Alpha-Klima releases its improved first dataset of EU Fire Probability

Alpha-Klima Wildfire Risk v1.0

2.5 km × 2.5 km annual probabilities that at least one wildfire occurs in every grid cell across Europe (2001-2050), modelled with an explainable, monotonic-constraint XGBoost approach and two emissions pathways (RCP 4.5 & RCP 8.5).

  • Coverage 36 countries • EPSG: 3035
  • Scenarios Historical (2001-22) + RCP 4.5 & 8.5 (2023-50)
  • License CC BY-NC 4.0 • see details

1. Introduction & Alignment with Previous Research

Alpha-Klima Wildfire Risk v1.0 provides continent-wide, high-resolution probabilities of wildfire occurrence from 2001 to 2050. The workflow builds on the single-step, monotonic-constraint methodology in ECB Statistics Paper No 49 and Burger et al. (2024).

  • Meteorological drivers — annual mean and maximum Fire Weather Index (FWI) from Copernicus.
  • Land-surface predictors — MODIS MCD12Q1 fractional land-cover classes.
  • Anthropogenic features — distance to roads, railways, and urban areas (HARCI-EU, Geoapify).

Key similarities to Burger et al. (2024)

  • Single-step occurrence modelling (no burned-area sub-model).
  • Gradient-boosted trees with monotonic constraints for physical consistency.
  • Europe-wide coverage at 2.5 km, including Turkey.
  • Historical calibration (2001-22) plus scenario projection (2023-50).

Key differences to Burger et al. (2024)

  • Urban layer— Geoapify urban extents for more significance of the distance to human settlements.
  • Automated pipeline — full reproducibility with IBM Research’s CLAIMED Framework.

2. Methods

2.1 Model Inputs & Pre-processing

  • Meteorology (FWI): Historical 2001-2005 plus RCP 4.5 / 8.5 projections 2005-2050; mean & max statistics (year resolution).
  • Land Cover: MODIS MCD12Q1 (500 m) aggregated to 2.5 km; forest, cropland, grassland, urban fractions.
  • Anthropogenic: Distance to nearest road, railway, urban area (shorter = higher ignition risk).
  • Fire observations: MODIS MCD64A1 converted to binary labels (2001-23).
  • Grid harmonisation: All layers at 2.5 km (EPSG 3035); missing values nearest-neighbour-filled.

2.2 Modeling Approach

  1. Target — binary flag: any wildfire in cell-year.
  2. Algorithm — XGBoost + monotonic constraints; spatial-temporal k-fold; isotonic calibration.

2.3 Bias Correction & Scaling

  • Static inputs after 2022 — land cover & infrastructure frozen.
  • Interpolation / resampling — bilinear (continuous) / majority or nearest-neighbour (categorical).

3. Scenario Outputs & Caveats

Historical (2001-22) — Model aligns with observed patterns from MCD64A1.

Forward-looking (2023-50) — RCP 4.5 & 8.5 projections indicate rising wildfire probability across the Mediterranean and parts of Central Europe, with decreases in wetter Atlantic zones.

Known limitations

  1. Static land cover and infrastructure beyond 2023.
  2. Only Historical, and RCP 4.5 & 8.5 scenarios supplied.
  3. Not designed for real-time emergency response.

4. Datasets Used

  1. Fire Weather Index (FWI) — Copernicus C3S
  2. Burned-area history — NASA MODIS MCD64A1
  3. Land-cover classification — NASA MODIS MCD12Q1
  4. Critical infrastructure & urban extents — HARCI-EU, Geoapify

5. Licensing & Attribution

License: CC BY-NC 4.0

Citation: Alpha-Klima (2025). Wildfire Risk for Europe under RCP 4.5 and RCP 8.5: 2001-2050 (v1.0). DOI 10.5281/zenodo.15123977

6. Contact

Alpha-Klima Team
C/ José Bardasano Baos 9, 7ºAB, 28016 Madrid, Spain
contacto@alpha-klima.com