Artificial intelligence for climate reconstruction: Spatiotemporal modelling of precipitation and temperature trends in Italy

Abstract

Understanding past climate dynamics is essential to address the current and future challenges of climate change, particularly in highly vulnerable areas such as the Mediterranean basin. However, the use of observational data is often limited by the fragmentation, heterogeneity, and discontinuity of historical time series. In this study, we present an innovative methodology based on deep learning models (LSTM and fully connected neural networks) for reconstructing monthly climate data on a regular grid (10 km × 10 km) across the entire Italian territory over the period 1950–2020. Using an extensive archive of observational series, the developed models were able to fill data gaps and generate spatially and temporally coherent climate fields, which were then validated against the ERA5 reanalysis dataset. The resulting correlations exceed 0.96 for temperature variables and 0.8 for cumulative precipitation, confirming the accuracy and reliability of the reconstructed product. Trend analysis revealed three key indicators of ongoing climate change: (i) widespread and persistent warming, with rates > +0.04 °C/year in mountainous regions; (ii) a significant decline in monthly cumulative rainfall; and (iii) an intensification of daily extreme rainfall events. This dual pattern, less widespread rainfall and more intense extremes, suggests a structural transformation of the Italian hydrological cycle, driven by thermodynamic processes and changes in synoptic-scale atmospheric circulation. The final dataset, accessible via the AIClimate platform (https://lca.dst.unipi.it/AIClimate), offers a concrete resource for climate change studies, hydrological modelling, and the planning of adaptation strategies in highly vulnerable regions.