Fire weather indices tailored to regional patterns outperform global models

Fire weather indices (FWIs) are widely used to assess wildfire risk, but are typically designed for specific regions and not adapted globally. Here, we present a systematic effort to generate country-specific FWIs that capture regional fire-weather patterns. We evaluate three widely used indices across countries, finding that the Canadian FWI performs best overall (ROC AUC of 0.69). Tailoring the index to each country with a Genetic Algorithm significantly improves its accuracy, raising the ROC AUC from 0.69 to 0.79. To further improve accuracy while maintaining interpretability, we develop a single Decision Tree model per country, achieving an ROC AUC of 0.86. Attempts to develop a single global Decision Tree yielded substantially lower accuracy, highlighting the limitations of universal models and the importance of capturing regional characteristics such as weather patterns, vegetation types, and topography for accurately predicting wildfire risk. Adapting FWIs regionally is crucial under accelerating climate change conditions.