FOREST FIRE SUSCEPTIBILITY ASSESSMENT WITH MACHINE LEARNING METHODS IN NORTH-EAST TURKIYE

Omer Kantarcioglu, Kaspar Schindler, Sultan Kocaman2023 · DOI

Abstract

Forest fires have devastating effects on biodiversity, climate, and humans. Producing detailed and reliable forest fire susceptibility maps is crucial for disaster management. This study assesses the performance of Random Forest (RF) and Artificial Neural Network (ANN) classifiers for producing forest fire susceptibility maps in north-east Türkiye (Trabzon, Gümüşhane, Rize, and Bayburt). Using twelve variables including topographic, anthropogenic, hydrologic, vegetation, and land use data, both models showed high predictive power (AUC 0.88-0.89). Random Forest performed better in terms of overall accuracy and F-1 score. Elevation was identified as the most influential factor.

Introduction

As climate change intensifies, forest fires are becoming more frequent and severe, particularly in Mediterranean and Black Sea regions. To manage these disasters, authorities require "susceptibility maps"—predictive tools that highlight areas most likely to experience ignition. The challenge lies in the local nature of fire risk; models must account for specific geographic complexities. This study focuses on the north-east provinces of Türkiye, characterized by the rugged Pontic Mountains, to determine which machine learning (ML) approach handles this terrain best and which environmental factors drive fire risk.

Methodology

The study follows a supervised learning pipeline applied to high-resolution geospatial data. Twelve input variables were curated across five categories: Topographic, Anthropogenic, Hydrologic, Vegetation, and Land Use. Two primary models were compared: Random Forest (RF), an ensemble method, and Artificial Neural Network (ANN), a multi-layer architecture for non-linear relationships.

Methodological workflow for forest fire susceptibility mapping.

The models produced spatial probability distributions resulting in susceptibility maps with a 25-meter resolution. Feature importance was evaluated using a "leave-one-out" approach to rank the influence of each variable.

Study Area and Datasets

The research covers the provinces of Trabzon, Gümüşhane, Rize, and Bayburt. The study integrated several high-quality datasets:

  • Fire Inventory: 368 fire ignition points (2013–2022) from the Trabzon Regional Directorate of Forestry.
  • Satellite Imagery: Sentinel-2 data for NDVI and MNDWI calculations.
  • Topography: EU-DEM v1.1 (25m resolution).
  • Land Cover: Copernicus Land Monitoring Service and ESA WorldCover.
  • Infrastructure: OpenStreetMap and World Bank data for settlements and power lines.
Study area location in Türkiye and distribution of fire ignition points.

Results and Discussion

Both models were highly effective, with the Random Forest achieving an Area Under the Curve (AUC) of 0.89 and the ANN achieving 0.88. However, RF outperformed ANN in secondary metrics, including Overall Accuracy (0.86 vs 0.80) and F1-Score (0.81 vs 0.75).

Key insights include:

  • Elevation as a Primary Driver: Elevation was the most influential factor, likely because it dictates local climate and vegetation types in the Pontic Mountains.
  • Model Characteristics: The ANN model tended to be more "pessimistic," predicting higher susceptibility across larger areas, which could lead to resource misallocation. The RF model provided a more granular and precise risk assessment.
  • Human Factors: While topography is vital, human infrastructure (like roads, noted in other studies) remains a critical "spark" for ignitions.

Conclusion

Random Forest proved to be a more reliable model for identifying specific high-risk zones in the complex terrain of north-east Türkiye. Future work should incorporate forest road networks to improve accuracy regarding human-caused ignitions and further investigate the impact of specific forest species on fire susceptibility.

Summary updated

Abstract

Forest fires have devastating effects on biodiversity, climate, and humans. Producing detailed and reliable forest fire susceptibility maps is crucial for disaster management. Data-driven machine learning methods can be applied for forest fire susceptibility mapping, and learning data required for this purpose can be obtained from high-resolution satellite imagery along with a fire inventory. In this study, we assessed the performances of Random Forest (RF) and artificial neural network (ANN) classifiers for producing forest fire susceptibility maps of a region in north-east Türkiye covering Trabzon, Gümüşhane, Rize, and Bayburt provinces using freely available Earth observation data and forest inventory provided by the regional directorate. Forest type, EU-DEM v1.1 (25 m), and tree cover density were retrieved from Copernicus Land Monitoring Service. Sentinel-2 images were utilized for calculating spectral indices such as normalized difference vegetation index and modified normalized difference water index to assess surface water and vegetation characteristics. Thus, a total of twelve variables including topographic, anthropogenic, hydrologic, vegetation and land use data were used as input. The RF and ANN illustrated similar prediction performances based on receiver operating characteristics (ROC) area under the curve (AUC) values, which were 0.89 and 0.88, respectively. The RF performed better in terms of overall accuracy and F-1 score. The susceptibility maps with 25 m resolution were also investigated visually. The ANN results predicted higher susceptibility levels and larger areas were found prone to wildfire. Leave-one-out analysis results indicated that elevation was the most influential factor based on the achieved OA.
BibTeX
@inproceedings{kantarcioglu2023forest,
  author = {Omer Kantarcioglu and Kaspar Schindler and Sultan Kocaman},
  title = {{FOREST FIRE SUSCEPTIBILITY ASSESSMENT WITH MACHINE LEARNING METHODS IN NORTH-EAST TURKIYE}},
  booktitle = {˜The},
  year = {2023},
  doi = {10.5194/isprs-archives-xlviii-m-1-2023-161-2023},
}
Figure