EO optical observations have proven highly accurate for snow monitoring e.g.,, yet their applicability is limited by sensor constraints and environmental factors . A primary challenge is the trade-off between spatial and temporal resolution. Low-resolution (LR) sensors, such as MODIS or VIIRS, provide daily estimates of snow cover fraction (SCF) but do not sample the process with the spatial detail required in complex alpine terrain . In contrast, high-resolution (HR) sensors, such as Sentinel-2, capture fine-scale spatial patterns but have infrequent revisit times, making them less suitable for tracking rapidly changing snow dynamics. Additionally, all optical sensors are affected by persistent cloud cover, which creates substantial data gaps–particularly in mountainous regions . To our knowledge, standard operational products rely on very basic gap-filling methods as multi-temporal filters, that do not consider the spatial context.

To address these limitations, we adopt the approach of , which integrates LR and HR optical satellite observations to produce a continuous, gap-free daily binary snow dataset (snow/snow-free) at HR (50 m). The method combines (i) LR daily SCF products (e.g., MODIS) and (ii) HR Sentinel-2 snow maps within an iterative gap-filling and downscaling framework, followed by a machine learning correction step. This framework relies on the assumption that inter-annual snow patterns are strongly influenced by local topography (elevation, slope, aspect) and meteorological conditions, resulting in recurrent spatial patterns of snow accumulation and ablation. By leveraging a long archive of HR acquisitions, the method reconstructs daily HR snow maps while correcting for known LR limitations, including errors from grain size variability, solar zenith angle, viewing geometry, and atmospheric effects. Details about the performance of the algorithm can be found in the original work.

The gap-filled HR binary snow maps are aggregated and resampled to the LISFLOOD model grid to derive daily pixel-wise SCF values, hereafter referred to as EO-SCF. Starting from HR products preserves spatial details, even after aggregation, which are crucial for accurate modeling and analysis. Specifically, using HR data enables better detection of small-scale snow patches and fractional snow cover – features often missed by LR products – thereby providing substantial improvements during the melting period, which is the main focus of this study.

A detailed description of the input datasets and a comparison with operational gap-filled products – which could provide users with faster and alternative access – is provided in Appendix . Evaluation against other existing snow cover products highlights the key strengths of this approach, notably the availability of a fully gap-filled daily HR dataset.

LISFLOOD is an open-source, spatially distributed hydrological model designed to simulate the complete water cycle in large river basins . Developed by the European Commission's Joint Research Centre (JRC) for the Copernicus Emergency Management Service (CEMS), it underpins both the European Flood Awareness System (EFAS) and the Global Flood Awareness System (GloFAS) , and supports the European and Global Drought Observatories (EDO and GDO) by providing key hydrological indicators (see https://drought.emergency.copernicus.eu/, last access: 23 January 2026). The current EFAS and EDO setup (v5.0) operates on a 1 arcmin grid covering the European Union, the European continent, and the Mediterranean coast, while GloFAS and GDO operate globally at 3 arc-minute resolution. The model includes multiple process modules, including surface runoff, infiltration, groundwater recharge, and snow dynamics. Human influences are also represented in the model. A water abstraction module simulates withdrawals from rivers and aquifers for irrigation, livestock, domestic, and industrial uses. Reservoirs are modelled using defined conservative, normal, and flood storage limits, along with corresponding discharge rules, thereby mimicking typical operational behavior within the hydrological routing scheme. LISFLOOD can be applied at multiple scales, from large river basins to global regions, supporting flood forecasting, water resource assessments, and analyses of water demand, river regulation, land-use changes, and climate change.

To compare LISFLOOD’s SWE output with EO-SCF, SWE must be converted into snow cover fraction (SCF). This conversion requires a parametrization that accounts for variations in snow depth and density within a pixel. Due to substantial intra-pixel variability, the relationship is influenced by factors such as topography and land cover . Several parameterizations have been proposed in the literature e.g.,. In this study, we adopt two approaches that balance model complexity with data availability. The first, proposed by and implemented in the Community Land Model (CLM), is relatively sophisticated. The second, introduced by , is a simpler empirical method requiring fewer inputs but has been shown to provide consistent results. For brevity, and given their similar performance, details and results from the second approach are provided in Appendix .

differentiates the accumulation and ablation periods. For the accumulation, the SCF after a precipitation event can be interpreted probabilistically. In a simple linear formulation, the probability of a pixel to be snow covered or the SCF is given by the following formula: 6SCF=min⁡(1,kaccum⋅SWE), where kaccum [1 mm⁻¹] is a scaling parameter that relates SWE to SCF. Accordingly, the probability that a pixel remains snow-free is 1-SCF. For multiple independent snowfall events, the cumulative probability that a pixel remains snow-free is the product of the individual probabilities. Therefore, after N+1 events, the cumulative SCF is 7SCFN+1=1-∏i=0N(1-SCFi).

A more refined, probabilistic formulation for accumulation uses a hyperbolic tangent function, which ensures that SCF asymptotically approaches 1 as SWE increases, to update SCF after each event: 8SCFn+1=1-[(1-tanh⁡(kaccumΔSWE))(1-SCFn)],

where ΔSWE is the amount of new SWE in mm. Although many state-of-the-art models assume a constant value for kaccum equal to 0.1, it likely varies pixel-wise due to topographic differences. The parameter kaccum can be estimated by measuring SCF and ΔSWE during precipitation events over an initially snow-free area, as suggested by . To identify such conditions, we selected the first day on which both EO-SCF observations and LISFLOOD SWE are greater than 0 (snow presence). Equation () is then inverted to obtain an estimated value of kaccum. The parameter is constrained to a maximum of 0.5, consistent with what reported in the original study. For pixels that never exhibit snow, the default value is retained. This procedure was repeated for all calibration seasons, and the final pixel-wise value of kaccum was taken as the average over time. The obtained parameters are reported in Appendix  in Fig. .

For the melting period, relate SCF to the dimensionless snow water equivalent scaled by the maximum SWE (SWEmax) through the following empirical equation: 9SCF=1-[1πarccos⁡2SWESWEmax-1]Nmelt where Nmelt is a parameter that depends on the topographic variability within the grid cell. It is calculated as follows: 10Nmelt=200max⁡(10,σtopo) with σtopo being standard deviation of the elevation within a grid cell calculated from the MERIT DEM with a spatial resolution of 90 m . When an accumulation follows a melting event, SWEmax needs to be updated to re-define a consistent depletion curve. This is done by inverting Eq. () and using the following equation: 11SWEmax=2SWEcos⁡(π(1-SCF)1/Nmelt)+1 Note that Eq. () would be circular with respect to Eq. (); however, after a new snowfall event, SWEmax is updated such that the previous SCF (prior to snowfall) is preserved along the depletion curve. In other words, SWEmax refers to SCF(t−1). For more details, refer to the original code at https://github.com/ESCOMP/CTSM/blob/master/src/biogeophys/SnowCoverFractionSwensonLawrence2012Mod.F90 (last access: 23 January 2026), which implements the snow cover fraction parameterization by .

The snowmelt coefficient is estimated by solving an optimization problem that minimizes the error between L-SCF and EO-SCF. This procedure aims to align model outputs with observations, thereby increasing the realism of the simulation results. We remark that L-SCF is derived from modelled SWE using Eq. () and the parametrization defined in Eqs. () and (), with Cm treated as a free parameter in this optimization phase. Specifically, we minimize the mean squared error (MSE) over time to identify the optimal coefficient that produces the smallest discrepancy. Accordingly, the following minimization problem is solved for each pixel to obtain the EO-based snowmelt coefficient: 12EO-Cm=arg⁡min⁡Cm∑t∈ThyL-SCF(t,Cm)-EO-SCF(t)2

The optimization is performed on a pixel-wise basis over all time steps within a given hydrological year Thy to obtain a single, time-invariant coefficient. To solve this optimization problem, we utilize the L-BFGS-B algorithm – a limited memory quasi-Newton, gradient-based optimization algorithm – provided in the SciPy library . Estimated values are constrained between 0.5 and 10, as values outside this range lack physical meaning and may result from errors, such as having too few days to perform a reliable optimization (e.g., ephemeral snow at low altitudes). Specifically, according to Eq. (), lower values of Cm would result in negative melting, which is not physically realistic. Subsequently, a mean value for each pixel is calculated using data from five seasons (2017/22) while one (2022/23) is used exclusively for a further independent assessment of the results. For pixels where the coefficient could not be estimated – such as those without snow during the year or where pixels were masked out due to the presence of water bodies – L-Cm is retained.

The obtained EO-Cm is first evaluated qualitatively against L-Cm. A correlation analysis is also performed against selected topographic, geographic, and land cover features. For a quantitative assessment, we compare the resulting snow-covered area (SCA), defined as the percentage of the basin covered by snow. To do this, SWE is simulated using the LISFLOOD snow module (Eq. ) with both L-Cm and EO-Cm. The simulated SWE is then converted to SCF (see Sect. ) and aggregated to the basin level to obtain SCA. The SCA obtained using L-Cm is denoted as L-SCA L-Cm, while that obtained using EO-Cm is denoted as L-SCA EO-Cm. These model-based estimates are then compared against EO-SCA, derived from EO-SCF, which serves as the benchmark. The results are reported in Sect. .

We also evaluate improvements at the pixel scale by computing metrics for SCF, using EO-SCF as the benchmark. Specifically, we calculate bias, root mean squared error (RMSE), and correlation between L-SCF (using both L-Cm and EO-Cm) and EO-SCF. Metrics are first computed per pixel over time and then averaged spatially across the basin. To focus on the model’s ability to detect snow, we exclude pixels where both modelled and observed SCF are zero (i.e., snow-free conditions). Including such pixels would artificially inflate performance metrics, particularly during periods when large portions of the basin are snow-free, thereby biasing the results. An analysis of the correlation with the errors versus physiographic and climatic features is also performed, together with an evaluation of the seasonal error trends. SCA results are reported in Sect. .

A SWE intercomparison is performed for basins where a reference dataset is available. To complete the assessment, we evaluate how the hydrological response changes when using EO-Cm. Specifically, we replace L-Cm with EO-Cm and run the LISFLOOD model from 1990 to 2023, including two warm-up years (1990–1991). The remaining 13 model parameters are kept unchanged, allowing us to isolate the impact of Cm on river discharge and assess whether the current EFAS5 setup can realistically capture snow accumulation and melt dynamics in the selected catchments. These results are reported in Sect. .

We compare the mean annual hydrographs of the two simulations for periods with observations to check if changes in SWE and snowmelt are reflected in the simulated discharge. In addition, we calculate the Kling-Gupta Efficiency (KGE) both over the entire period and aggregated by month, using daily simulated and observed river discharge data. This allows us to evaluate the overall performance of the simulations as well as their performance across different sections of the annual hydrograph, with particular attention to potential increases or decreases in performance during the snowmelt season. Finally we evaluate the spatial similarity of river discharge at pixel level to identify main difference between EO-Cm and L-Cm simulations upstream of the discharge station. The analysis focuses on upstream catchment areas larger than 100 km² to exclude very small headwater sections that may introduce noise in spatial comparisons. To do so, we extract daily river discharge time series from both model outputs and apply min-max normalization for each grid cell, standardizing discharge values across different magnitudes and catchments. We then compute the Euclidean distance between the normalized discharge values of the two models at each grid cell, quantifying the absolute dissimilarity in flow dynamics over time. The Normalized Euclidean Distance (NED) values range from 0 to infinity, where a value of 0 indicates that the two time series are identical or highly similar, and larger values signify greater dissimilarity between the time series. These results are reported in Sect. .

In this section, we present the snowmelt coefficients (Cm) obtained by calibrating against EO data (EO-Cm – see Fig. ), as described in Sect. , and compare them qualitatively with the coefficients obtained from the EFAS5 calibration (L-Cm – see Fig. ), as described in Sect. . L-Cm are lumped at the sub-catchment scale within the major basin, with sub-catchments defined according to the availability of streamflow data for calibration, as explained in Sect. . The corresponding histograms are also shown in the figures. Missing values (white areas) indicate masked regions, such as water bodies or pixels that did not experience snow during the analyzed period.

Notable differences are observed, as L-Cm represents a lumped coefficient, whereas EO-Cm is spatially distributed. EO-Cm exhibits very high values in basins or areas with ephemeral snow. Additionally, the parameter range in the standard LISFLOOD calibration (2.5–6.5) is narrower than in the EO-based calibration. The wider range in EO-Cm allows us to explore the potential of the new calibration to compensate for erroneous precipitation inputs and better represent the snowmelt phase (see next section). It is interesting to note that the coefficient EO-Cm assumes a value of 10 in those basins, or parts of basins, that exhibit ephemeral or short snow-cover periods, such as the Laborec and Mörrumsån, as well as the lower Adige. This value likely indicates a non-ideal optimization, possibly due to fluctuations in snow cover. We also evaluated the stability of this coefficient across different seasons. For brevity, results are reported only for the Adige River basin in Fig. . The figure shows that the coefficient is quite stable over years, confirming our choice of considering a temporal average.

To explore whether the spatialized snowmelt coefficient calibrated at the pixel scale captures meaningful patterns, we conducted a correlation analysis with topographic, geographic, and land cover features. The motivation was that such relationships could eventually support parameterizing snowmelt based on spatial characteristics alone, thereby reducing reliance on EO data, which are often labor-intensive to process. Table  reports the Pearson correlation coefficients between EO-Cm and several variables: elevation (DEM), elevation variability (DEM σ), slope, forest cover, mean snow cover duration (SCD), as well as mean precipitation (P) and temperature (T), averaged over the six analyzed hydrological seasons. However, the results indicate that no strong correlations emerge from this analysis. The most influential feature appears to be elevation, which exhibits a negative correlation in the Adige, Arve, Salzach, and Guadalfeo river basins. This indicates that lower values of EO-Cm are estimated for higher elevations, meaning snow persists longer at higher elevations, as expected. For the Laborec and Umeälven basins, which are relatively flat, the correlation with elevation becomes positive. The correlation with mean snow cover duration also aligns with the findings for elevation, confirming that snow lasts longer at higher elevations. The observed correlation, together with the temporal stability of the snowmelt coefficient, may also suggest a systematic elevation-dependent bias in the precipitation forcing.

To assess how well the standard multi-parameter calibration of streamflow captures snow dynamics and to quantify the improvements of an independently calibrated Cm, we conducted a comparative evaluation of the snow cover area (SCA). Specifically, we ran the LISFLOOD snow module using two different snowmelt coefficients: the standard, lumped value (L-Cm) and the new spatially-distributed value (EO-Cm). This produced two distinct model-derived SCA estimates (L-SCA), which were then compared against the observed SCA (EO-SCA). Note that, since the conversion from SWE to SCA is performed through a parameterized approach that relies on the parameter kaccum (see Sect. ), which is derived from EO-based SCF, the two SCA estimates compared in this evaluation are not fully independent variables.

In Fig. , the SCA trends with the two Cm are reported. These results offer a quick overview of the performances at basin level. Furthermore, to assess improvements at the pixel-scale we report the metrics for SCF, using EO-SCF as the benchmark. These are computed per pixel along the temporal dimension and then averaged spatially. To specifically assess the model's ability to detect snow cover, we exclude pixels where both the target and reference SCF are 0 (snow-free). While including these pixels would lead to better performances, the results might be biased especially when a large portion of the basin is snow-free.

In Table , we present the average metrics calculated across all analysed hydrological years for brevity. Similar metrics with a different SCF parametrization are reported in Table . Detailed metrics for each individual season are provided in Fig. .

Due to variations in the size and location of these basins, the SCA varies significantly and exhibits seasonal fluctuations influenced by how wet or dry the year was. For example, while most basins reach nearly 100 % snow cover at least briefly, others, such as Gállego and Guadalfeo, show very little snow cover. A generally good agreement can be observed between EO and LISFLOOD SCA, even with the EFAS5 calibration (L-Cm). However, important differences are evident in certain basins, including the Adige, Laborec, and Mörrumsån. In the Adige river basin, the LISFLOOD model shows a consistently larger SCA compared to satellite observations. These discrepancies, particularly noticeable during accumulation phases, such as in the 2020/21 season, may stem from an overestimation of snowfall and, consequently, precipitation fields over the basin, or from an inaccurate partition of solid/liquid precipitation. Despite these differences in magnitude, the timing of SCA variations appears to align well between the two datasets. Interestingly, the Arve basin shows that SCA remains above zero even during summer, as some pixels are either persistently snow-covered or correspond to glacierized areas. However, since LISFLOOD does not explicitly represent glaciers, these pixels are not simulated as permanently snow-covered in the model, thus resulting in lower SCA. On the other hand, the differences observed in the Laborec and Mörrumsån basins are more likely attributable to challenges in snow detection by the satellite. These include prolonged periods of missing acquisitions due to cloud cover, combined with ephemeral snowfalls in flat areas, as is the case in Mörrumsån. Additionally, significant forest coverage, particularly in the Laborec basin, further complicates accurate snow detection.

Overall, both the standard EFAS5 calibration and the EO-based optimization of the snow module yield satisfactory results at the basin scale. However, important differences are highlighted when considering SCF (Table ). Before dedicated EO-based calibration, the agreement is acceptable, with the highest bias observed at approximately 20 % for the Laborec basin but with generally moderate correlation. The highest root mean square error (RMSE) values are observed for Mörrumsån and Laborec, due to the reasons previously discussed. Additionally, Gállego also exhibits notable discrepancies, likely because most pixels remain snow-covered for only short periods complicating the analysis. The computed metrics indicate that, in general, the optimization (EO-Cm) improves bias, RMSE, and correlation, except for a slight increase in bias for Gállego. As shown in Fig. , these trends generally persist across individual seasons, with a few exceptions. Additionally, for the year not used in coefficient calibration (2022/23), EO-Cm still demonstrates an overall improvement in performance especially in reducing the bias and RMSE and increasing the correlation. However, the remaining errors stem from the fact that the new coefficient enhances SCF during the melting season only, while errors persist during the accumulation season. Moreover, the figure shows that the Alpine basins (Adige, Alpenrhein, Arve, Salzach) and especially the Umeälven basin, exhibit more stable behavior across seasons, with less scattered performance metrics. In contrast, the Laborec, Mörrumsån, and Gállego basins and partly the Guadalfeo display greater interannual scatter in the metrics, reflecting a more heterogeneous and ephemeral SCA.

We further investigate the dependency of model performance on catchment characteristics. Performance is evaluated in terms of BIAS (Fig. ) and RMSE (Fig. ) of SCF derived from EO-Cm, and these metrics are compared against a set of physiographic features (mean elevation, forest coverage, and slope) and climatic features (mean precipitation, mean temperature, and snowfall). We also distinguish between glacierized and non-glacierized basins.

The analysis shows that errors tend to be larger in lower-elevation and flatter catchments, with a similar increase associated with higher forest coverage. For the climatic features, we find an inverse relationship with RMSE: basins characterized by higher precipitation and snowfall generally exhibit lower errors. This aligns with expectations, as lower precipitation – particularly in the form of snowfall – typically results in more ephemeral snow cover, increasing the prevalence of fractional snow-covered areas, which are more prone to both detection and modeling errors. Glacierized catchments do not, on average, display substantially different performance relative to non-glacierized ones. The Umeälven basin always shows markedly lower RMSE. This is likely due to its prolonged and near-complete seasonal snow cover, which reduces snow-cover variability and associated modeling errors.

It is also interesting to analyse the monthly evolution of bias and RMSE, as shown in Fig. . Overall, model performance tends to deteriorate during the summer months across all basins, reflecting the snow depletion phase. This reduction in performance is partly driven by the smaller number of snow-covered pixels considered (snow-free pixels are excluded), which are often characterized by fractional snow cover and thus more prone to errors. Nevertheless, despite this seasonal decline, we generally observe an improvement when applying EO-Cm.

To fully assess the performance of the snow module before and after snow melt coefficient recalibration, spatialized SWE maps would ideally serve as a reference. However, a comprehensive analysis is not feasible due to significant challenges in data availability and usability. Given the spatial resolution of a LISFLOOD pixel, in-situ measurements are unreliable proxies, as they fail to capture intra-pixel variability – particularly in basins with complex topography. Additionally, SWE data is often missing, and when only snow height measurements are available, converting them to SWE requires assumptions about snow density, introducing further uncertainty. Remote sensing products also have limitations. The accuracy of SWE derived from microwave sensors is affected by factors such as vegetation cover, topography, and snow type . Furthermore, many available datasets, including those from the Copernicus Land Monitoring Service, have coarse spatial resolutions and often exclude mountain areas, further restricting their applicability . Therefore, we propose an intercomparison with SWE estimates from two different models in the Adige and in a subchatchment of the Alpenrhein, the Dischma Valley. For the Adige, we consider the IT-SNOW reanalysis dataset (see Fig. ). For Dischma, we compare our results with the Swiss Operational Snow-Hydrological (OSHD) model system, available for the first five seasons (see Fig. ). The metrics are reported in Table .

The results confirm a generally good agreement but also highlight LISFLOOD’s tendency to underestimate SWE compared to the other models in both basins. These differences are most likely linked to discrepancies in precipitation input data. Notably, IT-SNOW assimilates snow height measurements, which can enhance the accuracy of accumulation estimates. Discrepancies in the input forcings might explain the general low correlation for the Adige (0.44/0.43), while for Dischma we obtain a very high correlation (0.96) both previous and after Cm replacement. Notably, especially for the Dischma Valley the depletion curve is better captured when using EO-Cm. However, for both basins we obtain comparable or slightly improved metrics aligning with expectations based on the SCA comparison. Interestingly, we did not observe worse performance in representing different conditions, such as snow drought events, which appear to be reasonably well reproduced by both the standard and EO-Cm. This is particularly evident in the Adige basin during the 2021/22 and 2022/23 seasons. It should be noted that this analysis is not exhaustive, as it relies solely on intercomparison with other models with inherent limitations stemming from model parameterizations and the quality of forcing inputs. Therefore, the analysis lacks validation against reference data.

In this section, we evaluate the hydrological response to post-replacement of the snowmelt coefficient (EO-Cm), with results compared against benchmark simulations using the standard L-Cm. Note that, as explained in Sect. , all the other parameters are kept unchanged, following EFAS5. Results are shown as daily averages calculated only for periods with available observations (Fig. ). Reported variables include SWE, snowmelt, and discharge, all expressed in mm d⁻¹. Dashed lines in the figure denote the 10th and 90th percentiles of the time series, indicating variability in the modelled hydrological response.

In most catchments, the new EO-Cm affects the timing of both accumulation and melting phases, which in some cases also influences the timing and magnitude of runoff. In the Adige basin, the behavior of the snow cover is similar in both runs. The benchmark has a higher peak in SWE in March, but a higher depletion as well from April onward. This is reflected in the snowmelt, where the EO-run shows a slight shift in the timing of snowmelt, with lower snowmelt between March and May and higher in June, compared to EFAS5. This shift has minimal impact on the generation of discharge. Notably, the snowmelt coefficient assigned through the regionalization approach aligns closely with that derived from the EO-SCA.

In the Alpenrhein, Gállego and Salzach basins, the new EO-runs show an increase in snow accumulation, driven by reduced snowmelt before the peak. This results in a higher magnitude of snowmelt and a shift in its timing, with the peak occurring later compared to the benchmark. The effect of the newly calibrated snowmelt coefficient extends to discharge, leading to increased discharge during the snowmelt phase and reduced discharge during the snow accumulation phase. This shift does not affect the daily average discharge in the Gállego basin, but it does influence the Alpenrhein River, where it improves the agreement with observations during July and August. In case of Salzach basin the increased discharge between June and July is overestimating the observed discharge, with EFAS5 having a better match with observations.

The EO-run of the Arve basin shows a positive bias in SWE, which is only partially reflected in snowmelt since more SWE does not melt during the summer period. Snowmelt is significantly higher between June and August, and matching better observed river discharge for the same period.

In the Guadalfeo basin, a slight decrease in the snowmelt peak is observed, occurring in March, with no impact on total runoff. The model heavily underestimates river discharge compared to observations. This poor performance may be partly due to the regionalization of parameter assignment and/or the inclusion of the Rules reservoir in the model for the entire simulation period (1992–2023), despite it was opened in 2004. The observed mean daily values account for 13 years when the reservoir did not yet exist, as it was commissioned in 2004, whereas the simulation assumes the reservoir has been present throughout the entire period. In the Laborec basin, snow cover is reduced and snowmelt occurs earlier, peaking in February. Extreme values are reduced, with the 90th percentile decreasing from 60 to 52 mm for snow cover and from 100 to 80 mm per month for snowmelt. The discharge increases in January and February but decreases in March. While the flow regime is broadly similar to observations, notable differences remain with observed runoff being 30 % higher than both simulations. In the Swedish catchments Mörrumsån and Umeälven, no significant differences were found in mean daily values. The only noticeable deviation occurred in the 90th percentile of SWE in Mörrumsån, reflecting a period of high snow accumulation, followed by a sharp drop in SWE and increased snowmelt, likely driven by a temperature spike.

The performance statistic KGE, for the period with available observations, is reported in Table . Since the EO-Cm model was not recalibrated, a general decline in KGE and related metrics was expected. However, in most cases, the differences in metrics are not significant, and in some instances, there are improvements, which are highlighted in bold in Table . Across all catchments, the bias component remains virtually identical between the EO-Cm and L-Cm experiments, indicating that incorporating EO-Cm does not systematically increase or decrease the total volume of simulated discharge. A slight degradation is observed in correlation, particularly in the Laborec basin, suggesting that EO-Cm somewhat reduces the accuracy in capturing the timing of peak flows. This is particularly important in the operational context of EFAS5, where changes in correlation directly impact the system’s ability to anticipate or delay peak flows. The impact on variability is more heterogeneous: improvements are observed in the Adige, Guadalfeo, and Umeälven basins, while reductions occur in the Salzach and Alpenrhein. This pattern implies that EO-Cm can have a positive effect on representing interannual flow fluctuations in some basins, but not universally. Overall, model performance increases in the Adige and Mörrumsån basins following the introduction of EO-Cm, whereas noticeable declines occur in the Alpenrhein, Laborec, and Salzach.

Figure  presents monthly Kling-Gupta Efficiency (KGE) values for both EO-Cm (cyan) and L-Cm (orange) LISFLOOD runs. In most basins, such as Adige, Alpenrhein, Arve, Mörrumsån, and Salzach, both experiments achieve consistently high KGE scores (generally above 0.5), indicating robust model performance throughout the year. Conversely, in basins such as Gállego, Guadalfeo, and Umeälven, the simulations exhibit lower and more variable KGE scores, with several months showing negative values; this suggests poor performance and notable mismatches between simulated and observed river flows. The Gállego and the Guadalfeo basins in particular exhibit poor KGE in most months, which is reflected in their negative KGE in Table ; the EO-Cm did not yield significant improvements/worsening in these catchments. Performance differences between the two experiments are most pronounced in specific seasons. The EO-Cm model run outperforms the L-Cm run in the melting season of Arve and Alpenhrein, as well as in the snow accumulation season of the Adige and Mörrumsån basins. A decreased performance is observed for most months in the Salzach basin, while Laborec shows mixed results, with slight improvement in December and January and a decline in March and April. As stated in Sect. , the comparison is between a calibrated LISFLOOD run (L-Cm) and an uncalibrated one (EO-Cm). A deterioration in the performance of the EO-Cm simulation is therefore expected, since the remaining 13 parameters were not subject to calibration. Conversely, any improvement in KGE, its three components, or daily average performance would indicate that a calibrated EO-Cm configuration of LISFLOOD could potentially yield superior results. Moreover, because the current calibration strategy relies on a single parameter to represent snow processes, it is likely that the calibration compensates for structural or process-level deficiencies by adjusting other parameters, potentially masking the true contribution of the snow-related parameter.

Figure  presents the NED for each catchment. Darker colors indicate river sections where the two models produce similar daily discharge (low NED), while lighter colors highlight areas of stronger divergence.

The spatially heterogeneous EO-Cm shows a marked local impact in some river reaches with small upstream areas, where differences between model outputs are more evident. This influence decreases progressively downstream as localized effects are smoothed along the flow path. By the time discharge reaches the calibration points, typically located further downstream, the impact becomes negligible, as the calibration process compensates for or overrides local parameter variations. The Gállego and Mörrumsån basins exhibit the least effect from the EO-Cm implementation, with maximum NED values reaching only 1. Laborec and Umeälven follow, displaying peak NED values up to 3. In the remaining basins, most river reaches have NED values below 4, although some specific sections show higher localized NED values.

Reservoirs generally have minimal effect on the differences between the two simulated discharges across most basins. However, in the Gállego and Adige basins, distinct impacts are observed at the outflow of a single reservoir in each basin: in Adige, this occurs at the first upstream reservoir in the north-west area; in the Gállego, at the second reservoir moving downstream from the headwaters. In both cases, the grid cell immediately downstream of the reservoir shows greater differences than the grid cell directly upstream.