🔥🛰️ CanadaFireSat Model

This repository contains the code for training models on the benchmark CanadaFireSat available online. In this benchmark, we investigate the potential of deep learning with multiple sensors for high-resolution wildfire forecasting.

• 💿 Dataset on Hugging Face
  
• 📝 Paper on ArXiv
  
• 📊 Data repository on GitHub
• 🤖 Model Weights on Hugging Face

Summary Representation:

Model Architectures

In this repository, we train models following two different deep learning architectures, CNN-based using ResNet encoders and Transformer-based using ViT encoders.

Those models are trained across three data settings namely:

Setting	Source	Format	Type
SITS ONLY	Sentinel-2	Spatial	Multi-Spectral Images
ENV ONLY	MODIS	Spatial	Environmental Products
	ERA5-Land	Spatial	Climate Reanalysis
	CEMS	Spatial	Fire Indices
Multi-Modal	Sentinel-2	Spatial	Multi-Spectral Images
	MODIS	Tabular	Environmental Products
	ERA5-Land	Tabular	Climate Reanalysis
	CEMS	Tabular	Fire Indices

CNN-Based Multi-Modal Architecture

ViT-Based Multi-Modal Architecture

🛠️ Set-Up

• ⚠️ For best compatibility, use Python 3.8.x .

• In order to log the model training, you need to set-up a WandB profile or switch model loggers. You can specify your WandB information in global_config.yaml.

• Then, you also need to install the Python virtual environment:

python -m venv fire-env
source fire-env/bin/activate
pip install --upgrade pip setuptools wheel
pip install -r requirements/requirements.txt --extra-index-url https://download.pytorch.org/whl/cu117

• You can then download the data from Hugging Face 🤗 leveraging src.huggingface.download | Config: download.yaml.

• Specify the data paths in the global_config.yaml.

🏋️ Model Training & Evaluation

• Training: Run the src.train.segmentation_training script with your selected training config: ResNet_MULTI.yaml, ViT_MULTI.yaml, ...

• Evaluation: Run the src.eval.eval script with your selected evaluation config: eval.yaml, eval_tab.yaml, ... . The model config described in the evaluation should matched the one of the its training config.

📷 Results

📊 Performance Analysis: In this table, we describe the models' performances across data settings and architectures.

Encoder	Modality	Params (M)	Val		Test		Test Hard		Avg
			PRAUC	F1	PRAUC	F1	PRAUC	F1	PRAUC	F1
ResNet-50	SITS Only	52.2	45.9	49.4	54.0	59.9	26.2	36.7	42.0	48.7
	ENV Only	97.5	41.6	46.7	50.8	55.2	24.5	33.1	39.0	45.0
	Multi-Modal	52.2	46.1	51.2	57.0	60.3	27.1	37.4	43.4	49.6
ViT-S	SITS Only	36.5	45.2	50.6	51.2	51.9	25.7	33.8	40.7	45.2
	ENV Only	54.8	34.8	45.7	49.2	59.9	21.2	35.1	35.1	46.9
	Multi-Modal	37.7	43.9	50.0	56.2	59.2	24.7	35.6	41.6	48.3
Baseline (FWI)	ENV Only	-	20.0	32.7	43.1	50.3	21.1	32.7	28.1	38.6

🗺️ Use Cases on large ROI: We plot a large target area where a wildfire occurred in Québec in 2023, then the fire polygons corresponding to the wildfires, then our model predictions across the region.

Figure 1: Sentinel-2 tile from 2023/06/28 of size 14 km × 26 km before a large wildfire in Québec.

Figure 2: Fire polygons for the large wildfire on 2023/07/05 over the same tile.

Figure 3: Binary model predictions (in red) over the 2.64 km × 2.64 km center-cropped positive samples outlined in black.

🖋️ Citation

@article{porta2025canadafiresat,
  title={CanadaFireSat: Toward high-resolution wildfire forecasting with multiple modalities},
  author={Porta, Hugo and Dalsasso, Emanuele and McCarty, Jessica L and Tuia, Devis},
  journal={arXiv preprint arXiv:2506.08690},
  year={2025}
}