Fire-weather extremes over the Iberian Peninsula: Intensification and seasonal reorganization in deep-learning downscaled projections

Fire weather is projected to intensify across the Mediterranean, yet changes in the duration and timing of hazardous conditions remain poorly quantified, hampered by two limitations of general circulation models (GCMs): their coarse spatial resolution, which blurs regional fire-weather patterns, and their lack of the instantaneous noon variables required by the standard Fire Weather Index (FWI). Here we address both with a deep-learning (DL) downscaling framework: a U-Net maps coarse (1.5°) daily mean temperature, relative humidity, and wind speed to high-resolution (0.1°) 12 UTC FWI, while retaining the large-scale climate-change signal. Over the historical period (1979–2014), the DL downscaling improves the spatial structure, distribution, and extremes of FWI relative to GCM-based estimates, with the clearest added value for extreme fire-weather conditions. Applied to nine CMIP6 models across multiple SSPs and global warming levels (GWLs), the projections reveal robust, spatially coherent intensification of fire weather, strongest over Mediterranean and continental regions. Beyond this intensification, the results expose a systematic broadening of the fire-weather season: conditions historically confined to June–September increasingly emerge in late spring and persist into early autumn, shifting the ensemble-mean season to May–September at +3 °C GWL and, in some regions, extending toward May–October under stronger warming. This seasonal expansion shows substantial multi-model agreement and occurs even where absolute fire danger remains moderate, underscoring a change in the timing and duration of fire-weather extremes that GCM-resolution projections alone would miss.