Yel et al. (2026) Wildfire susceptibility mapping with multiple machine learning algorithms utilizing forest inventory and FIRMS data: A case study in Arsin, Trabzon, Türkiye

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-01-10
• Authors: Sude Gül Yel, Derya Mumcu Küçüker, Esra Tunç Görmüş
• DOI: 10.1016/j.jag.2026.105091

Research Groups

• Department of Land Registry and Cadastre, Artvin Coruh University, Artvin, Türkiye
• Faculty of Forestry, Karadeniz Technical University, Trabzon, Türkiye
• Department of Geomatics Engineering, Karadeniz Technical University, Trabzon, Türkiye

Short Summary

This study developed wildfire susceptibility maps for the Arsin Forest Sub-District in Trabzon, Türkiye, by integrating official fire records and FIRMS data with five machine learning models, identifying anthropogenic factors, particularly population density and proximity to hazelnut cultivation, as key drivers of fire risk.

Objective

• To develop wildfire susceptibility maps for the Arsin Forest Sub-District, Trabzon, Türkiye, using machine learning techniques.
• To identify the primary conditioning factors influencing wildfire occurrence in the region.
• To evaluate and compare the performance of multiple machine learning and deep learning models for wildfire susceptibility mapping under local conditions.

Study Configuration

• Spatial Scale: Arsin Forest Sub-District Directorate, Trabzon, Türkiye. The study area covers 557.83 km², with 147.46 km² forested, ranging from sea level to 3082 m elevation. All spatial layers were resampled to a 10 m resolution.
• Temporal Scale: Wildfire inventory data span 2001–2022 (FIRMS: 2001–2012; official records: 2013–2022). Climatic data from WorldClim v2.1 cover 2010–2021, averaged over the local fire season (November–April).

Methodology and Data

• Models used: Random Forest (RF), Artificial Neural Network (ANN), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN). Hyperparameter optimization was performed using RandomizedSearchCV. Model interpretation used Information Gain, Gini Index, and SHapley Additive exPlanations (SHAP).
• Data sources:
  • Wildfire occurrence data: Official fire records (Arsin Forest Sub-District Directorate, 2013–2022) and MODIS FIRMS MCD14DL active fire pixel data (2001–2012).
  • Conditioning factors (17 variables):
    • Topographic: Slope, aspect, elevation, Topographic Wetness Index (TWI) derived from a Digital Elevation Model (DEM).
    • Climatic: Temperature, precipitation, solar radiation, wind speed from WorldClim v2.1 (2010–2021).
    • Stand-related: Canopy closure, developmental stage, dominant tree species from Forest Management Plan.
    • Anthropogenic: Population density (TUİK, 2022), proximity to roads, settlements, rivers, agricultural areas, and hazelnut orchards (derived from Forest Plan and buffer zones).
  • Data processing: QGIS, Google Earth Engine, SMOTE for class imbalance, Python libraries (Geopandas, NumPy, Rasterio, Scikit-learn, Matplotlib, Statsmodels, TensorFlow, Seaborn, Scikeras, Pandas).

Main Results

• All 17 conditioning factors were included after multicollinearity analysis (maximum Variance Inflation Factor = 5.10908).
• Random Forest (RF) and Gradient Boosting Machine (GBM) models demonstrated the highest and nearly identical validation accuracy of 0.976 and AUC-ROC scores of 0.98, outperforming other models.
• The Deep Neural Network (DNN) model showed signs of overfitting, despite high training accuracy, with lower recall and F1-scores in validation.
• Very high fire susceptibility areas were identified, covering 4.62% (RF) and 4.86% (GBM) of the study area.
• Anthropogenic factors were the most influential group (50.9% of model influence), with population density being the most important individual predictor (14% importance score).
• Proximity to hazelnut orchards ranked as the fourth most important factor (10%), indicating a significant positive correlation with fire risk.
• SHAP analysis revealed strong positive correlations between fire occurrence and population density, proximity to settlements, roads, agricultural lands, hazelnut orchards, temperature, and wind speed.
• Distance to rivers and elevation showed strong negative correlations with fire occurrence, while precipitation had a suppressive effect.

Contributions

• Developed comprehensive wildfire susceptibility maps for the understudied Eastern Black Sea region of Türkiye, addressing regional specificities in fire dynamics and seasonality.
• Integrated diverse data sources, including official fire records and satellite-based FIRMS data, to overcome historical data limitations and create a robust wildfire inventory.
• Compared five advanced machine learning and deep learning algorithms, identifying Random Forest and Gradient Boosting Machine as the most suitable for the study area.
• Introduced and validated a region-specific anthropogenic variable, "proximity to hazelnut cultivation," demonstrating its significant influence on wildfire susceptibility.
• Utilized SHapley Additive exPlanations (SHAP) for a transparent and directional interpretation of conditioning factors' influence, enhancing the explainability of complex model predictions.

Funding

• Not explicitly stated in the paper. The authors acknowledged the Trabzon Regional Directorate of Forestry for providing official wildfire inventory records.

Citation

@article{Yel2026Wildfire,
  author = {Yel, Sude Gül and Küçüker, Derya Mumcu and Görmüş, Esra Tunç},
  title = {Wildfire susceptibility mapping with multiple machine learning algorithms utilizing forest inventory and FIRMS data: A case study in Arsin, Trabzon, Türkiye},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105091},
  url = {https://doi.org/10.1016/j.jag.2026.105091}
}