the noisy parameter, check Aleatoric Uncertainty section).

Testing

To evaluate a trained model, use the configuration in configs_test.

For example, to test a Variational Inference based Bayesian Neural Network:

python test.py --config configs_test/config_bnn.json

Before running any scripts, update the following paths in the configuration files:

• dataset_root: Path to your stored downloaded dataset.
• save_dir: Path where trained models will be saved.

Aleatoric Uncertainty

The noisy variable in the configuration files determines whether the model accounts for aleatoric uncertainty during training.

• Set noisy: true to include aleatoric noise in predictions.
• Set noisy: false to ignore aleatoric uncertainty.

Deep Ensembles

Deep Ensembles combine multiple trained models to improve predictive performance and quantify uncertainty. To run a Deep Ensemble:

1. Train the deterministic model multiple times using config_det.json.
2. Use all trained models together for ensemble predictions. For this, you have to specify the number of models you have trained in configs_test/config_des.json via the num_models variable.

---

Pretrained Models

All pretrained checkpoints for the models that have been used in the paper are available in the trained_models directory.

---

Citation

If you use our work, please cite: