IT-SNOW is a high-resolution, observation-constrained snow reanalysis for Italy, available from September 2010 onward, providing daily maps of SWE, snow depth, bulk snow density, and liquid water content at ∼500m spatial resolution (Avanzi et al., 2023). The reanalysis is generated retrospectively by rerunning the operational Snow Multidata Mapping and Modeling (S3M) chain for Italy over the historical period, constraining the estimates by using in situ and satellite snow data available for each date. S3M is a distributed cryospheric model developed by CIMA Research Foundation for the Italian Civil Protection Department (Avanzi et al., 2022).

Meteorological forcing is derived from a very dense network of automatic weather stations providing hourly precipitation, air temperature, relative humidity, and incoming shortwave radiation. Precipitation fields (∼1km resolution) are generated using a modified conditional merging approach that blends rain gauge observations with radar data through the GRISO algorithm (Bruno et al., 2021). Air temperature is spatialised through region-specific hourly linear regressions against elevation, fitted within meteorological homogeneous regions defined by the Italian Civil Protection. Relative humidity and incoming shortwave radiation are spatialised using inverse distance weighting. At each hourly time step, total precipitation is partitioned into rainfall and snowfall based on air temperature and relative humidity, which is more reliable than temperature alone (Zhang et al., 2017), following Froidurot et al. (2014).

S3M solves the snow mass balance using a single-layer scheme with hybrid temperature-index and radiation-driven melt, snow settling, liquid water outflow, and albedo evolution (Avanzi et al., 2022). The melt formulation follows Pellicciotti et al. (2005) in separating radiation- and temperature-driven components, but S3M modulates the coefficients of both components using mean air temperature over the preceding 10 d, providing a parsimonious proxy for snowpack cold content and thermal inertia (Avanzi et al., 2022). Glacier areas, identified from the Randolph Glacier Inventory v6.0 (Pfeffer et al., 2014), are handled separately within the model, so that the IT-SNOW SWE fields represent only seasonal snowpack.

The modelled snowpack is constrained by the daily assimilation of snow depth observations. These come from a network of ∼350 ultrasonic sensors that follows the distribution of Italian snow resources and is therefore densest across the northern regions, particularly within the Po river basin. The point measurements are spatialised each day through multilinear regressions against elevation, slope, and aspect, fitted separately over 10 subdivisions of the territory that group areas of comparable snow climatology. The resulting fields are then trimmed to the observed snow line, using a blended snow-covered area map that combines Sentinel-2 (Main-Knorn et al., 2017), MODIS/VIIRS, and EUMETSAT H SAF H10 retrievals: pixels observed as snow-free are removed, so that the regressions do not extend snow onto bare ground. Conversion of the masked fields to SWE relies on the modelled snow density, whose spatial variability is considerably lower than that of snow depth (Mizukami and Perica, 2008). The SWE fields obtained in this way are assimilated once per day through spatially distributed Newtonian relaxation, or nudging: at each grid cell the modelled SWE is corrected by a fraction of its departure from the observation-based estimate, and this fraction decreases with distance from the sensors. The procedure is confined to December–April, outside which the ultrasonic sensors may pick up the vegetation growing beneath them after melt-out and introduce spurious signals into the spatialised maps (Avanzi et al., 2023).

IT-SNOW has been evaluated against Sentinel-1-derived snow depth maps (C-SNOW; Lievens et al., 2019) over the period September 2016–April 2020, and against independent in situ snow depth and SWE measurements (Avanzi et al., 2023). The comparison against C-SNOW showed that IT-SNOW snow depth estimates were consistent with satellite-based snow data, with low mean bias across the evaluated mountain areas (mean bias: -10mm; mean RMSE: 220 mm), and comparison against ground measurements confirmed that the dataset reconstructs seasonal snow accumulation and melt dynamics (RMSE: 300–600 mm for snow depth; 90–300 mm for SWE).

The products evaluated in this study differ in both product architecture and the way observational information enters the modelling system. Because these structural differences directly affect how biases originate and propagate into the simulated SWE fields, understanding them provides the basis for interpreting the evaluation results. This requires distinguishing the role of the atmospheric component from that of the land-surface modelling component in each product.

Atmospheric reanalyses reconstruct past atmospheric conditions by reprocessing the historical observing record with a fixed numerical weather prediction model and data assimilation system (Baatz et al., 2021). The atmospheric model solves the governing equations for mass, momentum, and energy, advancing prognostic state variables such as air temperature, humidity, wind components, and surface pressure, which together describe the instantaneous state of the atmosphere. Data assimilation updates this model state, adjusting the prognostic variables to bring the analysed atmospheric state into closer agreement with available observations while maintaining consistency with the model dynamics (Kalnay, 2003). In conventional atmospheric reanalyses, precipitation and the turbulent fluxes of latent and sensible heat differ in nature from the prognostic state variables: they do not describe the atmosphere at an instant but represent transfers of mass and energy over a time interval, and are typically diagnosed from the evolving model state rather than analysed directly. They are generated by the model itself, with precipitation produced within the atmospheric column by the cloud microphysics and convection schemes, and the turbulent surface fluxes evaluated at the land–atmosphere interface from the model atmospheric state and the prescribed or simulated surface conditions (Betts et al., 1996).

Simulating SWE and other land-surface variables requires a land-surface model with dedicated schemes for snow, soil, and vegetation processes. Two main configurations are commonly used for running such a model: offline and coupled (Masson et al., 2013). In an offline configuration, the land-surface model is driven by meteorological forcing fields and does not return energy or moisture fluxes to the atmosphere. These forcing fields may come directly from a global atmospheric reanalysis or from a regional atmospheric model or reanalysis that dynamically downscales a larger-scale driving system to higher resolution (Giorgi and Mearns, 1999; Laprise et al., 2008). In a coupled configuration, the land-surface scheme is integrated within the atmospheric model, and the two components exchange energy, moisture, and momentum continuously (Betts et al., 1996). Such coupling allows land–atmosphere feedbacks to develop, but also permits biases in either component to propagate through the coupled system (Santanello et al., 2018). The main characteristics of the three evaluated products, including their atmospheric forcing source, land-surface model, and coupling configuration, are summarised in Table 1.

The fifth-generation ECMWF reanalysis ERA5 is the common upstream global reanalysis from which all three evaluated products derive their atmospheric information, though in different ways. ERA5 is produced within the Copernicus Climate Change Service based on the Integrated Forecasting System at approximately 31 km horizontal resolution (Hersbach et al., 2020). Its atmospheric state is constrained through a hybrid incremental four-dimensional variational data assimilation system (Courtier et al., 1994), in which observations are assimilated within 12 h windows using short-range forecasts as the background. The assimilated observing system includes surface stations, radiosondes, aircraft, ships, buoys, and a large volume of satellite radiances (Soci et al., 2024).

The differences in how each of the analysed products uses this common source are relevant for interpreting their SWE estimates: ERA5-Land is driven directly by ERA5 atmospheric forcing fields, CERRA-Land receives its forcing through the intermediate regional atmospheric reanalysis CERRA (Ridal et al., 2024), for which ERA5 provides lateral boundary conditions, and VHR-REA_IT uses ERA5 only as lateral boundary conditions within a fully coupled regional atmosphere–land system. The following sections describe each product in more detail, focusing on the atmospheric forcing chain and the snow scheme.

The comparison covers 14 snow-years from September 2010–August 2024. The start of this period is determined by the availability of the IT-SNOW reference dataset, while the end date is constrained by the SCIA observational dataset (Desiato et al., 2007, 2011), which is used for the complementary forcing evaluation described in Sect. 4.2 and, at the time of this analysis, is available through the end of 2024. The analysis is restricted to grid cells where IT-SNOW mean annual SWE exceeded 1 mm in at least one snow-year, thereby excluding cells with generally negligible SWE.

The SWE estimates were evaluated using three grid-scale indicators, each targeting a different aspect of snowpack representation:

mean annual SWE, computed by averaging daily SWE over each snow-year and then across the full study period, quantifies the overall snow mass estimated by each product.

The four datasets differ substantially in native resolution (IT-SNOW: ∼500m; VHR-REA_IT: 2.2 km; CERRA-Land: 5.5 km; ERA5-Land: ∼9km). Therefore, these indicators cannot be compared directly without spatial harmonisation. To enable a consistent quantitative intercomparison, the datasets were re-gridded to a common spatial framework (Lindsay et al., 2014; Mudryk et al., 2025). For mean annual SWE and SCD, annual indicator fields were first computed at each product's native resolution to preserve spatial gradients in complex terrain and to prevent the distortion of threshold-based statistics by prior spatial averaging (Matiu et al., 2021; Rajulapati et al., 2020). The resulting indicator fields were then re-gridded to the coarsest grid (ERA5-Land, ∼9km) through first-order conservative remapping (Jones, 1999), and annual bias was calculated at each common-grid cell by subtracting the IT-SNOW value from the corresponding product value. These annual differences were averaged across all the snow-years to obtain the mean bias. Percentage bias was also computed as a complementary analysis and the resulting maps are shown in Supplement. Daily correlation required a different approach: because the Pearson coefficient must be computed from paired time series at matching grid cells, daily SWE fields were first re-gridded to the common 9 km grid through first-order conservative remapping, and the correlation was then computed pixel-wise over the full study period.

For each product and indicator, the statistical significance of mean annual SWE and SCD biases was assessed at each common-grid cell by applying a two-tailed Student's t-test to the time series of annual product–reference differences, testing whether their mean differed from zero. As discussed by Wilks (2016), interpreting numerous grid-cell tests without accounting for multiple testing can overstate the spatial extent of significant differences because some results are significant by chance alone. The resulting p-values were therefore evaluated using the Benjamini–Hochberg procedure (Benjamini and Hochberg, 1995), which controls the false discovery rate (FDR), defined as the expected proportion of false rejections among all rejected null hypotheses. The procedure was applied separately across the evaluated grid cells for each product and indicator. An FDR control level of 0.10 was adopted following Wilks (2016), who recommends using twice the target global significance level to obtain an approximately 0.05 global test level in fields with moderate to strong spatial correlation. Statistical significance is shown in Supplement to preserve the legibility of the main bias maps.

Finally, to obtain further insight into regional differences in product performance, the snow regime of each grid cell was also classified as seasonal or ephemeral according to snow persistence in the IT-SNOW reference. For each cell and snow-year, the longest sequence of consecutive days with SWE of at least 5 mm (the same threshold used for computing SCD) was identified, and its duration was averaged across all the snow-years. Following the 60 d persistence criterion (Sturm et al., 1995; Petersky and Harpold, 2018), cells with a mean duration of at least 60 d were classified as seasonal and the remaining cells as ephemeral. Since the classification serves as an interpretative aid rather than an evaluation result, the resulting regime map is provided in Appendix Fig. A1.

Precipitation and temperature are the main atmospheric controls on snow accumulation and ablation, and biases in these fields can contribute to discrepancies in simulated SWE (Monteiro and Morin, 2023; Terzago et al., 2020). Evaluating the precipitation and temperature fields associated with the three SWE products against an independent observational reference therefore provides useful context for interpreting the corresponding SWE discrepancies, although these discrepancies may also reflect other factors related to model structure and process representation.

The meteorological observational reference used for this purpose is SCIA (Desiato et al., 2007, 2011), maintained by the Italian Institute for Environmental Protection and Research (ISPRA). SCIA provides daily gridded precipitation at 10 km resolution from 1961 onwards and daily minimum and maximum 2 m air temperature at 5 km resolution from 1981 onwards, both derived from a high-density network of regional meteorological stations. The dataset has been widely adopted as a reference in Italian hydrological studies (e.g., Moccia et al., 2025; Sarigil et al., 2024), and its availability through the end of 2024 at the time of this analysis defines the end date of the study period adopted in this work.

For both the SCIA reference and each evaluated product, annual and seasonal precipitation depths and mean 2 m air temperatures were first calculated at native resolutions and then re-gridded to the common 9 km grid following the same procedure described in Sect. 4.1. Long-term mean annual and seasonal bias maps of each product against SCIA were thus calculated and used as a complementary diagnostic to support the interpretation of the SWE evaluation.

Given that a comprehensive standalone assessment of gridded meteorological products over Italy is not the main scope of this study, the mean annual and seasonal bias fields are presented in the Appendices rather than in the main text: Appendix B reports the mean annual biases (Fig. B1 for precipitation, Fig. B2 for temperature), and Appendix C reports the seasonal biases in precipitation, temperature, and SWE for each product (Fig. C1 for ERA5-Land, Fig. C2 for CERRA-Land, Fig. C3 for VHR-REA_IT). These figures are referenced throughout Sect. 5 to support the interpretation of the SWE evaluation results.

In addition, the statistical significance of the mean annual precipitation and temperature biases was assessed using the same paired t test and false discovery rate (FDR) procedure described in Sect. 4.1, with SCIA as the reference. To preserve the legibility of the bias maps in Appendix Figs. B1 and B2, the results of the FDR-screening are shown in the corresponding expanded Supplement (Figs. S3 and S4).

The initial phase of the assessment is the comparison of the mean annual SWE estimates from each product. The results are presented in Fig. 3: figure's top row (panels a–d) displays the mean annual SWE from the IT-SNOW reference and the three land-surface products at their native resolution, allowing for a direct visualisation of their respective spatial distributions and magnitudes; in addition, to quantify the discrepancies from the reference data, the second row (panels e–g) presents bias maps calculated as the difference between each of the three land-surface products and the reference (Product – IT-SNOW), so that negative (red colour) values represent where the product underestimates the reference values and positive (blue colour) values represent an overestimation. For this comparison, the datasets resampled to the ERA5-Land resolution are used.

The three land-surface products (panels b–d) broadly reproduce the spatial distribution of SWE shown by the IT-SNOW reference (panel a), with the highest mean annual SWE values in the high Alps and along the Apennine ridge. The agreement, however, is largely confined to this overall pattern: the bias maps (panels e–g) show product–reference differences that are largest at high elevations, where mean annual SWE itself is the greatest, and that differ in sign among the products. Figure S1 presents an expanded version of this figure, adding percentage-bias panels and marking the cells whose biases are not significant after false discovery rate (FDR) screening. Where reference SWE is substantial, the percentage-bias panels largely mirror the absolute-bias pattern; where it is small, however, even small absolute differences can yield large percentage biases, which should therefore be interpreted with caution (Fig. S1, panels h–j). Across all products, cells with non-significant biases are generally confined to areas with the smallest bias values, whereas large biases are statistically significant (Fig. S1, panels e–g). Overall, this provides confidence in the validity of the mean SWE bias analysis, as the most prominent discrepancies prove to be statistically robust.

Considering the mean annual SWE estimates in more detail, we note that the coarse spatial resolution of ERA5-Land smooths out the high spatial variability of SWE (panel b). The bias map (panel e) shows large positive biases over much of the high Alpine region, alongside some localised negative values, whereas the Apennines and lowlands show widespread underestimation. The concentration of the overestimation in the highest mountain areas is consistent with what has been reported by Kouki et al. (2023), who attributed the ERA5-Land overestimation (150 %–200 %) of the spring SWE over the entire Northern Hemisphere mainly to excessively high values in the major mountain regions. The comparison of the meteorological forcing at seasonal scale (Fig. C1) shows that ERA5-Land exhibits positive winter precipitation biases together with negative winter temperature biases across the Alps, strongest over the north-western sector (Fig. C1, panels b and f), in agreement with Cavalleri et al. (2024a, b), Dalla Torre et al. (2024), and Sarigil et al. (2024). The co-occurrence of higher precipitation and lower temperatures in winter is consistent with enhanced snow accumulation and with positive SWE biases that persist into spring (Fig. C1, panels i–k). However, another potential explanation for such biases was given by Orsolini et al. (2019), who identified the discontinuation of snow-cover assimilation from the Interactive Multisensor Snow and Ice Mapping System (IMS) above 1500 m a.s.l. as a possible further contributor to high-elevation snow biases in ERA5 over the Tibetan Plateau, since the simulated snow cover at these elevations is no longer corrected by observed snow cover. The same restriction applies over the Alps, where excessive snow cover in ERA5 could increase surface albedo, artificially lowering near-surface temperatures; these cold biases could then be propagated into ERA5-Land, which relies on ERA5 meteorology as its atmospheric forcing. Moreover, snow–albedo interactions could also reinforce the positive SWE bias within ERA5-Land itself: within CHTESSEL, an excessive snowpack could prolong snow cover and maintain a higher surface albedo, reducing absorbed shortwave radiation and the energy available for melt, thereby further delaying snow depletion. In the Apennines, where snow is predominantly ephemeral and seasonal snow cells are largely confined to the highest elevations and most extensive over the central massifs (Fig. A1), the relationship between SWE biases and seasonal forcing biases is less straightforward. In the northern Apennines, the negative SWE bias is mainly consistent with a precipitation deficit rather than with temperature biases: both autumn and winter precipitation are underestimated, limiting snow accumulation from the onset of the season, whereas later anomalies are too weak or too late to offset the winter deficit (Fig. C1, panels a–d). In the central Apennines, however, the seasonal forcing biases do not readily explain the negative SWE bias: winter precipitation biases are near-neutral to weakly positive and winter temperature biases are negative (Fig. C1, panels b and f), conditions that would favour rather than suppress snow accumulation. This apparent inconsistency may reflect factors not resolved by the seasonal precipitation and temperature averages, including differences in precipitation-phase estimation, the limited representation of the high-elevation massifs on the 9 km grid, and differences in the representation of snow processes such as ablation. Ablation, in particular, can be relevant even during the accumulation season: in evaluations of ERA5-Land and other global products in High Mountain Asia (Liu et al., 2022) and in the western United States and the Andes (Fang et al., 2023), accumulation-season snow losses contributed to inter-product differences in SWE estimation.

CERRA-Land demonstrates an improvement in reproducing the spatial variability of mean annual SWE compared to ERA5-Land, thanks to its finer 5.5 km grid spacing (Fig. 3, panel c). However, despite this improved spatial detail, annual SWE remains negatively biased in most regions, with some localised exceptions (panel f). In the Alps, unlike ERA5-Land which produces marked overestimation due to combined, strong and spatially coherent precipitation and temperature biases, CERRA-Land SWE is influenced by an overall weak underestimation of temperature during the snow season (Fig. C2, panels e–g) and a much more varying pattern of precipitation biases (Fig. C2, panels a–c). In the central and eastern Alps, autumn precipitation bias is scattered and changes across the valleys (Fig. C2, panel a), while in winter it is overall weakly negative (Fig. C2, panel b). Such a forcing pattern only partially explains the strong negative SWE bias in the same region, since the weak cold bias would favour rather than suppress snow accumulation and retention. This suggests that the observed deficit may reflect not only forcing errors but also processes internal to the snow modelling chain, although dedicated process-level diagnostics would be required to identify the specific mechanisms involved. The main Alpine exception is the north-western Alps, where positive autumn–winter precipitation biases (Fig. C2, panels a–b) and colder conditions (Fig. C2, panels e–f) produce the only spatially coherent patches of positive SWE bias, although the cell-level pattern remains heterogeneous. However, interpretation of the SWE pattern in areas very close to the national border should nevertheless remain cautious since the reference datasets are less robust there: IT-SNOW assimilates fewer observations (Avanzi et al., 2023), and SCIA does not extend beyond the national border; in addition, gauge station density is also limited in such high-elevation areas. Across the Apennine chain, instead, the attribution is more direct: precipitation deficit is strongest during autumn and winter, especially in the north, leading to negative SWE biases during the main accumulation season despite the overall weak cold bias.

VHR-REA_IT exhibits the most spatially widespread negative SWE biases among the evaluated products, despite having the highest spatial resolution (∼2.2km) (Fig. 3, panels d and g). A strong negative SWE bias characterises much of the domain, although some high-elevation Alpine areas show locally positive anomalies during the early accumulation season (Fig. C3, panels i–j). The joint seasonal analysis reveals a seasonal asymmetry in the temperature bias, with a cold bias in winter transitioning to a strong warm bias in summer (Fig. C3, panels f and h), consistent with documented evaluations of the COSMO-CLM atmospheric fields over Italy (Adinolfi et al., 2023; Cavalleri et al., 2024a; Loprieno et al., 2024; Raffa et al., 2021). However, the seasonal meteorological forcing does not fully account for the SWE underestimation in several subregions. In the Alps, winter precipitation is generally overestimated and winter temperatures are colder than the reference (Fig. C3, panels b and f), conditions that should favour snow accumulation, yet SWE bias remains negative in all three Alpine subregions. Similarly, in the central Apennines, precipitation becomes weakly positive in winter while temperatures remain below the reference, yet SWE bias remains negative during the accumulation season (Fig. C3, panels i–k). Only in the northern Apennines do autumn–winter precipitation deficits provide a more direct explanation, although the persistence of negative SWE bias beyond winter suggests that precipitation deficit alone is insufficient to explain the full anomaly. The persistence of negative SWE biases under autumn–winter forcing conditions that would otherwise favour snow accumulation points to structural factors in the modelling chain. Unlike ERA5-Land and CERRA-Land, which are offline simulations driven by observationally constrained reanalysis forcing, VHR-REA_IT is a free-running dynamical downscaling in which COSMO-CLM and its land-surface scheme TERRA-ML operate as a fully coupled atmosphere–land system (Raffa et al., 2021), with no observations assimilated within the regional domain. Any initial negative SWE perturbation generated within this configuration, whether originating in the snow scheme structure, in model parameters, or in the internal forcing, can then be reinforced rather than corrected by the coupled atmosphere and land components: a reduction in simulated SWE lowers surface albedo, warms the lower boundary layer, and promotes further melt through land–atmosphere feedbacks (Santanello et al., 2018) that are not represented in the same two-way form in the offline configurations of the other two products. Finer resolution and two-way coupling therefore do not by themselves guarantee more accurate SWE estimates, underscoring the importance of observational constraints in high-resolution applications (Fang et al., 2023).

Beyond mean annual SWE magnitudes, snow cover duration (SCD), defined in Sect. 4.1 as the mean annual number of days with SWE at least 5 mm, provides complementary insights into snowpack temporal dynamics. SCD helps to better capture the integrated effects of meteorological and topographic factors, including temperature regimes, precipitation patterns, elevation, and energy balance processes, that collectively determine snow persistence. As a diagnostic variable, we used SCD for assessing the performance of the evaluated products across the complete seasonal cycle, as it integrates cumulative effects of biases in both accumulation and melting processes that may not be evident in mean SWE comparisons. Analogously to Fig. 3, Fig. 4 presents the spatial distribution of mean annual SCD for each dataset (panels a–d) at their native resolution and the corresponding biases relative to IT-SNOW (panels e–g) on the common resampled 9 km grid.

The SCD maps (Fig. 4, top panels a–d) confirm a common elevation-dependent spatial pattern across the country, consistent with mean annual SWE in Fig. 3. However, the bias analysis (bottom panels e–g) shows patterns that partially differ from the mean annual SWE biases (Fig. 3), indicating that biases in total snow mass do not always correspond to biases in snow cover persistence. Similarly to what showed for mean SWE bias, Fig. S2 complements these maps with percentage-bias panels and marking the cells whose biases are not significant after FDR screening. In the percentage-bias panels, the largest relative biases generally occur in lowland, southern and insular cells, where reference SCD is short (Fig. S2, panels h–j). As for mean annual SWE, the largest SCD biases are generally statistically significant, whereas non-significant cells mainly occur where biases are the smallest (Fig. S2, panels e–g).

ERA5-Land (panel e) displays a spatially heterogeneous SCD bias pattern that broadly mirrors the mean annual SWE biases (Fig. 3, panel e), with overestimation concentrated in the Alps and predominantly negative biases across the Apennines and southern Italy. Over the Alps, the positive SCD bias is strongest and most spatially coherent in the north-western sector, extending across the high-elevation areas where mean annual SCD is already longest, and weakens progressively towards the central and eastern sectors. This overestimation is consistent with the broader literature, which identifies prolonged snow cover as a characteristic bias of ERA5-Land over high mountain areas (Kouki et al., 2023; Monteiro and Morin, 2023). Over the European Alps specifically, Monteiro and Morin (2023) found ERA5-Land exhibiting the strongest and most spatially extensive SCD overestimation among all evaluated datasets when compared against MODIS satellite observations, with the largest discrepancies occurring at the melt-out date. Urraca and Gobron (2023), evaluating ERA5-Land against ground stations in the Northern Hemisphere, attributed the systematic positive SCD bias to the tendency of the CHTESSEL land-surface model to overestimate snow depth, likely due to excessive snowfall in the forcing. As discussed in Sect. 5.1, the combination of overestimated precipitation and persistent temperature underestimation over the north-western Alps inflates both snow accumulation and snow persistence, and the SCD overestimation confirms that these forcing biases translate into prolonged snow cover seasons. In the Apennines, SCD biases are predominantly negative but vary across subregions. The northern Apennines show the most widespread underestimation, consistent with the autumn and winter precipitation deficits identified for these sectors in Sect. 5.1, which limit the number of days on which SWE is at least 5 mm and thus shorten the simulated snow season. A similarly widespread negative SCD bias is observed across the southern Apennines. The central Apennines display a mixed pattern: negative SCD biases over the higher massifs occur alongside locally positive biases in lower-lying areas, where reference SCD is short and SWE is also slightly overestimated. The drivers of this mixed, elevation-dependent pattern are difficult to isolate. It may stem from the same combination of factors hypothesized in Sect. 5.1 for the mean SWE biases, such as the coarse grid representation of the topography and unresolved local snow processes.

CERRA-Land shows a more moderate Alpine SCD overestimation than ERA5-Land, while maintaining negative biases across the Apennines (Fig. 4, panel f). With the exception of some valleys in the central-eastern Alps, where the product underestimates SCD, most of the Alps show positive SCD bias, with the strongest and most spatially coherent overestimation occurring over the north-western portion of the chain, near the Swiss and French borders. This pattern aligns with the findings of Monteiro and Morin (2023), who reported that CERRA-Land particularly overestimates SCD over the western Italian Alps when compared against MODIS satellite observations. As discussed in Sect. 5.1, positive autumn–winter precipitation biases and colder conditions in this subregion favour both excessive accumulation and prolonged snow retention. Over the most eastern Alps, the positive SCD bias is weaker but still present, consistent with the comparatively modest annual SWE biases in these sectors. Across all three Apennine subregions and southern Italy, CERRA-Land underestimates SCD, with the strongest negative bias in the northern Apennines, consistent with the most severe precipitation deficits in the snow accumulation season identified in Sect. 5.1 (Fig. C2, panels a and b). In the central and southern Apennines, autumn and winter precipitation underestimation likewise coincides with negative SCD biases, whereas positive precipitation biases are concentrated in the warmer seasons and therefore provide little support for snow persistence.

VHR-REA_IT (panel g) underestimates SCD across all subregions in both the Alps and the Apennines. Unlike ERA5-Land and CERRA-Land, which show opposing SCD bias signs between the Alps and the Apennines, VHR-REA_IT is the only product with consistently negative SCD biases across all subregions. Within the Alps, the underestimation intensifies from the north-western sector towards the central and eastern sectors, where the SCD deficit is the largest among all subregions despite winter precipitation overestimation and cold temperature biases (Fig. C3, panels b and f), suggesting that forcing biases alone cannot explain the pattern and that the snow-model structure or parameterisation may have an important role, as discussed for mean annual SWE. Overall, VHR-REA_IT exhibits the poorest SCD performance among the three evaluated products.

To complement the assessment of annual snowpack characteristics, we evaluated how well each product reproduces the day-to-day variability of SWE through temporal correlation analysis. The Pearson correlation coefficient (r) was calculated at each grid cell between the daily SWE time series of each product and of the IT-SNOW reference, for the entire study period (September 2010–August 2024), following resampling to the common ERA5-Land 9 km spatial resolution. The results are presented in Fig. 5, which displays spatial maps of the correlation coefficients, providing a direct visual assessment of temporal correspondence between simulated and reference snowpack dynamics across Italy.

Overall, ERA5-Land demonstrates the highest daily correlations among the three products, with the strongest agreement in the Alpine region and weaker but still comparatively better performance in the Apennines. Despite the substantial overestimation of SWE magnitude at higher elevations documented in Sect. 5.1, the temporal variability of the simulated snowpack remains broadly consistent with the reference, particularly in the Alps. This is consistent with the offline product architecture described in Sect. 3.2.1: as ERA5's atmospheric data assimilation constrains the sequence and timing of synoptic weather events, the forcing inherited by CHTESSEL is expected to preserve realistic temporal patterns even where simulated mean annual SWE and SCD are biased at the land surface. Similar behaviour has been reported in other mountainous regions, where ERA5-Land maintains reasonable temporal consistency with the reference despite substantial magnitude biases: over parts of Iran, ERA5-Land preserves temporal consistency while overestimating snow depth (Majidi et al., 2025), and over parts of Canada it does so while underestimating SWE (Kanda and Fletcher, 2024).

CERRA-Land shows daily correlation patterns broadly comparable to ERA5-Land across the Alpine region, with similar spatial structure and only slightly lower coefficients in the north-western sector. In the Apennines, however, the correlation weakens more substantially than for ERA5-Land, particularly in the northern sector. This stronger degradation is consistent with the more severe autumn and winter precipitation deficits in CERRA-Land's forcing over the Apennines discussed in Sect. 5.1 (Fig. C2, panels a and b), which exceed those of ERA5-Land in the same subregions. In these lower-elevation Apennine regions, where snow cover is more intermittent, such deficits are more likely to reduce the simulated duration and continuity of snow cover during periods when the reference indicates snow presence, thereby affecting not only snow amount but also snow persistence and weakening the temporal correspondence of daily SWE with the reference time series.

VHR-REA_IT exhibits markedly poorer performance in capturing daily SWE dynamics, with substantially lower daily correlations throughout most of Italy, particularly in the Apennines and at lower elevations where correlations fall into the lowest class of the colour scale. This weak temporal correspondence is consistent with VHR-REA_IT's systematic underestimation of both mean annual SWE and SCD documented in Sect. 5.1 and 5.2. The severe underestimation of SWE produces frequent zero or near-zero values when a substantial snowpack should exist, thereby degrading the correlation. Therefore, the model fails to capture day-to-day variability for snow that it incorrectly simulates as absent.