Implementation of a multi-layer snow scheme in the GloSea6 seasonal forecast system: impacts on land–atmosphere interactions and climatological biases

Seo, Eunkyo; Dirmeyer, Paul A.; Tak, Sunlae

doi:10.5194/gmd-19-1261-2026

Articles | Volume 19, issue 3

https://doi.org/10.5194/gmd-19-1261-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-19-1261-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 19, issue 3

Model evaluation paper

|

11 Feb 2026

Model evaluation paper |

| 11 Feb 2026

Implementation of a multi-layer snow scheme in the GloSea6 seasonal forecast system: impacts on land–atmosphere interactions and climatological biases

Eunkyo Seo, Paul A. Dirmeyer, and Sunlae Tak

Download

Final revised paper (published on 11 Feb 2026)
Supplement to the final revised paper
Preprint (discussion started on 17 Jun 2024)
Supplement to the preprint

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2024-1066', Anonymous Referee #1, 16 Jul 2024
The paper by Seo and Dirmeyer, titled “Implementation of Multi-layer Snow Scheme in Seasonal Forecast System and Its Impact on Model Climatological Bias,” investigates the effects of implementing a multi-layer snow scheme on the climatological biases of a seasonal forecast system. Traditional single-layer snow schemes in land surface models often inadequately capture the insulating effects of snowpack, leading to warm and cold biases during winter and snow melting seasons. The study compares the performance of the Global Seasonal Forecast System (GloSea) versions 5 (single-layer) and 6 (multi-layer) over a 24-year period. Findings reveal that the multi-layer snow scheme in GloSea6 shifts the snow melting season by two weeks, improving surface temperature, permafrost extent, and overall model climatology. This enhancement mitigates near-surface warming bias and improves precipitation simulation over snow-covered regions.
However, it overlooks critical differences in vegetation treatment between the Noah and Noah-MP models. Suzuki and Zupanski (2018, doi: https://doi.org/10.1007/s11707-018-0691-2) provide a thorough examination of the uncertainties in solid precipitation and snow depth prediction, which is highly relevant to this study. The differences between the land surface models are notable: the Noah model uses a one-canopy layer with a simple canopy resistance and a linearized energy balance equation representing the combined ground-vegetation surface, considering seasonal LAI and green vegetation fraction. In contrast, the Noah-MP model includes snow interception features such as loading-unloading, melt-refreeze capabilities, and sublimation of canopy-intercepted snow, along with a detailed representation of radiation transmission and attenuation through the canopy, within- and below-canopy turbulence, and different options for representing the biophysical controls on transpiration. Therefore, the changes affect not only snow-covered areas but also the global vegetation albedo and surface temperature. In their results, they report that the snow depth changes, but the snow water equivalent does not. The reason for the longer period of snow cover is believed to be due to the more accurate representation of radiation and turbulent fluxes benath the vegetation canopy. Therefore, the multi-layer snow model is not the critical factor in this case.
To enhance the completeness of your study, it is crucial to discuss the impact of vegetation treatment in addition to the multi-layer snow scheme.
By addressing these points, the manuscript will provide a more holistic view of the improvements in seasonal forecast systems and their broader climate implications.
Specific Comments
Introduction and Background: Please include a discussion on the handling of vegetation in land surface models, specifically contrasting the Noah and Noah-MP models.

Methodology: Please provide detailed descriptions of the Noah and Noah-MP models, focusing on their treatment of vegetation and snow processes. In addition, please discuss how these differences might affect your results and the broader implications for climate modeling.

Results and Discussion: Please analyze the impact of vegetation treatment on your findings, especially in terms of global vegetation albedo and surface temperature.

Conclusion: Please emphasize the importance of considering both snow and vegetation processes in land surface models.
Citation: https://doi.org/10.5194/egusphere-2024-1066-RC1
- AC1: 'Reply on RC1', Eunkyo Seo, 27 Aug 2024
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-1066/egusphere-2024-1066-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-1066-AC1
RC2:
'Comment on egusphere-2024-1066', Anonymous Referee #2, 31 Jul 2024
The paper “Implementation of Multi-layer Snow Scheme in Seasonal Forecast System and Its Impact on Model Climatological Bias” evaluated the retrospective seasonal forecast performance of the Global Seasonal Forecast System (GloSea) version 5 (GloSea5, with a single-layer snow scheme) and version 6 (GloSea6, with a multi-layer snow scheme) over a 24-year period (1993-2016), focusing on the impacts of multi-layer snow scheme (GloSea6) versus single layer snow scheme (GloSea5) on the climatological biases of the seasonal forecast system. Results revealed that GloSea6 more accurately captures the snow phenology, elongating the snow melting season by two weeks, which improves the simulations of soil moisture, surface temperature, surface evaporation and subsequent land-atmosphere coupling regime in mid-to-high latitudes. This enhancement mitigates near-surface warming bias and improves precipitation simulation over snow-covered regions during late spring to summer. The authors attributed this improvement to the multi-layer snow scheme in GloSea6, yet more analyses are necessary to exclude other model physics updates (including atmosphere, ocean and sea ice) to support this conclusion.

Major points:
Snow cover is important in land surface modeling, besides the treatment of snowpack in single layer or multi-layer scheme, snow surface albedo and snow cover fraction (the percentage of a model grid that is covered by snow) are pivotal factors that influence the accumulation and melting of snow cover in climate models. How about the difference in these two factors in GloSea5 and GloSea6?

Why are surface soil moisture (SSM) simulated in GloSea5 are more than those in GloSea6 from October to March but the situation reversed (i.e., GloSea5 SSM less than GloSea6) after April, although the initial snow amount are close to each other on April 1^st (Figure1a,b)?

In line 270-271, the author claimed that the weaker insulating effect of the single-layer snow scheme leads to warmer surface temperature during thin (snow melting or freezing season) snow cover (figure 1c), in fact, during the freezing season from October to January when the air is colder than land surface, if the single-layer snow scheme provides a weaker insulating effect, it should lead to colder rather than warmer surface temperature. How to understand this contradiction?

Figure2a shows that GloSea6 provides more surface soil moisture in mid-to-high latitudes of the northern Hemisphere, in addition to the positive evapotranspiration-precipitation feedbacks suggested by the authors (line 355), is it possible that the update in atmospheric physics in GloSea6 rather than the update of snowpack scheme from single layer to multi-layer results in more precipitation than GloSea5 in these regions (Figure 6a,b)？

How to explain the deterioration of both bias and RMSE of Tmax and precipitation in GloSea6 with improved snow scheme in northeastern Eurasian continent (Figures 4e, 5f, 6f)?

Minor points:
Line 14 of the abstract, “permafrost extent” is not addressed in this study.

Line 114, “single-layer snow scheme allows the surface layer of the atmosphere to directly access the heat in the soil” is not true when the snowpack is thick.

Line 395, “(f)” should be “(c)”.

Line 412, “winch” should be “which”.

The title of this paper can be modified to be more appropriate for its content.
Citation: https://doi.org/10.5194/egusphere-2024-1066-RC2
- AC2: 'Reply on RC2', Eunkyo Seo, 27 Aug 2024
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-1066/egusphere-2024-1066-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-1066-AC2

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

AR by Eunkyo Seo on behalf of the Authors (27 Aug 2024) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (30 Aug 2024) by Jinkyu Hong

RR by Anonymous Referee #1 (05 Sep 2024)

RR by Anonymous Referee #3 (19 Sep 2024)

Suggestions for revision or reasons for rejection

The manuscript “Improving land-atmosphere coupling in seasonal forecast system by implementing a multi-layer snow scheme” evaluated the retrospective seasonal forecast performance of the Global Seasonal Forecast System (GloSea) version 5 (GloSea5) and version 6 (GloSea6) over a 24-year period (1993-2016) and tried attributing the improving retrospective seasonal forecasts in GloSea6 during winter and snow melting seasons to the implementation of multi-layer snow scheme in GloSea6. The comparison results indicated that the snow melting season shifts two weeks later in GloSea6, consequently improving the simulations of soil moisture in its climatology and memory, surface temperature, and land-atmosphere coupling regime as well. The authors thought that the subsequent improvements in surface temperature and precipitation over snow-covered regions were resulted from the implementation of multi-layer snow scheme in GloSea6. This probably holds when other model physics are same in GloSea5 and GloSea6. Therefore, the authors should find other ways to analyze the results. Here are some specific comments:

1. From Table 1, we could see that only the model resolution in two versions of GloSea are same. The model coupler and model physics are all updated in GloSea6. The surface air temperature and precipitation are impacted not only by the local effect but also the nonlocal effect. For a coupled system, the physics interact with each other. Therefore, it is It is unrealistic to talk about only one physical process while ignoring the influence of other physical processes.
2. From Table 3 in Kim et al. (2021), compare with GL6.0 in GloSea5, the GL8.0 in GloSea6 added the multi-layer snow scheme, improved the land surface albedo physics and modified the atmospheric rain fractions. Temporarily leaving aside the influences of modification in land surface albedo, how does the modification in partition between rain and snow affect the snow characteristics? Moreover, in around Line 120, how could the authors distinguish the albedo changes from the multi-layer snowpack or improved surface albedo physics?
3. In around Line 265, “The multi-layer snowpack also extends the area of snow cover, which leads to the increased surface albedo, where increasing snow amount leads to an increase of surface albedo at the land surface about 10 days later (SFs. 1a, b)”. The surface albedo was closely connected with the snow cover fraction over the snow-covered regions. However, the more snow amounts couldn’t always mean larger snow cover area. It would be good to compare the snow cover fraction and snow albedo in two versions. Otherwise, how to explain the similar snow amount in two versions from Oct to Jan but the different surface albedo?
4. In Line 25-30: “As the memory of initial land conditions can extend out to approximately 2 months, the importance of realistic land surface initialization in determining skill of the subseasonal forecast is paramount (Koster et al., 2011; Guo et al., 2011; Seo et al., 2019).” However, the two versions used different land initial conditions. How to exclude this effect from the effect of multi-layer snowpack?
5. Around Line 190, “reanalysis-based precipitation dataset with are available for 1979–present”. Not clear.
6. The label bar in Fig.8 was missed. It’s hard to follow the left-bottom-corner sub-figure.

Referee Report: PDF

Hide

RR by Anonymous Referee #4 (26 Sep 2024)

Suggestions for revision or reasons for rejection

I read the revised submission of the manuscript and the response to reviewers' comments. In my view the response to the reviewers is adequate, but both reviewers point out that additional differences between GloSea5 and Glosea6 (apart from the snow model), may obscure the interpretation. For a study into the impact of the snow model, it would have been best to have simulations with models that differ in the snow component only. However, to limit costs, the authors probably used standard output from back integrations. So no extra tailored information for diagnostics. This is fine, but it would be helpful to state clearly upfront that existing integrations are used, and that the other differences between the two systems may obscure the interpretation of the snow model changes. Please see below a few more suggestions to improve the paper.

Fig. 1a
Are the units correct for snow amount? An Eurasian average of 6m water equivalent sounds wrong.

Fig. 2
It is important to be clear about what is presented here. Are these averages of all the forecasts from May to August with a lead time of 60 days, i.e. verifying from July to October? The vertical title of 2a says 1-60 days, which suggests a 60day average of the forecasts? Please clarify.
For clarity, please add in Fig. 2 that most of the SM stations are over North America and Europe (as in line 316 of the main text).

Definition of SMM in section 2.2
I read the discussion about SMM several times, but still fail to understand what it means. The reason is probably that the observations contain a lot of noise and come with large uncertainty, so a lot of massaging is needed. Is it meant to be a daily average of time filtered SM? But what is the time scale of smoothing? It would be nice to have an intuitive definition of what SMM is supposed to represent, and then give a few details about the processing.

Lines 278-286 and Fig. 1d
The discussion of the response of air and soil temperatures to the multilayer snow pack is hard to understand. One reason is, that the profile of snow temperature is not mentioned. However, this is the key difference between a single slab model and a multilayer snow model. A multilayer scheme can represent a detailed temperature profile, and a single layer model can not. In fact a multi-layer model is not a requirement to represent insulation. A single slab can have variables for snow amount (or thickness) and a single temperature, and a conductivity controlled by snow amount. For clarity it would have been nice to show the temperature profiles from atmosphere to soil, including the snow layer for different seasons. I guess this information is not available from the integrations that are used.

In response to the reviewer's comments, the authors have changed the name of the single layer snow model to "zero"-layer snow model. It probably means that snow is only considered in the mass budget and not in the thermal budget, which is rather unusual. So, in the heat budget the atmosphere is directly coupled to the soil irrespective of the presence of snow? It would be good to spell this out very clearly. It is therefore not correct to conclude that a multi-layer snow scheme is needed to describe snow insulation. The only conclusion is that snow insulation is important. Insulation can also be simulated with a single layer snow scheme (although perhaps less accurately).

See for other implementations of single and a multi-layer snow schemes e.g.:
Arduini, G., Balsamo, G., Dutra, E., Day, J. J., Sandu, I., Boussetta, S., & Haiden, T. (2019). Impact of a multi‐layer snow scheme on near‐surface weather forecasts. Journal of Advances in Modeling Earth Systems, 11(12), 4687-4710.

Figs. 4,5
The improvement in temperature is impressive. The signals in minimum and maximum temperature are indeed typical for less thermal coupling of the atmosphere with the underlying soil. However, I am still confused (similar to figs. 2 and 3) about the verification season. Are these plots based on 60-day forecasts verifying from July to October? It is important for the interpretation to be very clear about it. I may have missed it somewhere, but it does not harm to say it again.

Figs. 6-10
These figures are certainly of interest, but are most likely only partially related to snow effects. The two modelling configurations are very different on many aspects (convection, turbulence, micro-physics, vegetation). For clarity, it would be good to be more outspoken about this in the manuscript.

Line 459
Please change "Most land surface models ..." to "Some land surface models ...".

Lines 463-465
This sentence again suggests that GloSea5 and GloSea6 differ mainly through the snow scheme. Could you please re-write by saying that back-integrations with GloSea5 and GloSea6 are analysed and that in spite of the differences in the atmospheric physics some of the signals can be traced back to the snow scheme.

Line 471
Please mention again what SMM is (e.g. soil moisture memory time scale?).

Hide

ED: Reconsider after major revisions (31 Oct 2024) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (18 Nov 2024) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (07 Dec 2024) by Jinkyu Hong

RR by Anonymous Referee #5 (31 Dec 2024)

Suggestions for revision or reasons for rejection

Conclusion

The manuscript, Improving land-atmosphere coupling in a seasonal forecast system by implementing a multi-layer snow scheme, fails to meet the standards required for publication in Geoscientific Model Development due to multiple critical issues. The title is unclear, as the manuscript does not clarify whether the multi-layer snow scheme was developed by the authors or implemented into the GloSea model by authors. Or authors here just to access its impacts. The introduction lacks detailed explanations of the mechanisms by which the multi-layer snow scheme addresses biases and fails to provide appropriate references to support the claims made. The data section is incomplete, providing insufficient detail on the differences between the single-layer and multi-layer snow schemes in GloSea5 and GloSea6 and their origins or physical basis. As model development paper, the origins, implementations, and development history of this ‘multi-layer’ snow scheme are the must have parts. After reading the manuscript and the responds to the reviewers, I still can’t get which multi-layer snow scheme is discussed in this study. Is there only one parameterization about multi-layer snow in land model community? What is the iteration of this kind of parameterizations?
The methodology is unclear, particularly, the lack of offline simulations makes it impossible to isolate the impact of the multi-layer snow scheme from other model changes, which significantly undermines the validity of the conclusions. Offline land model simulations, as demonstrated in studies like Arduini et al. (2019), could provide more robust insights into the impact of the snow schemes. Offline simulations for GloSea5 and GloSea6, even if not for long-term historical runs, would add significant value. The results section is weakened by inconsistent comparisons—such as snow water equivalent versus snow cover—and unsupported claims regarding the improvements attributed to the multi-layer snow scheme. Importantly, authors haven’t ruled out whether other changes in the atmospheric model could also influence the mid- to high-latitude regions.
Overall, the manuscript fails to provide the rigor and clarity required for a model development paper. To address these issues, the authors must (1) clarify what version of multi-layer snow scheme was discussed in this study, (2) provide detailed references and explanations for the scheme's physical basis, (3) conduct offline simulations to isolate the snow scheme's impacts, and (4) ensure consistent and meaningful comparisons of variables. Without these major revisions, the manuscript does not meet the publication standards of GMD.

Title
The title, “Improving land-atmosphere coupling in a seasonal forecast system by implementing a multi-layer snow scheme”, raises questions about its accuracy. Did the authors implement this scheme into the GloSea model, or did they develop the multi-layer snow scheme themselves? If not clarified, the title feels inappropriate and somewhat misleading.

1. Introduction
Line 50: What does "coupling strength" mean in this context? Is there a specific metric to quantify this "coupling strength"? Please clarify.
Line 56: In the introduction, it would be helpful to provide clear descriptions of the "warm and cold biases during winter and snowmelt seasons" caused by the absence of a multi-layer snowpack scheme. Since these biases are a major focus of the results section, a detailed explanation of their underlying mechanisms would strengthen the introduction.
Line 59: What does "Noah-MP" refer to?
Line 61: Is the "layered snowpack scheme" mentioned here the same as the "multi-layer scheme"?
Line 64: Please clarify the transition between Noah-MP and JULES. Are the authors providing examples of models using multi-layer schemes? Do these models employ the same "multi-layer scheme"? How many different multi-layer schemes exist within the land model community? The citation of Walters et al. (2017) is insufficient, as it is an overview paper on the JULES model rather than a specific reference for the multi-layer scheme.
Line 69: The names of the 13 S2S models or the study that characterizes these models are missing. Please provide a clear citation.
Line 70: The statement, “Hence, this study conducts a comparative analysis between GloSea5 (single-layer snowpack) and GloSea6 (multi-layer snowpack),” is misleading. As mentioned in the “Data” section, GloSea6 involves multiple changes, with snowpack treatment being just one of them. This distinction should be made clear early on.
Line 73: The primary and secondary objectives of the study are vague. Why does the primary objective receive only a single sentence, while the secondary objective is elaborated in more detail? Which of these objectives is the study's main focus?
In general, the introduction needs significant improvement. It lacks references detailing the multi-layer snowpack scheme and its physical or mathematical basis. Although it is possible that the multi-layer scheme performs better than the single-layer scheme, the mechanisms remain unclear. While summarizing the development of land models and snowpack treatments is challenging, Geoscientific Model Development (GMD) requires a more thorough and rigorous introduction to meet its high standards.

2. Data
Line 83: Using "Data" as the section title is not ideal. A more precise title would be beneficial.
Line 90: Additional clarification about JULES and GL would help readers understand the model structure. Are these the same land model?
Line 100: The citation of Kim et al. (2021) is inaccurate. The paper focuses on atmospheric improvements in GloSea6 and does not provide an overview of "all model components." Please revise this characterization.
Lines 113–115: There is confusion regarding the single-layer snow scheme in GloSea5. The authors state that it has constant thermal conductivity and density but later mention adjustments for layer thickness. Is the thermal conductivity constant or not in GloSea5? Additionally, a reference for the origin of this single-layer scheme is necessary.
Line 119: Walters et al. (2017) does not discuss snowpack treatment or the multi-layer snow scheme. This citation is inappropriate.
The data section should include a clear description of the snowpack treatments in GloSea5 and GloSea6. For the multi-layer snow scheme, its origin, physical or mathematical improvements over the previous treatment, and whether it was developed by the authors should be explicitly stated. These details are critical for a model development or assessment paper.

3. Methodology
Lines 209–216: The authors state that "a single-member run is used in this study" but later mention "ensemble mean values" for bias analysis. The term "ensemble" is used inconsistently throughout the manuscript. Please clarify whether the results are based on a single-member run or ensemble simulations.
In coupled ensemble simulations, it is challenging to identify which model components drive improvements in surface temperature, soil moisture, and other variables. Offline land model simulations, as demonstrated in studies like Arduini et al. (2019), could provide more robust insights into the impact of the snow scheme. Offline simulations for GloSea5 and GloSea6, even if not for long-term historical runs, would add significant value. While coupled model analysis is useful, comparing offline and coupled results would greatly strengthen the study.

4. Results
Line 277 and Figure 1: The comparison between GloSea5 and GloSea6 uses snow water equivalent and ERA5 snow cover percentage—two different variables. This mismatch should be clarified. Additionally, it is premature to conclude that GloSea6 snow water equivalent is superior to GloSea5 based on ERA5 snow coverage without further explanation or an "apples-to-apples" comparison.
Line 278: The discussion of albedo differences between GloSea5 and GloSea6 is significant and should be shown in the main part of the manuscript but not supplementary material. The larger initial albedo difference shown in SF.3b should be explained, particularly as it diminishes over time. Is it related to the surface albedo treatment changes in the land model between GloSea5 and GloSea6.
Line 286: The conclusion that GloSea6's improvements are due to the multi-layer snow scheme is unconvincing, especially since the largest differences occur in October and November when snow cover is minimal. Could these differences be attributed to changes in precipitation or large-scale circulation strength?
Line 309: A correlation coefficient of ~0.35 explains only ~10% of the variance between soil moisture (SM) and precipitation (PR). Correlation does not imply causation. The authors need to rule out other potential factors influencing precipitation and SM before concluding that SM changes driven by the multi-layer snowpack explain precipitation differences.
The change of precipitation can cause the change of the snow, which leads to albedo difference. The driver of the precipitation could be the snowpack treatment, but also could be the convection, aerosol, and cloud physics changes between GloSea5 and GloSea6.

Line 331: For the simulated SM differences between GloSea6 and GloSea5, it is true that there are large differences over the snow frontal region. However, there are also significant differences over the Amazon rainforest in South America, where snow is rare.
Line 334: Why do other model changes tend to impact tropical precipitation or SM but show no clear impact on northern mid- to high latitudes? While I do not deny that the snowpack treatment change contributes, there is no evidence ruling out whether other changes could also influence the mid- to high-latitude regions.
Line 359: The “significant improvement” in simulated SMM is unclear. A reduction from -3.7 to -3.3 days in SF.5d,e needs further explanation to justify it as a significant improvement.
Line 368: The authors claim that GloSea6 reduces the surface air temperature bias. Is this claim supported by previous studies, or is it based on the authors' analysis? Please clarify how the temperature bias data was derived and what control temperature data was used for comparison.
Line 371: Please clarify why decomposing Tmean into Tmax and Tmin helps identify the impacts of two major modifications in the LSM. I would appreciate further explanation on this reasoning.
Line 376: If the GloSea6 simulation tends to produce more snow than GloSea5, it could lead to similar temperature reductions. However, many factors could contribute to this outcome. While the multi-layer snowpack treatment might be one factor, other potential contributors must also be considered.
Line 391: In Figure 10d, values over Australia are predominantly negative, resembling those over the northern mid- to high latitudes. Therefore, it is difficult to conclude that the factors reducing biases over Australia are different from those affecting northern mid- to high latitudes.
Line 408: Please explain the assertion that "the improvement in the mid- and high-latitude regions of the Northern Hemisphere is likely due to the improved snow physics." In Section 2, the authors mention numerous changes to the atmospheric models used in GloSea5 and GloSea6 (Walter et al., 2017). Attributing these improvements solely to snow treatment, without considering the contributions of atmospheric changes, is not substantiated.

5. Summary and Conclusion
Line 485: The conclusion that differences between GloSea5 and GloSea6 in areas without snow (e.g., South and East Asia, Central Africa, South America, and Australia) are likely due to other factors arising from model version updates, while differences in snow-covered areas are attributed to the snowpack treatment, is interesting but insufficiently substantiated. This hypothesis could serve as the motivation for the study but cannot be presented as its conclusion without stronger evidence.
The authors must provide clear evidence demonstrating that the factors affecting non-snow regions do not influence snow-covered regions. In a global coupled model, atmospheric physics changes, such as those in convection, clouds, and aerosols, can have wide-reaching impacts. It is possible that both atmospheric physics updates and the snowpack treatment contribute to the observed results. However, it is the authors' responsibility to isolate the specific contributions of the snowpack treatment.
I recognize that isolating the impact of land modifications in a coupled simulation is challenging. However, accessing and analyzing offline simulations would provide a more robust approach to distinguish the effects of the snowpack treatment from other model changes. This would significantly strengthen the conclusions and the overall quality of the study.

Referee Report: PDF

Hide

RR by Anonymous Referee #4 (31 Dec 2024)

Suggestions for revision or reasons for rejection

This paper present a study on the seasonal evolution and climate of two models: GloSea5 and GloSea6. GloSea6 is the result of an a major upgrade of GloSea5 with changes predominantly in the physics package: among a lot of details a new aerosol climatology, changes to the gas optics of the radiation scheme, changes to the cloud overlap handling, new ice optical properties, new micro-physics, changes to the gravity wave scheme, cloud top entrainment and convection closure. Also the land coupling was substantially changed, through the replacement of a single layer snow scheme by a multilayer version, a new vegetation climatology, and introduction of a wavelength dependent albedo.

This is a major model change with impact ranging from local processes to general circulation. The authors did a series of 100-day ensemble forecasts with initial conditions from October to April covering many years. The seasonal evolution of parameters related to snow is evaluated with focus on snow processes, albedo, and vegetation. Unfortunately, it is not always clear whether other processes (e.g. radiation) play a role. The tentative conclusion is nicely summarised in Fig 11, with the multi-layer snow scheme having less soil moisture in the snow season more soil moisture in summer, higher soil temperatures in winter and lower soil temperatures in summer, and 2 weeks longer snow cover. Attribution is predominantly to conductivity of the total snow layer, where the multi-layer model has a much stronger insulating effect.

The paper is a clear illustration of how difficult model development is. The main difficulty is that verification at the process level is very limited. Verification relies heavily on data sets that are at best constrained to some extent by satellite observations, and calibrated with in situ observations. Turbulent latent flux at the surface from GLEAM is a clear example. It is a good product compared to others, but it relies on surface net radiation from satellite or re-analysis and vegetation activity from satellite in clear sky conditions. Furthermore it uses the simple Priestley-Taylor formulation, which puts all the emphasis on correlation with radiation and not on atmospheric humidity. This is not a criticism, it is state of the art. The issue is that the vegetation response to environmental conditions is not well understood.

In conclusion, I recommend publication after some revision. The paper is well written, a nice set of diagnostics is presented, and the results are presented in a balanced way. The conclusions are not really firm, but the reader can decide for her/himself how to interpret the results. There are a few points, I would like the authors to address.

Major points:
1. The paper suggests that the insulating effect cannot be properly represented with a single layer. I disagree, because the increase of thickness of a single layer will increase the insulating effect. In the implementation of GloSea5, it may not work that way, perhaps because snow is combined with the top soil layer. However, a single layer model could have been implemented such that insulation increases with the layer thickness. What a single layer cannot represent is a range of response time scales. With a thicker layer there is more inertia and slower response. For instance, the multi-layer scheme should show a much better diurnal cycle of temperature, something that is hardly discussed in the paper. Fig. 4 shows something about the diurnal cycle but all seasons are put together and therefore it is hard to see snow signals.

2.
Are there any offline simulations with forcing at say 10m above the surface from wind temperature, humidity, downward radiative fluxes and precipitation to test the snow scheme, albedo effects and vegetation changes? I know it is beyond the scope of the current paper, but something of this nature must have been done during the development of the new snow scheme. This would be extremely helpful in disentangling the impact of the main aspects that are believed to be responsible for the impact that is seen in the seasonal integrations.

3.
It is not clear why the snow stays 2 weeks longer on the ground with the multilevel snow scheme. It is suggested that the higher soil temperature is playing a role but I doubt it. The main source of heat for melting comes from the atmosphere. So it would be good to look at all components of the surface energy balance. A major mechanism for melt is from sensible heat flux in case of partial snow cover. Solar radiation heats the snow-free areas, which brings the air above zero. The warm air melts the snow over the snow covered fraction. This is one of the (perhaps very few) advantages of a tiled surface coupling. Perhaps the authors can comment on where the heat for snow melt is coming from.

4.
Snow cover parameterisation is a key process in snow evolution. I know it is uncertain and hard to come up with a sensible formulation. However, in a paper about snow related processes and a comparison between model versions where the vegetation data has changed with impact on snow cover, it deserves more discussion.

5.
Fig. 8 presents an interesting diagnostic where coupling regimes of evaporation with soil moisture or net radiation are identified. I have the feeling that using GLEAM as reference is misleading. GLEAM, GloSea5 and GloSea6 are all models, although constrained by observations in different ways. GloSea5 and GloSea6 are constrained by observations via the initial conditions (re-analysis) and the climatology for vegetation. For instance, GLEAM shows very strong coupling between evaporation and radiation, which is understandable given the use of the Priestly-Taylor model. It would be good to dedicate a few lines of discussion on this aspect.

Minor points:
1.
I could not find information on liquid water content of snow? Is it represented in the multi-layer snow model or both models?
2.
In Fig.1: What is standardised difference?
3.
Line 275-278: The later snow melt in GloSea6 is mentioned. The subsequent sentence refers to a lower snow albedo in GloSea6. This sounds contradictory. Please explain. I would expect a lower albedo during melt to speed up the melting.
4.
Line 300: "For the surface air temperature, GloSea6 is colder during the snow freezing season due to the energy loss from the air to the ground". The word ground confused me as you probably mean snow surface.
5.
Line 301: "In February and March, when the snow begins to melt, GloSea6 simulates higher air temperature because the snowmelt over warmer ground results in reduced cooling from below". I am not sure about this interpretation. Is the higher temperature not the result of lower snow fraction, so the increased snow-free fraction allows the air to be heated well above zero. With 100% snow cover, air temperature would not rise above zero by heating from the surface.

Hide

RR by Anonymous Referee #6 (15 Jan 2025)

Suggestions for revision or reasons for rejection

I see that several reviews have already been submitted for this paper. I wasn’t sure why it needed still another review, but I went ahead and read it fresh, not being influenced by the earlier reviews and the responses thereto. I did uncover a number of issues that I feel should be addressed prior to the paper’s publication. Because my reading was independent of the other reviews, it’s safe to say that if an earlier reviewer made some similar comments, the authors haven’t yet addressed properly addressed the issue within the paper itself.

1. I found the paper to be a bit unfocused regarding what it was addressing. According to the title and abstract, the idea was to examine the impacts of using a multi-layer snow model on seasonal forecasts and land-atmosphere feedbacks. However, GloSea5 and GloSea6 differ in more than their snow model, and the paper often digresses into talking about differences in non-snow areas and what might be causing them (e.g., discussion around Figure 3, lines 370-371, 398-402, 414-424, 451-454, 487-489). Much more focus is needed. In the context of this paper, perhaps the only real value of showing global maps for various quantities, not really discussed very much here, is indicating whether the changes over the snow areas are larger than those over the rest of the globe, which might be suggestive (though not proof) of snow impacts. In fact, for most of the global plots, there are changes seen everywhere, calling into question the ability to say that those over snow areas are necessarily due to the snow model. I found unconvincing the idea expressed in lines 484-489, in which the authors state that if the changes are seen over the snow areas, then they are due to the snow model changes, whereas if they are seen elsewhere, then they are caused by something else.
2. A missed opportunity was an analysis of the offline land runs used to generate the initial land conditions (lines 131-136); the meteorological forcing was unfortunately different, but the subsurface thermodynamics and insulating effects of the snowpack could be examined much more cleanly. Probably too late for this study, though.
3. Many of the arguments for why a particular change is related to the updated snow model and not to other GloSea5/GloSea6 differences seem to me speculative at best. I fully understand the difficulties involved with trying to isolate the snow impacts from all of the other differences in the two systems; the authors were forced to work with what was available. That doesn’t mean, though, that speculation can be presented as fact or even likelihood. Examples: lines 293-295; line 369; lines 390-391, lines 402-404, lines 407-409, lines 484-485, and more. The discussion of Figure 2bc is strange; why couldn’t the differences in soil moisture RMSE simply reflect precipitation changes (Figure 6) that have nothing to do with the snow code? I do agree that the 2-week delay in snowmelt for GloSea6 in Figure 1a is probably due to the snow model. It just seems that most of the other snow model impact statements highlighted in the paper are far more speculative and, again, not clearly labeled as speculation.
4. I have a lot of trouble with the time-lagged terrestrial coupling index. This concept has been around for decades, and those who use it seem to ignore the fact that a lagged correlation does not imply causality, given that both variables examined in the correlation could be controlled by some external forcing that has its own memory and spans the time period. That is, if factor A affects variables B and C over a time scale of a week, then variable B will be correlated with variable C a day later with no underlying causal connection. I don’t find the paper convincing at all when discussing the results of this index. Perhaps that’s just me. Certainly, though, the paper shouldn’t be implying to the reader that this index definitively indicates causality.
5. Even if I were to accept the concept of the time-lagged terrestrial coupling index, its application here seems questionable. Figure 1f suggests that the soil-moisture-leading-precipitation value is much higher than the precipitation-leading-soil-moisture value at lead 1 day, which seems impossible. Precipitation’s causal impact on soil moisture is unquestionable, whereas soil moisture’s impact on precipitation must be much more tenuous. How can the authors explain Figure 1f? I’m forced to wonder if there’s an error in the analysis.
6. Section 4.3. I’m familiar with the use of R(SSM,LH/R_n) to differentiate water- versus energy controlled processes; I would think the analysis of that would be enough. How does R(R_n,LH) work, though, given that LH is scaled by the net radiation *even when* the LH is controlled by soil moisture? That is, even in a water-controlled regime, a given amount of soil moisture should produce more LH if the R_n is increased. This idea would explain the positive values in 10abc, with the negative values for GLEAM in the desert probably just some artifact associated with incredibly low LH values there. If R(SSM,LH) and R(R_n,LH) actually do allow a distinction between water-limited and energy-limited processes, what then accounts for the overlap in the positive values for the maps? (And why are the two color bars reversed?) Overall, the discussion in this section (kernels, etc.) was lost on me. In any case, the connection to the impact of “multi-layer snow processes” is very weak, basically amounting to speculation about the fact that certain differences are seen in snow areas, a discussion that does not properly account for the fact that differences of comparable magnitude are seen elsewhere across the globe.
7. Speaking of GLEAM, some discussion is needed regarding the fact that GLEAM LH values are not true observations and have their own error, which may(?) be considerable in the context of the analyses performed.

Minor comments
-- Why does the abstract talk about reducing model error over South Asia? What would that have to do with a snow model?
-- Line 55 suggests that LSMs generally don’t have multi-layer snow schemes (which would surprise me), whereas Line 69 suggests that most do, which comes off as contradictory. Is there support for the statement on Line 55?
-- On line 129, I would change “is attributable to” to “is assumed herein to be largely attributable to”.
-- Lines 214-220 need a lot of work. I read through them several times and only have a vague sense for what they are saying.
-- Line 250: Are “source” and “target” reversed here?
-- Line 272 states that the initial snow amounts in GloSea5 and GloSea6 are essentially the same. Figure 1a, though, seems to say that for January and February initializations, there’s a few millimeters difference in SWE over the huge Eurasian area, and presumably locally the differences would be much higher in places. This doesn’t seem insignificant at all, not given the size of the averaging area.
-- Line 330: Change “indicating” to “suggesting”.
-- A little confusing is the focus on the “multi-layer” aspect of the snow scheme. Another change in the snow model between GloSea5 and GloSea6 is the amount of vegetation sticking out of the surface to affect the net snow albedo (lines 283-284), something that would have an impact on the same snow areas the authors sometimes focus on. Can the authors put this particular change in context?

Hide

ED: Reconsider after major revisions (28 Feb 2025) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (15 Apr 2025) Author's response Author's tracked changes Manuscript

ED: Reconsider after major revisions (04 Jul 2025) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (13 Aug 2025) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (23 Sep 2025) by Jinkyu Hong

RR by Anonymous Referee #7 (14 Nov 2025)

Suggestions for revision or reasons for rejection

The central question for this manuscript is whether one can robustly claim that the implementation of a multi-layer snow scheme has a substantial impact on seasonal forecasts in GloSea6. In this revision, the authors have added an additional experiment (G6single) in which the snow scheme is kept single-layer within the GloSea6 framework, so that other model updates from GloSea5 to GloSea6 are fixed except for the snow parametrization.
Below I summarise my main concerns.

Major comments
1. With the added G6single experiment, the natural way to quantify the impact of the multi-layer snow scheme on seasonal forecasts is to focus on the model differences that isolate the snow parametrization.
G6multi – G6single in the coupled system (Fig. 2) and JULESmulti – JULESsingle in the offline LSM experiments (Fig. 1).
In Fig. 1, the JULESmulti – JULESsingle differences in surface properties appear visually quite large. However, this is at least partly due to the very narrow vertical range used on the y-axes. For example, the differences in latent heat flux are generally smaller than about 1.2 W m⁻². It is not clear whether such small differences are statistically significant and/or physically meaningful for seasonal forecasting, especially given that typical summer LH values are several tens of W m⁻².

A similar issue arises for the coupled simulations in Fig. 2. The snow-scheme difference G6multi – G6singleis consistently smaller than the total difference G6multi – G5single for essentially all variables, including precipitation. This implies that a substantial part of the difference between GloSea5 and GloSea6 including meteorological conditions must arise from other updates (e.g. atmospheric, ocean, sea-ice, land-cover/albedo changes, stochastic physics and ensemble size), not from the snow-scheme change alone.

The same pattern appears in other figures. For example, in Fig. 5, visual differences in air temperature and precipitation between G6multi and G6single seem small. Accordingly it is not clear whether they are statistically or practically significant.

2. When comparing Fig. 1 (offline JULES) and Fig. 2 (coupled GloSea) the snow-scheme differences clearly show changes in both magnitude and timing (e.g. the peak period of the anomalies). In particular, the snow-scheme differences in the coupled system (i.e., G6multi – G6single) appear smaller than those in the offline LSM, and their peaks are shifted.
Taken together with the fact that G6multi – G6single is much smaller than G6multi – G5single, these results suggest that the majority of the differences between GloSea5 and GloSea6 are likely due to updates in ensemble size and atmospheric/ocean/sea-ice physics and other components, rather than to the snow-scheme change alone.

3. The critical p value for rejecting the null hypothesis in the Granger-causality analysis is 1−p > 0.5, i.e. p < 0.5. This threshold is much weaker than typical statistical analysis (e.g. p < 0.1 or p < 0.05) and does not provide strong evidence for causality.

4. The title of the manuscript explicitly refers to a “seasonal forecast system”, which naturally leads the reader to expect a quantification of seasonal forecast skill(e.g. correlation, RMSE, probabilistic scores) for the different model configurations. In the current version, however, the focus is primarily on mean-state differences and process diagnostics, and the quantitative assessment of seasonal forecast skill remains limited.

5. In several figures, gridpoint-wise significance tests are applied to the differences between model configurations (e.g. t-tests for mean differences, Granger causality p-values) and the results are displayed on spatial maps. Even though these comparisons are model–modelrather than model–observation, this still constitutes a multiple-testing problem, because many hypothesis tests are performed simultaneously across the spatial grid.
In such a setting, a certain fraction of grid points will appear “significant” purely by chance even if there is no true signal.

6. Before the G6single experiment was introduced, the correlation and causality diagnostics (e.g. R(SSM,LH), R(Rn,LH), Granger causality maps) were arguably the only way to infer the possible role of the snow scheme. With G6single now available, the primary evidence for the snow-scheme impact should come from the direct model differences(G6multi – G6single and JULESmulti – JULESsingle).
At present, it is not fully clear how much additional support the correlation and causality diagnostics provide for the claim that the multi-layer snow scheme has a substantial impact in GloSea6, beyond what can be inferred from the direct differences.

Minor comments
1) For clear comparison between the offline and coupled experiments, y-axis scales used in Fig. 1 and Fig. 2 for the same variables should be the same. Otherwise, small differences may appear exaggerated in one figure and muted in another.

2) The text around line 385 refers to Fig. S2 for evidence that differences in initial conditions are negligible. However, there are no relavant information to show show quantitative differences in initial conditions to drive the climate model. Please revise Fig. S2 (or add a new supplementary figure or table) to provide such quantitative evidence or adjust the text accordingly.

3) The analysis in this manuscript is limited to the Northern Hemisphere, not the global domain. Authors may want to adjust the title to reflect this spatial focus, or to state this limitation prominently in the Abstract and Introduction.

4) In Fig. 2(j), the “standardized difference” (G6multi–G6single) time series is shown, but its definition and interpretation remain unclear. It is not obvious that simply dividing the model difference by the model standard deviation provides a meaningful measure of the significance of the model differences.Please provide a precise mathematical definition (over what period and domain the standard deviation is computed) and explain what aspect of the physical behaviour this metric is intended to highlight. If the goal is to emphasise lead–lag relationships between variables, you may consider presenting or at least explicitly referring to lead–lag correlations instead.

Hide

ED: Reconsider after major revisions (14 Nov 2025) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (10 Dec 2025) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (17 Dec 2025) by Jinkyu Hong

RR by Anonymous Referee #4 (21 Dec 2025)

ED: Reconsider after major revisions (02 Jan 2026) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (03 Jan 2026) Author's response Author's tracked changes Manuscript

ED: Publish subject to minor revisions (review by editor) (15 Jan 2026) by Jinkyu Hong

1. Emphasize the G6multi–G6single contrast (snow-scheme-only attribution)Please focus your analysis and narrative on the differences between G6multi and G6single, as this pair most directly isolates the impact of the multi-layer snow scheme within the same forecast system. Your core claim concerns the effect of introducing a multi-layer snow scheme on seasonal forecast skill and land–atmosphere coupling; therefore, the primary attribution pathway should be consistently supported by evidence from G6multi − G6single, and—where appropriate—the corresponding offline LSM contrast JULESmulti − JULESsingle. At present, it is difficult to locate, within the figures and accompanying discussion, the key results that directly support the argument as stated.

2. Moderate the strength of the conclusions in light of methodological limitationsPlease consider toning down several arguments, given the following issues and limitations that may affect robustness and interpretation:
- Ensemble-size imbalance and implications for significance/robustness:The experimental design uses different ensemble sizes across key configurations (e.g., 7 members versus 4 members). This imbalance may influence the estimated mean responses and statistical power. If feasible, please provide an appropriate robustness assessment (e.g., a matched-size subset analysis, resampling/bootstrapping, or another clearly justified approach) and clearly state how this imbalance affects your inference.
- Sensitivity and limitations of the Granger-causality analysis:While Granger causality can be informative as a diagnostic tool, it is sensitive to assumptions and preprocessing choices. Please provide a clearer statement of its limitations with respect to establishing physical causality, and discuss the potential for confounding by common large-scale atmospheric drivers.
- Sensitivity of soil-moisture memory and coupling metrics:The soil-moisture memory metric and coupling indices are central to your mechanistic narrative (snow → melt timing → SM → LH → temperature/precipitation). However, these diagnostics can be sensitive to forecast drift and lead-time dependence. Please consider additional clarification and/or sensitivity analyses addressing these dependencies (e.g., lead-time stratification, drift treatment, or related robustness checks).

3. Title revisionPlease revise the title for concision and to improve compliance with GMD conventions.
- The current title contains redundant phrasing; please simplify it.
- Please include the relevant model name(s) and version number(s)(e.g., the seasonal forecast system and land model), consistent with GMD guidance for model evaluation papers and related categories.
4. Code and data availabilityPlease ensure that materials currently listed as “available upon request” are either deposited in a suitable permanent archive or replaced by a fully reproducible access pathway. This should include explicit versioning, retrieval scripts/queries, and a clear description of any licensing constraints. If open release is restricted for any component, please describe how confidential or controlled access can be provided to others upon request, consistent with journal policy.

5. Figure accessibility and color-vision-deficiency (CVD) robustnessSeveral figures are visually crowded, with too many lines and markers, which reduces readability. In addition, some key figures rely on color combinations that may be difficult for readers with common forms of color-vision deficiency (notably red–green distinctions in line plots). Because these figures convey central results, please revise them so that the information remains clear without reliance on color cues (e.g., by reducing visual complexity and using line styles, marker shapes, and adequate luminance contrast).

Hide

AR by Eunkyo Seo on behalf of the Authors (20 Jan 2026) Author's response Author's tracked changes Manuscript

ED: Publish subject to technical corrections (29 Jan 2026) by Jinkyu Hong

AR by Eunkyo Seo on behalf of the Authors (02 Feb 2026) Author's response Manuscript

Short summary

This study examines a multi-layer snow scheme in seasonal forecasts. Compared to a single-layer scheme, it better captures snow insulation, delaying spring snowmelt by 1–2 weeks. This postpones evaporation and slows soil moisture depletion, which promotes evaporative cooling due to increasing energy partitioning into latent heat flux and enhances precipitation occurrence. This leads to realistic land-atmosphere interactions and reduced biases across Northern Hemisphere mid- and high-latitudes.