Landscape-scale prediction of spruce budworm-induced host mortality using machine learning

Context Insect outbreaks are an important cause of tree mortality in North America. The recurrent outbreaks of spruce budworm ( Choristoneura fumiferana , Clem., SBW) can extend beyond millions of hectares and cause extensive mortality and growth reduction in host trees, with important ecological and economic consequences. Objectives While prior research has focused on stand-level impacts of outbreaks, fewer studies have focused on mortality at the landscape level. We aimed to assess the performance of machine learning (ML) algorithms to predict SBW-induced mortality, inferred from biomass loss, and to identify key contributing drivers. Methods The study was carried out across forest landscapes in Québec, Canada. We compared eight ML algorithms and developed an ensemble model from the best-performing models. Predictors included stand-scale variables (forest composition, climate, topography) and neighbourhood variables (host proportion, patch complexity and abundance) computed at 10, 18 and 26 km. Results Ensemble models achieved an Area Under the Precision-Recall Curve (AUC-PR) of 0.689–0.707 and a Kappa score of 0.535–0.548. Mortality probability increased with higher primary host proportion, patch complexity and summer temperature, but decreased with secondary host proportion at the local and neighbourhood scale. Using these models, we mapped mortality probability for the ongoing outbreak that started in 2006. Conclusions By identifying vulnerable areas, this study can be used in targeting suppression activities. Furthermore, it demonstrates the performance of ML in modelling ecological disturbances across broad environmental gradients, with potential applications to other regions and insect outbreak systems.