Trentino-South Tyrol is a region located in north-eastern Italy covering an area of around 13 000 km2 (Fig. 1). Its territory is entirely mountainous, including a large portion of the Dolomites and southern Alps, and it is located at the intersection region for various types of air masses: humid influences from the Atlantic north-west, dry air masses from the continental east, typically experiencing cold winters and warm summers, and warm contributions from the Mediterranean area bringing humid winters and dry summers (Adler et al., 2015). The geographical location and the complex topography of the region determine a strong climate variability and contribute to define small-scale effects in the spatial distribution of temperature and precipitation (Price, 2009). The territory is characterized by strong altitude gradients with very narrow valleys surrounded by steep slopes. Elevation range extends from 65 m a.s.l. (above sea level) in the areas close to Lake Garda in the south to 3905 m a.s.l. of the Ortles peak in the Stelvio National Park (north-west). The mean elevation of the region is around 1600 m with only 4 % flat areas, i.e. with slope steepness below 5 %. The main valley is the one of Adige River, the second longest Italian river, flowing from Reschen Pass in the north-western corner of the region, crossing Venosta Valley west to east, and then the entire region north to south towards the Adriatic Sea.

The database used for the present study was set up by retrieving the observation series of daily maximum and minimum temperature (Tmax⁡, Tmin⁡) and daily total precipitation (P) from the stations of the meteorological networks of the Trentino-South Tyrol region. More precisely, the data for the Autonomous Province of Trento were received from Meteotrentino (https://www.meteotrentino.it/, last access: 14 June 2021), while the meteorological series for the Autonomous Province of Bolzano were provided by the hydrographic office. In total 311 series of daily Tmax⁡ and Tmin⁡ and 243 series of daily P were collected (Fig. 2). The series were integrated with the records from a few close sites located in surrounding regions, especially in the north where the terrain is mainly mountainous, and the less dense regional coverage could not be enough to provide a suitable representation of climate gradients in the high-Alpine environment over the borders. With this aim, some daily series from Switzerland, Austria and Veneto were also included. Swiss sites were retrieved from MeteoSwiss (IDAWEB, https://gate.meteoswiss.ch/idaweb/, last access: 14 June 2021), Veneto data were provided by the Regional Agency of Environmental Protection (ARPA Veneto), and Austrian series were provided by the Zentralanstalt für Meteorologie und Geodynamik (ZAMG). In addition, data for some further Austrian locations were collected from the HISTALP database of daily homogenized records (Auer et al., 2007; http://www.zamg.ac.at/histalp/dataset/station/csvHOMSTART.php, last access: 14 June 2021).

The control of data quality and homogeneity represents a crucial preliminary step to improve the general accuracy of the collected observation series. The weather station records can be affected by a number of both random and systematic errors, including erroneous transcriptions of data, station malfunctions, sensor drifts and inhomogeneities due to non-climatic factors, such as station relocation, or changes in surroundings or in sensors employed. From around the 1990s, the manual measurements were replaced by automatic stations in most meteorological services. In this study we did not distinguish between manual and automatic records, and their consistency was assured by quality-check and homogenization procedures. In addition, the observation time and meteorological day definition slightly vary among the networks and could change within the years, especially in the shift from manual to automatic systems. For instance, daily precipitation records from the regional network generally referred to the cumulated precipitation from 08:00 UTC of the previous day to the 08:00 UTC of the current day, and also the recent automatic records are defined by following this definition. However, this information is not always reported, and other daily reading times could have been adopted in some cases such as 07:00 and 00:00 UTC. Clear shifts due to changes in time coding were checked and corrected, but no specific correction of the observation time was applied, and induced inhomogeneities were assessed through the quality-check procedures.

The collected database spanned the period from 1950 to present, and the whole interval was used for the quality and homogeneity checks in order to increase the statistics and the robustness of the analyses. The series in close proximity (horizontal distance < 3 km and vertical distance < 100 m) but covering different time periods were merged in order to improve the length and continuity of available records. The merging concerned mostly some stations for Trentino province, which were split due to slight relocations and/or after the transition from mechanical to automatic sensors.

The quality-check analyses were performed on the daily series of Tmax⁡, Tmin⁡ and P in order to detect outliers and to assess the spatial and temporal consistency of data. Implausible values, such as negative precipitation and out-of-range records, were scanned by setting fixed thresholds. In particular, P exceeding 500 mm, Tmin⁡ < -40 ∘C, Tmin⁡ > +40 ∘C, Tmax⁡ < -30 ∘C, Tmax⁡ > +50 ∘C and diurnal temperature range > 35 ∘C were considered and removed if not supported by metadata or by surrounding station records. In addition, cases of daily Tmin⁡ exceeding Tmax⁡ were invalidated, and periods of continuous null daily P over more than 1 month were detected and removed if a simultaneous dry period was not reported for nearby sites. Temporal consistency was also scanned in both Tmax⁡ and Tmin⁡ series by searching for differences in consecutive days greater than 20 ∘C (Durre et al., 2010).

In order to further assess the overall accuracy of the series, monthly records were computed, and each one was simulated over the whole spanned period by means of the surrounding station data (Crespi et al., 2018; Matiu et al., 2021). The agreement between monthly observed and simulated values was measured in terms of mean error (BIAS), mean absolute error (MAE) and squared correlation (R2) and allowed us to assess the spatial consistency of the series and to further detect stations affected by frequent malfunctions and periods of suspicious records. In particular, 11 temperature series and 13 precipitation series with low-quality data or duplicates of other series were identified and discarded from the following analyses. The final mean reconstruction BIAS over all Tmax⁡ and Tmin⁡ series was almost null, and mean MAE was 0.7 and 0.6 ∘C, respectively, with a mean R2 of 0.98 for both variables. For P series, the average relative BIAS was -1 %, the relative MAE 17 % and the mean R2 0.89.

After the quality check, all series with more than 30 years of valid records underwent homogeneity controls and the ones showing relevant breaks were homogenized. To this aim, the Craddock test (Craddock, 1979) was applied by following an approach similar to that adopted in previous studies (see, for example, Brunetti et al., 2006; Brugnara et al., 2012). The homogeneity test was applied to the monthly records by using as reference the five nearby stations with the highest amount of data in common. In some cases, the comparison with the available homogenized records from HISTALP and MeteoSwiss supported the identification of possible breaks. The T and P records showing relevant breaks were homogenized by deriving the correcting factors at monthly scale and applying then the adjustments to the daily values. More specifically, once an inhomogeneous period was identified in the test series, the adjustment for each month was estimated independently as the arithmetic average of the corrections derived from each reference series according to a common homogeneous period. In the case of precipitation, if no annual cycle was evident, a single yearly adjustment was computed as the average of the monthly factors; otherwise the monthly factors were used. The most recent years were always left unchanged in order to ease the update of station records. For precipitation series, the corrections derived from the monthly values were directly applied as multiplicative factors to the daily records in the corresponding inhomogeneous period. In the case of temperature series, in order to take into account the annual seasonality and extract a correction for each day of the year, the 12 monthly factors were interpolated at a daily resolution by means of a second-order trigonometric interpolation. The resulting 365 (366 for leap years) daily adjustments were then applied to the inhomogeneous daily records as additive corrections.

If Tmax⁡ and Tmin⁡ series were corrected, their internal consistency was checked to be preserved. In total 11 Tmax⁡ series, 13 Tmin⁡ series and 48 P series were homogenized. Short periods with strong breaks for which no robust correction factors could be computed were invalidated. A total of 20 and 22 breaks in Tmax⁡ and Tmin⁡ were identified with almost half of them concerning the years prior to the 1980s, which is outside the period of the computed spatial product. The average length of the inhomogeneous periods was about 7 years with an average annual adjustment of +1.2 and -1.1 ∘C for Tmax⁡ and Tmin⁡, respectively. As regards the P series, 84 breakpoints were identified in total, and the mean annual multiplicative factor applied was about 1.1. Similar to T series, the average extent of the corrected periods was about 11 years, and they mainly occurred before the 1980s. Due to the lack of specific metadata supporting the break interpretation in the series, we adopted a precautionary approach and no correction was performed if the breakpoint could not be identified clearly.

As the last step, in order to increase the data coverage and to maximize the temporal extent of the records, a gap-filling procedure was applied to reconstruct the missing daily data. The filling was performed by considering the data of the surrounding station with the highest correlation with the data of the series under evaluation and by rescaling its daily value for the ratio (in case of precipitation) and the difference (in case of temperature) of the daily averages of the two series over a common subset of data, which is defined by a window centred on the daily gap and extending over both years and days. The reconstruction was performed only if the test series contained at least 70 % of valid daily data in the selected window and a total of around 8000 daily entries were reconstructed over the period 1980–2018.

All processed series with less than 10 years of records were discarded from the database used for the interpolation since they could not provide robust long-term references from which the anomalies were computed. Daily mean temperature series were finally derived as the average of Tmax⁡ and Tmin⁡ values.

The resulting database contained 236 daily mean temperature and 219 daily precipitation series, 205 and 188 out of them located within the region, respectively (Fig. 2). The station distribution varies strongly with elevation. The best data coverage is between 500 and 1500 m for both temperature and precipitation databases, while the number of available stations gradually decreases for higher altitudes, and no precipitation sites are located above 3000 m. The portion of the domain above 1500 m is, in relative terms, underrepresented by the available observations, and thus a larger uncertainty needs to be assigned to the final gridded estimates for such areas (Fig. 3). The data coverage over higher altitudes can be improved in the future thanks to the integration of the most recent automatic weather stations when their temporal extent becomes suitable for climatological assessments.

The data availability also varies over time (Fig. 4). The first relevant improvement of data coverage occurs during the 1970s, while the greatest increase is observed after 1980 and, especially for temperature, after 1990 when the automatic stations started to operate in most areas. Due to the significantly lower station availability before the 1980s, which could reduce the general accuracy of the results for the early decades, the starting year for the computation of the gridded dataset for both temperature and precipitation was set to 1980. Although the coverage of temperature series is less dense than that of precipitation data before 1990, this is not expected to affect the general robustness of the results thanks to the greater spatial coherence typically shown by temperature records (Brunetti et al., 2006). The effects of the variability in data availability on the result accuracy is discussed more in detail in Sect. 3.1.

The 250 m resolution fields of daily mean temperature and daily precipitation over Trentino-South Tyrol were computed over the period 1980–2018 by applying the anomaly-based approach. In this framework, the final gridded output is derived from the superimposition of the gridded station daily anomalies and the interpolated climatologies, i.e. 30-year means, for the reference period 1981–2010. The interpolation grid and all topographic features were derived from the digital elevation model (DEM) Copernicus EU-DEM v1.1 (https://land.copernicus.eu/imagery-in-situ/eu-dem/eu-dem-v1.1, last access: 14 June 2021), which was aggregated from the original 25 m to the target 250 m resolution. The elevation values of the DEM used for precipitation interpolation were further smoothed by replacing the elevation of each point by a weighted average of the surrounding grid-cell elevations with weights halving at 2 km distance from the considered cell. The halving distance was chosen as the one minimizing the modelling errors (see Sect. 3.1 for a description of errors). As suggested in previous work, this procedure can avoid entering in the interpolation terrain details that are too fine with respect to the expected scales of interaction between atmospheric circulation and orography (Brunetti et al., 2012; Foresti et al., 2018). In the following the two steps of the interpolation procedure are described separately.

The uncertainty evaluation of the gridded datasets is essential to properly apply the products and interpret the results. One of the most important aspects to consider is that the grid-cell values obtained by the spatial interpolation represent areal mean estimates of temperature and precipitation. The punctual conditions at single station sites, especially the daily precipitation peaks, are smoothed after the spatialization so that the fine resolution of the daily grids does not correspond to the scales being effectively resolved, which are limited by the horizontal spacing of the station network.

The accuracy of the gridded dataset of daily mean temperature and precipitation was evaluated by applying the anomaly-based reconstruction scheme to simulate the daily records of all stations in Trentino-South Tyrol over the study period 1980–2018 in a leave-one-out approach, i.e. by removing the station data under evaluation in order to avoid self-influence. The same approach was also applied to assess the errors of the interpolated 1981–2010 monthly climatologies, separately. The reconstruction accuracy was computed by comparing the simulations and observations in terms of mean error (BIAS), mean absolute error (MAE), root mean square error (RMSE) and correlation. For daily temperature reconstruction, the resulting BIAS, as averaged over all stations and the whole time period, was almost zero, and an MAE (RMSE) of around 1.5 ∘C (1.9 ∘C) was obtained. The agreement between daily observations and simulations was high with a mean Pearson correlation coefficient of 0.97 (Fig. 5). As regards the extremes, the correlation was still high if only the temperature values above 95th percentiles were considered with a mean Pearson coefficient of 0.94, while a slight overestimation was observed for the simulation of temperature records below the 5th percentile (Fig. 5). The reconstruction errors and correlation by months were also evaluated (Table 1). MAE ranges from 1.1 ∘C in July to 1.8 ∘C in October, when the lowest correlation (0.80) was also obtained. BIAS is within -0.5–0.5 ∘C in all months. The simulated 1981–2010 monthly mean temperature climatologies at all station sites in the study region showed zero BIAS and MAE (RMSE) ranging from 0.5 (0.6) ∘C in May to 0.9 (1.1) ∘C in January (Table 2).

For precipitation, the reconstruction errors by averaging over all stations and daily records showed an MAE (RMSE) of 1.1 (3.2) mm, zero BIAS and a mean correlation coefficient of 0.91. By considering wet days only, i.e. daily precipitation values ≥ 1 mm, the mean correlation decreased to 0.87, and errors increased with mean BIAS of -0.6 mm and MAE (RMSE) of 3.4 (5.8) mm, suggesting a tendency to underestimate the higher precipitation records. The mean monthly errors ranged from 0.6 mm in January to 1.7 mm in July for MAE and from 2.0 mm in January to 4.1 mm in July for RMSE (Table 1). As regards the precipitation climatological reconstruction, BIAS is below 0.5 mm in all months, and MAE (RMSE) ranges from 4.1 (5.6) mm in February to 10.7 (14.2) mm in November (Table 2). It is worth noting that systematic errors in rain-gauge measurements can be attributed to the instrument type, site elevation and exposure to wind. For instance, a well-known source of uncertainty affecting precipitation records, especially in mountain environments, is the “rain-gauge undercatch” which could be particularly relevant during episodes of strong wind and solid precipitation and could account for up to several tens of percent of the measured values (Frei and Schär, 1998; Sevruk et al., 2009). Several approaches were developed to account for the undercatch at station sites (see, for example, Grossi et al., 2017); however the magnitude of correction is highly variable with locations, sensors and seasons, and unproper corrections could reduce the general accuracy of the dataset. For this reason, the contribution of rain-gauge undercatch was not quantified in the current analyses.

Due to the high variability in interpolation errors with the daily precipitation intensity, the comparison between simulated and observed precipitation on wet days during winter and summer was assessed by splitting the data for intensity intervals (Fig. 6).

Figure 6 confirms the general tendency of the daily reconstruction method to underestimate intense daily precipitation totals, especially in summer when the median underestimation in simulations for the highest quantile interval (0.98–0.999) is about 24 %. In contrast, overestimation of the lowest quantiles (0.1–0.2) in summer was depicted with a median exceedance of 30 %. The greater difficulty in simulating summer precipitation could be mostly ascribed to the higher spatial variability in summer precipitation events, mainly driven by convection.

In order to assess also the temporal variability in the dataset uncertainty over the study period and the influence of changes in data coverage over time, the daily simulations and observations were aggregated at monthly scale, and reconstruction errors were evaluated for each month over all available stations. The annual averages of monthly errors over 1980–2018 are reported in Fig. 7. It is worth noting that the values are almost stable over the whole period for both temperature and precipitation with MAEs around 0.7 ∘C and 13 mm, respectively. However, temperature errors slightly decreased after 1990 as a consequence of the increase in station density (see Fig. 4a). As regards precipitation, a greater error variability can be observed after 2000 probably due to the changes occurring in station networks during the transition from the manual to automatic rain gauges. The resulting reconstruction errors and their limited variability over the spanned period enable the application of the derived dataset for the assessment of climatological trends in temperature and precipitation. Such applications could be useful to compare and integrate previous existing studies analysing long-term climate trends and their spatial patterns over the region (e.g. Brugnara et al., 2012). A trend analysis was beyond the scope of the current study and thus not included in the present work, but it will be discussed in forthcoming studies.

The 250 m gridded dataset allows us to discuss and analyse the main features of the climate in the region. In the following, the 30-year annual climatologies of mean temperature and total precipitation over the period 1981–2010 are shown in Fig. 8. The derived features were largely in agreement with the findings of previous works focusing on the regional climate (e.g. Adler et al., 2015), and the 250 m grid spacing allows for a detailed visualization of the spatial patterns. The mean annual temperature ranges from +14 ∘C in the Garda Valley located in the southernmost part of the region to about -11 ∘C at the Ortles peak with an average value of around +5 ∘C over the entire study region. The thermal contrast between the inner valleys and the steep surrounding reliefs is well depicted. Mean annual values are around +12 ∘C in the main valley bottoms, while the isothermal of 0 ∘C occurs at about 2400 m. The annual cycle of mean temperature, as averaged over the region and based on the 1981–2010 normals, is characterized by the warmest conditions in July, while the coldest month is January at -3.2 ∘C (Fig. 9). The greatest warming occurs between April and May, when the thermal range between the coldest and the hottest locations also increases, reaching the maximum between May and June at almost 30 ∘C. The most relevant cooling is found in the transition from October to November when the isothermal of 0 ∘C drops from around 2700 to 1700 m and the occurrence of cold air pools in the valleys becomes more frequent.

Topographic effects are also evident in the precipitation distribution over the study region. Annual totals range from 530 mm in the inner Venosta Valley (north-western South Tyrol) to more than 1700 mm in the southernmost part of the region, where wet southerly flows are more relevant. The mean annual precipitation sum is around 1000 mm as a spatial average over the whole region, with drier conditions in the northern inner valleys of South Tyrol with annual totals around 900 mm. As regards the mean annual cycle, the driest conditions occur in late winter, with a minimum in February when the mean regional totals are below 40 mm. In contrast, precipitation significantly increases in May, with the growing contribution of local convection, and the annual maximum is reached in July, with a regional average of almost 120 mm. Regional mean precipitation values show a secondary minimum in September and increase again in October, when greater contrasts over the region, especially between north and south, are depicted with the wettest contributions in the Trentino valleys, which are more exposed to humid air masses coming from the Mediterranean sea.

By focusing on distinct portions of the study area, a relevant subregional variability can be observed (Fig. 10). In particular, a different annual precipitation cycle characterizes the northern and southern portions of the study region. In the northern and central parts the precipitation cycle has a maximum in summer, and the peak increases moving from west (Venosta Valley), where rain-shadow effects are more frequent, to east, where the Alpine ridge receives wet contributions from both northerly and southerly flows. In contrast, the annual cycle in the south of the region shows two precipitation maxima occurring in spring and autumn. The influence of subtropical high-pressure areas is particularly relevant for the southernmost valleys and contributes to the drying tendencies over summer months which defines here the local precipitation minimum.

Among the useful applications of high-resolution gridded climate variables at a daily timescale, they allow us to study and reconstruct past episodes of particular interest, such as intense precipitation events, and their temporal day-by-day pattern over the entire region. In particular, we reported as example cases the computed 250 m daily fields for two past episodes experiencing meteorological extremes over the region.

One of the most recent intense events occurred on 27–30 October 2018 when an intense storm affected a large portion of the eastern Italian Alps leading to severe impacts, such as floods, landslides, interruption of traffic and electricity supply, and severe forest damage due to intense wind gusts of up to 200 km h-1 (Dalponte et al., 2020; Davolio et al., 2020). During this event, exceptional amounts of precipitation were recorded in a few days at many locations. In Fig. 11, the 4 d precipitation sum over 27–30 October 2018 is reported, together with the daily anomalies with respect to the 1981–2010 monthly normals for the month. In some areas, especially the upper Isarco Valley in the north of South Tyrol, the south-western portion of Trentino and along its eastern border with the Veneto region, precipitation amounts were particularly intense, with 4 d totals locally exceeding 500 mm (Fig. 11a). Over the whole region, the 4 d precipitation totals were greater than the monthly normals for October, with 80 % of the grid points with precipitation amounts more than double the climatological totals for the month (Fig. 11b). As discussed in Sect. 3.1, the reported daily fields provide a comprehensive overview of the precipitation patterns over the region and the locations of the maxima; however the applied daily interpolation based on weighted spatial average has a smoothing effect so that the daily gridded fields are expected to underestimate the very localized peaks at individual stations.

As a second example case, the summer of 2003 was considered when exceptionally high temperature values were recorded in a large part of western Europe due to the influence of sub-tropical air masses and the presence of a wide high-pressure system over Europe. The high temperatures and the relatively scarce precipitation occurring in the previous winter and spring months fostered drought conditions in several regions. In Trentino-South Tyrol the mean temperature anomalies over the summer period were, as averaged over the region, around +3 ∘C with respect to the reference 1981–2010 normals (Fig. 12a). In particular, August was the hottest month, and the greatest temperature values were reached on 11 August with a mean temperature on a regional level greater than +21 ∘C and daily maximum temperature close to +40 ∘C in Bolzano and Trento. The distribution of the daily mean temperature on 11 August shows values above +30 ∘C along the main valley bottoms (Fig. 12b), and over almost the whole region the daily values exceed the climatological means with a mean anomaly of about +7 ∘C for that day.

In order to provide an example of the application of the gridded meteorological data in combination with other high-resolution products, the dataset of mean temperature and precipitation was compared to a 250 m resolution daily dataset of snow cover available over the region. The daily snow-cover maps used for the comparison are based on Moderate Resolution Imaging Spectroradiometer (MODIS) images and provide binary snow-cover information (snow, snow free) using an algorithm tailored to complex terrain (Notarnicola et al., 2013). In this product, clouds were interpolated in Terra-only images yielding almost cloud-free daily maps for the period 2000–2018 (Matiu et al., 2020). Starting from the snow-cover data, the snow-cover duration (SCD) over the winter season (December, January and February) was computed at each grid point of the study domain, and the Pearson correlation with the mean winter temperature and total winter precipitation was computed over all available winters of SCD data (2001–2018). Since SCD saturates at high elevations preventing the correlation coefficients from being computed, all points above 2500 m were excluded from the analysis. As expected, correlation coefficients with temperature are negative within the inter-quantile range (25th–75th percentiles) for all elevation intervals (Table 3). The anticorrelation decreases with increasing elevation, with median values ranging from -0.55 below 500 m to -0.41 above 1500 m. In contrast, the correlation with total winter precipitation is positive within the inter-quantile range in all cases, and a clear elevation dependency is evident, with correlation values increasing with altitudes, up to 0.55 as the median in the range of 1500–1750 m. In considering the behaviour of SCD correlation with temperature and precipitation over different elevation bands, the effect of winter SCD saturation with increasing elevation should be taken into account. This could partly explain the increase in the anticorrelation with temperature and the concurrent decrease in precipitation–SCD dependency in the upper elevation bands (above 1750 m). These outcomes are in agreement with other literature studies investigating the dependency of several snow parameters with temperature and elevation in the Alps (see e.g. Móran-Tejeda et al., 2013; Schöner et al., 2019; Matiu et al., 2021) and specifically in the Trentino region (Marcolini et al., 2017).

The elevation dependency of the winter SCD correlation with temperature and precipitation is also highlighted in the spatial distribution of the coefficients over the region (Fig. 13). The greatest anticorrelation with temperature is depicted along the main valley floors, especially the Adige and Isarco valleys, while the greatest dependency on precipitation is pointed out along the mid-elevation zones, decreasing towards both lower elevations, where winter precipitation occurs generally in liquid form, and higher altitudes, where snowfall usually occurs in the winter season, and, regardless of their magnitude, snow lasts on the ground due to the relatively low temperature.