On the importance of both climate and vegetation evolution when predicting long-term wildfire susceptibility

Background Predicting long-term wildfire susceptibility requires consideration of both climate and vegetation shifts. Aims This study models and projects future wildfire susceptibility, explicitly considering the interplay of climate and vegetation evolution. A Graph Convolutional Network (GCN) captures the non-linear spatial dependencies of wildfire susceptibility on landscape and climate variables. Methods The model is tested over Portugal, projecting three Shared Socioeconomic Pathways (SSPs) climate scenarios. To decouple the role of climate and vegetation evolution, the model considers both with static vegetation and with vegetation dynamics simulated using a presence-only species distribution model. Key results The GCN is validated against historical wildfire data, and achieves Receiver Operating Characteristic Area Under the Curve (ROC-AUC) scores of 0.88 for the training set, 0.83 for the test set and 0.85 for the validation set, together with F1-scores of 0.81, 0.79, and 0.79, respectively. While different SSPs have a moderate general impact with static vegetation, incorporating vegetation dynamics leads to significant local variations. Notably, the low-emission SSP causes the most pronounced local changes due to climate-driven habitat suitability shifts, particularly favoring eucalyptus expansion and increasing susceptibility. Conclusions Long-term wildfire risk assessments must integrate both climate change and vegetation evolution to capture localized susceptibility changes, which is crucial to develop anticipatory land management strategies.