We use the PCR-GLOBWB 2.0 model in the 30 arcsec setup developed by (approx. 1 km at the Equator and 650 m in the longitudinal direction in the Alps). The model runs at a daily time step. PCR-GLOBWB 2.0 is a global hydrological model and contains different modules, which represent both natural processes related to vegetation, snow, soil, groundwater, and river routing and anthropogenic processes such as human water use and irrigation.

Here, we use a regional model setup (latitude: 43–51°, longitude: 3–18°) covering the Alps and the upstream parts of the catchments of four major central European rivers (i.e. the Rhone, Rhine, Danube, and Po; Fig. ). We focus on the period 1990–2019 as all forcing datasets are available for this time period, and initial glacier volumes are often only available for around the year 2000.

We implement the different forcing datasets and model changes in a step-by-step manner and thus perform several model runs. A schematic overview and further details on the sequential model runs performed in this study are provided in Fig.  and Table , respectively.

Based on the initial regional model setup introduced in Sect. , we further develop the representation of cryospheric and soil processes to improve discharge simulations in mountain regions. We aim to find a regionally valid setup that works well for a larger domain and that is thus not directly fine-tuned for individual catchments, in line with the philosophy of many global hydrological models. The next few sections describe the structural changes made to the PCR-GLOBWB 2.0 model. The parameters used in the equations and any fixed values are listed in Table S1 in the Supplement. The calibration strategy for specific parameters is then further outlined in Sect. .

The aim of our evaluation is to assess the effect of the model development both on simulated discharge and the representation of individual processes such as snow accumulation and snowmelt. We thus evaluate the forcing datasets for precipitation and the model output for streamflow, SWE, glacier mass balance changes, glacier surface elevation evolution, and soil moisture. We generally evaluate the daily simulations, except for the glaciers for which we use annual mass balances. For the evaluation, we split the study period into three blocks with different annual mean air temperature characteristics: (1) the calibration period (2000–2009), (2) a colder evaluation period (evaluation I, 1990–1999), and (3) a warmer evaluation period (evaluation II, 2010–2019) (see Fig. S1a in the Supplement). The main evaluation is performed over evaluation II, which is also the period considered if a specific time interval is not specified. In the Supplement (Sect. S2), we use both evaluation periods to assess the transferability of the new model setup to time periods with different mean temperatures within the framework of a differential split sample test (DSST; ).

The evaluation of the meteorological forcing focuses on precipitation, which is more spatially heterogeneous than temperature. We try to analyse the differences between the datasets and their realism, acknowledging the large uncertainties around high-altitude precipitation. We resample all three forcing datasets (STANDARD, CHELSA, and CERRA; see Sect. ) to the grid points of the coarser reference APGD and calculate both the Pearson correlation and the absolute bias against the APGD.

We evaluate the performance of the daily discharge simulations by comparing simulated discharge against observed streamflow, i.e. the station data from the LHDA+. We match model grid cells to discharge stations by matching the catchment contours following . We then evaluate discharge simulation performance by comparing the simulated time series to the observed time series using the Kling–Gupta efficiency (KGE; ), using only stations that have at least 3 years of data over the considered evaluation period (number of valid stations: evaluation I – 1676, calibration period – 1869, evaluation II – 2105). The KGE is defined as follows: 6KGE=1-(r-1)2+μsimμobs-12+σsimσobs-12, where r is the Pearson correlation coefficient between observations and simulations, μobs is the mean over the observations, μsim is the mean over the simulations, σobs is the standard deviation over the observations, and σsim is the standard deviation over the simulations. KGE scores above -0.41 indicate that model simulations improve performance compared to assuming a constant flow corresponding to the average of the observed discharge time series . To assess whether and by how much a specific change in model structure improves model performance for discharge, we use the KGE skill score (KGESS) as used by and . The KGESS compares the KGE score of a model run with a new setup against a model benchmark: 7KGESS=KGEmodel-KGEbench1-KGEbench, where KGE_model is the KGE score of the new model run, and KGE_bench is the KGE score of a benchmark model run, which represents an intermediate step within the model improvement chain.

The Alpine region contains many reservoirs for water regulation, especially for hydropower production . Their presence and how they are represented can influence model performance (e.g. ). Therefore, we investigate how model performance for discharge varies between natural and regulated catchments. As not each catchment with a reservoir might be strongly regulated, we define a catchment's degree of regulation by dividing its total reservoir capacity by its mean discharge (in units of time). We only consider a catchment to be regulated when this degree of regulation exceeds 0.1 years (see Fig. S3 in the Supplement). Another way to separate regulated from natural catchments is the water balance signature (WB), which describes the normalized deviation from a closed water balance assuming no long-term storage effects . WB is defined as follows: 8WB=QP-1-EPP, where Q is the (observed) averaged discharge (mm d⁻¹), P is the precipitation (mm d⁻¹), and EP is the potential evapotranspiration (mm d⁻¹) averaged over the catchment and study period. Positive values of WB indicate that a catchment's discharge is higher than expected. Assuming that the meteorological components could be well-estimated, such positive values could suggest that a catchment gains more water than what comes in through precipitation. Negative values of WB indicate that a catchment loses more water than just the potential evapotranspiration. have shown that this metric can be used to identify catchments affected by hydropower production and water transfers. In addition, we performed a small evaluation which indicated that, indeed, there exists a relationship between WB and the degree of regulation (see Fig. S3 in the Supplement). However, other factors such as errors in the meteorological forcing or additional water input from glaciers due to imbalance can also lead to strong water balance deviations, which can also affect model performance with respect to discharge. We thus use WB as a general metric to study the effect of such deviations in the water balance on model performance. Furthermore, we also use WB as an indication of water transfers, hydropower production, and other water balance deviations in combination with catchment-based information on reservoirs.

In addition to discharge, we evaluate model performance for SWE by calculating spatial averages over different elevation bands (0–1000, 1000–2000, 2000–3000 m, or the entire country) over Switzerland. Additionally, we evaluate it over Austria, which has not been used for model calibration and can therefore provide insights into how well the model generalizes to other regions. Then, we compute the average seasonal SWE cycle over these zones and compare it to the seasonal SWE cycles derived from different snow reanalysis products (OSHD, SNOWGRID; see Table ). Furthermore, we calculate the KGE for SWE by comparing modelled SWE and SWE inferred from observations at specific measurement stations (NH–SWE; see Table ).

We evaluate glacier simulations in terms of both their mass balance and their geometry evolution. Simulated mass balances were summed over the hydrological year (starting in October to facilitate comparisons with the observations of WGMS and GLAMOS; see Table ) and compared to observed time series, as well as to observed average geodetic mass loss estimates derived from GECM. To evaluate glacier geometry evolution, we both visually compare spatial patterns of glacier surface elevation changes against observations (GECM; see Table ) and quantitatively evaluate long-term glacier changes. It is difficult to evaluate the long-term response of simulated glaciers to climate forcing given the relatively short study period. To address this problem, we repeat an experiment by , in which glaciers are continuously forced (for 300 years) with the modelled mean mass balance for the period before 2018. The glaciers respond to this forcing by changing their shapes, and they stabilize when they are in balance with the applied forcing. We then compare this to the estimates of .

We evaluate soil moisture against the ESACCI soil moisture dataset. This evaluation is more difficult than the evaluation of other variables since the satellite data are not directly comparable to our model output: they measure soil moisture in the top 5 cm of the soil, whereas we model moisture in the top 15 or 30 cm (see Table ). Furthermore, the resolution of the observations is much coarser than the one of the simulations. Therefore, we resample our simulations to the resolution of the satellite data using spatial averages and only select locations with at least 50 % of valid daily data over evaluation II. We calculate the Spearman rank correlation coefficient between the daily model simulations and the satellite observations to assess model performance because this metric should be applicable despite the different soil moisture depths.

Precipitation in all three meteorological forcing datasets used in this comparison correlates well with gridded precipitation station data from the APGD (see Fig. ). Precipitation from CERRA-CHELSA shows generally higher correlations with observed precipitation than the precipitation of the STANDARD and CHELSA input datasets (compare Fig. a and b with c). The correlation between the three datasets and the reference dataset varies across the Alps and is especially low in the Po Plain. All meteorological datasets show a positive bias in precipitation over the Alps and a negative bias over the Apennines (see Fig. d, e, and f). Around the Alps, the STANDARD and CHELSA forcings show slightly positive biases, whereas CERRA-CHELSA shows a slightly negative bias. Overall, CERRA-CHELSA and STANDARD have a smaller precipitation bias than the CHELSA dataset (mean absolute bias: CERRA-CHELSA – 0.4 mm d⁻¹, CHELSA – 0.5 mm d⁻¹, STANDARD – 4 mm d⁻¹).

The choice of meteorological forcing has a strong effect on simulated discharge, but the effect is not uniform across catchments. Using the STANDARD forcing, we see generally better model performance for discharge over the Alps than in the surrounding areas (see Fig. g). However, locally, there is poor model performance for discharge in certain catchments in the western Alps, southern Switzerland, and eastern Austria. Forcing the model with CERRA-CHELSA leads to improved performance in these regions, as well as in southern Germany (see Fig. i). In contrast, CERRA-CHELSA leads to reductions in model performance for discharge in eastern France, parts of Switzerland, and western Austria. Using CHELSA leads to a slight worsening of model performance for discharge compared to runs with STANDARD or CERRA-CHELSA forcing, except in parts of the Alps (see Fig. h).

In summary, we find that discharge simulations generated with the CERRA-CHELSA and STANDARD forcing datasets are generally better than those generated with the CHELSA dataset (see Fig. j). As the precipitation of the CERRA-CHELSA dataset aligns better with the reference precipitation dataset than the STANDARD dataset (see Fig. a, b, d, and e), we performed all further analyses with the CERRA-CHELSA dataset.

SWE representation benefits to some degree from the proposed adjustments in the snow module (see Fig. ). The introduction of the snow transport scheme only improved snow representation at the highest elevations (2000–3000 m), where unrealistic snow towers were a major issue (compare the transport run with the CERRA-CHELSA benchmark run in Fig. c and f). Here, the snow transport scheme ensures that the snow is redistributed to lower elevations, where it subsequently melts away. ERA5-Land and CERRA-Land also show very high SWE values, suggesting that these models also suffer from unrealistic snow build-up. Without calibration, further structural changes to the snow module (i.e. the seasonally varying DDF, exponential temperature dependence, and a rain-to-snowfall transition temperature range) did not improve performance against observational SWE stations (see Fig. g and h). Averaged over elevation zones, however, these changes lead to an improvement in the SWE representation at higher elevations (e.g. for elevation zone 2000–3000 m, the KGE compared to the reference products increases from 0.85 to 0.89 over Switzerland and from 0.53 to 0.57 over Austria). At lower elevations, these changes lead to too-low melt rates over Austria (for elevation zone 1000–2000 m, KGE increased from 0.87 to 0.88 over Switzerland but dropped sharply from 0.64 to 0.46 over Austria; compare uncalibrated snow and transport runs in Fig. a, b, d, and e). These differences between elevation zones suggest that the model structure leads to varying melt rates with elevation: this is to be expected as the temperatures at which snow starts to melt are reached later in the year at higher elevations, such that the same temperatures are combined with different values of the time-varying DDF which produces differing melt rates. Calibrating the DDF in the snow module against SWE shows the best performance over Switzerland (KGE averaged over Switzerland against OSHD: 0.91), though still with slightly too-high melt rates over Switzerland (see full model run in Fig. a, b, and c). In Austria, the calibrated snow run is, while slightly less accurate than in Switzerland, the most accurate of the presented model runs (KGE averaged over Austria against SNOWGRID: 0.80), even though no Austrian SWE data were used for calibration (see Fig. d, e, and f). This demonstrates that the snow module is generally transferable to other regions. The comparison of our model runs against SWE estimates at measurement stations (see Fig. g and h) confirms that SWE simulations profit from the calibration.

Changes in SWE performance are reflected in the performance of discharge simulations (see Fig. ). Snow transport improves discharge simulations in the highest parts of the Alps but has hardly any effect outside of the mountains (see Fig. b). Structural changes to the snow module lead to an improvement in discharge simulations in catchments at the highest elevations but a worsening in catchments at lower elevations (see Fig. c). Finally, calibrating SWE improves discharge representation in most of the catchments that worsened from the structural changes, with a few exceptions in the Alps (see Fig. d). Figure a, b, and c illustrate that the changes in the snow module mainly improve model performance for discharge in catchments with high snowfall fractions, whereas the changes are negligible or slightly negative in catchments with low snowfall fractions. Still, some catchments with high snowfall fractions experience decreases in performance for discharge: these decreases mostly happen in the presence of reservoirs and/or with negative values of WB (median KGESS for catchments with more than 30 % snowfall compared to CERRA benchmark: regulated 0.017, unregulated 0.080).

The new glacier module captures the general behaviour of glaciers. Both spatial patterns of glacier elevation changes and time mean patters of mass balances are roughly reproduced (see Fig. a, b, c, d, e, and f), even though the model shows biases for individual glaciers in terms of mass balance (see Fig. a, b, and g) or retreat (see rapid retreat of Mer de Glace in Fig. d). Generally, mean geodetic mass balances are underestimated (see Fig. g). Figure h shows the results from the equilibrium experiment based on . Most glaciers reach equilibrium over time, although there is significant committed mass loss. Overall, we end up with around 40 % committed mass loss in 2018, which was also found by . Our modelling scheme thus captures the general behaviour of long-term glacier responses, with glacier retreat adjusting to a new steady-state condition. However, we see deviations for individual glaciers, with a larger glaciers generally retreating slightly less and smaller glaciers retreating slightly more than the reference dataset. In conclusion, our model evaluation shows that the new glacier module works reasonably well for the total or groups of Alpine glaciers (see Fig. g and h), while it can be significantly biased at the scale of individual or small glaciers (see Fig. g, h).

The addition of the glacier module mainly improves discharge simulations in highly glacierized catchments (see Figs. e and f). The positive effect of glaciers is much less visible in catchments with a small glacier or snowfall fraction (see Fig. e). Furthermore, our results indicate that, in certain regions, especially around the heavily regulated Rhône River in southwestern Switzerland, discharge performance can decrease with the addition of glaciers (see Fig. e). Such a performance decrease generally occurs in catchments with a negative WB or reservoir regulation (see Fig. f; median KGESS in catchments with >5 % glacier cover compared to calibrated snow: regulated -0.19, unregulated 0.15).

Model performance for soil moisture varies over the domain (see Fig. a and b), with generally higher performance in flatter low-elevation areas (such as the Rhone Valley, the Po Plain, or the Rhine Valley) and lower performance over hilly or mountain areas. Implementing the soil changes has a mixed effect on performance: generally, it improves performance in areas where the model was already performing well (i.e. the lower flatlands) and worsens performance in locations with lower model performance (see Fig. c).

Over the entire domain, the changes made to the soil module, i.e. to the runoff partitioning, generally increase model performance for discharge (see Fig. f). The improvements in discharge performance are strongest in catchments with lower snowfall fractions (<0.3; see Fig. h) and are generally independent of the catchment area (see Fig. g). Discharge performance increases are stronger in unregulated catchments (see Fig. g, h, and i; median KGESS compared to snow and glaciers: regulated 0.03, unregulated 0.09).

Our new model setup, which includes updated snow, glacier, and soil modules, leads to general performance increases in streamflow simulations compared to the existing PCR-GLOBWB 2.0 setup, with performance depending on catchment characteristics (see Fig.  and the summary provided in Table  and Fig. ). Catchment area and WB are major controls of absolute model performance for discharge in terms of KGE, which is highest in large and natural catchments, in which the water balance is nearly closed (see Fig. a). Similarly, the model performs well in snow-covered and glacierized catchments, where the model additions lead to a substantial improvement in model performance for discharge (see Fig. b, c, e, and f). In contrast, the model adjustments can decrease performance in catchments with a negative WB (see Fig. d, e, and f; median KGESS of full model against CERRA benchmark: positive WB 0.09, negative WB 0.02).

Our new model setup generally led to increased model performance compared to the old setup: structural and parameter changes applied to the snow, glacier, and soil modules improved both SWE and discharge simulations (Figs. , , , and ). This highlights the importance of improving process representation in hyper-resolution modelling efforts.

Updating the runoff partitioning in the soil leads to a clear improvement in the simulation of discharge in natural catchments (Figs. f and g, h, and i; column “soil partitioning” in Table ), with rainfall-dominated catchments (≈ less than 30 % snowfall) benefiting the most (Fig. g and h). This major increase in model performance is caused by a modest change in the soil parameters, supporting the suggestion that the move to a hyper-resolution requires careful review of parameterizations in LHMs . Still, absolute performance in smaller rainfall-dominated catchments remains relatively limited (Fig. a and b). In addition, soil moisture representation did not improve in regions that already had poor performance (Fig. c and d). Please note that the reference product used for computing model errors can have its own biases and has a much coarser spatial resolution than our model, and so error estimates might not be entirely representative. Our results therefore suggest that there is a need for more representative ways to compare soil moisture simulations and observations. Still, the representation of soil moisture and fast discharge responses needs to be further improved if LHMs are to be applicable at smaller spatial scales. These further improvements can come from two directions: first, soil heterogeneity could be included in models in more detail. For example, suggested that sub-grid-scale land cover variability could still be important in hyper-resolution modelling, and we hypothesize that this would also improve the representation of spatial variations in discharge behaviour. However, including this information would come at the cost of increased computational demands. Second, more explicit consideration of hillslope processes such as preferential flow in the model structure (e.g. ) could lead to further improvements in the simulation of flashy runoff responses.

Discharge simulations mostly benefit from the structural changes made to the snow module in snow-dominated catchments (Figs. c and e; column “snow module” in Table ). Similarly, found, for the catchment model HBV, that exponential melt combined with seasonally varying DDFs improved discharge simulations, whereas the rain-to-snow transitions improved SWE representation but led to slightly poorer results for discharge. However, whereas our discharge simulations in most snow-dominated mountain catchments improved due to these structural changes, we noted a slight decrease in discharge performance at lower elevations, where snow contributions are less important. One reason for why the structural changes in the snow module are better suited for catchments in mountainous terrain might be differences in terms of the dominant snow processes between high and low elevations, such as the frequency of rain-on-snow events. Since built their snow model for alpine Switzerland, they might have prioritized representing melt patterns at higher elevations with thick snow cover over patterns in flatter terrain with only limited snowfall. For example, our scheme does not explicitly include melt due to liquid precipitation. Another reason for this slight decrease in discharge performance at lower elevations could be related to our choice of regionally averaged DDFs since, in reality, these DDFs show smaller-scale variability in space (e.g. with aspect, albedo, elevation, land cover, and vegetation) (e.g. ). Such variability is not accounted for in our model since we focused on regionally valid parameterizations instead of local solutions. However, ignoring this variability could lead to biases in SWE representation in specific locations. More elaborate snow module formulations, such as parameterizations that include aspect (e.g. ) or radiation (e.g. ), could increase our ability to capture more detailed spatial melt patterns. Such approaches can become feasible now that higher spatial resolutions can resolve slopes and represent vegetation cover and land use in a more detailed way. However, they come at the cost of increased model complexity and a larger number of input variables. In any case, the slight decrease in discharge performance at lower elevations was reduced via the calibration of modelled SWE against a regional SWE reanalysis product, leaving us with a good overall discharge and SWE simulation performance (compare Fig. c and d). The improved discharge representation after calibration highlights the performance gains that can be achieved by including more regional data in LHMs. However, regional calibration is only possible due to the comparably high quality of observational and reanalysis data in Switzerland. Many other regions around the globe continue to face a lack of observations of water balance components , which challenges accurate regional calibration. Furthermore, the slightly reduced performance in SWE representation over Austria (which was excluded from calibration) compared to Switzerland already indicates that parameters can vary between regions (compare Fig. d, e, and f with Fig. a, b, and c). Highly detailed calibration in one region might give a false sense of accuracy when applying the setup outside of the calibration region. Regionally valid datasets must thus be chosen with care.

Adding a glacier module led to general improvements in discharge simulations in glacierized catchments in the Alps in catchments with a near-zero or positive value for WB (i.e. more observed discharge than is “expected”; see Fig. f and column “glacier module” in Table ). The effect of glaciers on general discharge performance further downstream remains limited (see Fig. e), although this effect might be larger during certain months or seasons . Aside from discharge, glacier mass balances and spatial patterns in elevation changes are also reasonably well represented (see Fig. ). Still, we note that glaciers can show significant biases in the mass balance (see Fig. g) and in the responses (see Fig. j). Differences in long-term responses compared to previous experiments into which ice dynamics were incorporated might be partly explained by the Δh parameterization, which ignores potential delays in glacier response to mass changes . Further biases in both mass balance estimates and responses are likely to be related to the relatively coarse spatial resolution of our model compared to the size of individual glaciers, which makes it more difficult to accurately describe patterns in melt or snow accumulation for small glaciers consisting of only a few grid cells. Melt representation of individual glaciers could be improved by resolving glaciers at a higher spatial resolution than the resolution simulated here, for example, by including elevation zones . Resolving glaciers at a higher spatial resolution could also lead to even more realistic glacier retreat since the Δh parameterization was originally designed for higher spatial resolutions and is likely to be less accurate at the coarser 30 arcsec model grid. Further gains in glacier representation could be realized by improving glacier accumulation estimates by continuing to develop the snow component or by improving spatial melt patterns by varying the glacier DDF in space (analogously to what we suggested for snow).

While the improvements in process representation generally lead to an increase in model performance for discharge, there are regions where the performance of discharge simulations decreases after implementing the structural and parametric changes (see Fig. d, e, and f). Our analysis shows that such performance decreases are common in catchments with negative values of WB, which can be related to issues with the meteorological forcing, glacier melt estimates, or the representation of water abstractions or (hydropower) reservoirs in the model. Indeed, the strongest negative performance changes after the introduction of snow and glaciers often occur in catchments with negative WB or regulation (see Fig. b and f; row “regulated catchments” in Table ). In the Alps, significant glacier or snowmelt occurs above hydropower reservoirs (e.g. in Switzerland, 4 % of all hydropower is related to glacier mass loss; ). Hydropower severely changes flow seasonality , essentially decoupling observed streamflow from snowmelt and ice melt. Accurate representation of river regulation and hydropower reservoirs is thus important, but LHMs appear to have difficulties with modelling streamflow in these regulated catchments (e.g. ). These issues might become even more apparent at a hyper-resolution because (a) small rivers are now represented and might be affected by reservoirs that are not represented in the model, and (b) reservoir schemes in LHMs were developed to represent regulation behaviour on a coarse grid (e.g. they mimic the combined effect of all reservoirs within a 50 km × 50 km grid cell) and likely need an update when moving to higher spatial resolutions . However, even if more reservoirs were introduced into the model, accurate discharge modelling in regulated systems remains challenging due to limited data on operation strategies and regulations . Thus, both improving the representation of human water management in models and collecting new data should remain an active area of research.

Besides model structure and parameterization, the meteorological reanalysis forcing has a major influence on model performance for discharge (see Fig. g, h, and i; column “meteorological forcing” in Table ). This implies that improvements in simulation performance that can potentially be achieved by structural changes are bound by the quality of the meteorological forcing dataset, especially in a non-calibrated model setup. Replacing the STANDARD meteorological forcing product with CERRA-CHELSA led to an overall increase in correlation with the reference dataset (compare Fig. a and c) but showed mixed performance changes for discharge over the model domain (compare Fig. g and i). This apparent inconsistency could be related to uncertainties affecting our precipitation evaluation and the LHM. Our precipitation evaluation is likely to be affected by uncertainties in the reference APGD. Although the APGD is based on a particularly dense network of direct observations, it still has uncertainties related to precipitation undercatch (for which it was not corrected) and in regions with less dense station coverage . Our evaluation of the precipitation products should therefore be interpreted as an indication of consistency with interpolated surface measurements but not necessarily as an indication of absolute performance. Collecting new, better, and more observations of meteorological variables over mountainous regions should thus remain a priority. The limited performance for discharge in small flashy catchments, as was pointed out before, might further prevent the direct translation from more accurate precipitation input into improved discharge performance. The ever-present uncertainty in forcing datasets and their evaluation also highlights the advantage of the calibration strategy we proposed and applied here, which focuses on processes like snow or glaciers rather than on discharge. Calibrated models can compensate for meteorological uncertainty and model deficiencies . Including the representation of processes such as snow or glaciers into the calibration procedure avoids some parameter equifinality and preserves internal relationships (e.g. ), which facilitates identifying where models need improvements. Finally, as presented in the Methods section, the STANDARD and CHELSA meteorological forcing datasets were derived using statistical or physical and/or heuristic downscaling, whereas the CERRA-CHELSA dataset is based on a dynamically generated regional reanalysis product. None of these methods used to create the hyper-resolution forcing datasets led to general improvements in simulation performance across the full domain. Regional differences between the datasets, observational uncertainty, and deficiencies in hydrological model structure did have a pronounced effect on model performance for discharge.

Our new model setup generally shows a better representation of discharge than the existing hyper-resolution version of PCR-GLOBWB 2.0 (see Fig. d, e, and f). Absolute discharge performance is generally satisfactory, especially at larger spatial scales and in natural rivers (see Fig. a, b, and c), and spatial patterns can be resolved in high detail (see Fig. a and b). Applications of such hyper-resolution LHMs should thus make use of these strengths by studying spatial patterns at larger spatial scales. Still, the model shows poor discharge simulation performance in some catchments, especially in smaller catchments at lower elevations and in rivers with a negative WB. Users should thus be aware of such limitations when looking at individual catchments. While the hyper-resolution simulations might not be completely accurate in certain places, they can still be valuable in helping to complete the spatial picture and in enabling the study of spatial variability . This is especially true in regions where no observational data are available. There, these LHMs provide a first-order estimate of river discharge without requiring local calibration.

In addition to discharge, LHMs also provide information on other variables such as SWE. Although there exist many local or national snow reanalysis products that are more accurate than our SWE simulations, semi-global products are often unrepresentative for mountain regions since their coarse resolution makes it difficult to represent complex terrain . Our analysis shows that our higher-resolution model setup outperforms coarser reanalysis products such as CERRA-Land and ERA5-Land, even without calibration (Fig. ), but does not quite reach the performance of national-scale reanalysis products such as the OSHD product (Fig. g). Hyper-resolution LHMs could thus provide additional SWE products that bridge the quality gap between national and global datasets.

How suitable is the model for climate impact assessments, a typical application of hydrological models? While many larger-scale climate change impact assessments still use temperature-index-based approaches, including studies focusing on snow (e.g. ) or glaciers (e.g. ), there is discussion about the robustness of degree-day approaches under climatic change . A short evaluation, which explicitly addressed model transferability, shows that model performance for discharge and SWE remains mostly consistent over the warmer and colder evaluation periods compared to the reference period, although the temperature changes over the study period are limited (Fig. S1 in the Supplement). Furthermore, we specifically aimed for increased robustness of parameters by focusing on SWE or glacier melt instead of streamflow during calibration. showed that the stability of hydrological model parameters such as DDFs across periods with different climate conditions can be increased by emphasizing snow representation during calibration. All of this suggests a reduced sensitivity of our model to increases in temperature. Still, the general caveats of the model remain applicable when extrapolating into the future. For example, the glacier module shows a realistic dynamic response to climate over larger regions but not necessarily for each individual glacier (Fig. ). Users should be mindful of this uncertainty in individual glacier responses to climate change. Over time, continued glacier retreat will reduce glacier melt contributions to discharge since, globally, most glaciers have either already surpassed their peak in glacier melt contributions or will do so in the coming decades . This will lead rivers to shift towards more snow-dominated regimes . Further into the future (i.e. near the end of the century), the exact response of glaciers will thus be less important.

In summary, the presented model setup adds value to regional-scale hydrological studies, focusing on general patterns and on larger rivers which are nearly natural to moderately regulated. Caution should be used when interested in individual, small, or heavily regulated rivers, where catchment-scale models likely show better performance. Still, in regions with limited data availability, large-scale model runs with a global parameterization might remain the best option available, even for such small rivers.