AI Reconstructs 70 Years of Italian Climate, Revealing a Shifting Hydrological Cycle

Italy’s climate record has just been rebuilt by artificial intelligence. In a study published in Theoretical and Applied Climatology, Monica Bini and Marco Luppichini of the University of Pisa describe a deep learning framework that reconstructs monthly temperature and precipitation across the entire Italian territory on a fine 10 by 10 kilometre grid, spanning the years 1950 to 2020. The work tackles one of the most stubborn problems in climate science: historical weather observations are abundant but fragmented, heterogeneous, and riddled with gaps, particularly in regions like the Mediterranean where station networks changed repeatedly over decades. By training neural networks on an extensive archive of observational series, the researchers produced spatially and temporally coherent climate fields that fill those gaps and allow, for the first time at this resolution, a unified national picture of seven decades of climate change.

The Mediterranean basin is widely regarded as one of the planet’s most vulnerable climate hotspots, a region where warming is progressing faster than the global average and where water resources are already under strain. Understanding how temperature and rainfall have behaved in the past is essential for projecting what comes next, yet the observational record has always been the weak link. Individual weather stations open and close, instruments are replaced, stations move, and urban development alters local readings. These discontinuities mean that a simple average of available measurements can mislead, and the further back in time one looks, the sparser and less reliable the network becomes. Traditional approaches to this problem rely on homogenisation techniques, statistical procedures designed to detect and correct artificial shifts in the data, but these methods struggle when station density drops or when the inhomogeneities themselves are difficult to distinguish from genuine climate signals.

The Pisa team’s answer was to let machine learning do the heavy lifting. At the core of the methodology are two families of deep learning models: Long Short-Term Memory networks, known as LSTMs, and fully connected neural networks. LSTMs are a specialised form of recurrent neural network introduced in the late 1990s, designed specifically to learn patterns in sequential data. Unlike conventional networks, they carry information forward through time using internal memory cells and gating mechanisms, which allows them to capture dependencies that stretch across months and seasons. This makes them naturally suited to climate data, where the value of a variable in one month is shaped by a chain of preceding conditions. The fully connected networks complement this by learning nonlinear relationships between predictors at each location, drawing on spatial context to estimate values where direct observations are missing.

Trained on the extensive Italian observational archive, the models learned to reconstruct monthly climate variables across the regular 10 kilometre grid, effectively weaving thousands of patchy station records into a single seamless fabric. The output is a set of climate fields that are coherent both in space and in time, meaning that neighbouring grid cells tell a physically plausible story and that the evolution of any given cell through the decades is smooth rather than erratic. This is a crucial property, because naive gap-filling can produce artefacts, sudden jumps or dips that reflect the data rescue process rather than the actual atmosphere. The reconstructed product then had to prove itself against an independent benchmark, and the researchers turned to ERA5, the state-of-the-art global reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts, which blends observations with a numerical weather model to create a physically consistent record of the atmosphere.

The validation results are striking. Correlations between the reconstructed fields and ERA5 exceed 0.96 for temperature variables, an agreement so close that the two datasets are nearly indistinguishable in their broad patterns. For cumulative precipitation, correlations exceed 0.8, which is a strong result given that precipitation is notoriously harder to model than temperature, being patchier in space, more episodic in time, and heavily influenced by local topography. In a country like Italy, where the Alps and the Apennines create dramatic orographic effects and where a few tens of kilometres can separate a humid coastal plain from a rain-shadow valley, capturing precipitation realistically is a formidable challenge. The fact that a data-driven model achieves this level of fidelity suggests that the observational archive, once properly exploited, contains enough information to constrain the climate signal even where individual stations are silent.

With a validated reconstruction in hand, the team turned to the question that motivated the work: what has actually happened to Italy’s climate since 1950? The trend analysis revealed three key indicators of ongoing change. The first is widespread and persistent warming, with rates exceeding 0.04 degrees Celsius per year in mountainous regions. That figure, equivalent to more than four degrees per century, underscores the particular exposure of high-altitude environments, where snowpack, glaciers, and seasonal water storage are all sensitive to even modest temperature shifts. Mountain areas act as early warning systems for the wider climate system, and the Italian Alps and Apennines are clearly sounding the alarm.

The second and third indicators concern rainfall, and together they tell a more subtle and arguably more consequential story. The analysis found a significant decline in monthly cumulative rainfall across the country, meaning that the total volume of water delivered over a month has been shrinking. At the same time, it documented an intensification of daily extreme rainfall events, meaning that when it does rain, the downpours are becoming heavier. This dual pattern, less widespread rainfall but more intense extremes, is precisely the signature that physical reasoning predicts for a warming world. A warmer atmosphere holds more water vapour, roughly seven percent more per degree of warming, following the Clausius-Clapeyron relation, so any given storm has more moisture to draw upon. Meanwhile, changes in synoptic-scale atmospheric circulation, the large-scale patterns that steer weather systems across the Mediterranean, can reduce the frequency of rain-bearing conditions, so that dry spells lengthen between the increasingly violent events.

The authors interpret this combination as evidence of a structural transformation of the Italian hydrological cycle, driven both by thermodynamic processes, the direct physics of a warmer, moister atmosphere, and by shifts in the circulation patterns that govern when and where rain falls. The implications are far-reaching. A hydrological cycle that delivers the same or less total water in fewer, more concentrated bursts is a cycle that produces both more droughts and more floods. Reservoirs and soils that once received steady replenishment now face long dry intervals punctuated by runoff events so intense that much of the water escapes quickly to the sea rather than recharging aquifers. Agriculture, hydropower, civil protection, and urban drainage systems were all designed around a rainfall regime that the data show is no longer stable.

What elevates this study beyond a national climate audit is the accessibility of its output. The reconstructed dataset has been made publicly available through the AIClimate platform hosted by the University of Pisa, offering a concrete resource for climate change studies, hydrological modelling, and the planning of adaptation strategies in highly vulnerable regions. For hydrologists, a coherent 70-year gridded record is invaluable for calibrating rainfall-runoff models and assessing flood and drought hazard. For planners and policymakers, it provides the historical baseline against which future projections can be judged and against which adaptation measures, from reservoir management to land-use planning, can be designed. And for climate scientists studying the Mediterranean, it removes a long-standing obstacle: the fragmentation of national archives that has hampered regional synthesis.

The study also sits within a broader movement in the geosciences, one in which deep learning is increasingly deployed not to replace physical models but to extract maximum value from imperfect observations. From filling gaps in rainfall databases to predicting floods in fast-flowing watersheds and forecasting river flow in small catchments, the same family of techniques is proving adept at problems where data exist but are messy, incomplete, or nonlinear in ways that classical statistics handle poorly. Italy, with one of the longest and richest observational histories in the world yet also one of the most heterogeneous, is an ideal proving ground. The message of this research is twofold: the past seven decades of Italian climate can now be seen whole, and what they reveal is a country where warming is entrenched in its mountains, its rains are thinning, and its storms are growing fiercer. That combination, captured with unprecedented spatial detail and validated against independent reanalysis, turns a fragmented archive into a national climate memory, and it arrives at a moment when the Mediterranean can least afford to forget.

Subject of Research: Deep learning reconstruction of monthly temperature and precipitation trends across Italy from 1950 to 2020

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

Article References: Bini, M., & Luppichini, M. (2026). Artificial intelligence for climate reconstruction: Spatiotemporal modelling of precipitation and temperature trends in Italy. Theoretical and Applied Climatology, 157(10), Article 667. https://doi.org/10.1007/s00704-026-06599-9

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06599-9

Keywords: artificial intelligence, deep learning, LSTM, climate reconstruction, Italy, Mediterranean climate, precipitation trends, temperature trends, ERA5 reanalysis, extreme rainfall, hydrological cycle, climate change