the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Jul 2026
Modeling 21st century snow dynamics in Switzerland using high-resolution Climate CH2025 scenarios
Abstract. Snow is a key component of Alpine landscapes, providing numerous ecosystem and economic services that include hydropower production, winter tourism, groundwater recharge, and regulation of stream temperatures, with significant implications for aquatic ecosystems. In a warming climate, expected increases in winter precipitation do not necessarily lead to greater snow accumulation, as rising air temperatures shift precipitation from snowfall to rainfall and reduce the persistence of snow on the ground. This creates a need for future snow projections at locally relevant spatial scales to support adaptation for snow-dependent sectors.
Here, we present daily projections of snow water equivalent (SWE) for Switzerland at 1x1 km² resolution, based on the Climate CH2025 scenarios, which downscaled and bias-adjusted meteorological forcings from an ensemble of EURO-CORDEX regional climate models. SWE was simulated using a spatially distributed temperature-index snow model for 12 climate model chains, selected for their ability to accurately represent atmospheric forcings for modeling snow cover dynamics across Switzerland over the historical reference period (1991–2020). To reduce biases associated with the simplified degree-day-based snowmelt representation used here, simulated SWE was quantile mapped toward SPASS-CLQM, a new gridded climatological snow reference dataset for Switzerland.
We found that the univariate quantile mapping of meteorological forcings in Climate CH2025 models reduced precipitation under sub-freezing conditions, resulting in too little simulated snowfall in many model chains, which then propagated nearly linearly into biases in SWE. Simulated SWE in the final set of 12 models was additionally quantile mapped toward SPASS-CLQM, which substantially reduced SWE biases across elevation bands and was then applied to future simulations. By the end of century (2069–2099), projected SWE showed widespread declines across Switzerland relative to 1991–2020. Percentage September–May mean SWE losses exceeded 80 % below 1000 m a.s.l. for the highest emission pathway (RCP8.5), where snow became increasingly rare. At intermediate elevations (1000–2000 m a.s.l.), mean SWE declines of 50–90 % are projected, with the snowpack becoming increasingly ephemeral. Above 2000 m a.s.l., mean SWE reductions ranged from 20 to 70 %, indicating that even high-elevation seasonal snow storage would reduce substantially. These projections provide a spatially detailed basis for assessing future changes in Alpine snow resources and for informing the management of snow-dependent hydrological, ecological, and economic systems in a rapidly warming climate.
- Preprint
(10773 KB) - Metadata XML
-
Supplement
(2705 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-3478', Anonymous Referee #1, 16 Jul 2026
-
AC1: 'Reply on RC1', Harsh Beria, 10 Sep 2026
Reviewer 1
General Assessment:
This study evaluates simulations of snowfall fraction and snow water equivalent from 26 Climate CH2025 models across Switzerland over the period 1991–2020, and projects the variations in SWE during 2035-2099 for the RCP8.5 scenario. The overall framework is clear, and the results are generally robust. However, there is a lack of detailed explanations regarding research methods and data. Some suggestions require the authors to give further consideration.
Response: Thank you for the assessment. We agree that the methods were lacking in sufficient details. We propose to expand descriptions of precipitation-phase partitioning, snow modeling, description of quantile-mapping applied, and to include a more detailed workflow figure. We will also include a new section within Methodology describing the reference datasets used in the manuscript.
Major Comments:
Although this study evaluated the ability of each model chain to simulate snowfall and SWE, it did not provide a detailed description of the reference datasets used. Since the selected reference datasets are considered reliable, more comprehensive information about the data would be necessary. However, this study offered no further explanation on this matter. Please add a description of this section.
Response: In the revised manuscript, we will add a subsection “Description of reference datasets” within the “Study area and methods” section, describing the reference datasets used in this study (RhiresD, TabsD, SPASS-CL, SPASS-CLQM).
For each model chain, annual snowfall was derived from the bias-adjusted and downscaled daily precipitation and air temperature fields. Is the bias in snowfall produced by each model chain related to the biases in temperature and precipitation simulated by the models themselves? It is suggested to compare the simulated temperature and precipitation with the observed values.
Response: The negative snowfall bias is found after univariate quantile-mapping of precipitation and air temperature, rather than being inherited directly from the raw model output. To illustrate this, Figure 4 compares the precipitation and air temperature relationship in (1) raw regional climate model (RCM) output, (2) the bias-adjusted forcings, and (3) the RhiresD/TabsD reference. Daily precipitation averaged across Switzerland within 1°C temperature bins from −20°C to 30°C shows that the bias-adjusted forcings systematically contain too little precipitation at temperatures below 0°C. In contrast, the raw RCM outputs generally contain too much precipitation below 0°C.
Although snowfall is not calculated using a fixed temperature threshold (precipitation-phase partitioning will be described in more details in the revised manuscript), Figure 4 indicates that snowfall underestimation in the bias-adjusted model chains is primarily associated with insufficient precipitation under freezing conditions (compare red vs blue lines in Figure 4). The resulting snowfall biases propagate into simulated SWE, resulting in SWE underestimation for most model chains (Figure 6).
When determining snowfall, is the amount of snowfall determined based on the relationships between temperature and precipitation thresholds, or does the model already distinguish between rainfall and snowfall? If the former, please specify the criteria used for determination.
Response: Snowfall is computed from daily precipitation and air temperature and is not taken from a native RCM snowfall field. Snowfall (Ps) is calculated with total precipitation (P; mm/day) and air temperature (T; °C) using the following equation:
P_s=P/(1+e^((T-T_pbase)/m_p ) ),
where Tpbase (°C) is the temperature threshold below which precipitation falls as snow, set to 1°C, mp (°C) determines the temperature range for mixed precipitation, set to 1.24°C.
This was only briefly explained in L184-186 of the original manuscript: “Snowfall was computed using a continuous, temperature-dependent precipitation partitioning routine, similar to the approach used in OSHD (Magnusson et al., 2014).”
We will expand this description in Subsection 2.3 of the revised manuscript.
Minor Comments:
Fig.1: The legend should represent snowfall fraction rather than snow fraction.
Thanks for pointing it out, we will change the legend title to “Snowfall fraction” in the revised manuscript.
Section 2.2 Snow modeling: The simulation accuracy of SWE requires a detailed description.
Response: We will expand Section 2.2 to describe the model configuration, add a new subsection “Description of reference datasets”, describing the relevant SWE reference datasets used in this study (SPASS-CL, SPASS-CLQM). The evaluation subsection “2.3 Model sub-selection for future snow simulations and bias adjustment of SWE” will define the September-May mean and maximum SWE metrics and the bias calculations.
The comparison of modeled SWE against SPASS-CL is to ensure consistency with an observation-based implementation of the snow model for the 1991-2020 reference period. This will be further clarified in the revised manuscript.
Line 201: Need a detailed introduction to the SPASS-CLQM dataset
Response: We will add the subsection “Description of reference datasets” describing the SPASS-CLQM dataset, and explain how it differs from SPASS-CL.
Lines 295-296: This study selected 12 model chains to assess future SWE changes based on the criterion that SWE bias should be within 10%. What is the rationale behind choosing 10% as the threshold? Could this involve subjective judgment? Could additional quantitative criteria (such as standard deviation ranges) be introduced to enhance the objectivity and rationality of this decision?
Response: We agree that choosing a ±10% threshold of mean SWE bias is subjective, as any other threshold would be. Nevertheless, the ±10% threshold is justified by the fact that uncertainty of automatic SWE measurements are ~10% based on the WMO-No8 report. The key objective here is to choose models which show acceptable snow simulations for the 1991-2020 period, which is indeed the case (Figure 6). We will cite the WMO-No8 report in the revised manuscript.
Note that we also evaluate the effect of the ensemble sub-selection on the range of simulated winter temperature and precipitation signals in order to ensure a decent coverage of the spread of the full ensemble (see Figure 8).
The study reports simulation results for different elevation intervals but does not include a DEM map of the study area. It is recommended to add an elevation distribution map to clearly illustrate the spatial distribution across different elevations.
Response: We will add a DEM map for Switzerland as a second subplot in Figure 1.
Fig. 4: Place the legend on the first sub-figure.
Response: We will place the legend on the first subplot of Figure 4.
Why consider only the future changes in SWE under the RCP8.5 scenario?
Response: Restricting the ensemble to RCP8.5 was primarily a computational choice, and we will make this more explicit in the abstract and conclusions.
A key objective of this work is to establish the SWE modeling workflow within the framework of the Climate CH2025 scenarios, which express future changes in terms of Global Warming Levels (GWLs). In a subsequent study, the transient SWE simulations presented here will be mapped to GWL1.5°C, GWL2°C, GWL3°C, corresponding to 1.5, 2, and 3 °C global warming, respectively.
RCP8.5 was selected because it ensures that all RCMs in the ensemble reach 3°C of global warming by 2100, as several simulations under RCP2.6 and RCP4.5 do not. Given the computational effort that would have been required to run the high-resolution snow model for all RCMs and for all available emission pathways, RCP8.5 therefore provided a pragmatic way to retain all three target GWLs across the full model ensemble and maintain consistency with the Climate CH2025 framework. Although RCP8.5 provides an improbable emission pathway, it is important for future risk assessment and planning.
Importantly, our objective was also to preserve the spread in projected winter temperature and precipitation changes across climate models, which in our current modeling setup is indeed the case (Figure 8).
Why were the periods 2035–2065 and 2069–2099 selected for future assessment, and what is the basis for this division?
Response: The intended purpose of the two windows was to contrast changes in SWE around mid-century (near future, with immediate practical relevance) with conditions near the end of the century. Therefore, we included two windows, one centered around mid-century i.e. 2050 (2035-2065), and the other including simulations until the end of century 2100 (2069-2099).
Citation: https://doi.org/10.5194/egusphere-2026-3478-AC1
-
AC1: 'Reply on RC1', Harsh Beria, 10 Sep 2026
-
RC2: 'Comment on egusphere-2026-3478', Anonymous Referee #2, 20 Jul 2026
The authors provide a useful and well-organized study that fills a genuine gap in the climate impact field across Switzerland, by providing a 1x1 SWE (snow water equivalent) dataset based on the high-end emission scenario. While the study could be considered for publication, in its current form I recommend major revision. Several methodological choices are not explained, backed or justified in enough detail (e.g. the quantile mapping approach used, the GWL block-time-shift method, the degree-day snow model itself), which makes it hard to judge the results in their proper context and would make it difficult/ impossible to reproduce the study (FAIR). In addition, some of the figures could benefit from some adaptations.
Major comments
- Model exclusion (IPSL-WRF) contradicts the historical evaluation. Excluding IPSL-WRF because its future climate signal looks like an outlier in summer precipitation, after selecting models based on historical performance, seems like (and risks) narrowing the projected spread on purpose!? Please justify this with an independent criterion (e.g. from the CH2025 report itself or other studies flagging IPSL-WRF!?), or show results with/without this model to test sensitivity?
- SWE quantile-mapping validation (Section 3.5): stating that the post-QM agreement with SPASS-CLQM (Fig. 9) shows the correction is "reliable" overstates imho what this check can really reliably show. QM guarantees close agreement with the calibration target over the fitting period by construction, regardless of the underlying model's actual skill I would argue and thus this might not be an independent evidence of reliability?
- Which QM variant was used? Neither the meteorological forcing QM nor the SWE correction parts state whether this is standard quantile mapping or for example a trend-preserving variant (e.g. quantile delta mapping). In my opinion, this matters here, since naive QM is known to dampen or distort the projected change signal, and the remarkable numbers (>80% SWE loss) depend on this choice. Please clarify.
- SPASS products. Two versions of the same underlying snow model climatology are used for two different purposes: SPASS-CL for model screening (Section 2.3) and SPASS-CLQM, described as enhanced via data assimilation from 1998 onward, for the actual SWE bias correction. Please clarify why the (implicitly weaker) SPASS-CL is used for screening rather than SPASS-CLQM throughout, and how much the two datasets actually differ (i.e. before SPASS-CLQM's assimilation period begins and after), I think this is worth a supplement figure. If they are nearly identical before 1998, this may not matter much, but this isn't stated.
- Unclear whether RhiresD/TabsD is the actual bias-adjustment target inside CH2025. The paper uses RhiresD/TabsD as an evaluation benchmark (Fig. 3, Fig. 4), however, the reference dataset of CH2025's own QM is not clear, thus I am wondering if this was Rhires/Tabs as well? If it's the same, the "evaluation" of snowfall bias in Section 2.3/3.1 is a bit circular maybe?
- No equations given for any core component. The degree-day melt model, the precipitation-phase partitioning function, the snow-covered-area algorithm, and the QM transfer function are all described only in words (regardless of whether the equations are simple and straightforward). Since it is anyway central, at least the melt equations, the partitioning threshold function, and the QM equation should be given explicitly, even briefly.
- Fig 2 (flowchart) is too coarse. It hides many separate steps: the parallel obs-driven historical run, two different reference datasets (SPASS-CL vs SPASS-CLQM) used at different stages, the two non-criterion exceptions, and the fact there are two separate QM steps (forcing-level and SWE-level). Please expand this into a more complete, detailed flowchart.
- Section 4.3 (GWL mapping) is unclear and partly self-contradicting. It's not clear why the GWL framework "cannot be transferred" to the 12-model subset. Is this a sample-size problem, or because the subset was chosen for historical snow performance and therefore isn't representative of future warming trajectories? The next sentence (RCP8.5 guarantees high GWLs by 2100) doesn't really support the previous claim either. Section 3.4 already gives a concrete reason (the subset misses the warmest spring-warming chains) –> this might be linked here instead of a vague new justification. (See also next point)
- "Block-time-shift" approach is not explained. The term is used with no definition, but what is a "block", what gets "shifted", and why is linking CMIP5-driven RCMs to CMIP6 warming trajectories needed in the first place? Since this underlies the whole GWL framing (and the confusion in the point before), 2-3 explanatory sentences here would help a lot, not just a citation to the external report. (This is also true for many parts of the methodology)
Minor comments
- Section 3.2 (snowfall bias "propagating" into SWE bias): (a) is SPASS-CL really the same experiment as the obs-forced historical run, or a separately built product with its own assumptions? and (b) Fig. 3's snowfall bias looks like an absolute fraction difference (colorbar -0.4 to 0.4) while Fig. 5's SWE bias is explicitly a relative percentage. These aren't the same kind of number, so "larger" isn't really a fair comparison or am I wrong?
- RCM-dominance statement. "RCM parameterization together with downscaling and bias adjustment" lumps the raw dynamical model and the statistical correction applied afterward as one cause, but the paper's own later sections show these can behave very differently. Also, the claim that GCM choice barely matters seems a bit strong when then pointing out that the single best-performing chain is specifically the CCCMA-driven one (so differs significantly from the others!?).
- Fig. 4 needs a sentence of explanation. Since the QM was used, it isn't expected to reproduce the correct joint/conditional precip -temp relationships which leads to the mismatch under sub-zero temperatures and thus could be seen as an expected result of that choice, not an unexplained separate problem!?
- ±10% cutoff. Fig. 6 plots the quantity the selection is based on and marks retained models with an outline, but doesn't show the threshold band itself; the criterion is otherwise only supported by supplementary Fig. S1. Please add the threshold line to Fig. 6, or bring the relevant panel from S1 into the main text.
- Black outline in Figs. 6 & 8 is weak. Retained/excluded is one of the most important parts in both figures, but it's shown with the least visible channel (thin outline) on top of two other categorical channels (color = RCM, shape = GCM). Consider using solid vs. hollow/transparent fill, or two side-by-side panels instead, maybe much easier to read, also in grayscale.
- Fig. 12: I found it hard to see what are now no change and what not considered areas, maybe this could be visually done better!?
- Only RCP8.5 is used throughout. Given the paper's aim of supporting adaptation planning (abstract, conclusions), including at least one more moderate scenario (e.g. RCP4.5) would make the results considerably more useful!? RCP8.5-only risks a bit the framing of the "high-end" outcome as the outcome, especially since the abstract/conclusions don't always make the RCP8.5-only scope clear at first read.
Citation: https://doi.org/10.5194/egusphere-2026-3478-RC2 -
AC2: 'Reply on RC2', Harsh Beria, 10 Sep 2026
Reviewer 2
The authors provide a useful and well-organized study that fills a genuine gap in the climate impact field across Switzerland, by providing a 1x1 SWE (snow water equivalent) dataset based on the high-end emission scenario. While the study could be considered for publication, in its current form I recommend major revision. Several methodological choices are not explained, backed or justified in enough detail (e.g. the quantile mapping approach used, the GWL block-time-shift method, the degree-day snow model itself), which makes it hard to judge the results in their proper context and would make it difficult/ impossible to reproduce the study (FAIR). In addition, some of the figures could benefit from some adaptations.
Response: We thank reviewer 2 for their comment and have responded to their suggestions below.
Major comments
- Model exclusion (IPSL-WRF) contradicts the historical evaluation. Excluding IPSL-WRF because its future climate signal looks like an outlier in summer precipitation, after selecting models based on historical performance, seems like (and risks) narrowing the projected spread on purpose!? Please justify this with an independent criterion (e.g. from the CH2025 report itself or other studies flagging IPSL-WRF!?), or show results with/without this model to test sensitivity?
Response: Strictly speaking, based on percentage SWE biases, IPSL-WRF biases slightly exceeded the -10% to 10% range, which was the criterion used as the basis of models selected in this study (Figures S1, 6). However, we acknowledge that IPSL-WRF could have been part of the final ensemble based on other criteria, such as absolute SWE bias (Figure S1) or SWE bias in mm (Figure 6). The primary reason for excluding IPSL-WRF from our simulations was because of the unreasonable future changes in summer precipitation and air temperature. This has been acknowledged in the manuscript (L307-310):
“Second, the IPSL-WRF model chain driven by IPSL boundary conditions was excluded despite its reasonable reference-period snow simulation. Based on percentage biases, IPSL-WRF was only slightly less accurate to the next-best selected model chain, and its SWE bias in millimeters was well within the ensemble spread.”
As IPSL-WRF was also identified as an outlier model in the CH2025 report, we will cite the CH2025 report as a further justification for excluding IPSL-WRF model chain in the revised manuscript.
- SWE quantile-mapping validation (Section 3.5): stating that the post-QM agreement with SPASS-CLQM (Fig. 9) shows the correction is "reliable" overstates imho what this check can really reliably show. QM guarantees close agreement with the calibration target over the fitting period by construction, regardless of the underlying model's actual skill I would argue and thus this might not be an independent evidence of reliability?
Response: We agree that the comparison between quantile-mapped SWE and SPASS-CLQM over 1991-2020 is an in-sample calibration result and therefore Figure 9 does not provide independent evidence of model reliability. We will revise Sections 2.3 and 3.5 accordingly and avoid describing this agreement as a validation of model reliability.
Before SWE quantile mapping, SWE simulations were evaluated against the independent and like-to-like SPASS-CL reference dataset, which showed that most model chains underestimate SWE (Figures 5, 6), consistent with the snowfall underestimation in the input meteorological forcing (Figure 3). SWE quantile mapping was then applied as a post-processing step to reduce these historical SWE biases, similar to the approach used in operational SWE estimation in Switzerland (Marty et al., 2025; Mott et al., 2023). We already acknowledge in Section 4.1 that this quantile-mapping of SWE cannot compensate for missing process parameterization in our model (L513-516).
“To reduce systematic biases in simulated snow storage, we bias-adjusted SWE toward SPASS-CLQM, the best available observation-constrained climatological SWE dataset for Switzerland, which substantially reduced reference-period SWE biases across elevation bands. Nevertheless, the correction cannot fully compensate for missing physical processes resulting in additional uncertainties in snow accumulation and ablation dynamics”.
- Which QM variant was used? Neither the meteorological forcing QM nor the SWE correction parts state whether this is standard quantile mapping or for example a trend-preserving variant (e.g. quantile delta mapping). In my opinion, this matters here, since naive QM is known to dampen or distort the projected change signal, and the remarkable numbers (>80% SWE loss) depend on this choice. Please clarify.
Response: The CH2025 meteorological forcings were generated using univariate non-trend-preserving quantile mapping, applied independently to precipitation and air temperature. The SWE data were adjusted using a similar empirical quantile-mapping approach, as described in Michel et al., (2024). These details will be mentioned in the revised manuscript.
- SPASS products. Two versions of the same underlying snow model climatology are used for two different purposes: SPASS-CL for model screening (Section 2.3) and SPASS-CLQM, described as enhanced via data assimilation from 1998 onward, for the actual SWE bias correction. Please clarify why the (implicitly weaker) SPASS-CL is used for screening rather than SPASS-CLQM throughout, and how much the two datasets actually differ (i.e. before SPASS-CLQM's assimilation period begins and after), I think this is worth a supplement figure. If they are nearly identical before 1998, this may not matter much, but this isn't stated.
Response: SPASS-CL and SPASS-CLQM SWE datasets serve two different purposes.
SPASS-CL is used for screening models because it provides an unassimilated, observed temperature- and precipitation-driven SWE reference to assess the consequences of only changing the meteorological forcings from different RCMs, while keeping the snow model configuration unchanged. Using SPASS-CL therefore provides a like-for-like evaluation of the effects of meteorological forcings on SWE simulations. As SPASS-CLQM incorporates additional observational constraints through SWE data assimilation, differences between RCM-driven SWE simulations and SPASS-CLQM would reflect: (1) meteorological forcings, and (2) data assimilation, making it difficult to attribute differences specifically to the input meteorological forcing. We will make this distinction explicit in Section 2.3.
The purpose of the final SWE bias-adjustment against SPASS-CLQM is to constrain the historical SWE climatology toward the best available observation-constrained SWE estimate, for RCMs that show the most accurate SWE simulations over the 1991-2020 period. We will clarify this in the revised manuscript.
SPASS-CL and SPASS-CLQM have been evaluated in detail in previous publications (Marty et al., 2025; Michel et al., 2024), and do not warrant a supplementary figure in this paper.
- Unclear whether RhiresD/TabsD is the actual bias-adjustment target inside CH2025. The paper uses RhiresD/TabsD as an evaluation benchmark (Fig. 3, Fig. 4), however, the reference dataset of CH2025's own QM is not clear, thus I am wondering if this was Rhires/Tabs as well? If it's the same, the "evaluation" of snowfall bias in Section 2.3/3.1 is a bit circular maybe?
Response: Yes, RhiresD and TabsD were the reference datasets used for CH2025 bias-adjusted meteorological forcings. We will state this explicitly in the revised manuscript and will provide the related CH2025 references.
We agree that comparisons of bias-adjusted precipitation and air temperature fields against RhiresD/TabsD are not independent evaluations, these evaluations were already carried out in the CH2025 scientific report (MeteoSwiss and Zurich, 2025). Figures 3 to 5 instead connect how biases evolve through the snow modeling chain: from raw RCM meteorological forcings, through bias-adjusted snowfall diagnostics, and finally to SWE simulated with a high-resolution offline snow model.
This distinction is important because CH2025 bias-adjustment is applied independently (univariately) to precipitation and air temperature. Consequently, agreement with the reference marginal distributions does not guarantee that the joint temperature-precipitation relationship is well represented. Since snowfall is derived from both variables, alterations in this dependence structure can introduce snowfall biases. Figure 4 illustrates this effect: the raw RCMs generally produce excessive precipitation under cold conditions, whereas the bias-adjusted forcings produce too little, reversing the sign of the resulting snowfall bias.
The subsequent comparison of simulated SWE against SPASS-CL provides a more independent evaluation of whether these snowfall biases propagate into SWE (Figure 6).
- No equations given for any core component. The degree-day melt model, the precipitation-phase partitioning function, the snow-covered-area algorithm, and the QM transfer function are all described only in words (regardless of whether the equations are simple and straightforward). Since it is anyway central, at least the melt equations, the partitioning threshold function, and the QM equation should be given explicitly, even briefly.
- Fig 2 (flowchart) is too coarse. It hides many separate steps: the parallel obs-driven historical run, two different reference datasets (SPASS-CL vs SPASS-CLQM) used at different stages, the two non-criterion exceptions, and the fact there are two separate QM steps (forcing-level and SWE-level). Please expand this into a more complete, detailed flowchart.
Response: We will expand the “2.2 Snow modeling” section to include descriptions of precipitation-phase partitioning, snow modeling, description of quantile-mapping applied, and a more detailed workflow in the revised manuscript.
- Section 4.3 (GWL mapping) is unclear and partly self-contradicting. It's not clear why the GWL framework "cannot be transferred" to the 12-model subset. Is this a sample-size problem, or because the subset was chosen for historical snow performance and therefore isn't representative of future warming trajectories? The next sentence (RCP8.5 guarantees high GWLs by 2100) doesn't really support the previous claim either. Section 3.4 already gives a concrete reason (the subset misses the warmest spring-warming chains) –> this might be linked here instead of a vague new justification. (See also next point)
Response: The point in L551 was “the GWL framework cannot be transferred directly to the adjusted subset”.
The primary reason for lacking direct transferability is the substantially smaller ensemble size: our subset contains 12 models compared to 26 models in Climate CH2025. As this manuscript does not project SWE changes in GWL space, we prefer not to expand this aspect further here; the GWL mapping of SWE will be addressed in a subsequent article.
Importantly, the subset was selected based on historical snow performance, with additional verification that this selection does not significantly alter the range of future winter and spring warming trajectories (see Section 3.4 and Figure 8).
- "Block-time-shift" approach is not explained. The term is used with no definition, but what is a "block", what gets "shifted", and why is linking CMIP5-driven RCMs to CMIP6 warming trajectories needed in the first place? Since this underlies the whole GWL framing (and the confusion in the point before), 2-3 explanatory sentences here would help a lot, not just a citation to the external report. (This is also true for many parts of the methodology)
Response: We understand that introducing the block-time-shift approach without explaining it creates unnecessary confusion, particularly because this procedure is not used in the analyses presented here. It is a central element of the GWL-concept of the CH2025 scenarios and will be used in a follow-up study in which the transient snow simulations presented here will be transferred to the CH2025 GWL framework. We will therefore remove the following sentence from the revised manuscript.
“In contrast to earlier Swiss climate scenarios, Climate CH2025 expresses future changes primarily in the frame of Global Warming Levels, using a block-time-shift approach to link CMIP5-based EURO-CORDEX regional simulations with CMIP6 global warming trajectories (MeteoSwiss and Zurich, 2025; Michel et al., 2024).”
Minor comments
- Section 3.2 (snowfall bias "propagating" into SWE bias): (a) is SPASS-CL really the same experiment as the obs-forced historical run, or a separately built product with its own assumptions? and (b) Fig. 3's snowfall bias looks like an absolute fraction difference (colorbar -0.4 to 0.4) while Fig. 5's SWE bias is explicitly a relative percentage. These aren't the same kind of number, so "larger" isn't really a fair comparison or am I wrong?
Response: SPASS-CL simulates SWE using the same modeling structure as applied in this study, driven by observed meteorological forcings (RhiresD and TabsD grids). It does not include any additional data-assimilation step. It therefore provides a directly comparable SWE reference for evaluating the effect of meteorological forcings on simulated SWE. We will clarify this in the newly proposed subsection “Description of reference datasets”.
We agree that the different units used in Figures 3 and 5 may cause confusion as Figure 3 expresses biases in fractions whereas Figure 5 expresses biases in percentage. Values of −0.4 to 0.4 in Figure 3 translate to −40% to 40%. To make this clearer, we will use percentage biases in Figure 3 in the revised manuscript.
- RCM-dominance statement. "RCM parameterization together with downscaling and bias adjustment" lumps the raw dynamical model and the statistical correction applied afterward as one cause, but the paper's own later sections show these can behave very differently. Also, the claim that GCM choice barely matters seems a bit strong when then pointing out that the single best-performing chain is specifically the CCCMA-driven one (so differs significantly from the others!?).
Response: The clustering of snowfall and SWE biases by RCM is evident in Figures 3, 5, and 6. For example, four of the five CLMCOM-CCLM4 chains and all five MOHC-HADREM chains show comparatively good historical performance, whereas all four CNRM-ALADIN and all five SMHI-RCA chains exhibit pronounced negative snowfall and SWE biases. This has also been reported in other studies (Matiu et al., 2024; Meyer et al., 2019), and reported in the original manuscript (L212-214):
“This is consistent with previous evaluations of EURO-CORDEX simulations over the European Alps, which showed that bivariate model biases in temperature and precipitation at local scales are more influenced by RCMs (Matiu et al., 2024; Meyer et al., 2019).”
Our statement that the CCCMA-driven CLMCOM-CCLM4 chain has the lowest overall snowfall bias does not imply that CCCMA-driven simulations generally outperform those driven by other GCMs. Rather, this is the best-performing individual model chain within a broader pattern in which CLMCOM-CCLM4 chains generally show relatively small biases irrespective of the driving GCM. The original text states (L214-218):
“CLMCOM-CCLM4 model chains generally showed the smallest snowfall fraction biases across different driving GCMs, with the lowest overall bias found for the chain driven by CCCMA. In contrast, several CNRM-ALADIN and SMHI-RCA model chains showed pronounced negative biases in the high-elevation Alpine region, with snowfall fraction being underestimated by up to 40% (Figure 3).”
- Fig. 4 needs a sentence of explanation. Since the QM was used, it isn't expected to reproduce the correct joint/conditional precip -temp relationships which leads to the mismatch under sub-zero temperatures and thus could be seen as an expected result of that choice, not an unexplained separate problem!?
Response: We acknowledge that univariate quantile-mapping is not expected to reproduce the joint / conditional precipitation-temperature relationship. This has been explained in subsection 4.2 (limitations section), where the text positioning is more adept. We will link this better when explaining Figure 4 in the revised manuscript.
L524-L527: “However, univariate correction treats precipitation and temperature independently and therefore does not explicitly preserve their joint dependence, which is particularly important for snow applications, because snowfall is controlled by the co-occurrence of precipitation at near-freezing or sub-freezing temperatures.”
- ±10% cutoff. Fig. 6 plots the quantity the selection is based on and marks retained models with an outline, but doesn't show the threshold band itself; the criterion is otherwise only supported by supplementary Fig. S1. Please add the threshold line to Fig. 6, or bring the relevant panel from S1 into the main text.
Response: We agree that it will be beneficial to bring some components of Figure S1 within the main text, we will address this in the revised manuscript.
- Black outline in Figs. 6 & 8 is weak. Retained/excluded is one of the most important parts in both figures, but it's shown with the least visible channel (thin outline) on top of two other categorical channels (color = RCM, shape = GCM). Consider using solid vs. hollow/transparent fill, or two side-by-side panels instead, maybe much easier to read, also in grayscale.
Response: Thanks for this suggestion, we will try to improve the contrast between selected vs unselected models and modify Figures 6 and 8 appropriately in the revised manuscript.
- Fig. 12: I found it hard to see what are now no change and what not considered areas, maybe this could be visually done better!?
Response: The light blue shading reflects minor changes in SWE. We will clarify this in the figure caption.
- Only RCP8.5 is used throughout. Given the paper's aim of supporting adaptation planning (abstract, conclusions), including at least one more moderate scenario (e.g. RCP4.5) would make the results considerably more useful!? RCP8.5-only risks a bit the framing of the "high-end" outcome as the outcome, especially since the abstract/conclusions don't always make the RCP8.5-only scope clear at first read.
Response: Restricting the ensemble to RCP8.5 was primarily a computational choice, and we will make this more explicit in the abstract and conclusions.
A key objective of this work is to establish the SWE modeling workflow within the framework of the Climate CH2025 scenarios, which express future changes in terms of Global Warming Levels (GWLs). In a subsequent study, the transient SWE simulations presented here will be mapped to GWL1.5°C, GWL2°C, GWL3°C, corresponding to 1.5, 2, and 3 °C global warming, respectively.
RCP8.5 was selected because it ensures that all RCMs in the ensemble reach 3°C of global warming by 2100, whereas several simulations under RCP2.6 and RCP4.5 do not. Given the computational effort that would have been required to run the high-resolution snow model for all RCMs and for all available emission pathways, RCP8.5 therefore provided a pragmatic way to retain all three target GWLs across the full model ensemble and maintain consistency with the Climate CH2025 framework. Although RCP8.5 provides an improbable emission pathway, it is important for future risk assessment and planning.
Importantly, our objective was also to preserve the spread in projected winter temperature and precipitation changes across climate models, which in our current modeling setup is indeed the case (Figure 8).
Citation: https://doi.org/10.5194/egusphere-2026-3478-AC2
-
RC3: 'Comment on egusphere-2026-3478', Anonymous Referee #3, 31 Jul 2026
General comments
This study examines the anticipated reduction in snow resources across Switzerland during the 21st century using high-resolution climate modelling. The authors integrated regional climate simulations with a spatially distributed snow model to project daily snow water equivalent at a 1x1 km² resolution.
The authors have carried out an interesting study that is certainly scientifically relevant. The study is very well written, the results are generally robust, and the conclusions are supported by data. I see the novelty in providing the spatially continuous daily dataset for snow water equivalent (SWE) future projections, which remains very rare at such large spatial scales. Therefore, I believe the study has the potential to be published. However, I have some comments listed below that I would like to be addressed.
Major comments
One of the major limitations of the study is that the authors assessed only the SSP5-8.5 climate projection, which represents the most pessimistic scenario. Is there a specific reason for this, and why was, for example, the SSP2-4.5 scenario not used as well to make results more robust? From today's perspective, the SSP2-4.5 scenario seems to represent a more realistic future climate trajectory and, in my opinion, it would make sense to include it alongside SSP5. I am aware that extending the analysis to another SSP scenario may be too much for revisions to this manuscript; therefore, I would encourage the authors to expand the discussion section on this important issue.
The authors assessed future changes in mean and maximum SWE. While these two characteristics are certainly suitable indicators of the snow regime, I believe other characteristics would also be interesting to examine, such as snow cover duration or the dates of snow onset and meltout. One argument for using them is that they may respond differently to climate change than mean or maximum SWE. In particular, the day of meltout seems to respond strongly to climate change, whereas the day of snow onset is relatively less sensitive. I believe that including additional characteristics will provide a more comprehensive overview of future changes in snow in Switzerland; therefore, I would encourage the authors to extend the analysis in this respect.
Section 2.2: I believe that more information about the snow model should be provided. For example, how is LWC calculated, how do melt factors change over the course of the season, or what threshold temperature has been used for snow/rain identification or snowmelt onset? Consider also adding equations for individual snow model components. I understand that this is already a published approach, but a bit more consistent information would help readers avoid jumping between the literature.
Specific comments
Figure 1: The figure needs some formal improvements, such as adding a graphical scale and a north arrow. The colour key title should be “Snowfall fraction [-]” (instead of “Snow fraction”). For the colour scale, please consider a slight modification, as the colours are currently difficult to distinguish within the 0.6 to 1 interval.
Section 2.1. Although CH2025 is a published dataset, I would like to see a bit more information to avoid the need to consult the referenced literature. Which GCM-RCM models and SSPs are included? Although this information will appear later in the results and figures, a simple table of climate models with brief descriptions might help readers unfamiliar with this dataset get oriented.
Similar to above, please provide a bit more information regarding the RhiresD and TabsD datasets (L 177), the temperature-dependent precipitation partitioning routine (L 185), the SPASS-CL dataset (L 191), and the SPASS-CLQM (L 201). As noted above, I fully understand that the information can be found in the cited literature, but a few more sentences would help readers get oriented.
L 197: I understand and agree to limit model choice according to their performance in the 800-2600 m elevation band, but please provide some justification here.
L 209: The section describes biases in snowfall fraction, but the information on what the biases are assessed against is missing. The information is only in the Fig. 3 caption (against observed gridded data). Please clarify directly in the text.
Citation: https://doi.org/10.5194/egusphere-2026-3478-RC3 -
AC3: 'Reply on RC3', Harsh Beria, 10 Sep 2026
Reviewer 3
General comments
This study examines the anticipated reduction in snow resources across Switzerland during the 21st century using high-resolution climate modelling. The authors integrated regional climate simulations with a spatially distributed snow model to project daily snow water equivalent at a 1x1 km² resolution.
The authors have carried out an interesting study that is certainly scientifically relevant. The study is very well written, the results are generally robust, and the conclusions are supported by data. I see the novelty in providing the spatially continuous daily dataset for snow water equivalent (SWE) future projections, which remains very rare at such large spatial scales. Therefore, I believe the study has the potential to be published. However, I have some comments listed below that I would like to be addressed.
Response: We thank reviewer 3 for their positive evaluation of our manuscript. We have answered their comments below.
Major comments
One of the major limitations of the study is that the authors assessed only the SSP5-8.5 climate projection, which represents the most pessimistic scenario. Is there a specific reason for this, and why was, for example, the SSP2-4.5 scenario not used as well to make results more robust? From today's perspective, the SSP2-4.5 scenario seems to represent a more realistic future climate trajectory and, in my opinion, it would make sense to include it alongside SSP5. I am aware that extending the analysis to another SSP scenario may be too much for revisions to this manuscript; therefore, I would encourage the authors to expand the discussion section on this important issue.
Response: A key objective of this work is to establish the SWE modeling workflow within the framework of the Climate CH2025 scenarios, which express future changes in terms of Global Warming Levels (GWLs). In a subsequent study, the transient SWE simulations presented here will be mapped to GWL1.5°C, GWL2°C, GWL3°C, corresponding to 1.5, 2, and 3 °C global warming, respectively.
RCP8.5 was selected because it ensures that all RCMs in the ensemble reach 3°C of global warming by 2100, whereas several simulations under RCP2.6 and RCP4.5 do not. Restricting the ensemble to RCP8.5 was primarily a computational choice. This has been clarified briefly in Section “4.3 Mapping transient snow scenarios to global warming levels” and we will make this more explicit in the abstract and conclusions.
Importantly, our objective was to preserve the spread in projected winter temperature and precipitation changes across climate models, which in our current modeling setup is indeed the case (Figure 8).
The authors assessed future changes in mean and maximum SWE. While these two characteristics are certainly suitable indicators of the snow regime, I believe other characteristics would also be interesting to examine, such as snow cover duration or the dates of snow onset and meltout. One argument for using them is that they may respond differently to climate change than mean or maximum SWE. In particular, the day of meltout seems to respond strongly to climate change, whereas the day of snow onset is relatively less sensitive. I believe that including additional characteristics will provide a more comprehensive overview of future changes in snow in Switzerland; therefore, I would encourage the authors to extend the analysis in this respect.
Response: Thank you for your comment, we agree that including another snow indicator will be beneficial for future implications of this study. However, given the additional changes proposed to be made (adding subsection on data description, description of the snow model, etc.), the revised manuscript will be much longer than the original one. In a follow-up study mapping our transient simulations to global warming levels (see L545-557), a number of additional snow indicators like snow onset and meltout dates will be considered.
Section 2.2: I believe that more information about the snow model should be provided. For example, how is LWC calculated, how do melt factors change over the course of the season, or what threshold temperature has been used for snow/rain identification or snowmelt onset? Consider also adding equations for individual snow model components. I understand that this is already a published approach, but a bit more consistent information would help readers avoid jumping between the literature.
Response: We will expand “2.2 Snow modeling” section to include descriptions of precipitation-phase partitioning and more details (also equations) about snow modeling in the revised manuscript.
Specific comments
Figure 1: The figure needs some formal improvements, such as adding a graphical scale and a north arrow. The colour key title should be “Snowfall fraction [-]” (instead of “Snow fraction”). For the colour scale, please consider a slight modification, as the colours are currently difficult to distinguish within the 0.6 to 1 interval.
We will change the legend to “Snowfall fraction” and further improve Figure 1, including adding a DEM map as a second subplot in Figure 1.
Section 2.1. Although CH2025 is a published dataset, I would like to see a bit more information to avoid the need to consult the referenced literature. Which GCM-RCM models and SSPs are included? Although this information will appear later in the results and figures, a simple table of climate models with brief descriptions might help readers unfamiliar with this dataset get oriented.
Response: We will add a figure summarizing all available models, and the ones that were shortlisted in supplementary material of the revised manuscript.
Similar to above, please provide a bit more information regarding the RhiresD and TabsD datasets (L 177), the temperature-dependent precipitation partitioning routine (L 185), the SPASS-CL dataset (L 191), and the SPASS-CLQM (L 201). As noted above, I fully understand that the information can be found in the cited literature, but a few more sentences would help readers get oriented.
Response: In the revised manuscript, we will add a subsection “Description of reference datasets” within the “Study area and methods” section, describing the reference datasets used in this study (RhiresD, TabsD, SPASS-CL, SPASS-CLQM).
L 197: I understand and agree to limit model choice according to their performance in the 800-2600 m elevation band, but please provide some justification here.
Response: An earlier publication from Marty et al., (2025) showed the limitations of the reference SWE datasets (SPASS-CL and SPASS-CLQM) at very-low and very-high elevations. We will cite this reference as a further justification in the revised manuscript.
L 209: The section describes biases in snowfall fraction, but the information on what the biases are assessed against is missing. The information is only in the Fig. 3 caption (against observed gridded data). Please clarify directly in the text.
Response: We will expand L209 by adding information about the observed gridded dataset in the revised manuscript.
Citation: https://doi.org/10.5194/egusphere-2026-3478-AC3
-
AC3: 'Reply on RC3', Harsh Beria, 10 Sep 2026
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 398 | 143 | 36 | 577 | 66 | 20 | 16 |
- HTML: 398
- PDF: 143
- XML: 36
- Total: 577
- Supplement: 66
- BibTeX: 20
- EndNote: 16
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General Assessment:
This study evaluates simulations of snowfall fraction and snow water equivalent from 26 Climate CH2025 models across Switzerland over the period 1991–2020, and projects the variations in SWE during 2035-2099 for the RCP8.5 scenario. The overall framework is clear, and the results are generally robust. However, there is a lack of detailed explanations regarding research methods and data. Some suggestions require the authors to give further consideration.
Major Comments:
Minor Comments: