A new method for updating snow fields in the NWP models using satellite snow extent based &rsquo;Snow Barrel&rsquo; pseudo-observations as applied to HARMONIE-AROME cycle 43

Hasu, Mikael; Siljamo, Niilo; Kurzeneva, Ekaterina; Rontu, Laura

doi:10.5194/egusphere-2026-1408

Preprints

https://doi.org/10.5194/egusphere-2026-1408

Preprints

02 Jun 2026

| 02 Jun 2026

A new method for updating snow fields in the NWP models using satellite snow extent based ’Snow Barrel’ pseudo-observations as applied to HARMONIE-AROME cycle 43

Mikael Hasu, Niilo Siljamo, Ekaterina Kurzeneva, and Laura Rontu

Abstract. This paper introduces the "snow barrel" method, a new approach to integrate satellite-derived snow observations into NWP models. Snow is a key component of the environment, helping to regulate surface temperature, atmospheric and soil conditions, and playing a key role in the water cycle. However, assimilating satellite snow data into NWP models remains challenging due to resolution mismatches and the complexity of handling snow extent observations. The snow barrel method addresses these challenges by aggregating satellite pixel observations from the EUMETSAT H-SAF H32 intermediate product into 10x10 pixel areas and creating pseudo-observations that align with NWP model scales. Implemented within the HARMONIE-AROME model in the MetCoOp operational domain covering northern Europe, this approach selectively applies snow barrel observations where they conflict with the model's background field, particularly in regions with thin or patchy snow cover. The results demonstrate improved representation of snow cover during the transitional seasons without disrupting areas of solid cover or imposing a significant computational burden. The method effectively combines the spatial coverage advantages of satellite data with the precision of in situ measurements, particularly benefiting areas with sparse ground observations. Although constrained by cloud cover and lighting conditions inherent to optical satellite products, the snow barrel methodology offers a flexible framework that could be expanded to other satellite platforms and potentially adapted for additional surface parameters beyond snow cover.

Received: 13 Mar 2026 – Discussion started: 02 Jun 2026

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.

Download & links

Mikael Hasu, Niilo Siljamo, Ekaterina Kurzeneva, and Laura Rontu

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-1408', Anonymous Referee #1, 07 Sep 2026
General comments
This manuscript describes the use of a new “Snow Barrel” methodology to process satellite snow cover observations from the EUMETSAT H SAF H32 product for assimilation within the HARMONIE-AROME regional NWP system. The manuscript is well written and clearly presented, and the assimilation experiments demonstrate clear benefits to the analysed snow field relative to the use of in situ observations alone. Although it represents a potentially valuable contribution to the NWP and snow DA community, I remain unconvinced that the manuscript fully establishes (1) the novelty of the Snow Barrel concept relative to existing observation-processing approaches for satellite snow cover, (2) the added value of the Snow Barrel methodology itself relative to the underlying satellite observations, and (3) the operational benefit of the approach for NWP forecasting. These issues are discussed in the general comments below:
The manuscript would benefit from a clearer justification of the novelty of the Snow Barrel concept. As presented, the methodology appears largely to comprise established observation-processing steps: aggregation (superobbing) of high-resolution observations to the model scale, empirical conversion to a model variable, and quality-control and data-selection procedures. While the implementation is tailored to the target system, it is unclear what constitutes a substantive methodological advance beyond existing practice. The authors should more clearly distinguish the novel aspects of the approach from established operational processing frameworks.

While the manuscript demonstrates a positive impact from assimilating satellite snow-cover information, it does not isolate the contribution of the Snow Barrel methodology itself. Several aspects of the approach, including the observation-selection criteria, observation-to-SWE conversion, and prescribed error characteristics, appear to be based on heuristic choices, yet little evidence is provided that they have been optimised and demonstrate benefit compared to alternative approaches. A direct comparison between assimilation of Snow Barrels and assimilation of the original H32 product would be particularly valuable, as it would distinguish the impact of the satellite observations from the impact of the Snow Barrel processing. Without such evidence, it is difficult to determine how much added value the proposed approach provides.

Given that Snow Barrels are intended for use in a regional NWP system, I was surprised that the evaluation focuses almost exclusively on the snow analysis. The operational value of improved snow analyses ultimately depends on whether they translate into improved weather forecasts. I would therefore expect some assessment of forecast impacts on relevant variables such as 2 m temperature and humidity. Without such evidence, it is difficult to judge the practical benefit of the approach for operational NWP.

-----------------------------------------------
Specific comments
Line 59-60: “However, current snow data assimilation techniques in NWP remain limited, restricted to using in situ measurements” – this statement is overly restrictive as current snow DA approaches in NWP are not limited to use of in situ observations. Several operational centres also assimilate satellite-derived snow cover products, including the Met Office, ECMWF, NOAA, and potentially others. Please revise this statement and provide a more balanced overview of current operational practice.
Line 82: The manuscript would benefit from a brief explanation of why ground-based observations of snow depth need to be integrated with satellite retrievals in satellite-based snow water equivalent (SWE) products. A short addition would be useful highlighting the limitations of passive microwave retrievals alone, with respect to snowpack sensitivity and retrieval uncertainty under certain conditions, such as wet snow and forest cover.
Line 93: The mention of H43 alongside H31 and H34 is a little unclear. H43 is based on FCI observations from MTG, whereas the current wording could be interpreted as implying that it originates from MSG. Please clarify. In addition, it would be useful to comment on whether the higher spatial resolution of H43 (compared to its MSG-based predecessors) is actually sufficient for many NWP applications in northern Europe.
Line 154: “Additionally, snow cover enhances the surface emissivity.” Please specify the relevant wavelength range or part of the electromagnetic spectrum when discussing changes in emissivity. This statement is currently too general. For example, snow tends to increase emissivity in the thermal infrared, reduce emissivity in the visible spectrum, and can have varying effects in the microwave depending on frequency.
Line 179-180: Please clarify the meaning of “if more than 20% of pixels are classified as either ‘snow’ or ‘no snow’”. Does this mean that the combined total of snow and no-snow classifications must exceed 20% of pixels, or that a single category must individually exceed the 20% threshold (i.e. >20% snow or >20% no-snow)?
Line 183-184: The statement “They are only applied where they conflict with the model's background field...” does not appear to be fully consistent with the decision tree shown in Figure 3. If I have understood the flowchart correctly, observations indicating snow can be accepted when the background also contains snow, provided the background SWE is below 25 kg m⁻². Please clarify or rephrase.
Line 195 and Table 2: It is unclear whether these climatological snow-density values derived from a regional or global climatology? I also wonder whether use of the model background snow density might provide a more physically consistent approach, since it would be location-specific, meteorologically dependent, and evolve with the modelled snowpack state. I note that the manuscript discusses this as possible future work, but some additional justification of the current approach would be helpful.
Line 199-216 and Figure 3: The rationale for several of the observation selection criteria described in lines 199-216 (and illustrated in Figure 3) is not sufficiently convincing to me, and I would appreciate further justification for the chosen thresholds:
First, the use of snowy observations when the background snow water equivalent (SWE) is up to 25 kg m⁻² seems to imply that snow-presence observations can provide useful quantitative information about SWE in regions where the model already contains a shallow snowpack. However, the observations are only sensitive to the presence or absence of snow, rather than the amount of snow. I am therefore unsure why these observations should be expected to improve the analysed SWE. Furthermore, it is not clear that this criterion necessarily avoids erroneous reductions in snow amount. For example, if a deep snowpack exists in reality but the background contains only a shallow snowpack, a partially snow-covered Snow Barrel could result in an observation corresponding to an even lower SWE estimate, further reducing an already underestimated snowpack. Such an error may have more serious consequences than reducing an overly deep snowpack, since it could more rapidly lead to the erroneous removal of snow cover and the associated radiative impacts.

Second, I am not fully persuaded by the decision to reject snow-free observations when the background SWE exceeds 100 kg m⁻². The manuscript appears to assume that, in these circumstances, the model background is more trustworthy than the observation. However, one common model error is the misplacement of snow cover. If the overwhelming majority, or even all, of the pixels within a Snow Barrel are classified as snow-free, this would seem to provide strong evidence that the location is snow-free in reality and could therefore be particularly valuable for correcting the model state. I could more readily justify this criterion in cases where only a small fraction of the constituent pixels are classified as snow-free.

More generally, are these thresholds and selection criteria empirically optimised? If so, the manuscript would be strengthened by presenting evidence that they improve analysis or forecast performance relative to alternative choices. At present, the criteria appear somewhat heuristic, and it is difficult to assess whether they represent the best overall trade-off between accepting and rejecting observations.

Line 228: As noted above, I am not fully convinced that restricting the allowable background SWE range maximises the beneficial value of the observations in the melting zone. It seems possible that potentially informative observations are being excluded in situations where they could provide useful corrections to the model state.
Line 239: The statement “The magnitude of the difference….is approximately 30 kg m⁻²” requires clarification. Does this refer to the cumulative SWE difference across the entire domain, including both positive and negative increments, or some other measure?
Line 274-275: “A key strength of the snow barrels is…” This appears to be a general advantage of satellite snow-cover observations rather than a feature unique to the Snow Barrel methodology. The same benefit would presumably be obtained from the underlying H32 product. The wording should be revised accordingly.
Line 338-339: The specified observation errors appear surprisingly small. Given that the observations provide information primarily on snow presence or absence rather than snow quantity, I would have expected the observation error to be substantially larger relative to the background error. Additional justification of the chosen error values, or evidence supporting their calibration, would be helpful.
-----------------------------------------------
Technical corrections
Line 129: “developed in frames of ACCORD” should be revised to “developed in the framework of ACCORD”.
Line 241-243: The feature being discussed is difficult to identify in Figure 4. Please consider providing a zoomed panel or enlarged view of the relevant region.
Figure 7: Please improve the distinction between RMSE and bias curves. Different colours or line styles, or annotation of the lines, would make the figure easier to interpret.
Citation: https://doi.org/10.5194/egusphere-2026-1408-RC1
RC2:
'Comment on egusphere-2026-1408', Anonymous Referee #2, 10 Sep 2026
Thank you very much for the chance to review, “A new method for updating snow fields in the NWP models using satellite snow extent based ’Snow Barrel’ pseudo-observations as applied to HARMONIE-AROME cycle 43.” The paper presents a novel technique to update snow field in NWP models using a “snow barrel” approach, whereby the method identifies snow cover and non-covered pixels from optical satellites and compares those pixels to the model background fields and the H32 product. If there is a difference, the method proceeds to update the model.
I found the technique to be very interesting, and the figures well done. I believe there are a couple sections that could be better organized and framed for the reader, which I elaborate more on below. Additionally, I understand that the authors found most discrepancy between days of thin snow cover, but this seems inherent on the threshold for snow from the initial SWE prediction (which is currently unclear). More description and justification on this would be helpful.
Introduction
I found the introduction could be overall a bit tighter to the question at hand and could better highlight the gap that researchers are accomplishing. That is to say, why snow is currently tough for NWP models, how/why current methods are insufficient, and why new methods are needed. I found the current structure drifted a bit and it was hard to follow the research gaps they were addressing.

Line 30: I found the introduction to be lacking one critical paragraph that focused on what is a NWP, why accurate SWE is important for NWP, and why it is currently challenging. Paragraph 3 of the introduction does this somewhat, but I think this could be strengthen to justify why snow is important for NWP and why new methods are needed.

Line 37-55: I am not really sure what paragraph 4 is adding. I think Paragraphs 4 and 5 could be better framed around how these observations are or are not currently used by NWP models. This is touched on in Line 61, but I think these sections could be tighter and more direct to say why data assimilation of these observations is insufficient, hence why a barrel method is needed. The current method made it hard to follow some of the acronyms for the reader, so I feel this could be better organized.

Data
I found some of the information here redundant to the introduction (Line 66 and 109 are very similar) and a lot of the information in the data section also belonged into the introduction to strengthen the research gap (for example lines 79 – 89). I recommend the authors streamline the introduction to present the research gap and current challenges and use the data section to focus on the datasets and processing steps used for this analysis.

I noticed that the study period later appeared to be April 2024 to April 2025. Was there a reason for this study period? And the same thing for this study region. Both sets of information would be stronger in the data section.

Model and Methods
I appreciated figures 1 and 2 to guide the description of the analysis. I think some subsection headings would also benefit the reader as it was at times hard to follow the various steps to develop the model runs. I also found that some decisions appeared to be less justified, and it would benefit the reader to know the impacts of those decisions.

Line 161: typo for “similar”

Line 171: what does all members mean?

Line 179: Where did the 20% for snow or no snow come from? Other papers have cited 40% as a threshold for snow or no snow (Hall et al. 1995; 10.1016/0034-4257(95)00137-P). This is an example of a decision that was challenging to know how it was made and what impacts it might have on the analysis.

Line 184-188: I am bit confused. I understand that snow barrel values are only applied when the model prognostic variable disagrees with the snow barrel for snow or no snow, and the model variable is SWE, but this implies that there is a cut-off for SWE that the authors deem “snow.” Was this threshold anything above >0? Could really small SWE values (e.g., <1) also be considered no snow, in which case barrel values that indicate no snow for these respective values get assimilated into the model?

Line 190: It would be helpful to understand the breakdown of how the authors got to 0-10 cm snow depth. Is there any prior literature to back up this claim?

Line 214-215: Where do these standard deviations come from? Were they calculated or assigned or something else?

Results
Overall comment: do forested areas impacted utility of the barrel method? Given that optical cannot penetrate through the canopy these values would indicate no snow when the model would produce SWE. Could incorporating these values actually make the snow fields worse?

There are some “Cloud Gap Filled” optical products. Do the authors think these could provide utility?

For Figure 7 (Lines 259-265): Can you provide the RMSE and bias differences between Barrels and Reference across all months and then separated by season? It seems like the barrel method is particularly useful in the spring when clouds are scarce and snow is thin and patchy, but I didn’t see these explicit numbers, and I found it very hard to figure out the differences between barrel and reference in the plot because the points were very small.

I think some subsection headings for the Results would help here too.

Discussion
Line 278: The authors bring up “computationally efficient” multiple times, but I couldn’t find this comparison. Can the authors provide a metric for this efficiency compared to other methods?

Some of the topics in the Discussion belonged in the methods or the results. For example, Lines 296-300, for example. Lines 311-318 was also unclear how it related to interpretation of the study as it was mostly descriptive. I’d recommend the authors tighten the Discussion and work on reorganizing for the key takeaways, strengths, and limitations.
Citation: https://doi.org/10.5194/egusphere-2026-1408-RC2
RC3:
'Comment on egusphere-2026-1408', Anonymous Referee #3, 17 Sep 2026
My overall assessment is that the paper presents a potentially useful operational approach, but the evidence currently provided is insufficient to establish robust improvements in snow analyses or demonstrate benefits for NWP forecasts. My main scientific concern is that the method is intended primarily to improve shallow and partial snow cover, yet several of its assumptions may be particularly uncertain under these conditions. First, several influential methodological choices (including the 10 × 10 pixel barrel size, the 20% minimum classified-pixel threshold, the maximum pseudo-depth of 10 cm, and the background SWE thresholds of 25 and 100 kg m⁻²) are prescribed without sufficient justification or sensitivity analysis. Although these choices may be operationally reasonable, they directly influence the analysis increments and resulting snow fields. Their robustness therefore needs to be demonstrated. Second, shallow and partial snow cover can be difficult to identify using optical observations. A satellite “no snow” classification may indicate genuinely snow-free ground, but it may also reflect snow obscured by vegetation or unresolved within mixed pixels. The approximately 1 km resolution, followed by further aggregation into barrels, also raises questions about how adequately the observations represent shallow snow heterogeneity. Third, shallow snow can change substantially between the satellite overpass and assimilation the following morning, particularly during active melt or snowfall in shoulder seasons. The main quantitative evaluation covers one annual cycle and does not separately assess these rapidly changing conditions, limiting the evidence for the method’s performance during such events. Fourth, monthly climatological density values may poorly represent shallow, newly fallen snow, particularly during warmer seasons when the prescribed density reflects denser seasonal snow. The resulting uncertainty in the conversion from pseudo-depth to SWE should be assessed. Finally, for shallow snow, changes in surface albedo may be more consequential for near-surface weather than changes in SWE. Assimilating SWE alone may therefore be insufficient to address snow-related shortcomings in operational NWP systems. My other comments are listed below:

The Introduction does not clearly establish the novelty of the proposed approach relative to previous snow-assimilation efforts. The most relevant prior work is introduced only in the Methods section. Much of the Introduction explains basic snow variables, while insufficient attention is given to existing snow-cover assimilation approaches, their limitations, and the specific gap addressed by the proposed method.

Method and data section is insufficiently reproducible in its present form. Please make sure all the defined variables and processing step are defined clearly. For example:
Please define precisely how barrel snow fraction is calculated and how invalid pixels are handled. Now, the paper does not explicitly say whether the snow fraction is subsequently calculated relative to all 100 pixels or only the valid snow/no-snow pixels

Also, clearly state how the assimilation workflow handles the impacts of snow removal on other prognostic variables

The Results section would benefit from stronger quantitative support for the claims, particularly under thin or partial snow cover.

Although Figure 7 presents RMSE and bias, several key conclusions rely primarily on visual interpretation. For example, the closer agreement with satellite-observed snow coverage in southern Finland is assessed visually in the discussion of Figure 5 (lines 235–245). Similarly, the concentration of accepted barrels along the snowline in Figure 4 is interpreted as indicating their added value in these regions (lines 225–230). However, because barrel acceptance explicitly depends on background SWE and satellite classification, this spatial distribution partly reflects the selection rules and does not, by itself, establish where assimilation improves accuracy. I suggest supplementing these examples with quantitative comparisons against independent snow-cover observations, reporting both missed snow and false snow detection with and without barrel assimilation. Where suitable observations are available, separating performance under shallow versus deeper snow, or partial versus complete snow coverage, would help substantiate the claimed benefits under transitional conditions.

For Figure 7, please clearly identify the observations used to calculate RMSE and bias, explain how these statistics were calculated, and clarify whether these same observations were assimilated.

Figure 7 is particularly difficult to interpret because the absolute differences between experiments are small relative to the large event- scale RMSE peaks. A difference plot or relative improvement plot would be much more informative.

Please also report seasonal RMSE and bias values for both experiments, sample sizes, and uncertainty estimates for their differences to establish the magnitude and robustness of the improvements beyond the qualitative changes illustrated in the maps.

The primary quantitative evaluation presented in the manuscript is limited to a single annual cycle, from April 2024 through March 2025. Although, the manuscript briefly describes preliminary experiments conducted over the CARRA-East domain during 2017, but no quantitative results or sufficiently detailed methodology are presented to support the analysis.

Several advantages attributed to the snow-barrel approach in the Discussion are plausible but not directly demonstrated by the results presented. In particular, lines 282–285 state that aggregation reduces classification errors and ensures greater representativeness and reliability. However, the comparison between analyses with and without barrels evaluates the combined effect of adding satellite information through the proposed method; it does not isolate the benefit of aggregation itself. Please provide supporting evidence, such as an independent evaluation of aggregated versus unaggregated or alternatively aggregated H32 information, or qualify these statements as potential advantages. Similarly, lines 289–291 suggest that using barrels under partially cloudy conditions significantly enhances snow-data reliability. An evaluation of accuracy as a function of valid-pixel fraction would help support this claim. Otherwise, please distinguish improved availability under partial cloud from demonstrated improvements in reliability.

The discussion should engage more deeply with the central compromise of the proposed method (meaning it gains operational compatibility by turning a categorical areal observation into a point-like quantitative SWE pseudo-observation, but that transformation introduces substantial representativeness and structural uncertainty.)

The statement that current NWP snow assimilation is “restricted to using in situ measurements” is contradicted later by the discussion of prior satellite snow assimilation and should be corrected.

Line 30: Please clarify that both snow depth and density vary spatially and jointly determine SWE. The current wording emphasizes density variability without acknowledging spatial variability in depth.

Line 38: Please clarify which in situ snow observations are being discussed (for example, snow depth, SWE, or snow cover).

Line 46: Please explicitly identify the principal limitations of microwave-derived SWE estimates rather than referring generally to “certain limitations.”

Line 53- 55: seems out of place and break the flow

The sentence beginning “Refine the assimilation scheme…” is grammatically incomplete.

SWE units should be consistently formatted as kg m⁻²

Snow depth observations” should not be called “satellite snow depth observations” when the satellite measures snow extent and depth is empirically assigned. Same about SWE. Reframe these derived values as empirical pseudo-observations, not satellite-derived SWE or snow depth.

The text alternates between “snow field,” “snow cover field,” and “SWE field”; these are not interchangeable.

The computational efficiency claim should report actual runtime or percentage overhead.

“Significant” should be avoided unless statistical significance was tested.
Citation: https://doi.org/10.5194/egusphere-2026-1408-RC3

Mikael Hasu, Niilo Siljamo, Ekaterina Kurzeneva, and Laura Rontu

Viewed

Total article views: 361 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
211	116	34	361	22	20

HTML: 211
PDF: 116
XML: 34
Total: 361
BibTeX: 22
EndNote: 20

Views and downloads (calculated since 02 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	95	30	13	138
Jul 2026	29	31	3	63
Aug 2026	13	14	2	29
Sep 2026	71	39	14	124
Oct 2026	3	2	2	7

Cumulative views and downloads (calculated since 02 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	95	30	13	138
Jul 2026	29	31	3	63
Aug 2026	13	14	2	29
Sep 2026	71	39	14	124
Oct 2026	3	2	2	7

Viewed (geographical distribution)

Total article views: 328 (including HTML, PDF, and XML) Thereof 328 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 02 Oct 2026

Short summary

We studied a new way to use satellite images to update snow fields in numerical weather models. The idea is to group many small pixels into simple “snow barrels,” which fit the model scale better. The method does not affect areas with solid snow, since satellites cannot see true snow depth and each barrel can only add a thin 10-centimetre layer of snow. The results show that barrels capture thin and changing snow better than using only in-situ observations, especially in springtime.


Total:	0
HTML:	0
PDF:	0
XML:	0