Tonbul et al. (2026) Machine-learning wildfire susceptibility mapping with SHAP-based explainability in Türkiye’s fire-prone regions

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2026
• Date: 2026-01-01
• Authors: Hasan Tonbul, Sander Veraverbeke
• DOI: 10.1007/s00477-025-03136-4

Research Groups

• Department of Geomatics Engineering, Faculty of Civil Engineering, Istanbul Technical University, Istanbul, Türkiye
• Faculty of Science, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands
• School of Environmental Sciences, University of East Anglia, Norwich, UK

Short Summary

This study develops machine-learning models with SHAP-based explainable AI to map wildfire susceptibility in Türkiye's Mediterranean and Aegean regions, identifying key drivers and their interactions to enhance transparent decision-making for wildfire management.

Objective

• How do machine learning (ML) models predict wildfire susceptibility in the Mediterranean and Aegean part of Türkiye?
• What are the key drivers of wildfire susceptibility, and how do their interactions influence predictions in this region?
• How can model outputs be effectively interpreted and communicated to support decision-making?

Study Configuration

• Spatial Scale: Nine provinces in Türkiye's Mediterranean and Aegean regions (Antalya, Mugla, Mersin, Izmir, Hatay, Canakkale, Balıkesir, Aydin, and Adana), covering an area of approximately 116,176 square kilometers.
• Temporal Scale: Fire ignition data collected from 2013 to 2022 (10 years).

Methodology and Data

• Models used: Gradient Boosting Machine (GBM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) as machine learning classifiers. SHapley Additive exPlanations (SHAP) for explainable AI (XAI), Kernel Density Estimation (KDE) for spatial density analysis.
• Data sources:
  • Fire Ignitions: Visible and Infrared Radiometer Suite (VIIRS) active fire data (375 m spatial resolution, daily temporal repeat).
  • Land Cover: European Space Agency Climate Change (ESA CCI) land cover product (Forests, Shrubland, Grassland categories).
  • Climatic Factors: NASA Global Land Data Assimilation System (wind speed, soil temperature), TerraClimate (vapor pressure, soil moisture), ERA5 (precipitation, evapotranspiration, relative humidity, temperature, solar radiation), MOD11A1 (Land Surface Temperature - LST).
  • Vegetative Factors: Copernicus Land Monitoring Service (tree cover density), Normalized Difference Vegetation Index (NDVI), Fraction of Absorbed Photosynthetically Active Radiation (FPAR).
  • Topographic Factors: Elevation, slope, aspect.
  • Anthropogenic Factors: Proximity to settlements, proximity to roads, human population density, proximity to rivers.
  • All raster data accessed and processed via Google Earth Engine (GEE) Python API using the geemap package. Predictive models implemented with scikit-learn.

Main Results

• The Gradient Boosting Machine (GBM) model demonstrated the best overall performance with an Area Under the Curve (AUC) of 0.839, a recall of 0.765, precision of 0.835, and an F1-score of 0.813.
• Spatiotemporal analysis revealed peak fire activity in August (92 fires), July (75 fires), and September (65 fires) between 2013 and 2022. Hatay, Antalya, İzmir, and Mugla were identified as the regions with the highest fire activity.
• SHAP analysis identified temperature (mean absolute SHAP value: 1.22), Leaf Area Index (LAI, 1.13), and solar radiation (1.05) as the most influential conditioning factors for wildfire susceptibility.
• Partial dependence plots indicated a significant increase in fire susceptibility for temperatures exceeding 16 °C and solar radiation values above 3000 W/m². Higher LAI values also correlated with increased susceptibility.
• Elevation showed an inverse relationship with fire susceptibility, with values decreasing above 500 meters. Proximity to settlements and roads increased fire risk.
• Pairwise interaction analysis highlighted strong interdependencies, with the highest interaction strength observed between solar radiation and wind speed (0.30), followed by solar radiation and soil temperature. Temperature and solar radiation were involved in the highest number of significant interactions (18 each).
• The generated wildfire susceptibility map classified 13.9% of the study area under very high fire risk and 15.8% under high fire risk, primarily concentrated in Mugla, İzmir, Hatay, Antalya, and parts of Mersin.

Contributions

• Benchmarking of state-of-the-art ensemble machine learning models (GBM, RF, XGBoost) under explicit collinearity control for wildfire susceptibility mapping.
• Integration of SHAP-based explainable AI with threshold exploration and pairwise interaction analysis to illuminate non-linear and joint effects of climatic, vegetative, topographic, and human-access variables.
• Translation of model explanations into management-relevant guidance for a large, multi-provincial Mediterranean setting in Türkiye.
• Utilization of a multi-year (2013–2022) fire ignition record from remote sensing data and a broad range of 21 conditioning factors for a comprehensive and decision-oriented susceptibility assessment.

Funding

• The Scientific and Technological Research Council of Turkey (TÜBİTAK) under the 2219 International Postdoctoral Research Fellowship Program (Project No: 1059B192300251).

Citation

@article{Tonbul2026Machinelearning,
  author = {Tonbul, Hasan and Veraverbeke, Sander},
  title = {Machine-learning wildfire susceptibility mapping with SHAP-based explainability in Türkiye’s fire-prone regions},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2026},
  doi = {10.1007/s00477-025-03136-4},
  url = {https://doi.org/10.1007/s00477-025-03136-4}
}