AI for wildfire danger forecasting at different spatiotemporal scales

Wildfire danger reflects the interaction of processes acting across a wide range of spatial and temporal scales, from rapid weather-driven variability to slower fuel, hydrological, and climate-mediated controls. This contribution examines how recent advances in artificial intelligence, when combined with structured Earth System Data Cubes, can be used to improve wildfire danger forecasts and also to better understand the mechanisms that drive their variability across scales.

We build on two complementary datacube paradigms: (i) regional, high-resolution daily cubes (e.g., Mesogeos at 1 km × 1 day over the Mediterranean) to resolve local meteorology–fuel–human interactions, and (ii) global sub-seasonal to seasonal cubes (e.g., SeasFire at 0.25° × 8-day, integrating climate, vegetation, oceanic indices, and human factors) to represent large-scale context and teleconnections.

For short lead times, we show that deep learning models that jointly exploit meteorological forcing and surface state information (e.g., vegetation condition and wetness proxies) consistently outperform operational meteorology-only approaches such as the Fire Weather Index. Importantly, explainable AI methods help diagnose which drivers dominate different fire episodes, revealing physically plausible and event-dependent controls rather than fixed empirical relationships. At subseasonal-to-seasonal horizons, predictability increasingly depends on slow-varying land-surface conditions and remote climate signals. Here, we discuss multi-scale learning approaches that fuse local predictors with coarser global fields and climate indices, enabling skillful forecasts of burned-area patterns at multi-month lead times without assuming homogeneous predictability across regions or biomes.

Finally, we argue that improved accuracy alone is insufficient for operational use. We therefore emphasize uncertainty-aware modelling, drawing on Bayesian deep learning to quantify epistemic and aleatoric uncertainties, improve forecast calibration, and support decision-making under risk through interpretable predictions accompanied by explicit confidence information.