Lagrangian reconstruction of snow accumulation and loss on Antarctic sea ice

Campbell, Ethan C.; Riser, Stephen C.; Webster, Melinda A.

doi:10.5194/egusphere-2026-2089

Preprints

https://doi.org/10.5194/egusphere-2026-2089

Preprints

21 Apr 2026

| 21 Apr 2026

Lagrangian reconstruction of snow accumulation and loss on Antarctic sea ice

Ethan C. Campbell, Stephen C. Riser, and Melinda A. Webster

Abstract. Snow on Antarctic sea ice strongly influences the thermodynamics and freshwater balance of the coupled sea ice–upper ocean system. Yet understanding of its temporal and spatial variations remains limited by sparse observations, large uncertainties in remote sensing retrievals, and idealized model representations. We introduce a new open-source numerical model, the University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model, that simulates the mass and bulk density evolution of snow on sea ice in the Southern Ocean over 2003–2025. Hourly reanalysis snowfall is accumulated along Lagrangian sea ice drift trajectories determined from remotely sensed ice motion fields. The single-layer model incorporates physically and empirically informed parameterizations of key erosion and transformation processes, including surface and wind-blown snow sublimation, lead trapping, rain- and non-rain-related melt, compaction from wind and overburden pressure, and the large-scale effects of sea ice convergence and divergence. Model parameters are calibrated using snow buoy measurements from the Weddell Sea. The resulting reconstruction indicates that over one-third of annual snowfall intercepted by Antarctic sea ice is lost to the atmosphere, ocean, or to melt processes prior to complete sea ice melt, with blowing snow sublimation as the dominant sink. Comparison with satellite snow depth retrievals further suggests that widespread snow-ice formation consumes 49–60 % of the remaining snow. Overall, we infer an annual meteoric freshwater input to the Southern Ocean originating from snow on sea ice of 237 mSv, equivalent to more than half of the freshwater flux associated with circumpolar sea ice melt.

Received: 13 Apr 2026 – Discussion started: 21 Apr 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

Ethan C. Campbell, Stephen C. Riser, and Melinda A. Webster

Status: final response (author comments only)

Subscribe to comment alert

RC1: 'Review on egusphere-2026-2089', Anonymous Referee #1, 28 May 2026

Review of " Lagrangian reconstruction of snow accumulation and loss on
Antarctic sea ice" by Ethan Campbell et al.

This manuscript addresses an important and timely problem: how snow on Antarctic sea ice modulates the Southern Ocean freshwater budget and the sea-ice mass balance. The topic is clearly worth publication-quality study. However, in its present form, the paper overstates what its modeling framework can support. The central claim, that the reconstruction implies widespread snow-ice formation consuming roughly 49–60% of remaining snow and yielding an annual meteoric freshwater input of 237 mSv, more than half the circumpolar freshwater flux from sea-ice melt, is not robustly demonstrated given the model design, forcing uncertainty, observational representativeness issues, and the authors’ own stated omissions. The manuscript itself explicitly acknowledges that it does not represent snow-ice conversion, meltwater/liquid-water effects, or interannual memory on multiyear ice, yet those omissions directly target the same processes the paper later elevates into headline conclusions.
The most serious problem is a mismatch between the simulated quantity and the inferred quantity. The model is calibrated against snow-buoy surface-height-change records, which the manuscript treats as a closer analogue to snow accumulation because basal snow-ice conversion is not represented. But the paper then compares this reconstructed accumulation to satellite snow depth products and interprets the residual largely as snow-ice formation. That residual also contains biases from satellite retrieval error, annual re-initialization, multiyear-ice memory loss, buoy representativeness bias, the omission of liquid-water and saline-snow processes, and forcing uncertainty. In other words, the paper turns an underconstrained residual into a physical estimate of Antarctic-wide snow-ice conversion. That is not convincing as presently shown.
A second major weakness is that the regions most central to the manuscript’s calibration and strongest claims, especially the Weddell Sea, are precisely where perennial and multiyear processes matter most. The draft states that parcels are initialized every 15 February and then reset “without memory of the prior year,” and that this “limits the model’s influence on snow over multiyear sea ice.” It also initializes pre-existing snow density to a spatially uniform 320 kg m⁻³. Yet independent Weddell literature shows that older/perennial ice in the western Weddell carries distinctive snow and meteoric-ice signatures, that Snow Buoy records often represent second-year and even 2-to-4-year snow regimes, and that regionally adjusted densities differ materially between perennial and seasonal ice. This means the initialization strategy likely imprints Figure 8f, weakens autumn conditions, and undermines the western/northern Weddell comparisons.
Third, the forcing and validation strategy is too fragile for the paper’s quantitative claims. The manuscript uses ERA5 snowfall as the core depositional forcing, while also noting a spread of about 200 mm yr⁻¹ among Southern Ocean reanalyses and then tuning snowfall upward by 32% during calibration. ERA5 itself provides uncertainty information through an ensemble, but the draft does not propagate forcing uncertainty through to the headline freshwater-flux estimate. Similarly, the passive-microwave product used for both annual initialization and later evaluation is subject to dry-snow-only applicability and a 50 cm retrieval limit, while comparisons to Antarctic observations show substantial biases for older ice regimes. CryoSat-2 products are likewise affected by saline basal snow, radar-penetration uncertainties, and sparse validation. These are not minor caveats. They are central to the interpretation of Figure 8 and to the inferred snow-ice conversion.
My recommendation is reject in present form, with encouragement to resubmit only after substantial redevelopment. The work is promising, but the current paper makes quantitative claims that outstrip the demonstrated reliability of the method.

Major methodological weaknesses
The most important scientific omission is the absence of explicit meteoric-ice formation pathways from the prognostic snow budget: snow-ice formation from flooding/seawater infiltration, superimposed ice formation from meltwater refreezing, and the accompanying salinity effects. The manuscript acknowledges that snow-ice conversion is not represented explicitly and later estimates it from model-minus-satellite differences. That is a methodological shortcut, not a demonstrated retrieval. Independent Weddell Sea studies published in Arndt (2022), show that these omitted processes are not peripheral. In the northwestern Weddell, superimposed ice and snow ice averaged about 0.11 ± 0.11 m and 0.22 ± 0.22 m, respectively, together amounting to 3%–54% of total ice thickness. A broader Weddell Snow Buoy/SNOWPACK analysis found mean maximum snow-ice thicknesses of 16 cm across all analyzed buoys, 35 cm on snow-ice-bearing floes, and about 27% of snow accumulation converted to snow ice, with much larger values possible regionally. Those are order-of-magnitude effects for a snow-depth study, not second-order corrections.
This omission also matters specifically for the paper’s freshwater interpretation. Snow ice forms from seawater-soaked snow and is therefore saline; superimposed ice forms from internal melt/refreezing and changes both snow depth and snow thermal properties. The manuscript argues that the “mass budget of total meteoric freshwater” should remain robust despite uncertainties in conversion processes. I do not think that is shown. The partitioning, timing, vertical location, salinity signature, and ocean-coupling consequences of the freshwater pathway depend directly on whether snow remains snow, becomes snow ice, becomes superimposed ice, or is redistributed into ridges or leads. In Antarctic conditions, those pathways are central to the physical meaning of the freshwater flux, not a cosmetic post-processing issue.
A closely related weakness is the neglect of liquid-water effects, including melt metamorphism, infiltration, enhanced destructive metamorphism, reduced compactive viscosity in wet snow, and melt-freeze evolution. The manuscript explicitly states that these effects are not represented and that rainfall and snowmelt are assumed to refreeze rapidly. Yet Antarctic observations and process studies show that internal melt-freeze cycles, superimposed ice formation, saline basal layers, and wet-snow metamorphism materially affect snow depth, density, radar penetration, and surface energy balance. In the Weddell Sea, internal melt-freeze cycles and superimposed ice are latitude dependent, and recent observations show superimposed-ice thicknesses up to and above 0.1 m regionally, with associated changes in conductive properties and internal ice melt. Under those circumstances, using a dry, single-layer framework to infer continent-scale snow-ice conversion from residuals is too optimistic.
The annual re-initialization strategy is another major liability. The manuscript states that parcels are initialized every 15 February from AMSR-E/2 snow depth, with pre-existing snow density set to 320 kg m⁻³, and then reset the next year “without memory of the prior year.” The same section concedes that this limits the model’s influence on snow over multiyear sea ice, and later the paper acknowledges a residual budget drift reflecting net depletion of snow originating from multiyear sea ice. This is especially problematic because western Weddell multiyear/perennial ice is one of the few Antarctic sectors where older ice is common and thick, and because Snow Buoy analyses in the Weddell indicate that many records represent second-year ice and some represent 2-to-4-year snow regimes. A model that forgets prior-year snow and meteoric-ice history cannot be treated as fully credible in exactly the sector where its validation story is strongest.
That logic also likely explains the suspiciously homogeneous autumn snow-density field in Figure 8f. If snow density is initialized everywhere to 320 kg m⁻³ in mid-February and only later evolves through simplified wind and overburden parameterizations, then the relatively uniform autumn density structure is at least partly an initialization artifact. That would be particularly important in the Weddell Sea, where regionally adjusted perennial-versus-seasonal snow densities of 340 and 264 kg m⁻³ have already been used in the literature following Arndt (2022a), precisely because older/perennial and seasonal ice regimes differ materially. The paper should not present autumn density heterogeneity as an unconstrained emergent result when the initial condition is spatially uniform.

Data sources and validation
The ERA5 snowfall forcing is a critical uncertainty that the manuscript underplays. The draft itself notes that Southern Ocean reanalyses agree in broad spatial patterns but differ by roughly 200 mm yr⁻¹ in snowfall magnitude. By simple conversion, that corresponds to roughly 0.6–0.8 m of annual snowfall depth at densities of 340–264 kg m⁻³, before compaction and losses, a very large perturbation relative to the climatological snow depths discussed in the paper. The authors then tune snowfall deposition upward by 32%, which strongly suggests compensation between forcing bias and process tuning. ERA5 is an excellent product, but it is still a reanalysis with uncertainty information that should be exploited, not treated as essentially fixed. Studies comparing reanalysis snowfall forcings have shown that comparatively modest snowfall differences can materially alter modeled snowpack and sea-ice evolution.
The snow-buoy treatment is also not persuasive enough for a circumpolar calibration framework. The manuscript is aware that Snow Buoys measure changes in the surface interface relative to a deployment baseline, and it explicitly uses that fact to argue that buoy data are a better analogue for modeled snow accumulation than for true snow depth. But the paper does not adequately grapple with the representativeness cost of that choice. The AWI Snow Buoy methods paper warns that time series from level ice may not represent the overall snow-mass distribution, notes that snow in ridges can be substantially deeper than on level ice, and shows that snow-ice and superimposed-ice formation can change the pure-snow layer without much change in measured surface elevation. A later Weddell analysis goes further and states directly that Snow Buoys are usually deployed on level ice and may therefore underestimate snow accumulation and snow-ice formation relative to the broader floe and ridged pack. So the calibration data are neither true snow depth nor clearly representative of grid-cell snow mass. That is a serious observation-operator problem, not just random noise.
The satellite comparisons in Figure 8 are therefore meaningful only as qualitative consistency checks, not as quantitative proof of Antarctic-wide snow-ice conversion. For AMSR-E/2, the product documentation states that the snow-depth algorithm is applicable only under dry-snow conditions and has an upper retrieval limit of 50 cm because of limited penetration depth, while Antarctic passive-microwave evaluations note difficulties over multiyear ice and show substantial underestimation for older algorithms or comparable products. The manuscript also uses AMSR-E/2 snow depth to initialize the model every year, which makes later comparison with AMSR-E/2 only partly independent, especially in autumn and in multiyear-ice regions most sensitive to the initial condition. That circularity needs to be stated plainly.
CryoSat-2 is not a clean arbiter either. Kacimi and Kwok show that high-salinity basal snow can bias CryoSat-2 freeboards by at least a few centimeters, that the resulting Antarctic ice thickness/volume can shift materially under plausible corrections, and that validation remains difficult because seasonally and regionally diverse field datasets are lacking. Fons et al. also stress that Antarctic radar-altimetry snow-depth retrievals carry substantial uncertainty due to radar penetration, external snow assumptions, and sampling biases in ship-based comparisons. In a saline Antarctic snowpack, CryoSat-2 residuals cannot be straightforwardly mapped to snow-ice formation without stronger independent constraints.

Interpretation of figures and conclusions
Figure 3 does not by itself demonstrate robust generalization. It shows that the calibration procedure improves the fit to the chosen buoy split, but the validation set still carries a notable negative bias, and the split is only one random partition of 39 buoys from a single basin. That is not enough to establish that the tuned parameterization is transferable to the circumpolar Antarctic. A blocked cross-validation by sector, ice regime, and buoy type would be far more convincing than a single random 75/25 split. As shown, Figure 3 mainly demonstrates that the model can be tuned to Weddell buoy behavior, not that the inferred Antarctic-wide process partition is correct.
Figure 5b is one of the least convincing parts of the paper. The plotted cumulative process contributions imply very large compaction and very rapid melt-related losses over short intervals, yet those processes are not independently validated at the event scale. The text itself admits that the model likely produces an unusually large superimposed-ice amount for buoy 2014S9, about 32 cm , and that liquid-water effects are not explicitly represented. Observational work in the Weddell records mean superimposed-ice thicknesses closer to 0.11 ± 0.11 m, with regional increases but still much smaller typical values. I therefore do not find the partitioning in Figure 5b physically trustworthy enough to support detailed causal claims about what fraction was compaction, what fraction was warm-air melt, and what fraction was rain. This panel is presently more illustrative than evidentiary.
Figure 8 is the paper’s inferential centerpiece, but it is not strong enough for the claims placed on it. The manuscript subtracts satellite snow depth from reconstructed net snow accumulation and interprets the resulting 20–30 cm positive residual over much of Antarctica in spring as evidence of widespread snow-ice formation. That residual, however, includes at least six components: true snow-ice conversion, superimposed-ice effects, liquid-water metamorphism and infiltration, AMSR/CryoSat retrieval error, annual initialization artifacts, and representativeness mismatch between level-ice point observations and gridded satellite fields. The authors themselves acknowledge that the CryoSat-2 and AMSR-derived inferences are not perfectly consistent. Under those circumstances, “widespread snow-ice formation of that magnitude” should be stated as a hypothesis to be tested, not as a strong result.
The final quantitative conclusion, 237 mSv of annual meteoric freshwater input, “more than half” the sea-ice-melt flux, is therefore too definite. Even if the broad message that snow-on-sea-ice freshwater matters is almost certainly correct, the exact number depends on a chain of assumptions that has not been sufficiently stress-tested. The manuscript should at most present this as a provisional estimate with large uncertainty bounds, after propagating forcing, process, initialization, and observation uncertainty. In its current form, the headline number reads as far more secure than the evidence warrants.

Required revisions and additional analyses
To become publishable, the manuscript would need a stronger separation between what is directly simulated, what is empirically constrained, and what is only inferred by residual. In practice, that means the authors should either incorporate explicit snow-ice and superimposed-ice physics, including liquid-water and saline-snow effects, or substantially soften the inferential claims and reframe the snow-ice discussion as a bounded hypothesis. The present hybrid approach is too internally inconsistent for such strong conclusions.
The paper also needs explicit multi-year-memory experiments. At minimum, a second configuration should carry snow state, density, and meteoric-ice proxies across years, especially for perennial/multiyear sectors in the western and northern Weddell. Without that, the reader cannot assess how much of Figure 8 and the freshwater budget are artifacts of the annual reset.
The forcing should be tested as an ensemble, not a single tuned realization. The obvious first step is to bracket ERA5 with at least one or two alternative reanalyses and to use ERA5’s own uncertainty information where possible. The freshwater-flux headline should then be reported as an uncertainty range propagated from forcing, calibration, and retrieval uncertainty. A single deterministic number is not defensible at present.
Finally, the validation framework needs an explicit observation operator. Snow Buoys measure one thing, passive microwave another, radar altimetry another, and ship or ice-core observations another. Those cannot be interchanged casually. The authors should distinguish among surface-interface change, pure-snow depth, snow mass, meteoric-ice thickness, and total freshwater content, and only compare model output to an observation after transforming it into the observed quantity. That would immediately improve the logic of the paper.

Citation: https://doi.org/10.5194/egusphere-2026-2089-RC1
RC2: 'Comment on egusphere-2026-2089', Anonymous Referee #2, 02 Jun 2026

Review for “Lagrangian reconstruction of snow accumulation and loss on Antarctic sea ice”
This article presents a new model for snow on Antarctic sea ice: the University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model. The model uses a combination of reanalysis data (for atmospheric forcing) and observation-derived ice drift to estimate snow mass and bulk density over Southern Ocean sea ice. The model includes parameterized representations of various snow processes including snow compaction, sublimation, blowing snow loss to leads, changes due to ice motion, superimposed ice, and melt, but currently excludes snow-ice formation. Model processes are parameterized and the parameters are calibrated against a set of buoy snow observations. Snow budget terms from the model output are examined. Seasonal averages are compared to passive microwave and altimetry-derived estimates of snow depth.
I think the model is promising overall and represents a good step forward in Antarctic snow-on-sea-ice modelling, and I also commend the authors for the easy accessibility of the model source code and clear documentation. The article is generally clearly written and well-referenced, with good-quality figures. However, in my view, some of the model development methodology and scientific reasoning require expansion and/or clarification, particularly pertaining to the omission of snow-ice and some details of the model calibration process. I think some additional revisions/testing are required before the manuscript would be suitable for publication. I include my comments with suggestions below.
Major comments:
The authors take great care to make careful choices of parameterizations of several snow processes, which makes the exclusion of snow-ice formation somewhat puzzling to me. I understand that there are non-trivial difficulties with this (lack of knowledge of ice thickness being a key limitation), and that validation with buoys remains possible despite the exclusion of this process because of their tracking of snow surface height relative to their initialization baseline, but nevertheless it seems like a large omission given the prevalence of snow-ice conversion on Antarctic sea ice. It is not clear to me why the authors chose to omit this process rather than attempt to e.g. parameterize and calibrate it as has been done with other snow processes. The frequent attribution of budget inconsistencies and validation discrepancies to snow-ice formation seems hasty to me also. I think the authors need to take more caution in how this discussion is framed. I also urge caution at the interpretation of model-observation differences as evidence of snow-ice formation, since other processes/biases may factor into the differences. Snow-ice itself may also introduce further uncertainty into remote sensing measurements, so I think these inferences need to be stated with more caution than is currently employed.
I am sympathetic to the challenges of calibrating models with limited observational dataset availability, and the successive halving method is in my view a reasonable choice for this particular calibration problem, but I wonder about the reliance of the model parameters on a specific choice of buoys. Calibration is currently performed once for a single selection of calibration buoys; how much do the optimal parameters vary when different sets of buoys are (randomly) selected? I think more testing may be necessary to examine the robustness of this parameterization, and to guard against potential overfitting.
Minor comments:
Some of the figure text is small, please increase the size for legibility (e.g. Fig 7, Fig 8)
You mention some other modelling efforts (e.g. CASSIS); can you provide a sense of how well WASSAIL performs relative to other snow-on-sea-ice simulations? This could also help contextualize the impact of the missing snow-ice process.
If I’m interpreting things correctly, model output is interpolated to a regular lat/lon grid; would this not introduce artefacts at the pole? I’m curious of the motivation for using that grid rather than a polar grid (e.g. EASE). Also, is the final output resolution 0.75x0.75 degrees? I think this needs to be stated more explicitly earlier in the text for clarity; I was initially under the impression that the resolution was 0.25.
Line-by-line comments:
115: bias of -19 cm; is this for AMSR2 – OIB or OIB – AMSR2? Please specify
155-156: Snowfall includes a scaling parameter; is a scaling approach also used for the total precipitation?
Figure 1 looks good overall, but at 100% zoom it is very difficult to distinguish the calibration and validation sets in panel b. If possible, please consider making this panel larger or otherwise adjust how the lines are formatted for clarity.
Section 2.2.2: I think it would be worth mentioning how the CS2 retrieval can still have biases e.g. from melt; although it can often identify the snow-ice interface, it can still have difficulties
233: How is SIC treated in calibration mode? Are there times where remote sensing SIC disagrees with in situ? (or e.g. grid edge effects)
269: To clarify, does this mean that the inverse scaling factor (gamma_dens) is what is calibrated in the model, whereas k_n remains fixed? I.e. is the scaling factor essentially a way to calibrate k_n?
316: Consider specifying here that the “available snow depth divided by the time step” option is so that you do not remove more snow than the available snow depth
325: Is the factor of 12.5 prescribed from literature?
515-516: As phrased, this reads like the goodness-of-fit metrics are also used for the calibration when only RMSE is used, please clarify
557-558: Although the studies cited do have divergent conclusions on reanalysis biases, I think it’s worth mentioning that they examine different time periods (e.g. CloudSat is ~2006-2016, drifting station records are from ~1950s-1991) which may suggest different biases in different time periods (particularly also when there may be discontinuities in products assimilated in ERA5 across decades). Though regardless uncertainties are indeed high.
Figure 4: It’s difficult to see which lines are dashed; consider distinguishing buoys vs model in other ways (e.g. use a lighter shade of the line colour, different line weights, etc).
604: If the model were run for multiple years during calibration, could you constrain the model drift further and close the budget? Or are you concerned that this would cause the free parameters to take on unphysical values?
608-609: It’s not clear to me why snow-ice formation is grouped with snow release to the ocean if snow-ice is not simulated, please elaborate. Is it implicitly assumed to be part of the budget term?
Figure 7: How do you account for year-to-year differences in marginal ice cover for these averages? I.e. since these are multiyear averages, it’s possible that grid squares near ice margins may only include a few years of data in the average. I suggest checking this and either discussing possible impacts or applying a minimum threshold for the number of available years.
736: Are you comparing the 2003-2025 time period from the model with CryoSat-2 from 2010-2021? If so, some of the disagreement may be spurious; I suggest comparing 2010-2021 from the model and CS2.
750: The correspondence being less robust for CS2 makes me wonder how much of this is due to snow-ice formation vs how much of this is from other discrepancies (particularly if different time periods are being compared here).

Citation: https://doi.org/10.5194/egusphere-2026-2089-RC2

Ethan C. Campbell, Stephen C. Riser, and Melinda A. Webster

Data sets

University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model data, v1.0.0 (2003-2025) Ethan C. Campbell https://doi.org/10.5281/zenodo.19507962

Model code and software

University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model and analysis code, v1.0.0 Ethan C. Campbell https://doi.org/10.5281/zenodo.19509689

Ethan C. Campbell, Stephen C. Riser, and Melinda A. Webster

Viewed

Total article views: 362 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
225	110	27	362	28	25

HTML: 225
PDF: 110
XML: 27
Total: 362
BibTeX: 28
EndNote: 25

Views and downloads (calculated since 21 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	112	52	15	179
May 2026	89	50	6	145
Jun 2026	24	8	6	38

Cumulative views and downloads (calculated since 21 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	112	52	15	179
May 2026	89	50	6	145
Jun 2026	24	8	6	38

Viewed (geographical distribution)

Total article views: 358 (including HTML, PDF, and XML) Thereof 358 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 04 Jun 2026

Short summary

The processes controlling snow evolution on Antarctic sea ice are poorly constrained from observations and models. We developed a new model that parameterizes key snow erosion and transformation mechanisms and calibrated it with snow buoy data. The simulation shows that more than one-third of deposited snow is lost before sea ice melt. Snow on sea ice is found to be a larger source of freshwater to the Southern Ocean than previously estimated, with implications for water mass transformation.


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