the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 13 Mar 2026
Assessing VIIRS constellation seasonal snow cover over the French mountains with Sentinel-2
Abstract. Remote sensing observations of snow-covered area from moderate-resolution (100–1000 m) optical sensors are critical for water resource monitoring and climate-related applications. Many studies have relied on MODIS snow products, but NASA plans to stop science data collection from MODIS instruments in the next two years. The Visible Infrared Imaging Radiometer Suite (VIIRS), designed as follow-on instrument to MODIS, has similar characteristics and is currently onboard three satellite platforms with daily revisit. Several agencies now distribute operational snow cover products based on VIIRS data, yet independent evaluations of these products remain scarce, limiting their adoption by the snow science community. Here, we assess NASA VIIRS snow cover products over several mountain ranges in France using Sentinel-2 snow cover products as a high-resolution reference. We also evaluate a near-real-time VIIRS snow cover product developed by Météo-France over metropolitan France. The analysis covers an entire winter season and examines performance across contrasting topographic settings. Across the different products, we find an average bias close to zero and a root mean square error ranging from 10 to 15 % for snow cover fraction, while the corresponding binary snow classification achieves a F1-score of 92–93 %. However, uncertainties can reach 30 % for challenging observation conditions, including mixed pixels, forested areas, and north-facing slopes. Finally, we demonstrate that combining observations from multiple VIIRS platforms effectively reduces cloud cover without degrading snow cover retrieval quality.
Competing interests: At least one of the (co-)authors is a member of the editorial board of The Cryosphere.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(9420 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 24 Jun 2026)
- RC1: 'Comment on egusphere-2026-1122', Anonymous Referee #1, 07 Jun 2026 reply
-
RC2: 'Comment on egusphere-2026-1122', Anonymous Referee #2, 20 Jun 2026
reply
This manuscript provides a useful and timely evaluation of operational VIIRS snow cover products against a high-resolution Sentinel-2 reference, with a well-designed stratification by land cover, topography, and sensor geometry. The work is relevant given the impending decommissioning of MODIS. However, there are several points that I believe warrant attention before the manuscript is published.
Section 5.1: Table 3 suggests that the choice of cloud mask has an effect on product performance that is comparable to that of the retrieval algorithm itself. The influence of the cloud mask on snow retrieval is therefore an important consideration, and it is closely tied to the well-known problem of cloud/snow discrimination. Since the VIIRS products evaluated here originate from different sources and apply different cloud masks, and since the present study does not explicitly isolate the effect of the cloud mask, this represents a limitation of the analysis. Where feasible, I would encourage the authors to compare and evaluate the accuracy of the cloud masks across the different product sources and to assess their influence on the computed snow cover area. At a minimum, this issue should be discussed.
Section 5.3: The multi-platform compositing clearly improves data availability, and the reported reductions in cloud cover are substantial. However, given that the three platforms share the same orbital plane with overpasses only about 50 minutes apart, it would be helpful to disentangle the mechanisms responsible for this improvement. Specifically, how much of the gain arises from cloud advection over the roughly 50-minute window, and how much from differing viewing geometries or other factors? A more important concern relates to Figure 5 and the comparability of the products shown. Section 3.1 states that NASA does not currently distribute a JPSS-2 product, yet Figure 5 presents single-platform series labelled MF-FSC-VNP-L3, MF-FSC-VJ1-L3, and MF-FSC-VJ2-L3. As I understand it, these single-platform series are all derived from the Meteo-France pipeline using one platform at a time, and are therefore not the same products (VNP10A1, VJ110A1) evaluated in Section 5.1. I suggest that the authors clarify explicitly how comparability is ensured between the single-platform and multi-platform composites, and between these and the NASA products evaluated elsewhere.
Section 6.1: The statement that Landsat-based products "are only available in the United States" is inaccurate, as Landsat provides global coverage. I suspect the authors intend to say that the specific high-resolution reference datasets or studies they cite are concentrated in the United States. The sentence should be reworded accordingly. In addition, since Landsat data are in fact globally available, the authors should explain why Landsat-based reference data were not used in this study.
Section 6.2: I have a substantive concern with the MOD10A1 versus VNP10A1 comparison. Terra (MOD10A1) is a morning-orbit platform (descending node around 10:30), whereas VIIRS daytime overpasses occur in the early afternoon (around 13:30), as does Aqua (MYD10A1). On the grounds of overpass-time matching, VNP10A1 is more appropriately compared with MYD10A1. The comparison with MOD10A1 remains informative but should be framed accordingly. This bears directly on the interpretation in the manuscript. The authors attribute the finding that MOD10A1 yields more snow than VNP10A1 to a less restrictive cloud mask. This attribution is likely incomplete. Because the morning overpass precedes much of the intraday accumulation of positive degree-hours, the morning platform may simply observe more snow than the early-afternoon platform owing to within-day melt, independent of cloud masking. Notably, the authors themselves observe that studies reporting the opposite result used MYD10A1, which indicates they are aware of the Terra/Aqua distinction. The same overpass-time effect is discussed in the China-based VIIRS evaluations they cite. I recommend that the authors add a comparison with MYD10A1.
Section 6.4: As the authors themselves demonstrate in Figure 7, the NDSI-FSC relationship differs markedly between forested and open areas. Yet the methodology (Section 4.1) applies a single linear relationship to convert NDSI to FSC everywhere. Applying one relationship across land cover types necessarily introduces a systematic error in forested pixels. I encourage the authors to strengthen the associated uncertainty analysis, for example by fitting separate NDSI-FSC relationships for forested and open terrain (or finer land-cover classes) and reporting the resulting change in accuracy. This would substantiate the discussion in Section 6.4.
Section 6.5: Topography exerts a strong influence on moderate-resolution snow retrievals, particularly in high mountains, and the discussion in this section is somewhat brief relative to the importance of the effect. The manuscript's own results, such as omissions on northern slopes and declining performance with increasing slope, are consistent with prior scale-effect studies of moderate-resolution snow retrieval (e.g., work over the Tibetan Plateau) that demonstrate a strong dependence of retrieval accuracy on slope and aspect. Importantly, such studies also show that the locally optimal NDSI threshold varies with slope and aspect because of illumination conditions, which directly challenges the use of a single, fixed NDSI-FSC relationship and threshold in this study. This connection, and its implications for the fixed-threshold approach, should at least be discussed.Citation: https://doi.org/10.5194/egusphere-2026-1122-RC2
Data sets
Data of the article "Assessing VIIRS constellation seasonal snow cover on French mountains with Sentinel-2" Nicola Imperatore, Simon Gascoin, Matthieu Lafaysse, Marie Dumont, Adrien Mauss, Stéphane Guével, and Jean-Baptiste Hernandez https://doi.org/10.5281/zenodo.18157029
Model code and software
Code of the article "Assessing VIIRS constellation seasonal snow cover on French mountains with Sentinel-2" N. Imperatore https://doi.org/10.5281/zenodo.18773106
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 670 | 600 | 69 | 1,339 | 57 | 89 |
- HTML: 670
- PDF: 600
- XML: 69
- Total: 1,339
- BibTeX: 57
- EndNote: 89
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors analyze VIIRS snow cover product by leveraging a powerful approach of using coincidence Sentinel-2 snow cover products as a high-resolution reference dataset for error analysis. This approach is powerful because it enables a very large sample size for the analysis. The intent and overall architecture of the study is sound and a valuable contribution to the snow science community but, to recommend for publication I recommend major adjustments to the implementation details of the analysis and the structure of the paper.
More text needs to be included that defends the use of a NDSI based fractional snow cover truth data set over a physically based model. Evaluating a product that uses NDSI makes sense, but comparison to a NDSI based fractional snow cover "truth" product concerns me given the existence of physically based models, including open source versions, that outperform NDSI for calculating the fractional snow cover of multi spectral pixels. NDSI does not contain information regarding the spectral signature of snow and various non-snow mixed pixels can have the same NDSI values as snow-covered pixels. Various studies have shown that NDSI is less accurate than spectral unmixing techniques when estimating fractional snow cover. I’d prefer a physically based spectral unmixing truth data set for the highest fidelity truth set from Sentinel-2 and for the paper to have the most impact, but If the author wants to stick to a NDSI based truth dataset, they should better explain the choice. There are always trade offs with choice of validation data set, so I would like to see more discussion on the reason behind this choice and the known pros and cons (cons are omitted currently) so the reader can better contextualize the results.
Line 170 - Please explain and defend the choice to reproject and resample at the same spatial resolution as the original data set instead of coarsening to account for geolocation uncertainty. Coarsening is a typical approach to eliminate concern on geolocation accuracy that does not impact overall error analysis metrics (especially at the sample sizes of this paper) but can reduce the impact of geolocation errors on results. Also, in Section 4.4 I recommend only reprojecting the truth data from Sentinel-2 to the VIIRS projection. Please explain the reasoning to choose a third projection, a non-equal area projection to reproject all the data into for a fractional snow cover area analysis. The point is made that the errors (in terms of area size) introduced are small, but why not keep the VIIRS data in their original projection and only retroject the high spatial resolution truth data to the VIIRS projection for error analysis? This choice to reproject both datasets is error inducing. I strongly recommend keeping the dataset that is being evaluated in the same projection that it is delivered in, and coarsening prior to analysis.
Line 220 - “We binarize the FSC values into "snow" and "no-snow" classes using a threshold of 50% for both the reference and evaluation datasets.” I disagree with this choice. Pixels with substantive snow cover are classified as snow free pixels for analysis. Why was the binary snow cover threshold not set at FSC>0? It is possible to accurately detect snow in Sentinel data well below 50% snow cover and I’d expect the truth dataset to be a binary mask of all pixels that include snow. This is also a reason for such a high representation of TN’s in the overall dataset. It seems that the snow covered pixels are a significant portion of the TN dataset with the 50% cutoff. I disagree with this decision and would like to see the snow cover fraction threshold that binarizes between snow and now snow set at a FSC>0 or similar. This choice, alongside the choice to reproject both the truth and VIIRS data, concerns me on the quality of the output analysis. The results in the first few rows of Table 4 seem to invalidate the decision to use a threshold of 50% for snow / no snow cover as you see significantly worse performance in this range of valid snow covers from 1-99% FSC, which is an artifact of the choice to call snow covered pixels with less than 50% snow cover, snow free pixels, alongside the reprojecting and lack of coarsening to account for geolocation errors in the analysis.
Use of accuracy metirc - I disagree that this is a valuable measure of algorithm performance given the unbalanced nature of the data set and relative ease algorithms have in distinguishing snow free pixels from pixels with factional snow cover, FN and FP are far more challenging issues that are important for understanding the value of a snow cover product. I’d rather only see F1 score in the graphs in figure 4 and the detailed performance analysis.
Line 312 and use of VIIRS to validate MODIS - I disagree with this approach. The same truth data set (Sentinel) should be used to access MODIS and VIIRS and make determinations on the relative performance of the two. The current chained approach to error analysis introduces unnecessary error propagation into the analysis.
Discussion Section - Many results are presented directly in the discussion section. Recommend moving these important results into the results section and explaining the approach to these analysis in the methods section.
Line 411 - recommend removing accuracy as a main result metric.
Technical corrections:
- Suggest rewording sentence from line 18-20 for clarity.
- Line 35 - recommend changing “guaranteed” to “operationally planned” or similar, there is risk with any operational system that it does not meet expected mission lifetime and there can be a gap in coverage before a replacement is fielded.
- Line 67 - contradicts with Table 2 - reviewed and ASO snow depth data, not SWE was used for the validation. Suggest changing “which provides lidar based snow water equivalent (SWE) measure” to “which provides 3m spatial resolution lidar based snow depth measurements that were converted into binary snow masks.”
- Line 227 - after “we also report the F1 score “ suggest adding “which is the harmonic mean of omission and commission”.