| 08 Sep 2026
A long-term wintertime snow depth dataset on Arctic sea ice (1978–2025) derived from multisource passive microwave radiometer data
Abstract. Abstract. Snow on sea ice is a key component of the Arctic climate system, strongly regulating the surface energy and mass balances of ice-covered regions. This study presents a long-term wintertime snow depth (SD) dataset on Arctic sea ice spanning the period 1978–2025 derived from multi-channel brightness temperature (TB) observations acquired by the Scanning Multichannel Microwave Radiometer (SMMR), Special Sensor Microwave/Imager (SSM/I), and Special Sensor Microwave Imager/Sounder (SSMIS). We first analysed the relationship between the spectral gradient ratio (GR) of vertically polarized TBs at 19 and 37 GHz (GRV(37/19)) and altimetric snow depth derived from the Ice, Cloud and land Elevation Satellite-2 (ICESat-2) and CryoSat-2 missions. Based on this analysis, we proposed a new SD estimation algorithm using three predictors, including GRV(37/19), bulk snow density from the NASA Eulerian Snow On Sea Ice Model (NESOSIM) and a cumulative time variable to account for the effects of snow metamorphism on passive microwave retrievals. The resulting SYSU SnowDepth dataset was validated against various independent observations, categorized into point-scale and transect-based measurements. While the dataset demonstrates an overall good accuracy, the validation performance was strongly influenced by the spatial representativeness of the reference data. Specifically, root-mean-square error (RMSE) values ranged from ~3 to 7 cm against transect-based measurements from airborne and buoy array observations, but increased to 10 to 17 cm against point-scale measurements from individual buoys, aircraft landing sites and ship-based observations. Furthermore, retrieval accuracy was higher over first-year ice (FYI) than multi-year ice (MYI) and exhibited a seasonal variation with RMSE increasing from October to April as the snowpack thickens. To our knowledge, the SYSU SnowDepth dataset is the longest satellite-based SD record providing pan-Arctic coverage of both FYI and MYI throughout the full winter season (October–April). It is expected to significantly benefit altimetry-based sea ice thickness estimation, the assimilation of snow information into sea ice models, the assessment of light availability for under-ice biota, weather forecasting, and climate monitoring. The newly developed SYSU SnowDepth dataset is available at https://doi.org/10.5281/zenodo.21472646 (He et al., 2026).
This manuscript presents a potentially valuable 47-year wintertime snow-depth dataset for Arctic sea ice derived from passive microwave observations. In particular, the attempt to provide a temporally continuous record covering both first-year ice (FYI) and multiyear ice (MYI) from October to April is scientifically relevant. However, in its current form, I am not yet convinced that the proposed algorithm successfully retrieves snow-depth variability from passive microwave observations with sufficient accuracy, independence, and temporal consistency to support the construction of a long-term data record.
1. Contribution of GRV to the retrieval
My main concern is that the retrieval appears to be strongly controlled by season and sea-ice type, with additional information supplied by NESOSIM snow density, rather than primarily by the passive microwave signal itself. The proposed retrieval uses GRV(37/19), NESOSIM snow density, and a cumulative time variable, with the latter represented by a cubic polynomial. The manuscript's own feature-importance analysis indicates that the time variable is the dominant predictor over FYI, while GRV contributes substantially less.
This raises an important question regarding how much independent snow-depth information is actually provided by the microwave observations beyond the prescribed seasonal evolution and model-derived snow density. The authors should therefore demonstrate explicitly the added value of GRV, for example by comparing the full retrieval with appropriate baseline models with and without GRV. Such an analysis is necessary to establish whether the method retrieves actual snow-depth variability from PMW observations rather than mainly reconstructing broad seasonal and ice-type-dependent snow-depth patterns.
2. Independent validation results
The independent validation results also raise concerns. Although the manuscript reports relatively good agreement with spatially averaged airborne observations, the correlations become substantially weaker when the evaluation is separated by ice type, and notable positive biases occur over FYI in late winter. Validation against individual buoys yields much larger RMSE values, while the seasonal evaluation shows increasing errors toward March and April.
These results make me hesitant to accept statements such as "high accuracy" or "robust performance" without a clearer demonstration that the product resolves actual snow-depth variability rather than broad differences associated with season and ice type. In particular, the authors should distinguish between skill arising from FYI/MYI separation and skill in resolving snow-depth variability within each ice type.
3. Altimetric snow depth as the training target
Another major concern is the uncertainty associated with the altimetric snow-depth dataset used as the training target. The target snow depth is derived from combined ICESat-2 and CryoSat-2 freeboards and therefore depends on assumptions associated with CryoSat-2 retracking, radar penetration/scattering height, radar propagation through the snow layer, and the spatial and temporal matching of the two altimeters.
In particular, the retrieved CryoSat-2 freeboard can depend substantially on the choice of retracking algorithm and the associated radar-freeboard correction. These uncertainties are not negligible relative to the reported regression errors. The manuscript should therefore discuss the uncertainty of the training target more carefully and assess how it propagates into the fitted regression coefficients and final snow-depth product. Sensitivity to alternative CryoSat-2 freeboard/retracking products would also be valuable.
4. Long-term consistency of the brightness-temperature record
For a dataset intended to provide a 47-year climate-scale record, the long-term consistency of the passive microwave input data requires stronger justification. The current intercalibration relies on pairwise regressions between overlapping sensors, including some relatively short overlap periods, while SMMR and SSM/I/SSMIS also have different viewing geometries.
In addition, the open-water correction uses fixed brightness-temperature values attributed to Ivanova et al. (2015), although that study provides sensor-dependent tie points. The values adopted here appear to correspond to the Northern Hemisphere SMMR open-water tie points. The authors should therefore clarify how these fixed tie points are consistent with the intercalibrated TB record and whether they remain appropriate after the different sensors have been adjusted to a common reference.
More generally, the authors should demonstrate that transitions between SMMR, SSM/I, and SSMIS do not introduce artificial discontinuities in GRV or retrieved snow depth. It would also be useful to discuss whether an established fundamental climate data record, for example a CM SAF FCDR for the DMSP period, could be used or at least provide an independent benchmark for the intercalibration.
5. Consistency and traceability of auxiliary input datasets
I have a broader concern regarding the consistency of the auxiliary information used at different stages of the processing chain. SIC, sea-ice type, and snow density play important roles in the proposed retrieval, but different products, definitions, and assumptions appear to enter the calibration, construction of the training target, regression fitting, and final retrieval. For a data record extending over 47 years, these differences should be explicitly documented and their impact assessed.
For example, the final SYSU retrieval uses NOAA/NSIDC SIC, C3S sea-ice type, and NESOSIM snow density. However, the ICESat-2/CryoSat-2 snow-depth product used as the training target itself employs an independent snow-density parameterization in the CryoSat-2 radar propagation correction. Thus, snow density enters both the predictor and the construction of the training target, but through different datasets or parameterizations. The authors should discuss whether differences between these snow-density descriptions can influence the fitted relationship and whether part of the regression may compensate for inconsistencies between the auxiliary datasets rather than represent a physical PMW–snow-depth relationship.
Similar questions arise for SIC and sea-ice type. The proposed retrieval uses the NOAA/NSIDC SIC CDR for open-water correction and C3S sea-ice type to select the FYI or MYI retrieval equation, whereas other components of the processing chain and the altimetric reference product employ their own masks, classifications, or definitions. It is therefore not clear that FYI/MYI conditions, ice-covered pixels, and snow-density assumptions are defined consistently throughout the training and application of the algorithm.
I recommend that the authors provide a table describing, for each major processing step, the exact SIC, sea-ice-type, and snow-density source/version, spatial and temporal resolution, thresholds or class definitions, and resampling procedure used. More importantly, the sensitivity of the retrieved snow depth to reasonable alternative auxiliary products should be quantified. Without such an assessment, it is difficult to determine whether discontinuities or biases in the resulting 1978–2025 record originate from actual snow-depth variability or from inconsistencies among the auxiliary datasets.
6. Regression stability and temporal transferability
I am also concerned about the stability of the regression coefficients. In Table 3, several MYI coefficients vary substantially between the leave-one-year-out experiments, including changes in sign, despite similar reported fitting statistics. This raises questions about parameter stability, possible collinearity among GRV, snow density, and the time-dependent terms, and the temporal transferability of the model.
Moreover, the reported statistics appear to describe the data used for fitting rather than predictions for the excluded year. If so, this is not a genuine leave-one-year-out validation. I recommend that the authors train the model on two winters and evaluate it exclusively on the excluded winter, repeating this for all three winters. The resulting out-of-sample bias, RMSE, and correlation should be reported separately for FYI and MYI. This would provide a much more meaningful assessment of whether the regression can be transferred beyond the period from which its coefficients were derived.
7. Uncertainty characterization
Finally, the propagated uncertainty appears smaller than the discrepancies found in independent validation. The current Gaussian uncertainty analysis does not appear to account fully for structural model error, uncertainty in the altimetric training target, sensor intercalibration, open-water tie-point uncertainty, or covariance among regression coefficients and predictors.
Given that independent RMSE values can exceed 10 cm, while the reported product uncertainty is substantially smaller for much of the winter, the current uncertainty field may not represent the total uncertainty of the retrieved snow depth. The authors should revisit the uncertainty framework and clarify exactly which sources of uncertainty are included and excluded. A resampling or Monte Carlo approach incorporating regression covariance and major input uncertainties may provide a more realistic estimate.
Overall assessment
Overall, I consider the dataset concept potentially valuable, but several issues directly affect whether the resulting record can be regarded as a reliable, internally consistent, and quantitatively validated snow-depth dataset. In particular, the authors should demonstrate the independent information content of the passive microwave observations, assess true out-of-sample temporal stability, address uncertainty in the altimetric training target, establish consistency across the full set of auxiliary inputs, strengthen the long-term sensor-intercalibration analysis, and provide a more realistic uncertainty characterization.
I therefore recommend major revision before the dataset can be considered suitable for publication in Earth System Science Data.