Advancing fire weather forecasting: deep learning with Kolmogorov–Arnold and Fourier networks across Italy

Wildfires are among the most pressing environmental threats of the 21st century, with profound ecological, economic, and societal impacts. Southern Europe, and Italy in particular, faces escalating wildfire risk driven by the interplay of Mediterranean climate variability and accelerating climate change. This study presents the first documented application of Kolmogorov–Arnold Networks (KAN) and Kolmogorov–Arnold Fourier Networks (KAF) for forecasting the Fire Weather Index (FWI) across Italy, benchmarked against the conventional Long Short-Term Memory (LSTM) model. Using a 43-year ERA5-based dataset (1979–2021) of monthly mean FWI values, model performance was assessed across one- to six-month forecast horizons. All three architectures exhibited strong predictive skill, with KAN and KAF providing incremental yet consistent improvements over LSTM, particularly in regions with pronounced seasonal regularity. Beyond standard performance metrics, distributional diagnostics, including the Wasserstein Distance (WD) and Kernel Density Estimation (KDE), were applied to assess the models’ ability to reproduce the statistical structure of fire weather dynamics. The results confirm the feasibility, robustness, and operational potential of these next-generation architectures for integration into wildfire early warning systems. This work underscores how the synergy between advanced deep learning and climate intelligence can strengthen proactive fire risk management and long-term adaptation strategies under intensifying climate–fire interactions.