Preprints
https://doi.org/10.5194/egusphere-2026-5409
© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
https://doi.org/10.5194/egusphere-2026-5409
© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
15 Oct 2026
| 15 Oct 2026
Status: this preprint is open for discussion and under review for The Cryosphere (TC).
Machine-learning bias correction of passive-microwave sea ice concentration against a SAR-derived product: impact on Arctic sea ice forecasts
Abstract. Sea ice concentration (SIC) assimilation in operational systems relies predominantly on passive-microwave satellite retrievals, which present well-known limitations, such as smoothing of the ice edge and underestimation of SIC during the melt season. We develop a Random Forest model that bias-corrects AMSR2 SIC observations towards an existing deep learning, SAR-derived product (ML-SAR) in the Arctic. The bias-corrected AMSR2 observations, together with ML-SAR, are then assimilated into the 1/4° configuration of the Met Office Forecast Ocean Assimilation Model (FOAM) as the key experiment. Assimilating ML-SAR and bias-corrected AMSR2 produces a sharper ice edge during the ice growing season and reduces the 5-day forecast Integrated Ice Edge Error by 30 % in summer, relative to a control experiment assimilating only AMSR2 alongside the standard ocean observations. It also largely mitigates the SIC underestimation in the melt season, raising the Critical Success Index against the ice charts in all SIC categories and at all 5-day forecast lead times, by about 25 % on average, with respect to the control. Because the bias-corrected AMSR2 is assimilated wherever ML-SAR is absent, sea ice forecast improvements reach the whole Arctic, including regions of limited SAR coverage, and could not be obtained by assimilating ML-SAR alone. The additional ice volume produced by assimilating both ML-SAR and bias-corrected AMSR2, particularly in spring and summer, contributes to alleviating a known thin bias in the central Arctic in FOAM, but also leads to an overestimation of thickness in the Arctic marginal seas. Overall, the results demonstrate that, by combining the spatial completeness of passive-microwave observations with the fidelity of SAR-based data, this learned bias correction offers a computationally inexpensive, observation-based pathway towards improved Arctic sea ice prediction.
How to cite. Mignac, D. and Martin, M. J.: Machine-learning bias correction of passive-microwave sea ice concentration against a SAR-derived product: impact on Arctic sea ice forecasts, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-5409, 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
- Preprint
(4148 KB) - Metadata XML
-
Supplement
(659 KB) - BibTeX
- EndNote
Status: open (until 26 Nov 2026)
Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor
| : Report abuse
-
RC1: 'Comment on egusphere-2026-5409', Anonymous Referee #1, 04 Nov 2026
reply
This manuscripts tests the assimilation of different Arctic sea ice concentration (SIC) products in the coupled ocean-sea ice Forecast Ocean Assimilation Model (FOAM) by the UK Met Office. Primarily, the authors evaluate the transition from assimilating daily Level 3 (L3) SIC from OSI SAF to swath-level L2 SIC and a combination of L2 SIC with a novel SAR-/AMSR2-based L3 SIC estimate from the Danish Meteorological Institute (DMI) named ML-SAR. To account for biases between OSI SAF L2 and ML-SAR L3 SIC, a random forest bias correction is applied to the OSI SAF L2 SIC using ML-SAR SIC as target. Improved consistency between both products is achieved after applying the bias correction, mainly due to the correction for SIC underestimation in the L2 SIC in summer in the presence of melt ponds. Forecasts for different lead times are validated using products of SIC (from ice charts), ice edge location (from MASIE), and sea ice thickness (from upward-looking sonar and satellite products). Results show improvements in SIC, ice edge position (summer only), and partly also in ice thickness when assimilating the corrected OSI SAF L2 SIC data alongside ML-SAR L3 data instead of OSI SAF L3 SIC data. The work is presented well and provides novel insights into assimilation of newly developed SIC retrievals into a coupled ocean-sea ice system. I recommend publication of this work subject to the comments below, which aim at further clarifying the work.General comments1) The final SIC product assimilated in the "AMSR2-BC-ML-SAR" experiment is a highly complex blend of different datasets with each having specific biases and limitations (AMSR2 TB, AMSR2 SIC, SAR backscatter, SAR-ML SIC, historical sea ice data, and ice charts with its various input data) and methods (AMSR2 tie point SIC retrieval, ML-SAR machine learning retrieval, random forest bias correction, and rather subjective human-generated ice charts to train the ML-SAR product). The AMSR2 TB is even used three times during this process in different ways in a potentially suboptimal way: (1) in the AMSR2 tie point SIC retrieval with an atmosphere correction and land correction at 18 and 36 GHz, (2) in the ML-SAR retrieval as uncorrected TB at all frequencies input alongside the SAR data, and (3) in the ice charts potentially as TB and/or SIC estimates. While this approach might yield an accurate SIC estimate available in near-real time, it is almost impossible to track down the assumptions and limitations of the final SIC product. I am wondering if the authors have any suggestions on how the SIC estimation can be simplified to make optimal use of the information content of different sensors (AMSR2, SAR, ...). Did the authors consider assimilating AMSR2 SIC as the area that is not covered by open melt ponds in the SI3 model (see Geer et al., 2026, egusphere-2026-3421)? Assuming that the SAR product is less sensitive to melt ponds, although it is not clear to me why the microwave backscatter at C-band from water in melt ponds should be different to water in leads, this could potentially provide new information to the model about melt pond fraction considering the model surface state (skin temperature, refrozen melt ponds, etc.).2) The definition of sharpening of the sea ice edge means increasing the SIC for grid points with high SIC and decreasing SIC for grid points with low SIC. This process would revert the smoothing effect of a large satellite footprint. Sharpening tools require spatial context to identify edges and distinguish them from homogeneous parts of the SIC field. Here, the authors use a tool that acts on single pixels only and is therefore by design unsuitable to sharpen any edges. However, the authors state that the main purpose of the random forest bias corrections is to propagate a sharper ice edge from ML-SAR into the L2 SIC product. As can be seen in Figure 6, the tool only increases SIC and never reduces SIC, i.e., it does not revert any smoothing effects. The authors might have added latitude and longitude predictors to help locate ice edges, but these are not stationary at a given time of the year. In its current form, I assume that the model applies the same SIC bias to a pixel within a homogeneous SIC field and a pixel at an ice edge. While this can lead to changes in existing SIC gradients if the estimated bias varies with SIC (see Figure 7), it should not be confused with a sharpening of ice edges as stated throughout the manuscript.3) Sensitivity tests of the random forest model would be helpful to understand the way it corrects the OSI SAF L2 SIC as a function of time of year, uncorrected SIC, and location. If regional differences are small, I suggest to perform sensitivity tests only for contrasting regions like Fram Strait and Central Arctic (time on horizontal axis, uncorrected SIC on vertical axis, SIC bias correction as shading). In Figure 4, it appears that there are changes up to +60% for low SIC, which can only be explained by very large melt pond fractions only. Together with snow or ice temperature and melt pond fraction from SI3, this would allow to gain more confidence into the bias correction applied. Especially considering potential confirmation bias when ice chart data is used to train the ML-SAR product and used for evaluation of SIC in this work. I would also appreciate if the authors could publish their trained random forest model for others to use.4) For clarity, I suggest to rename "ML-SAR" into "DMI-ASIP" as in Wulf et al. (2026), since the model is not solely based on SAR. It also uses AMSR2 TB from 6.9 to 89.0 GHz and historical sea ice records as input, and we do not know how much information about SIC gets extracted from SAR backscatter alone. Please refer to it as a "SAR-AMSR2" and not just as a "SAR" product throughout this manuscript. Also, DMI-ASIP is trained on regional ice chart SIC of the Greenland and Canadian Ice Services and therefore reproduces biases in this data described as "better-safe-than-sorry" approach in Wulf et al. (2026). The impact of these biases on the sea ice concentration assimilation should be discussed.Specific comments- line 157: Please provide further information on how AMSR2 data is used in ML-SAR, how both datasets are matched, and how it incorporates SIC climatologies mentioned in Wulf et al. (2026).- line 169: Ice charts are used as training target in the SAR-AMSR2 product. The statement that ice charts are "more likely to overestimate than to underestimate ice cover" questions their suitability. Could you explain why they are still considered a good SIC reference?- line 220: Why is the threshold of 15% chosen? And is it applied to both datasets? Based on Fig. 3, I assume it is removed from the target only. If the goal of this approach is to improve the sea ice edge, it should help to also include the "ocean side" of the ice edge in the training.- line 224: Please explain why the 80/20 split applied to three years of data (September 2021 to August 2024) still leaves a full annual cycle for testing (9 July 2023 to 30 August 2024). Is the data amount not balanced between different years?- line 225: "Because the AMSR2 underestimation relative to ML-SAR grows with concentration": I am not sure if I understand Fig. 3 correctly, but it appears to me that the underestimation is largest for a SIC target of about 60%. In these conditions, the peak for AMSR2 SIC is about 20%. For a target of 100%, the AMSR2 SIC is also mostly near 100%. See also general comment 3.- lines 291-295: What kind of information are used to generate the ice charts? What information is used for continuous SIC estimates? Also, if any of these ice charts overlap, it might be interesting to get information on their uncertainties.- line 299: Please check the concentration bounds, there seems to be a gap.- line 419: Is there any physical explanation for improved skill with respect to ice charts along the entire marginal ice zone in December to May?- line 544: Based on the results, the improvement in ice edge in summer is directly linked to the melt pond issue. The ice edge improves by increasing sea ice concentration of broken ice from below to above 15% east of Greenland. Other changes to the ice edge in winter have no impact on the IIEE.- line 551: The effect is seen only during summer (July and August), therefore "most notably" can be removed.- Figure 1: Comment on the difference between ML-SAR and the ice charts. Ice charts are used for training, so it is surprising to see large differences in June east of Greenland and near Siberia.- Figure 7: Information on how the gradients are computed is missing.- Figure 12: How sensitive is the IIEE result to the threshold used for the sea ice edge in AMSR2 SIC? MASIE uses a threshold of 40%, while a threshold of 15% is used for the AMSR2 SIC (see also Meier et al., 2015, 10.3189/2015AoG69A694). Similarly, error bars for sea ice thickness could be useful to see if the change in RMSE of less than 0.1 m is statistically significant.- Figure 13: Is it possible to show the sea ice edge as a line instead of dots? It is difficult to visualise its contours near Greenland.ReplyCitation: https://doi.org/
10.5194/egusphere-2026-5409-RC1
Viewed
Total article views: 390 (including HTML, PDF, and XML)
Cumulative views and downloads
(calculated since 15 Oct 2026)
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 254 | 112 | 24 | 390 | 54 | 36 | 34 |
- HTML: 254
- PDF: 112
- XML: 24
- Total: 390
- Supplement: 54
- BibTeX: 36
- EndNote: 34
Viewed (geographical distribution)
Total article views: 390 (including HTML, PDF, and XML)
Thereof 390 with geography defined
and 0 with unknown origin.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
1
Latest update: 04 Nov 2026
Davi Mignac
CORRESPONDING AUTHOR
Met Office, Exeter, EX1 3PB, United Kingdom
Matthew J. Martin
Met Office, Exeter, EX1 3PB, United Kingdom
Short summary
Satellites track Arctic sea ice daily, but the widely used microwave sensors blur the ice edge. We used machine learning to sharpen the satellite observations by teaching them from more detailed radar images, then fed the improved observations into a global ocean and sea-ice forecasting model. The corrected observations produced a clearer, better-placed ice edge and more accurate short-range forecasts. This could help make Arctic shipping and weather forecasts more reliable.
Satellites track Arctic sea ice daily, but the widely used microwave sensors blur the ice edge....