Mesogeos: A multi-purpose dataset for data-driven wildfire modeling in the Mediterranean

Spyridon Kondylatos, Ioannis Prapas, Gustau Camps-Valls, Ioannis PapoutsisNeurIPS 2023

Wildfires are a growing global threat, particularly in the Mediterranean, where climate change is intensifying fire frequency and extremity. While Machine Learning (ML) and Deep Learning (DL) offer powerful tools to model these complex events, the field has historically been held back by a "data bottleneck."

Existing datasets are often fragmented, limited to small regions, or lack the temporal and spatial resolution needed for operational use. Furthermore, many studies rely on private datasets, making it impossible for the broader research community to reproduce or build upon their results.

To solve this, researchers have introduced Mesogeos, the first large-scale, high-resolution "datacube" specifically designed for machine learning wildfire modeling in the Mediterranean. Spanning 17 years of data at a 1km resolution, it harmonizes 27 variables—from meteorology to human activity—to provide a robust benchmark for forecasting fire danger and predicting burned areas.

Methods and Dataset Construction

The authors developed a sophisticated pipeline to create a spatio-temporal datacube. A datacube is essentially a multi-dimensional array where data is perfectly aligned across longitude, latitude, and time.

Pipeline for the Mesogeos Datacube

To build Mesogeos, the team employed several key technical approaches:

  • Harmonization: Disparate data sources (both raster and vector) were aligned into a uniform 1km x 1km grid with a daily time step.
  • Parallel Processing: Handling over 3 TB of data in memory required robust engineering. The team used the xarray and Zarr frameworks to process daily "chunks" in parallel, optimizing the dataset for cloud-based ML workflows. Researchers can now stream only the data they need without downloading terabytes of files.
  • Refined Ground Truth: Instead of relying solely on raw satellite detections, the authors cross-referenced the European Forest Fire Information System (EFFIS) with MODIS Active Fire data to pinpoint highly accurate ignition dates and locations.

The resulting dataset spans from 2006 to 2022 and integrates meteorology (temperature, wind, humidity, precipitation), vegetation and drought metrics (NDVI, LAI, soil moisture), topography (elevation, slope, aspect), and fire records.

Key Insights: The "Human Factor" and Multi-Purpose Utility

Unlike many physical models that focus strictly on weather and fuel, Mesogeos explicitly includes anthropogenic variables, such as population density and distance from roads. This acknowledges a critical reality: humans are the primary cause of ignitions in the Mediterranean.

Furthermore, the authors established two distinct "tracks" (benchmarks) to prove the dataset's versatility:

  1. Track A (Forecasting): Predicting the probability of a large fire (>30ha) occurring the next day.
  2. Track B (Estimation): Predicting the final shape and size of a fire once an ignition is detected.

Key Results and Baselines

The authors established strong baselines using state-of-the-art deep learning architectures to demonstrate the dataset's readiness.

Predicted fire danger map across the Mediterranean

Wildfire Danger Forecasting (Track A): The team compared LSTMs, Transformers, and Gated Transformer Networks (GTN). The attention-based models performed best, with the GTN and Transformer achieving an Area Under the Precision-Recall Curve (AUPRC) of 0.858 and 0.856, respectively. This demonstrates that attention mechanisms are highly effective at capturing the complex temporal "lead-up" to a fire event.

Burned Area Prediction (Track B): Using a U-Net architecture (common for image segmentation), they found that a model utilizing all 27 variables (AUPRC 0.418) outperformed a baseline that only knew the ignition point (AUPRC 0.394). While the improvement is modest, it quantitatively confirms that broad environmental context significantly aids in predicting fire spread.

Limitations and Future Work

While Mesogeos is a massive leap forward, the authors are transparent about remaining challenges. The most notable is the "Suppression Gap." The dataset currently lacks data on firefighting efforts; a fire might stay small not because of weather or fuel, but because of a rapid response by firefighters—a factor current models cannot "see."

Additionally, cloud cover can sometimes obscure satellite observations, leading to potential inaccuracies in ignition dates or burned area perimeters.

Looking ahead, the authors suggest the dataset could be used for novel tasks like Self-Supervised Learning (SSL)—where models learn general representations of Earth data before fine-tuning—or exact fire size regression. By releasing the code and the datacube publicly, the Mesogeos team has provided vital, open-access infrastructure for the next generation of AI-driven wildfire modeling.

Summary updated

BibTeX
@inproceedings{kondylatos2023mesogeos,
  author = {Spyridon Kondylatos and Ioannis Prapas and Gustau Camps-Valls and Ioannis Papoutsis},
  title = {{Mesogeos: A multi-purpose dataset for data-driven wildfire modeling in the Mediterranean}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
}
Figure