Terrain-driven wildfire susceptibility assessment with explainable AI: A case study from Alberta

Wildfires are becoming increasingly frequent and severe across boreal landscapes, posing significant threats to environmental stability, public health, and economic security, where climate change has intensified both their frequency and scale. While recent advances in remote sensing and predictive modeling have improved wildfire susceptibility assessment, existing approaches differ in how environmental interactions are represented across spatial scales and modeling frameworks. In this context, the present study applies a scalable and interpretable wildfire susceptibility analysis framework to the boreal forests of Alberta, Canada. The objective of this study is to evaluate predictive performance and model transparency using established ensemble tree-based machine learning algorithms (LightGBM, XGBoost, Random Forest, Gradient Boosting, and AdaBoost), combined with exhaustive feature selection and Bayesian optimization. Using a dataset comprising sixteen environmental, meteorological, topographic, and human-related variables, 65,451 model configurations were evaluated to identify optimal input subsets for predicting wildfire occurrence. Model performance was validated through confusion matrices and various classification-based indices, and feature importance was quantified using SHAP values. Results show that LightGBM achieved the highest predictive performance among the tested models, with overall accuracy exceeding 98% usingonly eight input variables, which include latitude, longitude, plan curvature, profile curvature, temperature, precipitation, wind direction, and distance to roads, with the two curvature metrics jointly accounting for more than 76% of the total normalized mean absolute SHAP attribution. These findings highlight the dominant influence of terrain shape on wildfire susceptibility in Alberta's boreal region and demonstrate that high-resolution susceptibility assessment can be achieved using a reduced targeted set of predictors. The results provide region-specific, interpretable insights that may support wildfire risk assessment and decision-making in applications such as early warning, zoning, and adaptive climate resilience strategies.