Spatially Consistent Fire Weather Index Predictions using Convolutional Neural Networks in Diverse Iberian Locations

The accurate prediction of the Fire Weather Index (FWI) is vital for effective wildfire management and climate-resilient planning. Multisite fire hazard forecasts are crucial for resource allocation, early intervention in high-risk areas, and identifying potential “megafire” threats from multiple simultaneous fire spots. Therefore, it is very important to account for the spatial consistency of these forecasts. This study examines the performance of Convolutional Neural Networks (CNNs) as a Statistical Downscaling (SD) technique for predicting FWI in different locations in the Iberian Peninsula. We contrast CNNs with two conventional SD methods: Generalized Linear Models and analogs. Using daily observed FWI data as predictands and ERA-Interim fields as predictors under a cross-validation setup, we discover that the CNN-Multi-Site-Multi-Gaussian (CNN-MSMG) model outperforms in daily FWI forecasting. This model integrates the covariance structure of the predictands into the CNN design, producing spatially consistent FWI forecasts. Furthermore, CNN-MSMG shows desirable features for estimating fire hazard in the climate change scenario, such as strong spatial consistency of extreme events and the capacity to generalize to new climate situations. These findings have important implications for improving FWI forecast accuracy and strengthening wildfire risk evaluation under climate change.