This article presents an evaluation of a SWE reanalysis generated by assimilation of the Snow CCI snow covered area product derived from MODIS data. The lack of accurate spatially distributed SWE in mountain regions is a well-known issue in snow hydrology. This work is a contribution to fill this gap.

The article reads very well and the study is well conducted. The utilization of the view angle as a weighting factor of the MODIS data is (to our best knowledge) new and interesting.

Our main concern is related to the motivation of the study. There are several snow covered area dataset available (https://lpvs.gsfc.nasa.gov/producers2.php?topic=snow). In particular NASA’s MOD10 products provide similar information as the Snow CCI product and are available globally in near real time. The Snow CCI daily SCF version 2 dataset used in this study is available over the period 2000-2020 only and it seems that it is not updated (version 3 extends to 2022). We can think of some reasons but we recommend that the authors explain why they have chosen the CCI product among others.

In addition, this study shows that a Landsat-derived SWE reanalysis largely outperforms the MODIS-CCI-derived reanalysis. Therefore, we are tempted to conclude that a global mountain snow reanalysis should be performed with Landsat fSCA. But the authors seem to implicitly consider that this is not an option. We believe that this should also be clearly stated and justified in the introduction.

Minor comments

Several acronyms were not defined (L14 WY, L69 fSCA, L11 CCI, L249 DOWY)

Fig 2: Because the tiles are defined in lon/lat angles, Fig. 2e merges tiles of different areas, giving more weight to tiles close to the equator.

L200 In this earlier study MODSCAG algorithm was used to retrieve fSCA and not SCAmod. Therefore there is no reason to specifically refer to this study to justify the 15% value. Other evaluations of MODIS-based snow products should be considered.

L206: “the weighting function 𝑤(𝜃) varies within (0,1] by its definition” Yet maximum MODIS scan angle is 55° hence w will never reach 0. It is difficult to understand how this weighting factor w was defined by Dozier et al. 2008. It would be useful to plot w as a function of the MODIS scan angle. In addition, from a more practical perspective, how were obtained the MODIS zenith angle values? It seems that the Snow CCI product does not provide such information.

L274. Cite the Vionnet et al. paper instead of the URL.

L280. The interpolation method is first an “aggregation” and then a nearest neighbor interpolation. What means aggregation (average?). Why not resampling directly to the target grid in a single operation? Why a nearest neighbor interpolation?

L295. Why was the evaluation limited to peak SWE? There are many other ASO SWE products in the Tuolumne (e.g. 49 SWE products between 2012 and 2019, Sourp et al. 2024 https://doi.org/10.5194/egusphere-2024-791, data available online https://nsidc.org/data/aso_50m_swe/versions/1).

L309-311. We find a bit confusing to use the Landsat posterior SWE as reference in section 3.1 especially in Figure 6 (where the colors indicate the residuals with respect to Landsat reanalysis). We could suggest to replace the right panel with another scatterplot showing the prior SWE instead of the Snow CCI posterior as a y-axis.

However, we find Figure 7 very informative and well designed. “30 out of 59 sites-year show improvement relative to prior” Does it suggest that assimilating the Snow CCI product was not beneficial on average?

Fig. 9: 1%-99% percentiles are usually taken to represent large sample size, here there are only 20 values.

L470. Figure 11 suggests that the thresholds of cloudiness and w discussed earlier in the paper could be revisited. This could be discussed and ideally a sensitivity analysis to these thresholds would be useful (but it may be a lot of computation to ask).

L526. Fig14 How to interpret the poor performance (i.e. the large difference with Landsat posterior estimates) in Bow domain for forest cover 0-10% in comparison with 10-50%?

L568. 0.01°

Reviewers: S Gascoin, L Sourp, N Imperatore

Summary and Recommendation

The manuscript uses a well-established Bayesian snow reanalysis / data assimilation framework to assess the utility of assimilating 0.01 deg (~1 km) fractional snow cover data (MODIS-based) from the ESA Snow Climate Change Initiative (CCI). The manuscript details the processing steps related to cloud screening and weighting for view angle in the reanalysis framework. The resulting SWE estimates are then evaluated in sub-regions in the western U.S. and western Canada with SWE time series at snow pillow and SWE maps derived from airborne lidar surveys. A separate reanalysis based on Landsat fSCA aggregated to 0.01 deg resolution is also evaluated as a reference.

Overall, I think this is an interesting paper that conducts a novel and logical analysis with the CCI snow cover dataset. As detailed below, I have some concerns about the potential impact of this paper and how the key results will benefit the community. Additionally, I have several more general/minor comments on the paper.

Major comments

1. My biggest critique of the paper is that it does not appear to have a significant or well-understood result. For those who are interested in the CCI snow dataset, the performance of the snow reanalysis are mixed, showing less accurate results than the Landsat-based snow dataset. The reasons are not fully explored here but several ideas are speculated (see Conclusions section, L. 557-566). I think it can be useful to have papers that show “negative results”, but it seems the paper stops short in showing why those poorer results are achieved. Exploring one or more of those hypotheses would add more substance and could yield a broader result that goes beyond the nuances of the CCI dataset. For instance, addressing the question of retrieval algorithm vs. spatial resolution (L. 560) would benefit a wider audience (e.g., those who are interested in CCI data as well as those interested in other remotely sensed snow cover datasets). Alternatively, testing the impact of the weighting scheme (versus using no weighting) on the SWE reanalysis might be another contribution that could be made here within the scope of the analysis, and this could have broader appeal.
2. I have concerns about interpretations if of differences in R correlation, particularly in Figures 7. A positive difference does not definitively mean that the correlation has improved, because R ranges from -1 to +1. For instance, a difference of +0.5 is not meaningful if the two R values are 0.0 and -0.5, as this indicates going from a weak negative relationship to no relationship at all. Hence, a positive difference in R is only a valid indicator of improved correlations when both are greater than 0. Why not just use R^2 here to avoid any of this potential ambiguity?

General comments

• The manuscript uses two different acronyms to describe the same variable – namely, “fSCA” and “SCF” both refer to “fractional snow cover”. This could cause confusion, so I suggest picking one convention and using it exclusively.
• The paper could use additional description on the land surface model and PBS approach. The description (L. 129-132) is rather meager. This expanded description does not have to be highly detailed but should provide enough context and explanation to help any readers who are unfamiliar with the approaches applied in Fang et al. (2022), etc. Additionally, the paper should confirm/clarify whether there are any other differences in the methods other than using a different snow cover dataset (see L. 283-286).
• Figures 4 and 5 (and corresponding text) seem to be out of order. I think the authors should consider swapping the order for both figures and their text. If I understand the process correctly, you would first do the screening (Fig. 5) and then do the assimilation (Fig. 4).
• More description is needed on whether and how quality control was done for the snow pillow data (Section 2.4.1). The California snow pillows are infamously noisy compared to NRCS SNOTEL (and note there are zero NRCS SNOTEL sites in this study). One of the sites used (Dana Meadows) also has known data issues during the study period (e.g., a tree growing in the snow pillow in 2007, see Lundquist et al., 2015).
• In a few places in the “Results and Discussion” section, the possible impacts of wildfire are speculated to be a factor (L.362-364, 439-441) but without compelling evidence. In general, I think the discussion and interpretation of results could be improved through more direct/substantive connections to other relevant studies.

Line Comments

• L. 52-58: Here, I think a sentence is clarify that this is not the same dataset as the UCLA Western U.S. Reanalysis daily snow dataset (published at NSIDC). The methods appears to be the same, but the source snow cover dataset is changed here to support the CCI effort. I did not fully understand that these were different datasets until I got deeper into section 2. I suspect others who are familiar with the UCLA WUS dataset may also experience some confusion.
• L. 93: You could note here that not only are Lajoie and Bow River basins at different latitudes, but also different snow climates.
• L. 98: Remove “The” before “Aspen”.
• L. 109-113: This is somewhat redundant with what is described earlier at L. 59-63. Why repeat this information about the goal? Consider some merging/reorganization of this text with the earlier text.
• L. 127: Why “mostly”? What else is used other than Landsat?
• L. 133: I am not sure “scatters” is the right word. Please consider rephrasing.
• L. 144: Replace “at” with “in”.
• L. 157: Please add a little more description about how this cloud mask is produced in snow CCI.
• L. 184, 199: The Rittger et al. (2020) citation is relevant and could be included at this parts of the text.
• L. 210 Consider replacing “reliability” with “quality”.
• L.214-220 and Figure 3: What is theta angle for w(theta) ~= 0.2? It might help to add a second (non-linear) x-axis with the theta values, as this might be easier for some to interpret (i.e., satellite view angle).
• L. 216: Please provide more information on where/when these calculations were performed.
• L. 245: Again, it is hard for a reader to know what theta angle corresponds to weights of 0.2 or less. Describing in terms of theta is more straightforward, in my opinion.
• L. 262: Consider adding some citations here on spatial representativeness, such as Meromy et al. (2012) or Herbert et al. (2024).
• L. 275: This would read better if rephrased as “in the Aspen domain”.
• L. 277-278: How did you handle the aggregation at the basin boundaries of the ASO data? In other words, some 0.01 deg pixels did not have complete ASO coverage and likely have missing data at the finer scale. Please describe in more detail your methods and assumptions here.
• L. 284: Add “(~1 km)” after “0.01 deg”.
• L. 287: You could add that the reason why it is not well understood over Canada is because it has not been produced there before (e.g., UCLA Western US reanalysis does not include Canada).
• L. 300: Add “more” before “forested”.
• L. 301: Add “(Fig. 1)” after “WUS”.
• L. 307: Specify the type of correlation (e.g., pearson, spearman, …)
• L. 318: I think this reads better if rephrased to say “… peak SWE better matches the Landsat reference …”
• L. 332: Suggest replacing “sites” with “pixels” because you are talking about the gridded dataset.
• L. 357: Suggest using more precise language here: replace “less snowy” and “more snowy” with “lower SWE” and “higher SWE”.
• L. 360: Replace “in Colorado” with “domain”.
• L. 404: Remove “ASO” as it is not relevant here.
• L. 435: Double check. Should this be “magenta” rather than “blue”?
• L. 529-531: Another complicating factor here with the thermal temperature screening (L. 393-396) is the mixed pixel problem with temperatures from multiple sources (snow, trees, etc.). Consider including this in your discussion, and see Lundquist et al. (2018).
• L. 551: Consider adding “, but not canopy correction (Rittger et al., 2020).” at the end of this sentence.
• L. 555-556: Another factor that could be discussed is higher forest cover in the Canadian domains. This should not be neglected.
• L. 567-571: I think this text may be too optimistic considering the results in Canada were not very skillful.

Figures and Tables

• Figure 2: Why use the full water year for these distributions? I am not sure why summer (i.e., snow-free months) are relevant, especially given that Aspen may have cloudier summers than the other locations due to regular convective thunderstorms that occur over the Rockies. How would this figure (and thresholds) change if October-June was used instead of the full WY?
• Figure 2e: I am not sure why this CDF has a step function appearance, and the text does not adequately explain it.
• Figure 4: What is “informative” here? This is described in the main text, but a brief description in the caption could be convenient for readers.
• Figure 6 caption: Add “in” after “(circles)”. Also state the units for “relative differences”.
• Figure 8: Are these biases consistent with those from other ASO surveys?
• Figures 9 and 14: I have trouble interpreting the stacked shading here. Is there a better way to convey the ranges?
• Figure 10: Looking at this figure and caption alone, it is impossible to discern what the blue and magenta shading represents. This comes later in Figure 11. Need to include the information in both places so your readers can understand.
• Figure 10c: Suggest stacking these two panels vertically rather than horizontally so their common axis (Difference in peak SWE) is shared/aligned.
• Table 1: It is incorrect to say that the Tuolumne snow pillow data come from the NRCS. These data are managed by the California Department of Water Resources via CDEC.
• Table 1: It is odd to list “NRCS SNOTEL” as the source on the Aspen lines when there are zero SNOTEL sites in that domain. Suggest removing “NRCS SNOTEL” and just adding “--” in this row.
• Table 3: Suggest adding “Bias” column for both Landsat and CCI at the end.

References

• Herbert, J. N., Raleigh, M. S., & Small, E. E. (2024). Reanalyzing the spatial representativeness of snow depth at automated monitoring stations using airborne lidar data. The Cryosphere, 18(8), 3495–3512. https://doi.org/10.5194/tc-18-3495-2024
• Lundquist, J. D., Wayand, N. E., Massmann, A., Clark, M. P., Lott, F., & Cristea, N. C. (2015). Diagnosis of insidious data disasters. Water Resources Research, 51, 3815–3827. https://doi.org/10.1002/2014WR016585
• Lundquist, J. D., Chickadel, C., Cristea, N., Currier, W. R., Henn, B., Keenan, E., & Dozier, J. (2018). Separating snow and forest temperatures with thermal infrared remote sensing. Remote Sensing of Environment, 209, 764–779. https://doi.org/10.1016/j.rse.2018.03.001
• Meromy, L., Molotch, N. P., Link, T. E., Fassnacht, S. R., & Rice, R. (2012). Subgrid variability of snow water equivalent at operational snow stations in the western USA. Hydrological Processes. https://doi.org/10.1002/hyp.9355