Artificial Intelligence for Climate Reconstruction: Spatiotemporal Modelling of Precipitation and Temperature Trends in Italy

Artificial Intelligence for Climate Reconstruction: Spatiotemporal Modelling of Precipitation and Temperature Trends in Italy | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Artificial Intelligence for Climate Reconstruction: Spatiotemporal Modelling of Precipitation and Temperature Trends in Italy Marco Luppichini, Monica Bini This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8702295/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version 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. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1 Introduction The study of historical climate series is of fundamental importance for understanding the patterns that govern climatology based on observational data. Such understanding is more crucial than ever in the current context of global climate change (Brönnimann et al. 2008 ; Hawkins and Sutton 2009 ; Vicente-Serrano et al. 2025 ). Climate change, already underway, increasingly demands the development of tools for territorial planning and environmental policies grounded in both past and projected climatology (IPCC 2023). However, effective planning requires reliable data: these are the foundation for developing strategies to mitigate and adapt to the impacts of ongoing climate change. Regions bordering the Mediterranean Sea, identified as a climate hotspot (Giorgi 2006 ; Lazoglou et al. 2024 ), are called upon to act swiftly and effectively. It is also worth emphasizing that climatology is deeply rooted in the cultural heritage of Mediterranean peoples: this region hosts some of the oldest and most extensive climate observation archives (Lundstad et al. 2023 ; Morbidelli et al. 2025 ). Italy is no exception, on the contrary, it stands out as one of the countries offering some of the longest and most continuous instrumental climate series in both the European and Mediterranean contexts (Luppichini and Bini 2025 ). In particular, cities such as Padua (Camuffo 1984 , 2002 ), Milan, Florence, and Pisa (Camuffo et al. 2025 ) boast systematic meteorological observations dating back to the 18th century, and in some cases even to the late 18th century, representing a scientific heritage of extraordinary importance. Nevertheless, meteorological and climate databases present several well-known critical issues, including: the discontinuity and limited temporal span of time series, missing data, relocation of measurement sites, and concerns regarding the reliability of sensors (Li-Juan and Zhong-Wei 2012 ; Venema et al. 2012 ; Ribeiro et al. 2016 ; Gubler et al. 2017 ; Coscarelli et al. 2021 ). This leads to a paradoxical situation: although we now possess an unprecedented volume of data in the history of climate research, it is often difficult to utilize. As a result, analyses tend to rely on a limited number of series deemed representative, often selected for their accessibility or apparent quality (New et al. 2000 ; Aalto et al. 2016 ; Peng et al. 2019 ). This causes a loss of geographic information, particularly concerning in complex contexts like Italy, where its central position in the Mediterranean basin and the presence of numerous mountain ranges create a highly articulated and locally variable climatology (Cantù 1977 ; Fratianni and Acquaotta 2017 ). Overcoming these limitations is a top priority, also to avoid losing a valuable legacy of information. In this study, we propose an innovative methodology for managing and analysing climate databases based on the use of deep learning models. Deep learning, through architectures such as convolutional neural networks (CNN) (Skansi 2018 ), recurrent neural networks (RNN) (Bustreo et al. 2018 ), and long short-term memory (LSTM)(Hochreiter and Schmidhuber 1997 ) networks, enables the capture of the spatiotemporal complexity of climatic phenomena, identifying hidden patterns even in the presence of significant discontinuities or noise. The integration of these approaches makes it possible to reconstruct continuous and coherent historical series on a regular grid (Sha and Guha 2023 ; Lupi et al. 2023 ; Guraka 2024 ), both spatially and temporally, thereby enabling new forms of high-resolution climate analysis and facilitating the creation of harmonised datasets on a national scale. The importance of adopting deep learning techniques in this context lies not only in their predictive capabilities but also in their flexibility in addressing the structural complexities typical of climate data. These models allow for the integration of multivariable information, the handling of heterogeneous time series, and the generalization to under-observed geographic areas (Luppichini et al. 2022a , 2024 ; Lupi et al. 2023 ). What was hardly feasible just a few years ago is now possible thanks to access to powerful hardware infrastructures, such as GPUs and cloud-based systems, which allow for the training of deep neural networks on large-scale datasets (LeCun et al. 2015 ; Goodfellow et al. 2016 ). Additionally, the development of open-source frameworks like TensorFlow (Abadi et al. 2015 ) has lowered the barrier of entry for the scientific community, making these technologies more accessible and adaptable in climatological research. In the current context, marked by the acceleration of climate change and the growing availability of data, the use of such techniques is not merely an opportunity but a necessity for addressing present and future environmental challenges with scientific rigour. The climate variables analysed are on a monthly scale and include: cumulative precipitation, maximum daily precipitation, mean temperature, minimum temperature, and maximum temperature. The analysis was conducted on a large national archive of Italian data, comprising tens of thousands of historical series. Deep learning models were applied to generate climate data on a regular grid with a spatial resolution of 10 km x 10 km, covering the time span from 1950 to 2020. The reconstructed data are available for consultation and free download through a web application accessible at the following link: https://lca.dst.unipi.it/AIClimate/ 2 Materials and Methods 2.1 Data time series The historical time series used in this work originate from a census promoted by the Istituto Superiore per la Protezione e la Ricerca Ambientale (ISPRA), within the project "Sistema Nazionale per l’Elaborazione e Diffusione di Dati Climatici" (SCIA; https://scia.isprambiente.it/ ). This portal gathers climate data collected over decades by various institutions dedicated to this activity (Morbidelli et al. 2025 ). We collected monthly data on total rainfall, maximum daily rainfall, mean temperature, maximum temperature, and minimum temperature from a national meteorological data portal. The cumulative rainfall dataset includes 11,189 stations and 3,918,202 records spanning the period 1860–2023. However, a preliminary inspection revealed that the majority of stations (11,024) became active only after 1950. Consequently, our analyses primarily focus on the 1950–2020 period, which comprises data from 10,995 stations (Fig. 1 ). The maximum daily rainfall dataset consists of 10,556 stations and 1,239,717 records for the period 1860–2023, with 10,363 stations active during 1950–2020 (Fig. 1 ). The mean temperature dataset includes 4,961 stations and the same number of records (1,239,717) from 1860 to 2023, with 4,927 stations active between 1950 and 2020 (Fig. 1 ). Similarly, the minimum temperature dataset comprises 4,961 stations and 1,239,891 records across the full temporal range, with 4,930 stations active during 1950–2020 (Fig. 1 ). Finally, the maximum temperature dataset contains 4,963 stations and 1,257,690 records for the period 1860–2023, of which 4,927 stations were operational between 1950 and 2020 (Fig. 1 ). The time series present several issues in terms of continuity and duration, which is a common limitation in many meteorological monitoring networks. Specifically, the average temporal coverage of rainfall time series spans approximately 36 years, but the actual average number of recorded years is only 31. For temperature series, the average temporal span decreases to 24 years, with an average of 21 years of actual data. This highlights the fact that only a limited number of stations cover the full reference period (1950–2020): approximately 120 temperature stations and 650 rainfall stations, with an average data gap of 16% and 30%, respectively (Fig. 2 ). Figure 1 and Fig. 2 illustrate the trends in station activity over time. Rainfall series show a gradual decline in the number of active stations, with a sharp drop at the end of the 1980s. In contrast, the number of active temperature stations increased steadily until recently, although a slight decline is observed over the past decade. 2.2 AI applications The dataset was used to develop deep learning models aimed at predicting monthly rainfall values at individual stations, using information from surrounding stations as input. A total of 32 models were constructed, differing in several aspects: the number of neighbouring stations considered, the inclusion of additional climate data, whether or not input standardisation was applied, and the type of neural network architecture adopted. Each record in the database contributed to building the input matrix for the neural networks. Specifically, the monthly rainfall value at a given station was predicted based on the same month's values from n neighbouring stations, where n was set to 5, 10, 15, or 20. For selected model configurations, the input matrix was enhanced with climate teleconnection indices covering the previous 12 months. These indices include: Atlantic Multidecadal Oscillation (AMO) East Atlantic (EA) Mediterranean Oscillation Index (MOI) North Atlantic Oscillation (NAO) Western Mediterranean Oscillation (WeMO) Sea Surface Temperatures (SSTs) from the Gulf of Genoa (GGSST), the broader Mediterranean Sea (MSST), and the North Atlantic (NASST) The SST data were sourced from the Extended Reconstructed Sea Surface Temperature (ERSST) database. All data inputs were standardised to a [0,1] scale to reduce dimensional imbalance and improve model training efficiency, in line with best practices for optimising neural network convergence (He et al. 2023 ). To perform the regression task, two distinct deep learning architectures were developed and tested: a fully connected feedforward neural network (FCNN) a recurrent neural network based on Long Short-Term Memory (LSTM) units with an encoder–decoder structure. Both models were implemented using the Keras Sequential API and trained to minimise the Mean Absolute Error (MAE), selected for its robustness against outliers. The FCNN architecture consists of four hidden layers with progressively smaller sizes (4096, 1024, 256, and 64 neurons), each using ReLU activation. To prevent overfitting and promote generalisation, L2 regularisation (λ = 0.01), batch normalisation, and dropout (30%) were applied at every layer. The LSTM model leverages the temporal nature of the input data. It follows an encoder–decoder framework, with the encoder composed of two LSTM layers (64 and 32 units, respectively). The first layer returns the full sequence, while the second compresses it into a final state. This state is then repeated using a RepeatVector and passed to the decoder, which also consists of two LSTM layers (32 and 64 units) that output a full sequence. The output is further processed by a TimeDistributed Dense layer, followed by a Flatten layer and two Dense layers to produce a final scalar prediction. Model training was conducted using the Adam optimiser with an initial learning rate of 0.001. Two callbacks were employed to monitor performance and avoid overfitting: EarlyStopping : halts training if the validation loss does not improve over a specified number of epochs. ModelCheckpoint : saves the model weights with the best validation performance. The dataset was split into three subsets: training (60%), validation (20%), and testing (20%). This partitioning ensures fair and unbiased model evaluation: the training set is used for learning, the validation set for tuning during training, and the test set provides an independent evaluation of predictive performance. A detailed overview of the model configurations is presented in Table 1 . Once trained, the models were applied to a regular 10 km × 10 km grid to reconstruct a high-resolution, continuous climatic dataset spanning from 1950 to 2020. This spatial resolution was chosen as a conservative balance between detail and representativeness, considering the uneven distribution of meteorological stations across Italy (Fig. 1 ). A finer grid could produce artificial variability due to local oversampling, whereas 10 km spacing is consistent with the typical climatic gradients observed over the Italian peninsula (Fratianni and Acquaotta 2017 ). All model outputs are available for consultation and download through the AI Climate web application ( https://lca.dst.unipi.it/AIClimate/ ) Table 1 Summary of the models Model Data Input Standardization Neural Network DL1 Rainfall No Dense DL1STD Rainfall Yes Dense DL2 Rainfall, Atmospheric Teleconnections No Dense DL2STD Rainfall, Atmospheric Teleconnections Yes Dense LSMT1 Rainfall No LSTM LSMT1STD Rainfall Yes LSTM LSTM2 Rainfall, Atmospheric Teleconnections No LSTM LSMT2STD Rainfall, Atmospheric Teleconnections Yes LSTM 3 Results 3.1 Time series reconstruction Figure 3 presents a detailed evaluation of the performance of a climate model across five meteorological variables: cumulative rainfall, maximum daily rainfall, maximum temperature, mean temperature, and minimum temperature. For each variable, three statistical indicators are reported: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²). The results are divided by dataset, distinguishing between training (in red), validation (in blue), and test (in green), thus allowing the assessment of both the learning phase and the model’s generalisation capability. In the case of cumulative rainfall, the model shows consistent behaviour across the three datasets, with low MAE and RMSE values and generally high R² values, above 0.9. This suggests good predictive performance even on independent data, that is, data not included during training, confirming the model’s ability to generalise to unseen distributions. In contrast, for maximum daily rainfall, higher errors and greater variability among datasets are observed, with lower R² values (ranging from 0.78 to 0.83), highlighting the model’s difficulty in accurately predicting extreme events, which are notoriously more complex to model due to their high variability. Turning to temperatures, the performance is very satisfactory. For maximum, mean, and minimum temperature, both MAE and RMSE remain low across all datasets, and the R² values are very high, often above 0.95 and in some cases close to 0.99. This indicates that the model is able to accurately represent thermal trends and generalizes well even on test data. This is therefore a very robust behaviour in temperature forecasting. It is noted that the use of the LSTM network tends to produce slightly better MAE values compared to other approaches, suggesting a greater ability to capture the main patterns of the signal. However, this accuracy in training appears to be accompanied by greater instability in RMSE and R² metrics, with a more marked difference between the training dataset and the validation and test sets, especially in the latter. This behaviour suggests a greater tendency toward overfitting, as the model performs significantly better on the training set than on independent data. In other words, the network may fit the training data very well but lose accuracy when applied to new situations, a dynamic that would require further measures to improve generalisation. Overall, the model appears very effective in forecasting temperatures and fairly reliable for cumulative rainfall, while it faces greater challenges in accurately reproducing maximum daily rainfall. The results obtained indicate a solid modelling framework. While we fully acknowledge that the seasonal cycle can inflate correlation-based metrics such as R², our evaluation also relies on MAE and RMSE, which quantify absolute errors and are not affected by the presence of a seasonal cycle. The consistently low MAE and RMSE values across stations and variables indicate that the model is not only reproducing the annual cycle but also providing accurate month-to-month estimates in absolute terms. For this reason, even though R² may reflect the contribution of seasonality, MAE and RMSE provide robust, season-independent evidence that the reconstruction skill is genuinely high. Figure 4 illustrates the application of the models developed in this study. It presents the annual mean of the predictions produced by the 32 climate models over six grid cells representative of the study domain. Each line represents the average value estimated by the models, while the uncertainty band reflects the variability among the simulations (90% confidence interval). This representation highlights how the adopted modeling approach allows the generation of continuous and spatially consistent climate time series. The joint analysis of the six sites shows strong internal consistency among the different variables and across various geographic regions, confirming the stability of the simulations even in the presence of heterogeneous climatic conditions. 3.2 Application of AI-derived products The evaluation of the results produced by the AI models also includes the analysis of trends and spatial relationships, aimed at enhancing the ability of the data to describe and characterize the study area. 3.2.1 Trend Analysis Figure 5 shows the trends obtained using the Mann-Kendall test for five climatic variables from 1950 to 2020. The slopes represent the average annual variation observed during the study period, expressed in specific units for each variable (mm/year or °C/year). Positive values indicate increasing trends, while negative values indicate decreasing trends. White areas denote either lack of data or non-significant results. For monthly cumulative rainfall, marked negative trends (− 0.4 mm/year) are observed over large portions of Northwestern and Southern Italy, especially in Sicily and Calabria. This trend is clearly evident, with some exceptions, along the Apennine chain. However, some Alpine and pre-Alpine areas show slight increases in precipitation. This highlights a growing spatial differentiation in available water resources and a complex pattern in cumulative rainfall trends. For maximum daily rainfall, widespread positive trends are evident in both the frequency and intensity of extreme precipitation events, especially in Northeastern Italy and along the northern Apennine chain (up to + 0.2 mm/year). This increase in extreme rainfall is consistent with climate model projections that foresee greater frequency and intensity of such events; this entails a potential rise in hydraulic risk, especially in already vulnerable areas such as Alpine valleys and densely populated regions. Monthly mean temperature shows a widespread warming across the entire country, with positive trends sometimes exceeding + 0.03°C/year. The Alpine areas and central Apennine chain are particularly affected, indicating accelerated warming in mountainous regions. Monthly minimum temperature also shows significant positive trends over much of Central and Southern Italy (up to + 0.04°C/year). However, some areas in Northwestern and central inland Italy show negative trends, suggesting the persistence of cold nights or increased thermal variability. Finally, monthly maximum temperature confirms a generalized warming of maximum temperatures, with increases up to + 0.04°C/year, especially intense in southern and insular regions. Overall, the results indicate a clear warming trend across the entire national territory, accompanied by a significant increase in extreme rainfall in the North and a decrease in average annual precipitation in several regions, including the Apennine chain. 3.2.2 Relationship between temperature and precipitation In an effort to characterize the relationships between temperature and rainfall variables, the Spearman correlation (Spearman 1904 ) was calculated between temperatures and precipitation indicators. To carry out this analysis, a seasonal decomposition of the time series was first performed, from which only the trend components were extracted. This approach allowed the removal of the seasonal component, which is relevant in both types of variables, thereby enabling a more focused analysis of the long-term signal. The results obtained are summarized in the six maps shown in the figure, which illustrate the spatial distribution of these correlations across the entire Italian territory. In the first group of panels (top row), the relationships between the three types of temperature and monthly rainfall are observed. Overall, a predominant negative correlation emerges: areas where temperatures, particularly maximum temperatures, tend to increase correspond to a decrease in monthly precipitation. This pattern is particularly evident in the northern regions and along the Adriatic coast. The thermal component thus appears to be inversely correlated with the monthly rainfall regime, suggesting that, as temperatures rise (Fig. 5 ), a progressive reduction in precipitation is observed. The situation changes significantly when considering the correlation between the temperature variables and the intensity of daily maximum rainfall. In this case, the signal is distinctly positive and well distributed across almost the entire study area. The highest correlations are especially observed in central-northern areas and along the Apennine ranges, where the maximum daily rainfall recordings increase with rising temperatures. This may indicate an active role of warming in driving intense convective processes. Minimum temperature also shows a positive association, although with greater spatial discontinuity, suggesting a connection with warmer and more humid nighttime conditions, which are favourable to atmospheric instability. Overall, these analyses highlight an interesting and climatically relevant dynamic: while temperature is inversely associated with the total amount of precipitation, it is positively associated with the intensity of the most extreme events. Figure 6 Relationship between temperature and rainfall quantified using Spearman correlation coefficients. 4 Discussion 4.1 Comparison against ERA5 To further strengthen the validity of the developed dataset, a direct comparison was carried out with the ERA5 dataset (Hersbach et al. 2020 ), extracted using the Google Earth Engine API and covering the 1979–2020 interval. The goal was to assess the data's consistency through statistical correlation analysis, using both the Spearman (Spearman 1904 ) and Pearson (Kirch 2008 )coefficients to quantify the intensity and direction of associations between the main climate variables analysed. Figure 6 provides a summary overview using boxplots of correlation values calculated across the spatial domain. The results show very high correlations for all variables, with particularly compact distributions and medians close to one for temperature variables. This confirms that the models are robust and their outputs are comparable to data obtained from other modelling approaches. Although precipitation shows greater dispersion in correlation values, the central tendency remains high, indicating strong alignment of the modelled results with the ERA5 reference even for this more variable component. Figure 7 extends this assessment spatially, presenting gridded maps of Spearman and Pearson correlations for each climate variable. The results highlight a strong and coherent spatial structure in agreement with ERA5. Correlations are especially high for temperature variables, with most of the domain exceeding values of 0.96. For precipitation, despite more evident spatial variability, most of the territory still presents values above 0.8. These patterns further confirm the models’ ability to reproduce both the magnitude and temporal behaviour of ERA5-derived climate fields. The analysis of modelled products against consolidated literature references is in fact an essential step for evaluating the dataset’s consistency with known climate dynamics. If the models faithfully replicate signals recognised in previous studies, this strengthens the validity of the adopted methodology and confirms the dataset’s reliability. Although ERA5 shows good performance when compared with our reconstructed fields, the use of reanalysis data cannot replace the need for a homogeneous, observation-driven climatic dataset. ERA5 is a modelling product that integrates observations through data assimilation, and therefore its spatial patterns and temporal variability remain influenced by the underlying numerical model and by the density of assimilated stations, which is highly heterogeneous across Italy. As a consequence, ERA5 may reproduce large-scale atmospheric dynamics well, but it can still exhibit local biases, especially in complex physiographic contexts such as mountain areas or coast–inland transitions, where fine-scale processes play a key role. This limitation is particularly relevant for variables such as precipitation and temperature extremes, for which reanalyses often underestimate variability and magnitude. Our reconstruction instead relies solely on in situ observations and is specifically designed to preserve the climatological characteristics of the Italian network, producing fields that are independent of model assumptions. Furthermore, our dataset provides a substantially finer spatial resolution (10 × 10 km) compared to ERA5 (0.25°), allowing a more detailed representation of local climatic gradients and physiographic structures. This level of detail is essential for regional climate studies, long-term trend analyses, and impact assessments, which require datasets that reflect the true heterogeneity of the territory. In addition, only a very small fraction of the Italian observational network (approximately 250 stations) is actually assimilated into ERA5, ensuring that the comparison between our fields and ERA5 does not involve circularity and confirming the need for a fully observation-based reconstruction. For these reasons, even though ERA5 performs well in terms of general correspondence, a reconstruction based on real observations remains necessary to obtain high-resolution, consistent, and physically grounded climate information. 4.2 Hydroclimatic Responses to Warming over Italy Trend analysis for the period 1950–2020 shows a generalised increase in monthly mean, minimum, and maximum temperatures, with growth rates exceeding + 0.04°C/year in some mountainous regions. This is in agreement with Brunetti et al. ( 2006 ), who reported a warming of about + 1.2°C across Italy between 1865 and 2003, with a marked acceleration since the 1980s. In parallel, a statistically significant reduction in monthly cumulative precipitation is observed, particularly in southern regions and along the Apennine ridge, with decreases exceeding − 0.4 mm/year. This seasonal and annual pattern confirms the findings of Brunetti et al. ( 2006 ) and Luppichini and Bini ( 2025 ), who observed a reduction in annual precipitation along with an increase in daily rainfall intensity across Italy (Cislaghi et al. 2005 ; Marani and Zanetti 2015 ; Guo and Montanari 2023 ). The statistical correlation between temperature and precipitation variables reveals a negative relationship between temperature and monthly rainfall, and a positive one between temperature and extreme daily rainfall. These results align with international studies that describe warming as having a dual effect: reducing the frequency of ordinary rainfall events (partly due to atmospheric stabilisation), while promoting short, intense convective episodes (Schröer and Kirchengast 2018 ). This is consistent with the literature on the central Mediterranean (Zittis et al. 2021 ), where temperature acts both as a suppressor of widespread rainfall, due to greater atmospheric stability, and as a driver of convective processes responsible for more intense and localised events (Schröer and Kirchengast 2018 ). These results are compatible with the Clausius-Clapeyron relationship, which describes the atmosphere’s increased capacity to hold moisture with rising temperatures (Hardwick Jones et al. 2010 ; Berg et al. 2013 ; Blenkinsop et al. 2015 ; Pumo et al. 2019 ). This empirical evidence represents one of the clearest signals of the hydrological cycle’s response to global warming in the Mediterranean context. The thermodynamic component’s influence on extreme rainfall intensification is further confirmed by the positive relationship observed between temperature and maximum daily rainfall, consistent with Clausius-Clapeyron scaling reported in the Mediterranean(Pumo et al. 2018 ; Pumo and Noto 2021 ; Luppichini et al. 2023 ; Noto et al. 2023 ; Haslinger et al. 2025 ) and in other parts of the world (Lenderink and Van Meijgaard 2008 ; Hardwick Jones et al. 2010 ; Lenderink and Attema 2015 ; Visser et al. 2021 ). In this sense, the analysis confirms that rising temperatures are associated with a higher probability of intense events. The observed increase in frequency and intensity of extreme daily rainfall fits within a broader context of hydrological cycle intensification, strongly influenced by synoptic-scale atmospheric circulation and the rise in Mediterranean SST. These changes are associated with the weakening of key cyclogenetic structures such as the Icelandic Low and the Gulf of Genoa Low (Trigo et al. 2002 ; Frankignoul et al. 2003 ; Luppichini et al. 2021 , 2022b ), accompanied by a strengthening of the positive NAO phase and more persistent anticyclonic conditions. Therefore, although clear signs of changing precipitation regimes are observed in Italy, particularly in the intensification of extreme events, these results must be interpreted within a broader framework shaped by local and regional factors. The association between higher temperatures and more intense daily rainfall is especially evident in contexts where atmospheric warming coincides with increased persistence of convective patterns or greater air mass instability (Haslinger et al. 2025 ). For this reason, the management of climate databases using modern technologies proves essential for establishing a solid and reliable base of high-resolution spatial climate data. 5 Conclusion This work presents a novel approach to reconstructing climate series with high spatial resolution and temporal continuity, applied across the Italian territory for the period 1950–2020. By leveraging deep learning models based on LSTM and feedforward architectures, it was possible to integrate fragmented historical series, fill observational gaps, and generate homogeneous monthly climate fields on a regular grid. Results indicate that the models accurately reproduce observed climate dynamics: modelled temperatures show strong agreement with ERA5 data (correlations > 0.96), and precipitation values—despite higher variability—exceed 0.8 on average across the domain. The national-scale statistical validation confirms the robustness and reliability of the reconstructed dataset. Trend analysis reveals three key indicators of ongoing climate change: A widespread and persistent warming, with rates exceeding +0.04 °C/year in mountainous regions; A significant decline in monthly cumulative rainfall; An intensification of extreme daily rainfall events, in line with trends observed throughout the Mediterranean. This dual behaviour—less widespread rainfall and more intense extremes—suggests a structural shift in the Italian hydrological cycle, driven by both thermodynamic processes and changes in synoptic-scale atmospheric circulation. In summary, this study demonstrates how artificial intelligence tools can overcome the limitations of traditional observation networks, producing climate products that are coherent, validated, and suitable for scientific and operational use. The resulting dataset offers a concrete resource to support hydrological modelling, climate risk assessment, and adaptation policy in a highly vulnerable Mediterranean context. Declarations Data available The dataset is accessible via the AIClimate platform (https://lca.dst.unipi.it/AIClimate/) Author contributions statement Marco Luppichini: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing Monica Bini: Conceptualization, Project administration, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing References Aalto J, Pirinen P, Jylhä K (2016) New gridded daily climatology of Finland: Permutation‐based uncertainty estimates and temporal trends in climate. Journal of Geophysical Research: Atmospheres 121:3807–3823. https://doi.org/10.1002/2015JD024651 Abadi M, Agarwal A, Barham P, et al (2015) TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems Berg P, Moseley C, Haerter JO (2013) Strong increase in convective precipitation in response to higher temperatures. Nat Geosci 6:181–185 Blenkinsop S, Chan SC, Kendon EJ, et al (2015) Temperature influences on intense UK hourly precipitation and dependency on large-scale circulation. Environmental Research Letters 10:54021 Brönnimann S, Luterbacher J, Ewen T, et al (2008) Climate Variability and Extremes during the Past 100 Years Brunetti M, Maugeri M, Monti F, Nanni T (2006) Temperature and precipitation variability in Italy in the last two centuries from homogenised instrumental time series. International Journal of Climatology 26:345–381. https://doi.org/https://doi.org/10.1002/joc.1251 Bustreo M, Beltrán-González C, Murino V, Camozzi F (2018) Recurrent Neural Networks. 59–79. https://doi.org/10.1007/978-3-319-75304-1_5 Camuffo D (1984) Analysis of the series of precipitation at Padova, Italy. Clim Change 6:57–77. https://doi.org/10.1007/BF00141668 Camuffo D (2002) History of the Long Series of Daily Air Temperature in Padova (1725–1998). Clim Change 53:7–75. https://doi.org/10.1023/A:1014958506923 Camuffo D, Becherini F, della Valle A (2025) Recovery of the Long Series of Precipitation in Pisa, Italy: Trend, Anomaly and Extreme Events. Climate 13:. https://doi.org/10.3390/cli13040073 Cantù V (1977) The climate of Italy. In: Wallen CC (ed) Climate of central and southern Europe. Elsevier, pp 127–184 Cislaghi M, De Michele C, Ghezzi A, Rosso R (2005) Statistical assessment of trends and oscillations in rainfall dynamics: Analysis of long daily Italian series. Atmos Res 77:188–202. https://doi.org/https://doi.org/10.1016/j.atmosres.2004.12.014 Coscarelli R, Caroletti G, Joelsson M, et al (2021) Validation metrics of homogenization techniques on artificially inhomogenized monthly temperature networks in Sweden and Slovenia (1950–2005). Sci Rep 11:. https://doi.org/10.1038/s41598-021-97685-7 Frankignoul C, Friederichs P, Kestenare E (2003) Influence of Atlantic SST anomalies on the atmospheric circulation in the Atlantic-European sector. Annals of Geophysics 46: Fratianni S, Acquaotta F (2017) The Climate of Italy. In: Soldati M, Marchetti M (eds) Landscapes and Landforms of Italy. Springer International Publishing, Cham, pp 29–38 Giorgi F (2006) Climate change hot-spots. Geophys Res Lett 33:. https://doi.org/https://doi.org/10.1029/2006GL025734 Goodfellow I, Bengio Y, Courville A (2016) Deep Learning. MIT Press Gubler S, Hunziker S, Begert M, et al (2017) The influence of station density on climate data homogenization. International Journal of Climatology 37:. https://doi.org/10.1002/joc.5114 Guo R, Montanari A (2023) Historical rainfall data in northern Italy predict larger meteorological drought hazard than climate projections. Hydrol Earth Syst Sci 27:2847–2863. https://doi.org/10.5194/hess-27-2847-2023 Guraka HB (2024) CLIMATE CHANGE FORECASTING USING DEEP LEARNING. EPRA International Journal of Climate and Resource Economic Review. https://doi.org/10.36713/epra18552 Hardwick Jones R, Westra S, Sharma A (2010) Observed relationships between extreme sub-daily precipitation, surface temperature, and relative humidity. Geophys Res Lett 37:. https://doi.org/https://doi.org/10.1029/2010GL045081 Haslinger K, Breinl K, Pavlin L, et al (2025) Increasing hourly heavy rainfall in Austria reflected in flood changes. Nature 639:667–672. https://doi.org/10.1038/s41586-025-08647-2 Hawkins E, Sutton R (2009) The Potential to Narrow Uncertainty in Regional Climate Predictions. Bull Am Meteorol Soc 90:1095–1108. https://doi.org/https://doi.org/10.1175/2009BAMS2607.1 He L, Liu Z, Bucci A (2023) Combining dimensionality reduction methods with neural networks for realized volatility forecasting. Ann Oper Res. https://doi.org/10.1007/s10479-023-05544-7 Hersbach H, Bell B, Berrisford P, et al (2020) The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society 146:1999–2049. https://doi.org/https://doi.org/10.1002/qj.3803 Hochreiter S, Schmidhuber J (1997) Long Short-Term Memory. Neural Comput 9:1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 IPCC (2023) IPCC, 2023: Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, H. Lee and J. Romero (eds.)]. IPCC, Geneva, Switzerland. Kirch W (ed) (2008) Pearson’s Correlation Coefficient. In: Encyclopedia of Public Health. Springer Netherlands, Dordrecht, pp 1090–1091 Lazoglou G, Papadopoulos-Zachos A, Georgiades P, et al (2024) Identification of climate change hotspots in the Mediterranean. Sci Rep 14:29817. https://doi.org/10.1038/s41598-024-80139-1 LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444. https://doi.org/10.1038/nature14539 Lenderink G, Attema J (2015) A simple scaling approach to produce climate scenarios of local precipitation extremes for the Netherlands. Environmental Research Letters 10:85001. https://doi.org/10.1088/1748-9326/10/8/085001 Lenderink G, Van Meijgaard E (2008) Increase in hourly precipitation extremes beyond expectations from temperature changes. Nat Geosci 1:511–514 Li-Juan C, Zhong-Wei Y (2012) Progress in Research on Homogenization of Climate Data. Advances in Climate Change Research 3:59–67. https://doi.org/10.3724/SP.J.1248.2012.00059 Lúcio P, Conde F, Ramos A (2006) A BAYESIAN APPROACH FOR RECOVERING AND HOMOGENISING METEOROLOGICAL TIME SERIES Lundstad E, Brugnara Y, Pappert D, et al (2023) The global historical climate database HCLIM. Sci Data 10:44. https://doi.org/10.1038/s41597-022-01919-w Lupi A, Luppichini M, Barsanti M, et al (2023) Machine learning models to complete rainfall time series databases affected by missing or anomalous data. Earth Sci Inform 16:3717–3728. https://doi.org/10.1007/s12145-023-01122-4 Luppichini M, Barsanti M, Giannecchini R, Bini M (2022a) Deep learning models to predict flood events in fast-flowing watersheds. Science of The Total Environment 813:151885. https://doi.org/https://doi.org/10.1016/j.scitotenv.2021.151885 Luppichini M, Barsanti M, Giannecchini R, Bini M (2021) Statistical relationships between large-scale circulation patterns and local-scale effects: NAO and rainfall regime in a key area of the Mediterranean basin. Atmos Res 248:105270 Luppichini M, Bini M (2025) Evolution of rainfall in Italy over the last 200 years: Interactions between climate indices and global warming. Atmos Res 326:. https://doi.org/10.1016/j.atmosres.2025.108276 Luppichini M, Bini M, Barsanti M, et al (2022b) Seasonal rainfall trends of a key Mediterranean area in relation to large-scale atmospheric circulation: How does current global change affect the rainfall regime? J Hydrol (Amst) 612:128233. https://doi.org/https://doi.org/10.1016/j.jhydrol.2022.128233 Luppichini M, Bini M, Giannecchini R, Zanchetta G (2023) High-resolution spatial analysis of temperature influence on the rainfall regime and extreme precipitation events in north-central Italy. Science of The Total Environment 880:163368. https://doi.org/https://doi.org/10.1016/j.scitotenv.2023.163368 Luppichini M, Vailati G, Fontana L, Bini M (2024) Machine learning models for river flow forecasting in small catchments. Sci Rep 14:26740. https://doi.org/10.1038/s41598-024-78012-2 Marani M, Zanetti S (2015) Long-term oscillations in rainfall extremes in a 268 year daily time series. Water Resour Res 51:639–647. https://doi.org/https://doi.org/10.1002/2014WR015885 Morbidelli R, Flammini A, Echeta O, et al (2025) A reassessment of the history of the temporal resolution of rainfall data at the global scale. J Hydrol (Amst) 654:132841. https://doi.org/10.1016/j.jhydrol.2025.132841 New M, Hulme M, Jones P (2000) Representing Twentieth-Century Space-Time Climate Variability. Part II: Development of 1901-96 Monthly Grids of Terrestrial Surface Climate. J Clim 13:2217–2238. https://doi.org/10.1175/1520-0442(2000)0132.0.CO;2 Noto L, Cipolla G, Pumo D, Francipane A (2023) Climate Change in the Mediterranean Basin (Part II): A Review of Challenges and Uncertainties in Climate Change Modeling and Impact Analyses. Water Resources Management 1–17. https://doi.org/10.1007/s11269-023-03444-w Peng S, Ding Y, Liu W, Li Z (2019) 1 km monthly temperature and precipitation dataset for China from 1901 to 2017. Earth Syst Sci Data. https://doi.org/10.5194/ESSD-11-1931-2019 Pumo D, Carlino G, Arnone E, Noto L V (2018) Relationship between extreme rainfall and surface temperature in Sicily (Italy). EPiC Series in Engineering 3:1718–1726 Pumo D, Carlino G, Blenkinsop S, et al (2019) Sensitivity of extreme rainfall to temperature in semi-arid Mediterranean regions. Atmos Res 225:30–44. https://doi.org/https://doi.org/10.1016/j.atmosres.2019.03.036 Pumo D, Noto L V (2021) Exploring the linkage between dew point temperature and precipitation extremes: A multi-time-scale analysis on a semi-arid Mediterranean region. Atmos Res 254:105508. https://doi.org/https://doi.org/10.1016/j.atmosres.2021.105508 Ribeiro S, Caineta J, Costa A (2016) Review and discussion of homogenisation methods for climate data. Physics and Chemistry of The Earth 94:167–179. https://doi.org/10.1016/J.PCE.2015.08.007 Schröer K, Kirchengast G (2018) Sensitivity of extreme precipitation to temperature: the variability of scaling factors from a regional to local perspective. Clim Dyn 50:3981–3994 Sha R, Guha T (2023) Climate Time Series Prediction with Deep Learning and LSTM. 2023 4th International Conference on Smart Electronics and Communication (ICOSEC) 1631–1637. https://doi.org/10.1109/ICOSEC58147.2023.10276117 Skansi S (2018) Convolutional Neural Networks. 121–133. https://doi.org/10.1007/978-3-319-73004-2_6 Spearman C (1904) The proof and measurement of association between two things. Am J Psychol 15:72–101. https://doi.org/10.2307/1412159 Trigo IF, Bigg GR, Davies TD (2002) Climatology of Cyclogenesis Mechanisms in the Mediterranean Venema V, Mestre O, Aguilar E, et al (2012) Benchmarking homogenization algorithms for monthly data. Climate of The Past. https://doi.org/10.5194/CP-8-89-2012 Vicente-Serrano S, Tramblay Y, Reig Gracia F, et al (2025) High temporal variability not trend dominates Mediterranean precipitation. Nature 639:658–666. https://doi.org/10.1038/s41586-024-08576-6 Visser JB, Wasko C, Sharma A, Nathan R (2021) Eliminating the Hook in Precipitation Temperature Scaling. J Clim 34:9535–9549. https://doi.org/10.1175/JCLI-D-21-0292.1 Zittis G, Bruggeman A, Lelieveld J (2021) Revisiting future extreme precipitation trends in the Mediterranean. Weather Clim Extrem 34:100380. https://doi.org/https://doi.org/10.1016/j.wace.2021.100380 Additional Declarations No competing interests reported. Supplementary Files supplementary.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 17 May, 2026 Reviews received at journal 15 Apr, 2026 Reviewers agreed at journal 30 Mar, 2026 Reviewers agreed at journal 27 Mar, 2026 Reviewers invited by journal 27 Mar, 2026 Editor assigned by journal 26 Jan, 2026 Submission checks completed at journal 26 Jan, 2026 First submitted to journal 26 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8702295","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":614339412,"identity":"c4b92695-8148-4f5c-8309-f8359c394f46","order_by":0,"name":"Marco Luppichini","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABEUlEQVRIiWNgGAWjYPACCyBmbGBmMGDg4WdIAAvxMDAzMBzArUUCoUWyIYGxAa4Ftx4JMMkMIgwOQLRAABYtuu1nH374wCCRuF26ufFzQcEdGePj6c8fMLbdkTFnZ2A8/AFTi9mZdGPJGUAtO+ccbJaeYfCMx+zMG8MGxrZnPJbN2B1mdiCNjZkHqGXDjcQ2Zh6DwzxmN3IYgVoOA9k4tJx/hqbFeEb6Q/xabqDbYiCRYEhAyzNmyRkGEsZgv4C0SAD9MiPh3GGgXxgbDpzB5rA0xg8fKmxkt0u3P/zM8+ewPX97+oMPH8oO25vzHz78oQJTCwQYAJEEskACWBApfrDqksAQwat+FIyCUTAKRhAAANxFal7GVpN5AAAAAElFTkSuQmCC","orcid":"","institution":"University of Pisa","correspondingAuthor":true,"prefix":"","firstName":"Marco","middleName":"","lastName":"Luppichini","suffix":""},{"id":614339413,"identity":"a7527f39-5f07-488c-8667-01c04b58ae31","order_by":1,"name":"Monica Bini","email":"","orcid":"","institution":"University of Pisa","correspondingAuthor":false,"prefix":"","firstName":"Monica","middleName":"","lastName":"Bini","suffix":""}],"badges":[],"createdAt":"2026-01-26 16:23:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8702295/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8702295/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105833820,"identity":"37d203f5-015b-419c-bb4a-01455cf8885a","added_by":"auto","created_at":"2026-03-31 15:12:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1712265,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution and temporal segmentation of meteorological stations used in the study.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/4bd3bdbe985a078191583f45.png"},{"id":105833837,"identity":"d46f0fe1-5ba7-498d-b8ba-3abf49ff401e","added_by":"auto","created_at":"2026-03-31 15:12:59","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":377645,"visible":true,"origin":"","legend":"\u003cp\u003eTemporal coverage, continuity, and data quality of the meteorological station archive for five climate variables.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/8979dfb3fae27f2d426a1ed7.png"},{"id":105833819,"identity":"25c68f42-3681-45ce-9654-28e1868fbed3","added_by":"auto","created_at":"2026-03-31 15:12:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":444801,"visible":true,"origin":"","legend":"\u003cp\u003eDeep Learning Models errors: MAE (Mean Absolute Error), RMSE (Root Mean Square Error), Coefficient of Determination (R\u003csup\u003e2\u003c/sup\u003e).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/9c7a730dc9f4ef6eac0f3471.png"},{"id":105833907,"identity":"d5122e13-287f-49c1-9fc7-3508ce5df3c9","added_by":"auto","created_at":"2026-03-31 15:13:13","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":774404,"visible":true,"origin":"","legend":"\u003cp\u003eApplication of the models in a regular grid 10x10 km and six examples of time series. The red line represents the median value, while the shaded red area indicates the 90% confidence interval.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/2cc1c2c3df44320cb49c3f72.png"},{"id":105833853,"identity":"6419c9b7-de08-4a60-be63-06af9c7b7ec3","added_by":"auto","created_at":"2026-03-31 15:13:02","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1316362,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution of trends identified through the Mann-Kendall test.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/8160a127a98740c3f5c498a2.png"},{"id":105833821,"identity":"ffcc0386-6e74-47a3-b1df-2bc983915729","added_by":"auto","created_at":"2026-03-31 15:12:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":813905,"visible":true,"origin":"","legend":"\u003cp\u003eRelationship between temperature and rainfall quantified using Spearman correlation coefficients.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/45de2e695b3c80abc41d6779.png"},{"id":105833851,"identity":"d0b46854-b6df-47fa-9b1c-bc05c00e4603","added_by":"auto","created_at":"2026-03-31 15:13:02","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":49745,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 6 Statistical correlations between the reconstructed dataset and ERA5 data.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/1b0bb09b86ece01ce3b6e3a0.png"},{"id":105833850,"identity":"aa9db651-7298-42f1-8e02-9b5848c0136e","added_by":"auto","created_at":"2026-03-31 15:13:01","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":789485,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 7 Spatial assessment of statistical correlations between the reconstructed dataset and ERA5 data.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/46eb55932d24b2c65a8f7c53.png"},{"id":105904922,"identity":"c13e7b4c-31b9-4d89-b62d-61848ff75c78","added_by":"auto","created_at":"2026-04-01 10:11:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6517635,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/8c4c2551-fede-4948-b9a5-2ed618c49e1b.pdf"},{"id":105833852,"identity":"e1089a26-4dd1-4d81-8d5d-ae7a02947840","added_by":"auto","created_at":"2026-03-31 15:13:02","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1127923,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8702295/v1/a651ec334407990e08c3e48c.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial Intelligence for Climate Reconstruction: Spatiotemporal Modelling of Precipitation and Temperature Trends in Italy","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eThe study of historical climate series is of fundamental importance for understanding the patterns that govern climatology based on observational data. Such understanding is more crucial than ever in the current context of global climate change (Br\u0026ouml;nnimann et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Hawkins and Sutton \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Vicente-Serrano et al. \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Climate change, already underway, increasingly demands the development of tools for territorial planning and environmental policies grounded in both past and projected climatology (IPCC 2023).\u003c/p\u003e \u003cp\u003eHowever, effective planning requires reliable data: these are the foundation for developing strategies to mitigate and adapt to the impacts of ongoing climate change. Regions bordering the Mediterranean Sea, identified as a climate hotspot (Giorgi \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2006\u003c/span\u003e; Lazoglou et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), are called upon to act swiftly and effectively. It is also worth emphasizing that climatology is deeply rooted in the cultural heritage of Mediterranean peoples: this region hosts some of the oldest and most extensive climate observation archives (Lundstad et al. \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Morbidelli et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Italy is no exception, on the contrary, it stands out as one of the countries offering some of the longest and most continuous instrumental climate series in both the European and Mediterranean contexts (Luppichini and Bini \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In particular, cities such as Padua (Camuffo \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1984\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), Milan, Florence, and Pisa (Camuffo et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) boast systematic meteorological observations dating back to the 18th century, and in some cases even to the late 18th century, representing a scientific heritage of extraordinary importance.\u003c/p\u003e \u003cp\u003eNevertheless, meteorological and climate databases present several well-known critical issues, including: the discontinuity and limited temporal span of time series, missing data, relocation of measurement sites, and concerns regarding the reliability of sensors (Li-Juan and Zhong-Wei \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Venema et al. \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Ribeiro et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Gubler et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Coscarelli et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This leads to a paradoxical situation: although we now possess an unprecedented volume of data in the history of climate research, it is often difficult to utilize. As a result, analyses tend to rely on a limited number of series deemed representative, often selected for their accessibility or apparent quality (New et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2000\u003c/span\u003e; Aalto et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Peng et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This causes a loss of geographic information, particularly concerning in complex contexts like Italy, where its central position in the Mediterranean basin and the presence of numerous mountain ranges create a highly articulated and locally variable climatology (Cant\u0026ugrave; \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1977\u003c/span\u003e; Fratianni and Acquaotta \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOvercoming these limitations is a top priority, also to avoid losing a valuable legacy of information. In this study, we propose an innovative methodology for managing and analysing climate databases based on the use of deep learning models. Deep learning, through architectures such as convolutional neural networks (CNN) (Skansi \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), recurrent neural networks (RNN) (Bustreo et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), and long short-term memory (LSTM)(Hochreiter and Schmidhuber \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1997\u003c/span\u003e) networks, enables the capture of the spatiotemporal complexity of climatic phenomena, identifying hidden patterns even in the presence of significant discontinuities or noise. The integration of these approaches makes it possible to reconstruct continuous and coherent historical series on a regular grid (Sha and Guha \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Lupi et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Guraka \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), both spatially and temporally, thereby enabling new forms of high-resolution climate analysis and facilitating the creation of harmonised datasets on a national scale.\u003c/p\u003e \u003cp\u003eThe importance of adopting deep learning techniques in this context lies not only in their predictive capabilities but also in their flexibility in addressing the structural complexities typical of climate data. These models allow for the integration of multivariable information, the handling of heterogeneous time series, and the generalization to under-observed geographic areas (Luppichini et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2022a\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Lupi et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). What was hardly feasible just a few years ago is now possible thanks to access to powerful hardware infrastructures, such as GPUs and cloud-based systems, which allow for the training of deep neural networks on large-scale datasets (LeCun et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Goodfellow et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Additionally, the development of open-source frameworks like TensorFlow (Abadi et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) has lowered the barrier of entry for the scientific community, making these technologies more accessible and adaptable in climatological research. In the current context, marked by the acceleration of climate change and the growing availability of data, the use of such techniques is not merely an opportunity but a necessity for addressing present and future environmental challenges with scientific rigour.\u003c/p\u003e \u003cp\u003eThe climate variables analysed are on a monthly scale and include: cumulative precipitation, maximum daily precipitation, mean temperature, minimum temperature, and maximum temperature. The analysis was conducted on a large national archive of Italian data, comprising tens of thousands of historical series. Deep learning models were applied to generate climate data on a regular grid with a spatial resolution of 10 km x 10 km, covering the time span from 1950 to 2020.\u003c/p\u003e \u003cp\u003eThe reconstructed data are available for consultation and free download through a web application accessible at the following link: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://lca.dst.unipi.it/AIClimate/\u003c/span\u003e\u003cspan address=\"https://lca.dst.unipi.it/AIClimate/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data time series\u003c/h2\u003e \u003cp\u003eThe historical time series used in this work originate from a census promoted by the Istituto Superiore per la Protezione e la Ricerca Ambientale (ISPRA), within the project \"Sistema Nazionale per l\u0026rsquo;Elaborazione e Diffusione di Dati Climatici\" (SCIA; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scia.isprambiente.it/\u003c/span\u003e\u003cspan address=\"https://scia.isprambiente.it/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). This portal gathers climate data collected over decades by various institutions dedicated to this activity (Morbidelli et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWe collected monthly data on total rainfall, maximum daily rainfall, mean temperature, maximum temperature, and minimum temperature from a national meteorological data portal. The cumulative rainfall dataset includes 11,189 stations and 3,918,202 records spanning the period 1860\u0026ndash;2023. However, a preliminary inspection revealed that the majority of stations (11,024) became active only after 1950. Consequently, our analyses primarily focus on the 1950\u0026ndash;2020 period, which comprises data from 10,995 stations (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The maximum daily rainfall dataset consists of 10,556 stations and 1,239,717 records for the period 1860\u0026ndash;2023, with 10,363 stations active during 1950\u0026ndash;2020 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The mean temperature dataset includes 4,961 stations and the same number of records (1,239,717) from 1860 to 2023, with 4,927 stations active between 1950 and 2020 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Similarly, the minimum temperature dataset comprises 4,961 stations and 1,239,891 records across the full temporal range, with 4,930 stations active during 1950\u0026ndash;2020 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Finally, the maximum temperature dataset contains 4,963 stations and 1,257,690 records for the period 1860\u0026ndash;2023, of which 4,927 stations were operational between 1950 and 2020 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe time series present several issues in terms of continuity and duration, which is a common limitation in many meteorological monitoring networks. Specifically, the average temporal coverage of rainfall time series spans approximately 36 years, but the actual average number of recorded years is only 31. For temperature series, the average temporal span decreases to 24 years, with an average of 21 years of actual data.\u003c/p\u003e \u003cp\u003eThis highlights the fact that only a limited number of stations cover the full reference period (1950\u0026ndash;2020): approximately 120 temperature stations and 650 rainfall stations, with an average data gap of 16% and 30%, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrate the trends in station activity over time. Rainfall series show a gradual decline in the number of active stations, with a sharp drop at the end of the 1980s. In contrast, the number of active temperature stations increased steadily until recently, although a slight decline is observed over the past decade.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 AI applications\u003c/h2\u003e \u003cp\u003eThe dataset was used to develop deep learning models aimed at predicting monthly rainfall values at individual stations, using information from surrounding stations as input. A total of 32 models were constructed, differing in several aspects: the number of neighbouring stations considered, the inclusion of additional climate data, whether or not input standardisation was applied, and the type of neural network architecture adopted.\u003c/p\u003e \u003cp\u003eEach record in the database contributed to building the input matrix for the neural networks. Specifically, the monthly rainfall value at a given station was predicted based on the same month's values from n neighbouring stations, where n was set to 5, 10, 15, or 20.\u003c/p\u003e \u003cp\u003eFor selected model configurations, the input matrix was enhanced with climate teleconnection indices covering the previous 12 months. These indices include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eAtlantic Multidecadal Oscillation (AMO)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEast Atlantic (EA)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eMediterranean Oscillation Index (MOI)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eNorth Atlantic Oscillation (NAO)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWestern Mediterranean Oscillation (WeMO)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSea Surface Temperatures (SSTs) from the Gulf of Genoa (GGSST), the broader Mediterranean Sea (MSST), and the North Atlantic (NASST)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe SST data were sourced from the Extended Reconstructed Sea Surface Temperature (ERSST) database. All data inputs were standardised to a [0,1] scale to reduce dimensional imbalance and improve model training efficiency, in line with best practices for optimising neural network convergence (He et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo perform the regression task, two distinct deep learning architectures were developed and tested:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003ea fully connected feedforward neural network (FCNN)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ea recurrent neural network based on Long Short-Term Memory (LSTM) units with an encoder\u0026ndash;decoder structure.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eBoth models were implemented using the Keras Sequential API and trained to minimise the Mean Absolute Error (MAE), selected for its robustness against outliers.\u003c/p\u003e \u003cp\u003eThe FCNN architecture consists of four hidden layers with progressively smaller sizes (4096, 1024, 256, and 64 neurons), each using ReLU activation. To prevent overfitting and promote generalisation, L2 regularisation (λ\u0026thinsp;=\u0026thinsp;0.01), batch normalisation, and dropout (30%) were applied at every layer.\u003c/p\u003e \u003cp\u003eThe LSTM model leverages the temporal nature of the input data. It follows an encoder\u0026ndash;decoder framework, with the encoder composed of two LSTM layers (64 and 32 units, respectively). The first layer returns the full sequence, while the second compresses it into a final state. This state is then repeated using a RepeatVector and passed to the decoder, which also consists of two LSTM layers (32 and 64 units) that output a full sequence. The output is further processed by a TimeDistributed Dense layer, followed by a Flatten layer and two Dense layers to produce a final scalar prediction.\u003c/p\u003e \u003cp\u003eModel training was conducted using the Adam optimiser with an initial learning rate of 0.001. Two callbacks were employed to monitor performance and avoid overfitting:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eEarlyStopping\u003c/b\u003e: halts training if the validation loss does not improve over a specified number of epochs.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eModelCheckpoint\u003c/b\u003e: saves the model weights with the best validation performance.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe dataset was split into three subsets: training (60%), validation (20%), and testing (20%). This partitioning ensures fair and unbiased model evaluation: the training set is used for learning, the validation set for tuning during training, and the test set provides an independent evaluation of predictive performance.\u003c/p\u003e \u003cp\u003eA detailed overview of the model configurations is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Once trained, the models were applied to a regular 10 km \u0026times; 10 km grid to reconstruct a high-resolution, continuous climatic dataset spanning from 1950 to 2020. This spatial resolution was chosen as a conservative balance between detail and representativeness, considering the uneven distribution of meteorological stations across Italy (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). A finer grid could produce artificial variability due to local oversampling, whereas 10 km spacing is consistent with the typical climatic gradients observed over the Italian peninsula (Fratianni and Acquaotta \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAll model outputs are available for consultation and download through the AI Climate web application (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://lca.dst.unipi.it/AIClimate/\u003c/span\u003e\u003cspan address=\"https://lca.dst.unipi.it/AIClimate/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of the models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eData Input\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStandardization\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNeural Network\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDL1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDense\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDL1STD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDense\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDL2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall, Atmospheric Teleconnections\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDense\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDL2STD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall, Atmospheric Teleconnections\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDense\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSMT1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSMT1STD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall, Atmospheric Teleconnections\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSMT2STD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRainfall, Atmospheric Teleconnections\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Time series reconstruction\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents a detailed evaluation of the performance of a climate model across five meteorological variables: cumulative rainfall, maximum daily rainfall, maximum temperature, mean temperature, and minimum temperature. For each variable, three statistical indicators are reported: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R\u0026sup2;). The results are divided by dataset, distinguishing between training (in red), validation (in blue), and test (in green), thus allowing the assessment of both the learning phase and the model\u0026rsquo;s generalisation capability.\u003c/p\u003e \u003cp\u003eIn the case of cumulative rainfall, the model shows consistent behaviour across the three datasets, with low MAE and RMSE values and generally high R\u0026sup2; values, above 0.9. This suggests good predictive performance even on independent data, that is, data not included during training, confirming the model\u0026rsquo;s ability to generalise to unseen distributions. In contrast, for maximum daily rainfall, higher errors and greater variability among datasets are observed, with lower R\u0026sup2; values (ranging from 0.78 to 0.83), highlighting the model\u0026rsquo;s difficulty in accurately predicting extreme events, which are notoriously more complex to model due to their high variability.\u003c/p\u003e \u003cp\u003eTurning to temperatures, the performance is very satisfactory. For maximum, mean, and minimum temperature, both MAE and RMSE remain low across all datasets, and the R\u0026sup2; values are very high, often above 0.95 and in some cases close to 0.99. This indicates that the model is able to accurately represent thermal trends and generalizes well even on test data. This is therefore a very robust behaviour in temperature forecasting.\u003c/p\u003e \u003cp\u003eIt is noted that the use of the LSTM network tends to produce slightly better MAE values compared to other approaches, suggesting a greater ability to capture the main patterns of the signal. However, this accuracy in training appears to be accompanied by greater instability in RMSE and R\u0026sup2; metrics, with a more marked difference between the training dataset and the validation and test sets, especially in the latter. This behaviour suggests a greater tendency toward overfitting, as the model performs significantly better on the training set than on independent data. In other words, the network may fit the training data very well but lose accuracy when applied to new situations, a dynamic that would require further measures to improve generalisation.\u003c/p\u003e \u003cp\u003eOverall, the model appears very effective in forecasting temperatures and fairly reliable for cumulative rainfall, while it faces greater challenges in accurately reproducing maximum daily rainfall. The results obtained indicate a solid modelling framework.\u003c/p\u003e \u003cp\u003eWhile we fully acknowledge that the seasonal cycle can inflate correlation-based metrics such as R\u0026sup2;, our evaluation also relies on MAE and RMSE, which quantify absolute errors and are not affected by the presence of a seasonal cycle. The consistently low MAE and RMSE values across stations and variables indicate that the model is not only reproducing the annual cycle but also providing accurate month-to-month estimates in absolute terms.\u003c/p\u003e \u003cp\u003eFor this reason, even though R\u0026sup2; may reflect the contribution of seasonality, MAE and RMSE provide robust, season-independent evidence that the reconstruction skill is genuinely high.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the application of the models developed in this study. It presents the annual mean of the predictions produced by the 32 climate models over six grid cells representative of the study domain. Each line represents the average value estimated by the models, while the uncertainty band reflects the variability among the simulations (90% confidence interval).\u003c/p\u003e \u003cp\u003eThis representation highlights how the adopted modeling approach allows the generation of continuous and spatially consistent climate time series. The joint analysis of the six sites shows strong internal consistency among the different variables and across various geographic regions, confirming the stability of the simulations even in the presence of heterogeneous climatic conditions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Application of AI-derived products\u003c/h2\u003e \u003cp\u003eThe evaluation of the results produced by the AI models also includes the analysis of trends and spatial relationships, aimed at enhancing the ability of the data to describe and characterize the study area.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Trend Analysis\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the trends obtained using the Mann-Kendall test for five climatic variables from 1950 to 2020. The slopes represent the average annual variation observed during the study period, expressed in specific units for each variable (mm/year or \u0026deg;C/year). Positive values indicate increasing trends, while negative values indicate decreasing trends. White areas denote either lack of data or non-significant results.\u003c/p\u003e \u003cp\u003eFor monthly cumulative rainfall, marked negative trends (\u0026minus;\u0026thinsp;0.4 mm/year) are observed over large portions of Northwestern and Southern Italy, especially in Sicily and Calabria. This trend is clearly evident, with some exceptions, along the Apennine chain. However, some Alpine and pre-Alpine areas show slight increases in precipitation. This highlights a growing spatial differentiation in available water resources and a complex pattern in cumulative rainfall trends.\u003c/p\u003e \u003cp\u003eFor maximum daily rainfall, widespread positive trends are evident in both the frequency and intensity of extreme precipitation events, especially in Northeastern Italy and along the northern Apennine chain (up to +\u0026thinsp;0.2 mm/year). This increase in extreme rainfall is consistent with climate model projections that foresee greater frequency and intensity of such events; this entails a potential rise in hydraulic risk, especially in already vulnerable areas such as Alpine valleys and densely populated regions.\u003c/p\u003e \u003cp\u003eMonthly mean temperature shows a widespread warming across the entire country, with positive trends sometimes exceeding\u0026thinsp;+\u0026thinsp;0.03\u0026deg;C/year. The Alpine areas and central Apennine chain are particularly affected, indicating accelerated warming in mountainous regions.\u003c/p\u003e \u003cp\u003eMonthly minimum temperature also shows significant positive trends over much of Central and Southern Italy (up to +\u0026thinsp;0.04\u0026deg;C/year). However, some areas in Northwestern and central inland Italy show negative trends, suggesting the persistence of cold nights or increased thermal variability.\u003c/p\u003e \u003cp\u003eFinally, monthly maximum temperature confirms a generalized warming of maximum temperatures, with increases up to +\u0026thinsp;0.04\u0026deg;C/year, especially intense in southern and insular regions.\u003c/p\u003e \u003cp\u003eOverall, the results indicate a clear warming trend across the entire national territory, accompanied by a significant increase in extreme rainfall in the North and a decrease in average annual precipitation in several regions, including the Apennine chain.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Relationship between temperature and precipitation\u003c/h2\u003e \u003cp\u003eIn an effort to characterize the relationships between temperature and rainfall variables, the Spearman correlation (Spearman \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e1904\u003c/span\u003e) was calculated between temperatures and precipitation indicators. To carry out this analysis, a seasonal decomposition of the time series was first performed, from which only the trend components were extracted. This approach allowed the removal of the seasonal component, which is relevant in both types of variables, thereby enabling a more focused analysis of the long-term signal. The results obtained are summarized in the six maps shown in the figure, which illustrate the spatial distribution of these correlations across the entire Italian territory.\u003c/p\u003e \u003cp\u003eIn the first group of panels (top row), the relationships between the three types of temperature and monthly rainfall are observed. Overall, a predominant negative correlation emerges: areas where temperatures, particularly maximum temperatures, tend to increase correspond to a decrease in monthly precipitation. This pattern is particularly evident in the northern regions and along the Adriatic coast. The thermal component thus appears to be inversely correlated with the monthly rainfall regime, suggesting that, as temperatures rise (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), a progressive reduction in precipitation is observed.\u003c/p\u003e \u003cp\u003eThe situation changes significantly when considering the correlation between the temperature variables and the intensity of daily maximum rainfall. In this case, the signal is distinctly positive and well distributed across almost the entire study area. The highest correlations are especially observed in central-northern areas and along the Apennine ranges, where the maximum daily rainfall recordings increase with rising temperatures. This may indicate an active role of warming in driving intense convective processes. Minimum temperature also shows a positive association, although with greater spatial discontinuity, suggesting a connection with warmer and more humid nighttime conditions, which are favourable to atmospheric instability.\u003c/p\u003e \u003cp\u003eOverall, these analyses highlight an interesting and climatically relevant dynamic: while temperature is inversely associated with the total amount of precipitation, it is positively associated with the intensity of the most extreme events.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e \u003cem\u003eRelationship between temperature and rainfall quantified using Spearman correlation coefficients.\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Comparison against ERA5\u003c/h2\u003e \u003cp\u003eTo further strengthen the validity of the developed dataset, a direct comparison was carried out with the ERA5 dataset (Hersbach et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), extracted using the Google Earth Engine API and covering the 1979\u0026ndash;2020 interval. The goal was to assess the data's consistency through statistical correlation analysis, using both the Spearman (Spearman \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e1904\u003c/span\u003e) and Pearson (Kirch \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2008\u003c/span\u003e)coefficients to quantify the intensity and direction of associations between the main climate variables analysed.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e provides a summary overview using boxplots of correlation values calculated across the spatial domain. The results show very high correlations for all variables, with particularly compact distributions and medians close to one for temperature variables. This confirms that the models are robust and their outputs are comparable to data obtained from other modelling approaches. Although precipitation shows greater dispersion in correlation values, the central tendency remains high, indicating strong alignment of the modelled results with the ERA5 reference even for this more variable component.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e extends this assessment spatially, presenting gridded maps of Spearman and Pearson correlations for each climate variable. The results highlight a strong and coherent spatial structure in agreement with ERA5. Correlations are especially high for temperature variables, with most of the domain exceeding values of 0.96. For precipitation, despite more evident spatial variability, most of the territory still presents values above 0.8. These patterns further confirm the models\u0026rsquo; ability to reproduce both the magnitude and temporal behaviour of ERA5-derived climate fields.\u003c/p\u003e \u003cp\u003eThe analysis of modelled products against consolidated literature references is in fact an essential step for evaluating the dataset\u0026rsquo;s consistency with known climate dynamics. If the models faithfully replicate signals recognised in previous studies, this strengthens the validity of the adopted methodology and confirms the dataset\u0026rsquo;s reliability.\u003c/p\u003e \u003cp\u003eAlthough ERA5 shows good performance when compared with our reconstructed fields, the use of reanalysis data cannot replace the need for a homogeneous, observation-driven climatic dataset. ERA5 is a modelling product that integrates observations through data assimilation, and therefore its spatial patterns and temporal variability remain influenced by the underlying numerical model and by the density of assimilated stations, which is highly heterogeneous across Italy. As a consequence, ERA5 may reproduce large-scale atmospheric dynamics well, but it can still exhibit local biases, especially in complex physiographic contexts such as mountain areas or coast\u0026ndash;inland transitions, where fine-scale processes play a key role. This limitation is particularly relevant for variables such as precipitation and temperature extremes, for which reanalyses often underestimate variability and magnitude.\u003c/p\u003e \u003cp\u003eOur reconstruction instead relies solely on in situ observations and is specifically designed to preserve the climatological characteristics of the Italian network, producing fields that are independent of model assumptions. Furthermore, our dataset provides a substantially finer spatial resolution (10 \u0026times; 10 km) compared to ERA5 (0.25\u0026deg;), allowing a more detailed representation of local climatic gradients and physiographic structures. This level of detail is essential for regional climate studies, long-term trend analyses, and impact assessments, which require datasets that reflect the true heterogeneity of the territory. In addition, only a very small fraction of the Italian observational network (approximately 250 stations) is actually assimilated into ERA5, ensuring that the comparison between our fields and ERA5 does not involve circularity and confirming the need for a fully observation-based reconstruction. For these reasons, even though ERA5 performs well in terms of general correspondence, a reconstruction based on real observations remains necessary to obtain high-resolution, consistent, and physically grounded climate information.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Hydroclimatic Responses to Warming over Italy\u003c/h2\u003e \u003cp\u003eTrend analysis for the period 1950\u0026ndash;2020 shows a generalised increase in monthly mean, minimum, and maximum temperatures, with growth rates exceeding\u0026thinsp;+\u0026thinsp;0.04\u0026deg;C/year in some mountainous regions. This is in agreement with Brunetti et al. (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2006\u003c/span\u003e), who reported a warming of about\u0026thinsp;+\u0026thinsp;1.2\u0026deg;C across Italy between 1865 and 2003, with a marked acceleration since the 1980s.\u003c/p\u003e \u003cp\u003eIn parallel, a statistically significant reduction in monthly cumulative precipitation is observed, particularly in southern regions and along the Apennine ridge, with decreases exceeding\u0026thinsp;\u0026minus;\u0026thinsp;0.4 mm/year. This seasonal and annual pattern confirms the findings of Brunetti et al. (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2006\u003c/span\u003e) and Luppichini and Bini (\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), who observed a reduction in annual precipitation along with an increase in daily rainfall intensity across Italy (Cislaghi et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; Marani and Zanetti \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Guo and Montanari \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe statistical correlation between temperature and precipitation variables reveals a negative relationship between temperature and monthly rainfall, and a positive one between temperature and extreme daily rainfall. These results align with international studies that describe warming as having a dual effect: reducing the frequency of ordinary rainfall events (partly due to atmospheric stabilisation), while promoting short, intense convective episodes (Schr\u0026ouml;er and Kirchengast \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis is consistent with the literature on the central Mediterranean (Zittis et al. \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), where temperature acts both as a suppressor of widespread rainfall, due to greater atmospheric stability, and as a driver of convective processes responsible for more intense and localised events (Schr\u0026ouml;er and Kirchengast \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). These results are compatible with the Clausius-Clapeyron relationship, which describes the atmosphere\u0026rsquo;s increased capacity to hold moisture with rising temperatures (Hardwick Jones et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2010\u003c/span\u003e; Berg et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Blenkinsop et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Pumo et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This empirical evidence represents one of the clearest signals of the hydrological cycle\u0026rsquo;s response to global warming in the Mediterranean context. The thermodynamic component\u0026rsquo;s influence on extreme rainfall intensification is further confirmed by the positive relationship observed between temperature and maximum daily rainfall, consistent with Clausius-Clapeyron scaling reported in the Mediterranean(Pumo et al. \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Pumo and Noto \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Luppichini et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Noto et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Haslinger et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) and in other parts of the world (Lenderink and Van Meijgaard \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Hardwick Jones et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2010\u003c/span\u003e; Lenderink and Attema \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Visser et al. \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In this sense, the analysis confirms that rising temperatures are associated with a higher probability of intense events.\u003c/p\u003e \u003cp\u003eThe observed increase in frequency and intensity of extreme daily rainfall fits within a broader context of hydrological cycle intensification, strongly influenced by synoptic-scale atmospheric circulation and the rise in Mediterranean SST. These changes are associated with the weakening of key cyclogenetic structures such as the Icelandic Low and the Gulf of Genoa Low (Trigo et al. \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Frankignoul et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2003\u003c/span\u003e; Luppichini et al. \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2022b\u003c/span\u003e), accompanied by a strengthening of the positive NAO phase and more persistent anticyclonic conditions.\u003c/p\u003e \u003cp\u003eTherefore, although clear signs of changing precipitation regimes are observed in Italy, particularly in the intensification of extreme events, these results must be interpreted within a broader framework shaped by local and regional factors. The association between higher temperatures and more intense daily rainfall is especially evident in contexts where atmospheric warming coincides with increased persistence of convective patterns or greater air mass instability (Haslinger et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFor this reason, the management of climate databases using modern technologies proves essential for establishing a solid and reliable base of high-resolution spatial climate data.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis work presents a novel approach to reconstructing climate series with high spatial resolution and temporal continuity, applied across the Italian territory for the period 1950\u0026ndash;2020. By leveraging deep learning models based on LSTM and feedforward architectures, it was possible to integrate fragmented historical series, fill observational gaps, and generate homogeneous monthly climate fields on a regular grid.\u003c/p\u003e\n\u003cp\u003eResults indicate that the models accurately reproduce observed climate dynamics: modelled temperatures show strong agreement with ERA5 data (correlations \u0026gt; 0.96), and precipitation values\u0026mdash;despite higher variability\u0026mdash;exceed 0.8 on average across the domain. The national-scale statistical validation confirms the robustness and reliability of the reconstructed dataset.\u003c/p\u003e\n\u003cp\u003eTrend analysis reveals three key indicators of ongoing climate change:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003eA widespread and persistent warming, with rates exceeding +0.04 \u0026deg;C/year in mountainous regions;\u003c/li\u003e\n \u003cli\u003eA significant decline in monthly cumulative rainfall;\u003c/li\u003e\n \u003cli\u003eAn intensification of extreme daily rainfall events, in line with trends observed throughout the Mediterranean.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThis dual behaviour\u0026mdash;less widespread rainfall and more intense extremes\u0026mdash;suggests a structural shift in the Italian hydrological cycle, driven by both thermodynamic processes and changes in synoptic-scale atmospheric circulation.\u003c/p\u003e\n\u003cp\u003eIn summary, this study demonstrates how artificial intelligence tools can overcome the limitations of traditional observation networks, producing climate products that are coherent, validated, and suitable for scientific and operational use. The resulting dataset offers a concrete resource to support hydrological modelling, climate risk assessment, and adaptation policy in a highly vulnerable Mediterranean context.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eData available\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe dataset is accessible via the AIClimate platform (https://lca.dst.unipi.it/AIClimate/)\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAuthor contributions statement\u003c/p\u003e\n\u003cp\u003eMarco Luppichini: Conceptualization, Data curation, \u0026nbsp;Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing \u0026ndash; original draft, \u0026nbsp;Writing \u0026ndash; review \u0026amp; editing\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMonica Bini: Conceptualization, Project administration, Resources, Supervision, Validation, Visualization, Writing \u0026ndash; original draft, \u0026nbsp; Writing \u0026ndash; review \u0026amp; editing\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAalto J, Pirinen P, Jylh\u0026auml; K (2016) New gridded daily climatology of Finland: Permutation‐based uncertainty estimates and temporal trends in climate. Journal of Geophysical Research: Atmospheres 121:3807\u0026ndash;3823. https://doi.org/10.1002/2015JD024651\u003c/li\u003e\n \u003cli\u003eAbadi M, Agarwal A, Barham P, et al (2015) TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems\u003c/li\u003e\n \u003cli\u003eBerg P, Moseley C, Haerter JO (2013) Strong increase in convective precipitation in response to higher temperatures. Nat Geosci 6:181\u0026ndash;185\u003c/li\u003e\n \u003cli\u003eBlenkinsop S, Chan SC, Kendon EJ, et al (2015) Temperature influences on intense UK hourly precipitation and dependency on large-scale circulation. Environmental Research Letters 10:54021\u003c/li\u003e\n \u003cli\u003eBr\u0026ouml;nnimann S, Luterbacher J, Ewen T, et al (2008) Climate Variability and Extremes during the Past 100 Years\u003c/li\u003e\n \u003cli\u003eBrunetti M, Maugeri M, Monti F, Nanni T (2006) Temperature and precipitation variability in Italy in the last two centuries from homogenised instrumental time series. International Journal of Climatology 26:345\u0026ndash;381. https://doi.org/https://doi.org/10.1002/joc.1251\u003c/li\u003e\n \u003cli\u003eBustreo M, Beltr\u0026aacute;n-Gonz\u0026aacute;lez C, Murino V, Camozzi F (2018) Recurrent Neural Networks. 59\u0026ndash;79. https://doi.org/10.1007/978-3-319-75304-1_5\u003c/li\u003e\n \u003cli\u003eCamuffo D (1984) Analysis of the series of precipitation at Padova, Italy. Clim Change 6:57\u0026ndash;77. https://doi.org/10.1007/BF00141668\u003c/li\u003e\n \u003cli\u003eCamuffo D (2002) History of the Long Series of Daily Air Temperature in Padova (1725\u0026ndash;1998). Clim Change 53:7\u0026ndash;75. https://doi.org/10.1023/A:1014958506923\u003c/li\u003e\n \u003cli\u003eCamuffo D, Becherini F, della Valle A (2025) Recovery of the Long Series of Precipitation in Pisa, Italy: Trend, Anomaly and Extreme Events. Climate 13:. https://doi.org/10.3390/cli13040073\u003c/li\u003e\n \u003cli\u003eCant\u0026ugrave; V (1977) The climate of Italy. In: Wallen CC (ed) Climate of central and southern Europe. Elsevier, pp 127\u0026ndash;184\u003c/li\u003e\n \u003cli\u003eCislaghi M, De Michele C, Ghezzi A, Rosso R (2005) Statistical assessment of trends and oscillations in rainfall dynamics: Analysis of long daily Italian series. Atmos Res 77:188\u0026ndash;202. https://doi.org/https://doi.org/10.1016/j.atmosres.2004.12.014\u003c/li\u003e\n \u003cli\u003eCoscarelli R, Caroletti G, Joelsson M, et al (2021) Validation metrics of homogenization techniques on artificially inhomogenized monthly temperature networks in Sweden and Slovenia (1950\u0026ndash;2005). Sci Rep 11:. https://doi.org/10.1038/s41598-021-97685-7\u003c/li\u003e\n \u003cli\u003eFrankignoul C, Friederichs P, Kestenare E (2003) Influence of Atlantic SST anomalies on the atmospheric circulation in the Atlantic-European sector. Annals of Geophysics 46:\u003c/li\u003e\n \u003cli\u003eFratianni S, Acquaotta F (2017) The Climate of Italy. In: Soldati M, Marchetti M (eds) Landscapes and Landforms of Italy. Springer International Publishing, Cham, pp 29\u0026ndash;38\u003c/li\u003e\n \u003cli\u003eGiorgi F (2006) Climate change hot-spots. Geophys Res Lett 33:. https://doi.org/https://doi.org/10.1029/2006GL025734\u003c/li\u003e\n \u003cli\u003eGoodfellow I, Bengio Y, Courville A (2016) Deep Learning. MIT Press\u003c/li\u003e\n \u003cli\u003eGubler S, Hunziker S, Begert M, et al (2017) The influence of station density on climate data homogenization. International Journal of Climatology 37:. https://doi.org/10.1002/joc.5114\u003c/li\u003e\n \u003cli\u003eGuo R, Montanari A (2023) Historical rainfall data in northern Italy predict larger meteorological drought hazard than climate projections. Hydrol Earth Syst Sci 27:2847\u0026ndash;2863. https://doi.org/10.5194/hess-27-2847-2023\u003c/li\u003e\n \u003cli\u003eGuraka HB (2024) CLIMATE CHANGE FORECASTING USING DEEP LEARNING. EPRA International Journal of Climate and Resource Economic Review. https://doi.org/10.36713/epra18552\u003c/li\u003e\n \u003cli\u003eHardwick Jones R, Westra S, Sharma A (2010) Observed relationships between extreme sub-daily precipitation, surface temperature, and relative humidity. Geophys Res Lett 37:. https://doi.org/https://doi.org/10.1029/2010GL045081\u003c/li\u003e\n \u003cli\u003eHaslinger K, Breinl K, Pavlin L, et al (2025) Increasing hourly heavy rainfall in Austria reflected in flood changes. Nature 639:667\u0026ndash;672. https://doi.org/10.1038/s41586-025-08647-2\u003c/li\u003e\n \u003cli\u003eHawkins E, Sutton R (2009) The Potential to Narrow Uncertainty in Regional Climate Predictions. Bull Am Meteorol Soc 90:1095\u0026ndash;1108. https://doi.org/https://doi.org/10.1175/2009BAMS2607.1\u003c/li\u003e\n \u003cli\u003eHe L, Liu Z, Bucci A (2023) Combining dimensionality reduction methods with neural networks for realized volatility forecasting. Ann Oper Res. https://doi.org/10.1007/s10479-023-05544-7\u003c/li\u003e\n \u003cli\u003eHersbach H, Bell B, Berrisford P, et al (2020) The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society 146:1999\u0026ndash;2049. https://doi.org/https://doi.org/10.1002/qj.3803\u003c/li\u003e\n \u003cli\u003eHochreiter S, Schmidhuber J (1997) Long Short-Term Memory. Neural Comput 9:1735\u0026ndash;1780. https://doi.org/10.1162/neco.1997.9.8.1735\u003c/li\u003e\n \u003cli\u003eIPCC (2023) IPCC, 2023: Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, H. Lee and J. Romero (eds.)]. IPCC, Geneva, Switzerland.\u003c/li\u003e\n \u003cli\u003eKirch W (ed) (2008) Pearson\u0026rsquo;s Correlation Coefficient. In: Encyclopedia of Public Health. Springer Netherlands, Dordrecht, pp 1090\u0026ndash;1091\u003c/li\u003e\n \u003cli\u003eLazoglou G, Papadopoulos-Zachos A, Georgiades P, et al (2024) Identification of climate change hotspots in the Mediterranean. Sci Rep 14:29817. https://doi.org/10.1038/s41598-024-80139-1\u003c/li\u003e\n \u003cli\u003eLeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436\u0026ndash;444. https://doi.org/10.1038/nature14539\u003c/li\u003e\n \u003cli\u003eLenderink G, Attema J (2015) A simple scaling approach to produce climate scenarios of local precipitation extremes for the Netherlands. Environmental Research Letters 10:85001. https://doi.org/10.1088/1748-9326/10/8/085001\u003c/li\u003e\n \u003cli\u003eLenderink G, Van Meijgaard E (2008) Increase in hourly precipitation extremes beyond expectations from temperature changes. Nat Geosci 1:511\u0026ndash;514\u003c/li\u003e\n \u003cli\u003eLi-Juan C, Zhong-Wei Y (2012) Progress in Research on Homogenization of Climate Data. Advances in Climate Change Research 3:59\u0026ndash;67. https://doi.org/10.3724/SP.J.1248.2012.00059\u003c/li\u003e\n \u003cli\u003eL\u0026uacute;cio P, Conde F, Ramos A (2006) A BAYESIAN APPROACH FOR RECOVERING AND HOMOGENISING METEOROLOGICAL TIME SERIES\u003c/li\u003e\n \u003cli\u003eLundstad E, Brugnara Y, Pappert D, et al (2023) The global historical climate database HCLIM. Sci Data 10:44. https://doi.org/10.1038/s41597-022-01919-w\u003c/li\u003e\n \u003cli\u003eLupi A, Luppichini M, Barsanti M, et al (2023) Machine learning models to complete rainfall time series databases affected by missing or anomalous data. Earth Sci Inform 16:3717\u0026ndash;3728. https://doi.org/10.1007/s12145-023-01122-4\u003c/li\u003e\n \u003cli\u003eLuppichini M, Barsanti M, Giannecchini R, Bini M (2022a) Deep learning models to predict flood events in fast-flowing watersheds. Science of The Total Environment 813:151885. https://doi.org/https://doi.org/10.1016/j.scitotenv.2021.151885\u003c/li\u003e\n \u003cli\u003eLuppichini M, Barsanti M, Giannecchini R, Bini M (2021) Statistical relationships between large-scale circulation patterns and local-scale effects: NAO and rainfall regime in a key area of the Mediterranean basin. Atmos Res 248:105270\u003c/li\u003e\n \u003cli\u003eLuppichini M, Bini M (2025) Evolution of rainfall in Italy over the last 200 years: Interactions between climate indices and global warming. Atmos Res 326:. https://doi.org/10.1016/j.atmosres.2025.108276\u003c/li\u003e\n \u003cli\u003eLuppichini M, Bini M, Barsanti M, et al (2022b) Seasonal rainfall trends of a key Mediterranean area in relation to large-scale atmospheric circulation: How does current global change affect the rainfall regime? J Hydrol (Amst) 612:128233. https://doi.org/https://doi.org/10.1016/j.jhydrol.2022.128233\u003c/li\u003e\n \u003cli\u003eLuppichini M, Bini M, Giannecchini R, Zanchetta G (2023) High-resolution spatial analysis of temperature influence on the rainfall regime and extreme precipitation events in north-central Italy. Science of The Total Environment 880:163368. https://doi.org/https://doi.org/10.1016/j.scitotenv.2023.163368\u003c/li\u003e\n \u003cli\u003eLuppichini M, Vailati G, Fontana L, Bini M (2024) Machine learning models for river flow forecasting in small catchments. Sci Rep 14:26740. https://doi.org/10.1038/s41598-024-78012-2\u003c/li\u003e\n \u003cli\u003eMarani M, Zanetti S (2015) Long-term oscillations in rainfall extremes in a 268 year daily time series. Water Resour Res 51:639\u0026ndash;647. https://doi.org/https://doi.org/10.1002/2014WR015885\u003c/li\u003e\n \u003cli\u003eMorbidelli R, Flammini A, Echeta O, et al (2025) A reassessment of the history of the temporal resolution of rainfall data at the global scale. J Hydrol (Amst) 654:132841. https://doi.org/10.1016/j.jhydrol.2025.132841\u003c/li\u003e\n \u003cli\u003eNew M, Hulme M, Jones P (2000) Representing Twentieth-Century Space-Time Climate Variability. Part II: Development of 1901-96 Monthly Grids of Terrestrial Surface Climate. J Clim 13:2217\u0026ndash;2238. https://doi.org/10.1175/1520-0442(2000)013\u0026lt;2217:RTCSTC\u0026gt;2.0.CO;2\u003c/li\u003e\n \u003cli\u003eNoto L, Cipolla G, Pumo D, Francipane A (2023) Climate Change in the Mediterranean Basin (Part II): A Review of Challenges and Uncertainties in Climate Change Modeling and Impact Analyses. Water Resources Management 1\u0026ndash;17. https://doi.org/10.1007/s11269-023-03444-w\u003c/li\u003e\n \u003cli\u003ePeng S, Ding Y, Liu W, Li Z (2019) 1\u0026thinsp;km monthly temperature and precipitation dataset for China from 1901 to 2017. Earth Syst Sci Data. https://doi.org/10.5194/ESSD-11-1931-2019\u003c/li\u003e\n \u003cli\u003ePumo D, Carlino G, Arnone E, Noto L V (2018) Relationship between extreme rainfall and surface temperature in Sicily (Italy). EPiC Series in Engineering 3:1718\u0026ndash;1726\u003c/li\u003e\n \u003cli\u003ePumo D, Carlino G, Blenkinsop S, et al (2019) Sensitivity of extreme rainfall to temperature in semi-arid Mediterranean regions.\u0026nbsp;Atmos Res 225:30\u0026ndash;44. https://doi.org/https://doi.org/10.1016/j.atmosres.2019.03.036\u003c/li\u003e\n \u003cli\u003ePumo D, Noto L V (2021) Exploring the linkage between dew point temperature and precipitation extremes: A multi-time-scale analysis on a semi-arid Mediterranean region. Atmos Res 254:105508. https://doi.org/https://doi.org/10.1016/j.atmosres.2021.105508\u003c/li\u003e\n \u003cli\u003eRibeiro S, Caineta J, Costa A (2016) Review and discussion of homogenisation methods for climate data. Physics and Chemistry of The Earth 94:167\u0026ndash;179. https://doi.org/10.1016/J.PCE.2015.08.007\u003c/li\u003e\n \u003cli\u003eSchr\u0026ouml;er K, Kirchengast G (2018) Sensitivity of extreme precipitation to temperature: the variability of scaling factors from a regional to local perspective. Clim Dyn 50:3981\u0026ndash;3994\u003c/li\u003e\n \u003cli\u003eSha R, Guha T (2023) Climate Time Series Prediction with Deep Learning and LSTM. 2023 4th International Conference on Smart Electronics and Communication (ICOSEC) 1631\u0026ndash;1637. https://doi.org/10.1109/ICOSEC58147.2023.10276117\u003c/li\u003e\n \u003cli\u003eSkansi S (2018) Convolutional Neural Networks. 121\u0026ndash;133. https://doi.org/10.1007/978-3-319-73004-2_6\u003c/li\u003e\n \u003cli\u003eSpearman C (1904) The proof and measurement of association between two things. Am J Psychol 15:72\u0026ndash;101. https://doi.org/10.2307/1412159\u003c/li\u003e\n \u003cli\u003eTrigo IF, Bigg GR, Davies TD (2002) Climatology of Cyclogenesis Mechanisms in the Mediterranean\u003c/li\u003e\n \u003cli\u003eVenema V, Mestre O, Aguilar E, et al (2012) Benchmarking homogenization algorithms for monthly data. Climate of The Past. https://doi.org/10.5194/CP-8-89-2012\u003c/li\u003e\n \u003cli\u003eVicente-Serrano S, Tramblay Y, Reig Gracia F, et al (2025) High temporal variability not trend dominates Mediterranean precipitation. Nature 639:658\u0026ndash;666. https://doi.org/10.1038/s41586-024-08576-6\u003c/li\u003e\n \u003cli\u003eVisser JB, Wasko C, Sharma A, Nathan R (2021) Eliminating the Hook in Precipitation Temperature Scaling. J Clim 34:9535\u0026ndash;9549. https://doi.org/10.1175/JCLI-D-21-0292.1\u003c/li\u003e\n \u003cli\u003eZittis G, Bruggeman A, Lelieveld J (2021) Revisiting future extreme precipitation trends in the Mediterranean. Weather Clim Extrem 34:100380. https://doi.org/https://doi.org/10.1016/j.wace.2021.100380\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"theoretical-and-applied-climatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"taac","sideBox":"Learn more about [Theoretical and Applied Climatology](https://www.springer.com/journal/704)","snPcode":"704","submissionUrl":"https://submission.nature.com/new-submission/704/3","title":"Theoretical and Applied Climatology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8702295/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8702295/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eUnderstanding 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 \u0026times; 10 km) across the entire Italian territory over the period 1950\u0026ndash;2020.\u003c/p\u003e \u003cp\u003eUsing 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.\u003c/p\u003e \u003cp\u003eTrend analysis revealed three key indicators of ongoing climate change: (i) widespread and persistent warming, with rates\u0026thinsp;\u0026gt;\u0026thinsp;+\u0026thinsp;0.04\u0026deg;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.\u003c/p\u003e \u003cp\u003eThe final dataset, accessible via the AIClimate platform (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://lca.dst.unipi.it/AIClimate/\u003c/span\u003e\u003cspan address=\"https://lca.dst.unipi.it/AIClimate/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), offers a concrete resource for climate change studies, hydrological modelling, and the planning of adaptation strategies in highly vulnerable regions.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence for Climate Reconstruction: Spatiotemporal Modelling of Precipitation and Temperature Trends in Italy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 15:12:06","doi":"10.21203/rs.3.rs-8702295/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-17T12:27:55+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-15T08:51:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"43085702334955895827521503571786297210","date":"2026-03-30T07:01:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"327448752843931660414481502209441572191","date":"2026-03-27T18:02:19+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-27T08:25:46+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-26T22:37:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-26T22:37:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"Theoretical and Applied Climatology","date":"2026-01-26T16:10:49+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"theoretical-and-applied-climatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"taac","sideBox":"Learn more about [Theoretical and Applied Climatology](https://www.springer.com/journal/704)","snPcode":"704","submissionUrl":"https://submission.nature.com/new-submission/704/3","title":"Theoretical and Applied Climatology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"1ab314f5-26e0-4444-8fe6-b2eea94d0670","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-17T12:27:55+00:00","index":13,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-31T15:12:07+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 15:12:06","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8702295","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8702295","identity":"rs-8702295","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}