Metaheuristic optimizer combined with advanced machine learning algorithm for accurate wildfire susceptibility in Western Canada: A novel approach

Increasing wildfire activity in a warming climate is a threat to ecosystems, communities and infrastructure, requiring accurate, interpretable and spatially robust susceptibility mapping. Existing hybrid wildfire susceptibility studies tend to focus on predictive correctness and pay little attention to the pseudo-absence uncertainty and consistency of the cross-strategy map. To address this gap, this study developed a province-wide 500 m wildfire susceptibility framework for Alberta, Canada through the use of Extreme Gradient Boosting with four swarm-based metaheuristic optimisers, i.e. Atom Search Optimisation, Particle Swarm Optimisation, Whale Optimisation Algorithm, Grey Wolf Optimisation and Shapley Additive exPlanations. Using 1733 historical wildfire events larger than 100 ha between 2000-2023, and 15 topographic, climatic, vegetation and anthropogenic predictors, the framework compared Random, Buffer and environmentally stratified pseudo-absence strategies and integrated outputs of these for more reliable and reduced spatial variability of susceptibility maps using an ensemble (a median approach). GWO-XGBoost had the best overall performance with mean Accuracy, AUC, AP and RMSE values of 0.910, 0.958, 0.944 and 0.292, respectively. Its superiority is probably due to a more stable and effective search of the hyperparameter space in XGBoost, leading to a stronger discrimination and a more consistent generalisation of the method on alternative pseudo-absence designs. SHAP analysis revealed that the annual temperature, wind speed and relative humidity were the most important controls on wildfire susceptibility. The final GWO-XGBoost map showed the persistence of a hot spot belt in north central and northeastern Alberta. The proposed framework can be used for preseason planning, targeted fuel management, early warning and spatial prioritisation of mitigation resources.