Adapting Global Wildfire Danger Models to High-Stakes Regional Fire Regimes

Sub-seasonal-to-seasonal (S2S) forecasting is increasingly used to support management decisions, from resource pre-positioning to seasonal risk communication. Yet global burned-area forecasting models, are trained on data dominated by high-frequency, seasonally regular fire regimes. This raises questions on their trustworthy deployment on regions of dense population exposure and high operational relevance which often exhibit irregular wildfire activity that is under-represented in training. Using a monthly forecast setup, we study this question across three such regions (Greece, California, New South Wales) and two architectures (FireCastNet, TeleViT). We compare a range of fine-tuning strategies, including the implementation of spatially dense graphs over the regions of interest (Local Area Modeling, LAM). LAM-adapted FireCastNet proves the most reliable and best-performing configuration throughout, consistently improving AUPRC over the global model across all regions. In absolute terms, skill remains modest in Greece and California, reaching [14.48--16.91]\% AUPRC in Greece, [16.24--18.21]\% AUPRC in California across forecast horizons. Notably New South Wales shows both the largest absolute AUPRC reaching [47.25--48.99]\% and the largest margin over its climatology baseline (24.06\% AUPRC) of the three regions, with roughly twice the performance ($\Delta$AUPRC $\approx$ +101\%). Integrated Gradients attribution suggests that for the region of New South Wales, the fine-tuned model has learned to rely on high-memory drivers such as vegetation state and accumulated drought, while for the regions of Greece and California it remains dominated by solar radiation, a shorter-memory signal, suggesting that the model identifies distinct fire regimes with differing degrees of predictability.