Susceptibility Mapping of Wildfires Using XGBoost, Random Forest and AdaBoost: A Case Study of Mediterranean Ecosystem

Abstract

The Mediterranean Region of Turkey has simultaneously witnessed intense wildfires throughout the summer seasons, resulting in many fatalities and injuries, linked to global warming and climate change. In the region, Antalya is one of the most vulnerable provinces to forest fire activities due to its climatic and anthropogenic conditions. In this chapter, three ensemble-based machine learning algorithms, including Extreme Gradient Boosting (XGBoost), Random Forest (RF), and AdaBoost (AB) were applied to model wildfire susceptibility of the Antalya province situated in the southern coastal zones of Turkey. For this aim, an inventory map was initially constructed by using the publicly available MODIS data retrieved from NASA’s Fire Information for Resource Management System between 2001 and 2020. Also, a total of 13 causative factors (temperature, precipitation, wind speed, elevation, slope, aspect, plan curvature, profile curvature, curvature, TRI, NDVI, LULC, and TWI) were selected by considering the main characteristics of the study area and previous studies conducted in this region. According to the results, the XGBoost algorithm produced the highest accuracy of 85.4%, followed by the RF (84.6%) and AB (78.9%). Considering the estimated AUC values, XGBoost and RF outperformed the AB algorithm by about 6%. McNemar’s statistical significance test results showed that RF and XGBoost algorithms produced similar performances, but their results significantly differ from that of the AB algorithm. According to the Shapley additive explanations (SHAP) strategy, elevation was found the most effective parameter for the occurrence of wildfire events, while curvature was the least effective one. SHAP analysis also showed that topographic (e.g., elevation and aspect) and climatic (e.g., temperature, precipitation, and wind speed) parameters had a larger impact on wildfire susceptibility. When all susceptibility maps were thematically interpreted, approximately 20% of the study area was covered by a very high fire susceptibility category. To be exact, the highest potential forest fire risk was mainly situated around the center and shoreline zones of the Antalya province. Taking all the results into consideration, producing a reliable wildfire susceptibility map could be used as an early warning system by decision-makers to manage, prevent, and mitigate potential fire risks and protect wildlife. © 2024 Elsevier B.V., All rights reserved.