Reconstruction of Antarctic sea ice thickness from sparse satellite laser altimetry data using a partial convolutional neural network

Hao Luo; Jiping Liu; Petteri Uotila; Qian Shi; Qinghua Yang; Wen Shi; Xiaoran Dong; Yafei Nie; Yue Xu; Ziqi Ma

REVIEW 4 major objections 4 minor 3 references

Reconstruction of Antarctic sea ice thickness from sparse satellite laser altimetry data using a partial convolutional neural network

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A neural network turns sparse Antarctic laser altimetry into complete 5-day sea ice thickness maps.

desk verdict A genuinely useful new Antarctic SIT reconstruction dataset, but the headline accuracy claim rests on a narrow Weddell-only ULS validation from the ICESat era, inside the training window, and the extension to ICESat-2 is not independently grounded. read the letter →

arxiv 2505.06255 v1 pith:7VJIDCEI submitted 2025-05-02 physics.geo-ph physics.ao-ph

Ziqi Ma , Qinghua Yang , Yue Xu , Wen Shi , Xiaoran Dong , Qian Shi , Hao Luo , Jiping Liu

show 2 more authors

Petteri Uotila Yafei Nie

This is my paper · ORCID

classification physics.geo-phphysics.ao-ph

keywords AntarcticseaicethicknesspartialconvolutionalneuralnetworkICESatICESat-2satellitelaseraltimetryreconstructionupward-lookingsonarvalidationEASEgrid

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a partial convolutional neural network can fill the gaps between sparse satellite laser tracks to produce a pan-Antarctic sea ice thickness map every five days at 12.5 km resolution, using ICESat (2003–2009) and ICESat-2 (2018–2024) along-track retrievals. Against independent upward-looking sonar measurements in the Weddell Sea, the reconstruction matches better than two ocean reanalyses and two radar-altimeter products. If right, it gives the first sub-monthly, spatially complete view of Antarctic sea ice thickness, needed to track mass balance and to validate models. The paper also argues the method can be updated near-real-time for operational monitoring.

What carries the argument

The central object is a partial convolutional neural network (PCNN) with a U-net architecture, where each partial convolutional layer updates only grid points with valid observations in the current mask, enabling the network to learn how to fill regions with no data from surrounding spatial patterns. The input is the equivalent SIT, defined as the product of along-track SIT and daily sea ice concentration (AMSR-E/AMSR2), mapped to a 12.5 km EASE grid and averaged over a 5-day moving window. The network is trained on daily SIT fields from the C-GLORSv7 and GLORYS12v1 ocean reanalyses using mean squared error loss, with binary masks telling it where observations exist.

What would settle it

A direct comparison of the 2018–2024 reconstruction against any in situ thickness record (mooring, ship-based, or electromagnetic) in the Weddell Sea, using the same monthly averaging as the ICESat-era validation, would settle whether the accuracy claim extends; the paper states that no such overlapping ULS data exist, so this comparison remains an open test.

Watch

Extended reading notes

Core claim

The central claim is that sparse along-track laser altimetry SIT retrievals, when combined with daily sea ice concentration and trained on reanalysis SIT fields through a partial convolutional U-net, yield a reconstructed pan-Antarctic SIT field at 5-day and 12.5 km resolution whose uncertainty is quantified and whose accuracy exceeds that of GLORYS12v1, C-GLORSv7, LEGOS, and SICCI against ULS observations (MAE 0.238 m versus 0.297, 0.371, 0.543, and 1.276 m). The temporal evolution, seasonal cycles, and intra-seasonal tendencies are validated by ULS and ICESat-2, including winter thinning regions attributed to polynyas.

Load-bearing premise

The accuracy claim for the ICESat-2 period assumes that the reconstruction skill demonstrated against upward-looking sonar in the ICESat era carries over unchanged, because no ULS or other in situ thickness measurements overlap the ICESat-2 years.

Editorial extensions

If this is right

A physically consistent 5-day pan-Antarctic SIT record now exists for 2003–2009 and 2018–2024, enabling study of sea ice volume variability on sub-monthly scales.
The quantified per-grid uncertainty (from 720 ensemble members) allows users to weight the data appropriately in model validation or assimilation.
The near-real-time updating capability could support operational monitoring and forecasting of Antarctic sea ice state.
The identification of winter-thinning regions, tied to polynya activity, offers a new observational constraint for ocean–atmosphere interaction studies.
The PCNN approach could be retrained on other reanalysis pairs or extended to additional altimetry missions to extend the record.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the reconstruction generalizes, the same partial-convolution approach could fill gaps in other polar geophysical fields, such as snow depth or surface temperature, where along-track sampling is sparse.
The ULS validation covers only the Weddell Sea, so the paper's pan-Antarctic accuracy claim rests on cross-validation consistency; users should expect regionally varying skill until denser in situ data are available.
The 5-day moving average smooths day-to-day signals, and a testable extension would be to compare daily reconstructions against withheld ICESat-2 tracks to quantify the smoothing error.
Because the network learns spatial patterns from reanalyses, systematic biases in those training fields could imprint on the reconstruction in sectors with few satellite observations, especially the Pacific and Indian Ocean sectors.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know: this is a data descriptor releasing a 5-day, 12.5 km pan-Antarctic sea ice thickness reconstruction for ICESat (2003–2009) and ICESat-2 (2018–2024), built with a partial convolutional network trained on two reanalyses. The dataset itself is new, and the uncertainty quantification is more thorough than most papers in this space.

Credit where due: the ULS validation is a genuine external check, and the reconstruction does beat the four comparators (GLORYS12v1, C-GLORSv7, LEGOS, SICCI) on MAE over the 23 Weddell Sea cases. The 720-member ensemble uncertainty, built from along-track scenarios, training periods, and model initializations, is a real strength. The regression analysis tying reconstruction RMSE to local SIT variability and observation count is a nice diagnostic and gives the reader a feel for where the method struggles. Code and data are on Figshare, which matters.

Now the soft spots, in proportion. The ULS validation is geographically narrow (Weddell Sea, monthly scale, 23 cases) and it sits entirely inside the 1993–2011 training window used for the reanalyses. That does not make the ULS comparison invalid, but it does mean the network's spatial patterns for 2003–2009 may be artificially familiar. The paper's key step is the claim that cross-validation consistency lets the ICESat-era ULS results be "reliably extended" to the ICESat-2 period. That extension is load-bearing because the product's main value is the 2018–2024 near-real-time coverage, and the cross-validation uses ICESat-2 SIT retrievals that come from the same OLMi algorithm used to generate the model input. Shared retrieval biases are invisible to that test. So the headline "higher accuracy than four datasets" is really "higher accuracy in the Weddell Sea during the ICESat era," and the abstract overreaches.

Also minor: the code is on Figshare but not version-controlled, and the comparison omits SMOS with a plausible but not fully satisfying justification. On the other hand, the paper itself is candid about several limitations—the lack of ULS overlap, the caution about daily variations, the 25 km land buffer—which makes me think the authors know the edges of their claim but let the abstract run ahead.

Bottom line: this deserves a serious referee. The dataset is useful for model validation, process studies, and potentially operational monitoring. The flaws are addressable: reframe the accuracy claim to the region and period actually validated, add an explicit statement that no independent thickness validation exists for the ICESat-2 period, and ideally retune the model on a training window that excludes 2003–2009 to see if the Weddell performance holds. I'd take it to review rather than desk reject.

Referee Report

4 major / 4 minor

Summary. The manuscript describes the development of a pan-Antarctic sea ice thickness (SIT) dataset at 5-day and 12.5 km resolution for 2003–2009 and 2018–2024, reconstructed from sparse ICESat/ICESat-2 along-track laser altimetry retrievals using a partial convolutional neural network (PCNN) trained on two ocean reanalyses (GLORYS12v1 and C-GLORSv7). The authors report that the reconstructed SIT outperforms four existing satellite-derived and reanalyzed Antarctic SIT datasets against upward-looking sonar (ULS) observations, and they validate the temporal evolution of the reconstruction against ULS climatologies and ICESat-2 along-track data. The data and code are made available via Figshare.

Significance. If the central accuracy claim holds, the dataset would be a valuable new resource for Antarctic sea ice research, providing sub-monthly, spatially complete SIT fields with near-real-time updating capability. The paper includes several strengths: the use of independent ULS observations as an external benchmark, a documented ensemble-based uncertainty estimate (720 members, Eq. 2 and Table 4), a clear description of the PCNN architecture and training workflow, and public release of both data and code. The independent ULS comparison in the Weddell Sea is a genuine check and gives the ICESat-era reconstruction some external grounding. However, the scope of the validation is more limited than the abstract's claim of 'higher accuracy' suggests, and the extension of that validation to the ICESat-2 period is not supported by the evidence presented.

major comments (4)

[Technical Validation, 'Validation of reconstructed sea ice thickness using ULS data'] The ULS validation period (2003–2009) falls entirely within the PCNN training window (C-GLORSv7 1989–2011 and GLORYS12v1 1993–2011, as described in Methods, 'Deep learning framework'). The model was trained to map masked reanalysis SIT fields to full reanalysis fields for exactly these years, so the reconstructed fields for 2003–2009 may be artificially familiar to the model, potentially inflating the reported MAE advantage (0.238 m vs. 0.297 m for GLORYS12v1) relative to what would be achieved on an unseen period. To support the 'higher accuracy' claim, the authors should either retrain the PCNN on data excluding 2003–2009 and repeat the ULS validation, or clearly restrict the accuracy claim to the ICESat-era reconstruction and rephrase the abstract accordingly.
[Technical Validation, 'Validation of reconstructed sea ice thickness using ULS data'] The statement 'This evidence allows the validation results established with ULS during the ICESat era (2003–2009) to be reliably extended to the ICESat-2 period (2018–2024)' is not supported by the cited cross-validation experiments. Those experiments (§Evaluation of PCNN via cross-validation) use in-box ICESat/ICESat-2 SIT retrievals as ground truth, but those retrievals are produced by the same OLMi algorithm used as model input, so shared retrieval biases are invisible to the comparison. The paper itself acknowledges this by noting the independence of ULS data, yet the extension inference is load-bearing because the primary value of the dataset is the near-real-time 2018–2024 product. The authors should either remove or substantially soften this extension claim, or provide independent thickness observations overlapping the ICESat-2 period.
[Abstract and Technical Validation, ULS comparison] The claim that the reconstructed SIT shows 'higher accuracy than the other four satellite-derived and reanalyzed Antarctic SIT datasets' is based on 23 ULS validation cases that are confined to the Weddell Sea and to the ICESat era. This is a regional, single-basin test, not a pan-Antarctic evaluation. The abstract and Data Records sections should be reworded to state that the reconstruction outperforms the four datasets in the Weddell Sea during 2003–2009, without implying pan-Antarctic or ICESat-2-era superiority.
[Methods, 'Data preparation and pre-processing' and 'Deep learning framework'] The reconstruction is partly circular with respect to the reanalyses: the loss function is mean squared error against GLORYS12v1/C-GLORSv7, the normalization statistics are derived from GLORYS12v1, and the internal testing set (2016–2020) evaluates the model against the same reanalysis family. Thus the reported internal metrics (R=0.89, RMSE=0.21 m over 3,288 samples) demonstrate the PCNN's ability to reproduce reanalysis fields, not independent observational accuracy. The manuscript should explicitly label these internal metrics as reanalysis-reconstruction consistency checks and separate them from the independent ULS-based evaluation.

minor comments (4)

[Figure 3 and 'Deep learning framework'] The text states that two daily samples from February 15, 2016 and September 15, 2016 were selected, but the results paragraph reports correlation and RMSE for February 15, 2019 and September 15, 2019; the years are inconsistent and should be corrected.
[Figure 11 caption and text] In the Technical Validation section, 'AWI212 (Fig. 11e), AWI217 (Fig. 11f) and AWI233 (Fig. 11-l)' uses an ambiguous lowercase 'l'; it should read 'Fig. 11l' for consistency with panel labels.
[Uncertainty estimates] The sentence 'Uncertainty estimates were unavailable for the colder periods of 2019 and 2024' is ambiguous; it presumably means ice-freezing (colder) months, but the phrasing could be misread as referring to entire years and should be clarified.
[Data Records] The description of the ICESat reconstruction states that it 'contains temporal discontinuities (see Table 2)'; it would be helpful to state explicitly that these discontinuities result from the ICESat mission's laser operation campaigns, as shown in Table 2, rather than from the reconstruction method.

Circularity Check

2 steps flagged · score 6.0 of 10

ICESat-2 accuracy claim is bridged by cross-validation whose 'ground truth' is the same OLMi retrieval product used as model input; the ULS-validated ICESat-era result is therefore extended through a self-consistency loop rather than independent evidence.

self definitional [Technical Validation > Validation of reconstructed sea ice thickness using ULS data (final paragraph), supported by Evaluation of PCNN via cross-validation]
"It is important to note that the cross-validation experiments demonstrate that the PCNN-reconstructed SIT exhibits consistent performance across various missions and remains temporally stable. This evidence allows the validation results established with ULS during the ICESat era (2003–2009) to be reliably extended to the ICESat-2 period (2018–2024)."

The cross-validation 'ground truth for accuracy assessment' is not independent thickness: it is the same OLMi ICESat/ICESat-2 SIT retrieval used as reconstruction input, merely held out spatially ('ICESat/ICESat-2 SIT observations were partitioned into two mutually exclusive subsets... reserved as ground truth'). Because ULS data are unavailable during the ICESat-2 period, the paper's only bridge to ICESat-2 accuracy is this internal consistency check. The bridge assumes that OLMi retrieval accuracy transfers across missions, and then confirms that assumption against OLMi retrievals themselves, making the extension a self-consistency loop rather than an independent validation of the 2018–2024 product.
self definitional [Technical Validation > Validation of reconstructed SIT's temporal evolution (after Eq. 3 and Fig. 13)]
"To validate this unexpected winter thinning, ICESat-2 along-track observations in these three regions are utilized to calculate the winter regional mean SIT and its temporal evolution using Eq. 3 (Fig. 13). ... These results confirm that the intra-seasonal variability in the reconstructed SIT fields predominantly originates from ICESat/ICESat-2 observational constraints, thereby validating their physical plausibility."

The reference used to validate the reconstructed winter thinning is the same ICESat-2 along-track SIT retrieval product that is the primary input to the PCNN reconstruction. The paper admits the variability 'originates from ICESat/ICESat-2 observational constraints,' i.e., the reconstruction is built from these observations. Confirming that the reconstruction captures the input observations is therefore expected by construction, not an independent test of physical plausibility. The validation reduces to checking that the output reproduces its own input.

full rationale

Two genuinely circular validation steps appear in Technical Validation. First, the extension of ULS-based accuracy from the ICESat era to the ICESat-2 era relies on cross-validation whose 'ground truth' is the same OLMi ICESat/ICESat-2 retrieval used as model input; this makes the bridge a self-consistency check of the retrieval product, not independent evidence for the 2018–2024 product. Second, regional winter thinning in the reconstructed field is 'validated' with the same ICESat-2 along-track observations that were used to create the reconstruction; the paper's own wording concedes the variability originates from those observations, so agreement is expected by construction. The ULS comparison for the ICESat period is genuinely independent and gives the method real external grounding, which prevents the paper from being fully circular. The overlap of the ULS validation period (2003–2009) with the reanalysis training window (1989–2011) is a limitation of the ICESat-era check, but by itself it is not a definitional circularity. Self-citations to OLMi (ref 19, co-authored) and PCNN applications (ref 35) exist but are not the main mechanism of the circularity found here.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The reconstruction rests on the suitability of reanalysis training fields, the transfer of learned spatial patterns to real altimetry data, and the validity of the OLMi retrieval and ULS conversion used for inputs and validation. No invented physical entities are introduced.

free parameters (3)

OLMi empirical ice-to-snow ratio and associated densities = Inherited from Xu et al. 2021, not restated
Along-track SIT, the primary input, is computed with OLMi; errors in this empirical ratio propagate into every reconstructed field.
PCNN hyperparameters and trained weights = Learning rate 2e-4, batch size 50, 11,000 iterations; weights learned from reanalysis training data
The reconstruction is the output of a network trained on GLORYS12v1 and C-GLORSv7; training configuration and learned weights determine the filled fields.
5-day moving average window and 25 km land exclusion buffer = 5 days; 25 km
Chosen to reduce sparse-sampling noise and land contamination; these choices affect temporal smoothing and coastal coverage.

assumptions (4)

domain assumption GLORYS12v1 and C-GLORSv7 reanalyses are suitable training data because their SIT spatial distributions are comparable to ICESat-corrected benchmarks.
Invoked in the section 'Ocean reanalysis data for training PCNN'; if false, the learned spatial covariance is biased and the reconstruction inherits that bias.
domain assumption Spatial covariance patterns learned from reanalysis fields transfer to real ICESat and ICESat-2 altimetry retrievals.
The PCNN is trained on masked reanalysis and applied to satellite observations; the paper does not directly measure transfer under the retrieval-error characteristics of real laser altimetry.
domain assumption The OLMi one-layer retrieval with snow depth estimated from an empirical ice-snow ratio gives sufficiently unbiased along-track SIT.
This is the fundamental thickness retrieval used for all satellite inputs; it is cited from Xu et al. 2021 and not re-derived in this paper.
domain assumption The ULS draft-to-thickness conversion h = 0.028 + 1.012*d is a valid representation of sea ice thickness.
Used for independent validation; taken from Behrendt et al. 2015; any bias in the conversion affects the validation results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reconstruction of Antarctic sea ice thickness from sparse satellite laser altimetry data using a partial convolutional neural network." pith.science (2026). https://pith.science/paper/7VJIDCEI

@misc{pith2026250506255,
  author       = {Pith},
  title        = {Pith review of: Reconstruction of Antarctic sea ice thickness from sparse satellite laser altimetry data using a partial convolutional neural network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7VJIDCEI}},
  note         = {Machine review of arXiv:2505.06255}
}

read the original abstract

The persistent lack of spatially complete Antarctic sea ice thickness (SIT) data at sub-monthly resolution has fundamentally constrained the quantitative understanding of large-scale sea ice mass balance processes. In this study, a pan-Antarctic SIT dataset at 5-day and 12.5 km resolution was developed based on sparse Ice, Cloud and Land Elevation Satellite (ICESat: 2003-2009) and ICESat-2 (2018-2024) along-track laser altimetry SIT retrievals using a deep learning approach. The reconstructed SIT was quantitatively validated against independent upward-looking sonar (ULS) observations and showed higher accuracy than the other four satellite-derived and reanalyzed Antarctic SIT datasets. The temporal evolution of the reconstructed SIT was further validated by ULS and ICESat-2 observations. Consistent seasonal cycles and intra-seasonal tendencies across these datasets confirm the reconstruction's reliability. Beyond advancing the mechanistic understanding of Antarctic sea ice variability and climate linkages, this reconstruction dataset's near-real-time updating capability offers operational value for monitoring and forecasting the Antarctic sea ice state.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

& Chen, D

1 Liu, J.,Zhu,Z. & Chen, D. LowestAntarctic Sea IceRecord Broken for the SecondYear in a Row. Ocean-Land-Atmosphere Research 2,0007,doi:10.34133/olar.0007(2023). 2 Purich,A. & Doddridge, E.W.Record lowAntarctic sea ice coverage indicates a new sea ice state. Communications Earth & Environment 4, 314, doi:10.1038/s43247-023-00961-9 (2023). 3 Luo, H., Yang,...

work page doi:10.34133/olar.0007(2023 2023
[6]

44 Meier, W

doi:10.5067/ATLAS/ATL10.006(2023). 44 Meier, W. N., Markus, T. & Comiso, J. C. NASA National Snow and Ice Data Center DistributedActiveArchiveCenter,Boulder,ColoradoUSA,(2018). 45 Nie, Y. et al. Southern Ocean sea ice concentration budgets of five ocean-sea ice reanalyses. Climate Dynamics 59,3265-3285,doi:10.1007/s00382-022-06260-x(2022). 46 Behrendt, A....

work page doi:10.5067/atlas/atl10.006(2023 2023
[2023]

5 Kurtz, N

Journal of Geophysical Research: Oceans 129,e2023JC020848,doi:https://doi.org/10.1029/2023JC020848(2024). 5 Kurtz, N. T. & Markus, T. Satellite observations of Antarctic sea ice thickness and volume. Journal of Geophysical Research: Oceans 117, doi:https://doi.org/10.1029/2012JC008141 (2012). 6 Wang, J. et al. A comparison between Envisat and ICESat sea i...

work page arXiv 2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

& Chen, D

1 Liu, J.,Zhu,Z. & Chen, D. LowestAntarctic Sea IceRecord Broken for the SecondYear in a Row. Ocean-Land-Atmosphere Research 2,0007,doi:10.34133/olar.0007(2023). 2 Purich,A. & Doddridge, E.W.Record lowAntarctic sea ice coverage indicates a new sea ice state. Communications Earth & Environment 4, 314, doi:10.1038/s43247-023-00961-9 (2023). 3 Luo, H., Yang,...

work page doi:10.34133/olar.0007(2023 2023

[2] [6]

44 Meier, W

doi:10.5067/ATLAS/ATL10.006(2023). 44 Meier, W. N., Markus, T. & Comiso, J. C. NASA National Snow and Ice Data Center DistributedActiveArchiveCenter,Boulder,ColoradoUSA,(2018). 45 Nie, Y. et al. Southern Ocean sea ice concentration budgets of five ocean-sea ice reanalyses. Climate Dynamics 59,3265-3285,doi:10.1007/s00382-022-06260-x(2022). 46 Behrendt, A....

work page doi:10.5067/atlas/atl10.006(2023 2023

[3] [2023]

5 Kurtz, N

Journal of Geophysical Research: Oceans 129,e2023JC020848,doi:https://doi.org/10.1029/2023JC020848(2024). 5 Kurtz, N. T. & Markus, T. Satellite observations of Antarctic sea ice thickness and volume. Journal of Geophysical Research: Oceans 117, doi:https://doi.org/10.1029/2012JC008141 (2012). 6 Wang, J. et al. A comparison between Envisat and ICESat sea i...

work page arXiv 2024