Comparative Assessment of Machine Learning Algorithms for Wildfire Susceptibility Mapping in South Wales, Australia

Abstract. Wildfire susceptibility mapping is an essential tool for proactive fire management in regions prone to wildfires. This study aims to determine areas of South Wales, Australia, that are susceptible to wildfires using a comprehensive set of environmental parameters and multiple machine learning (ML) models. A geospatial data set of topographical (digital elevation model, slope, aspect), climatic (temperature, precipitation, wind speed, soil moisture), vegetative (normalized difference vegetation index, forest cover, land use), and anthropogenic (distance to roads and rivers) attributes was created. Seven ML classifiers were developed: Random Forest, Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), CatBoost, Extreme Gradient Boosting (XGBoost), and Natural Gradient Boosting (NGBoost). Four-fold cross-validation was used to test the models, with area under the receiver operating characteristic (ROC) curve (AUC) being the primary model performance metric. Results indicate that ensemble tree-based models were superior to other approaches in performance. CatBoost, LightGBM, and XGBoost were the best performers, with maximum mean AUC values higher than 0.9. The least effective model among those tested was the SVM model. Across all the models tested, NDVI was determined to be the top predictor of wildfire susceptibility.