Summary of the paper

Cluzet et al. present an evaluation of the CrocO snow data assimilation system which is based on a Particle Filter technique that propagates information spatially from areas with observations (e.g., height of snow, HS) to unobserved areas. Using leave-one-out validation, the performance of this system is compared to that of the ensemble (i.e., open-loop) and operational model counterpart at hundreds of snow depth stations in the French Alps, Pyrenees, and Andorra over 10 years. Different localization schemes (rlocal vs. klocal) and radii are tested, and showed only minor changes in skill. All Particle Filter configurations showed improvements relative to the open-loop but no clear improvement over the operational model. Using the klocal scheme at 35 km for the rest of the study, the authors show it tends to mitigate the negative bias in the open-loop at certain elevations (1500-2500 m) and locations (e.g., Central-East Pyrenees). The klocal scheme shows the highest skill at locations where the open-loop has negative bias, with reduced skill when the open-loop has positive bias. The skill of the klocal PF was found to be insensitive to the density of HS observations, but the magnitude of open-loop bias decreased with increasing HS observations (a peculiar result). The study suggests there is complementarity in that areas with high errors in meteorological data (due to low station data) may be mitigated by a Particle filter with spatial propagation of snow information.

Recommendation

Overall, I find that this paper is of good quality and provides additional application and evaluation of the CrocO Particle Filter developed in recent work by the authors. It is conducted at a more coarse model scale (and less dense observation network) than previous studies (e.g., Magnusson et al 2014; Winstral et al., 2019) and is hence a unique, complementary contribution. I think there are some aspects that could be strengthened in the discussion and possibly the presentation of results, depending on what information is available about the HS stations. The issue is that the stations may benefit both the open-loop and oper (through providing meteorological data to the reanalysis) as well as the PF (through providing HS observations for assimilation). Disentangling these competing contributions to the model configurations could be useful in a revised manuscript.

Comments

1. A result highlighted in the paper is that the open loop has lower performance in areas with fewer stations and decreased bias as station density increases (L. 299-301 and Fig. 10), but this seems coincidental, as the open loop does not integrate HS observations. Is this more suggestive of reduced quality in the reanalysis forcing data, given the lower availability of surface stations? Alternatively, I wonder how many of the HS stations include meteorological observations that are integrated into the reanalysis data (see lines 92-93)? Overall, the authors should carefully consider the two sources of data offered by stations (meteorology and HS). It might seem that a lack of stations would influence both the open loop and DA for that reason. The authors seem to understand this later in the paper (L. 458-459), so a consistent viewpoint should be included through the paper. Is it possible to identify which HS stations have met data that are being integrated into the reanalysis forcing data? If so, it may be necessary to differentiate the results based on whether the HS stations have met data that are contributing to the forcing data or not.
2. The leave-one-out validation (L. 63-65, 166-169) should be clarified whether this is done spatially (e.g., one station is removed) or temporally (e.g., individual HS data are removed at a station). I am still unsure after re-reading these parts.
3. The study highlights a range of elevations (1500-2500 m) where is a strong negative bias in the open loop and oper simulations. The role of gauge undercatch is discussed as the most likely cause of this bias (L. 351-353). However, I do not find this convincing, as gauge undercatch is likely a universal problem for measuring precipitation and snowfall in mountain environments. Is there another factor that might explain the bias in this specific elevation zone? Are the wind speeds higher, and hence larger undercatch errors at these elevations? Or is this suggestive of an oddity in the HS station data, for example, more ski area observations in these zones, which may be biased toward higher snow depths?
4. In the discussion, the authors discuss variability of snow redistribution by wind in the context of error compensation with the DA system (L. 423-432). One aspect that is missing here is the impact of scale – both the process scale and the model scale – and this is something that should be identified and discussed. The resolution of the model in this study is quite large, such that a process like wind redistribution cannot be represented explicitly, and ultimately the variability must be handled in other manners (e.g., as a sub-grid parameterization, see Clark et al., 2011 for example; or improving the model resolution to a finer scale). Please comment more on this issue.

 General comments

-  The argument that weather stations provide better estimates of surface meteorology than NWP may not be a widely supported viewpoint, for instance see Lundquist et al. 2019. Please comment.

- There are many instances where papers that are “in prep” or “in review” are cited, for instance Vernay et al., Deschamps-Berger et al., etc. Please check with the TC journal guidelines on whether these types of papers (not yet published) are permitted to be cited.

Technical corrections

- Line 13: Add “of” after “strategy”.

- Lines 53: Add “the” before “Météo-France”.

- Line 71: Replace “Does” with “Can”. Also, delete “manage to”.

- Line 84: Add “Alps” after “Northern”.

- Line 104: Replace “less” with “fewer”.

- Line 137-142: Can you include a brief summary of the key elements of the ensemble generation technique? This would be useful for someone who has not read Cluzet et al. (2020, 2021) or who needs a reminder about the methods.

- Line 151: Replace “radius” with “radii”.

- Line 152: Should be “size” instead of “sizes”.

- Line 156: Suggest replacing “remind” with “note”.

- Line 157: Should be “rlocal”.

- Line 166: Should be “requires use of independent data”.

- Line 202: Add “of” at the end of this line.

- Line 203: Remove “If”.

- Line 218: Replace “Despite” with “Although”.

- Line 228: A final sentence is needed here to offer an explanation for interpreting the new formulation of the skill score.

- Line 236-237: This argument is not reliable, as the 2012 year looked to be snowier but the bias was lower (0.11 vs. 0.16). Please remove example or provide better justification/explanation.

- Lines 239-240 and Table 1: Suggest adding a column with SS so the reader does not have to do the simple math.

- Line 242: Add “negative” before “biases”.

- Line 250: Replace “radius” with “radii”.

- Line 292: Remove “Then”.

- Line 295: Please revise the phrasing “pointed as”, this is unclear to me.

- Line 309: Please clarify which simulation is being referred to here when discussing the “simulation bias”.

- Line 316: Replace “inferior to” with “less than”.

- Line 335: The phrasing is awkward here and I suggest rewording “enables to satisfactorily account for”.

- Line 402: Replace “Despite” with “Although”.

- Line 423: Delete “Then,”.

- Line 431: Delete “of course”.

TABLE AND FIGURE COMMENTS

- Figure 2 could be simplified into two panels: 2009-2015 and 2016-2018, since 2016 seems to be the most major change in the station network in the study period. Otherwise, the histograms appear to be similar.

REFERENCES

Clark, M. P., Hendrikx, J., Slater, A. G., Kavetski, D., Anderson, B., Cullen, N. J., et al. (2011). Representing spatial variability of snow water equivalent in hydrologic and land-surface models: A review. Water Resources Research, 47(7), W07539. https://doi.org/10.1029/2011WR010745

Lundquist, J., Hughes, M., Gutmann, E., & Kapnick, S. (2019). Our Skill in Modeling Mountain Rain and Snow is Bypassing the Skill of Our Observational Networks. Bulletin of the American Meteorological Society, 100(12), 2473–2490. https://doi.org/10.1175/BAMS-D-19-0001.1

Major comments

In the paper titled “Propagating information from snow observations with CrocO ensemble data assimilation system: a 10-years case study over a snow depth observation network” Bertrand Cluzet et al. assimilated snow depth observations from an in-situ network of 295 stations covering the French Alps, Pyrenees, and Andorra over the period 2009–2019. They attempted to demonstrate how in-situ observations of snow depth can help contain intermediate and large-scale modeling errors by means of data assimilation. However, while the results of snow depth data assimilation are closer to observations than the non-operational open-loop counterpart (open-loop), which does not use data assimilation, they are not as good as the operational deterministic snow cover modeling system (oper). It is natural that the results of snow depth data assimilation are better than those of open-loop, but it is necessary to show that snow data assimilation outperforms oper. Moreover, I believe that snow depth data assimilation not only improves the reproducibility of the same snow depth but also affects other parameters of the snow model such as snow water equivalent, snow density, snow surface temperature, and water and energy balances above the snow; hence, I would like to see the discussion on the impact and verification of these factors. The procedure of the data assimilation method is difficult to understand, especially the need to clarify the control variables and the analysis variables. Overall, the paper is not accepted in the current form and needs to be revised in accordance with the following comments. My main concerns are as follows:

(1) Although the method of data assimilation has already been published in papers such as Cluzet et al. (2020, 2021), I feel that it is necessary to describe the flowchart of data assimilation and to clarify what the control variables and analysis values are, so that the data assimilation method can be understood to some extent from this study independently. I feel that since there is a limit to the extent of explanation that can be expressed only in text, I would really like to see a flowchart of data assimilation. I think that data assimilation can be corrected for other elements of the model using cross-covariance and error covariance with snow depth. It should be clarified whether snow depth is the only control variable. Furthermore, the authors conducted two types of experiments for data assimilation using particle filters, but I did not understand the difference between rlocal and klocal. Please explain this in detail

(2) I would like to see an explanation of the difference between Oper and open-loop as a reference. I would also like to see a more detailed explanation of the SAFRAN massif, with expanded abbreviation, and the convincing data provided by SAFRAN to the snow model.

(3) Advantage of data assimilation is that by assimilating snow depth data, the information is correlated to other factors such as snow water content and precipitation, which will improve the accuracy of model estimation other than direct snow depth improvement. If you have observational data on snow density, snow surface temperature and snow water content, please add a discussion on whether the assimilation of snow depth affects other model parameters.

(4) I feel that there is a lack of literature in the introduction. As a pioneering work in satellite data assimilation, I feel it is necessary to mention the data assimilation study of MODIS snow cover and AMSR-E snow water content by Andreadis and Lettenmaier (2006).

Andreadis, K. M.; Lettenmaier, D. P. Assimilating remotely sensed snow observations into a macroscale hydrology model. Adv. Water Resour. 2006, 29, 872–886.

Line 42–43: "Their potential to improve local simulations is unambiguous as demonstrated by many studies.” After this sentence, please consider adding references such as Liston and Heimstra (2008) and Suzuki et al. (2015).

Liston, G. E.; Hiemstra, C. A. Simple data assimilation system for complex snow distributions (SnowAssim) J. Hydrometeorol. 2008, 9, 989–1004.

Suzuki, K.; Liston, G. E.; Matsuo, K. Estimation of continental-basin-scale sublimation in the Lena River Basin, Siberia. Adv. Meteorol. 2015, 2015.

Line 61–62: "These variants are used in a localised framework, in which only observations coming from a certain radius around the considered location are assimilated." After this, please consider adding references such as Zupanski (2021).

Zupanski, M. The Maximum Likelihood Ensemble Filter with State Space Localization. Mon. Weather Rev. 2021.

(5) Line 389: There are two reasons why the operational run could not be beaten by the assimilation in terms of RMSE.

Line 454–455: In other words: the assimilation can not beat the open-loop in the most densely observed areas (e.g. in the Northern Alps, 455 where the observation density is similar to the studies of Magnusson et al. (2014) and Winstral et al. (2019)) because the open-loop performance is already high there.

In the above two sentences, the verb "beat" is used, but I feel it is not appropriate. I would like to see this correction with  more appropriate words.

Line 267–269: The text needs to be revised and made more readable.

(6) The results of data assimilation in Fig. 6 should be shown in the form of Table 1 with numerical values. Moreover, there is no data for oper in RMSE and SS in Fig. 6; please resolve this.

Propagating information from snow observations with CrocO ensemble data assimilation system: a 10-years case study over a snow depth observation network

by: Bertrand Cluzet1, Matthieu Lafaysse1, César Deschamps-Berger1, 2, Matthieu Vernay1, and Marie Dumont1

The paper demonstrates that assimilation of in-situ sown depth observations does not provide significant RMSE improvements over the open-loop and operational simulations. It reduces bias in snow depth estimates and outperforms the open-loop simulations in specific elevation bands in locations with a lower observation density.

The methodology and the presentation of results are clear, and I agree that this approach seems to be relevant for the estimations of snow depth and SWE. For all these reasons, I think the paper should be published with a minor revision.

The paper is interesting and provides insight into analyzing ensemble data assimilation approaches. However, there is too much emphasis on the Continuous Ranked Probability Score (CRPS).  It would be more insightful to show snow depth maps or scatter plots of model simulations (open loop, oper, DA vs. observations).

L137-140: Please provide perturbation statistics (additive/multiplicative, mean, standard deviations/correlation coefficients, spatial, and temporal correlations) for each forcing).

L165: a summary about the updating step and how the PF updates the snow profile would improve the clarity of the setup section.

L229:  Please provide some information about the resolution of simulations.

How are the bias and RMSE computed (i.e., point vs. gridded simulations)? What is the uncertainty of this comparison?

The authors can elaborate more on the representative error. Gridded simulations are compared with ground-based point measurements. There is a scale gap in this comparison. It will benefit the paper if the authors discuss this source of uncertainty.

L300: What is the reason for presenting figure 10b?

If possible, please improve the quality of the figures.