the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Jul 2025
Enhanced neural network classification for Arctic summer sea ice
Abstract. Lead/floe discrimination is essential for calculating sea ice freeboard and thickness (SIT) from radar altimetry. During the summer months (May–September) the classification is complicated by the presence of melt ponds. In this study, we develop a neural network to classify CryoSat-2 measurements during the summer months, building on the work by Dawson et al. (2022) with various improvements: (i) we expand the training dataset and make it more geographically and seasonally diverse, (ii) we introduce an additional thinned floe class, (iii) we design a deeper neural network and train it longer and (iv) we update the input parameters to data from the latest publicly available CryoSat-2 processing baseline. We show that both the expansion of the training data and the novel architecture increase the classification accuracy. The overall test accuracy improves from 77 ± 5 % to 84 ± 2 % and the lead user accuracy increases from 82 ± 10 % to 88 ± 5 % with the novel classifier. When used for SIT calculation, we observe minor improvements in agreement with the validation data. However, as more leads are detected with the new approach, we achieve better coverage especially in the marginal ice zone. The novel classifier presented here is used for the Summer Sea Ice CryoTEMPO (CryoSat Thematic Product).
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(15650 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(15650 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2025-2789', Sanggyun Lee, 24 Aug 2025
- AC1: 'Reply on RC1', Anne Braakmann-Folgmann, 22 Sep 2025
-
RC2: 'Comment on egusphere-2025-2789', Anonymous Referee #2, 25 Aug 2025
General Comments
In this manuscript, the authors build on the summer CryoSat-2 lead/floe classification convolution neural network (CNN) of Dawson et al. (2022) by expanding and improving the training dataset and introducing a new thin-ice classification. They then evaluate Arctic summer sea ice thickness and volume using this improved lead/floe classification. This is an important challenge for the sea ice community as melt ponds on summer sea ice complicate classification of the sea surface in radar altimetry, and historically, sea ice thickness estimates excluded summer for this reason. The authors find an increase in accuracy when using the new dataset, which is promising for improving current summer sea ice thickness retrievals.
Overall, this manuscript is well-written and relevant to the wider science community, where desire for year-round Arctic sea ice observations is high. However, I do have some general concerns, particularly regarding the thin ice classification.
- My biggest concern arises from Fig 3., where different waveform parameters are plotted for leads, good floes, noisy floes, and the new thinned floes classification. At present, I am not convinced by the characteristics used to determine the thinned floe classification. I would not expect a decrease in pulse peakiness (PP) over thinned floes, but rather an increase compared to good floes. We tend to see a reduction in peak power over thick floes and thin floes are likely to return a more specular waveform and thus have a higher PP (e.g., Rinne and Similä, 2016; Laxon, 1994). Zygmuntowksa et al. (2013) find that PP values from thin ice and leads are similar. I think further evaluation and clarification of this new class is required.
- The lack of leads in the central Arctic in July and August (Fig 7.) is noted as likely an artefact caused by the conservative treatment of leads in the classifier design, which I understand. However, this does make me doubt its performance in the peak of summer, as there are areas of almost 0% leads during these months, but almost 100% good floes. Does this mean sea ice freeboard/thickness cannot be calculated at all in these grid cells? What happens if the classifier is relaxed for leads? Have you assessed against imagery in these cases, for example? What are the classifier performance statistics during these months specifically?
- I think an opportunity has been missed in Fig. 6 to provide a more in-depth visual comparison to imagery. You have included one SAR image and a qualitative assessment in Section 4.2., but ~3 examples would be useful here so we can get a better understanding of how the new classification performs particularly if the ice is more complex or during July/August given the poorer lead detection in those months, rather than fairly consolidated ice as it is in this example.
Specific Comments
Figures – I note specific cases below, but generally the text and symbols on the figures are too small, and at times the labels are overlapping/not visible. Please fix.
Line 90 – Is this limit based on the CryoSat-2 footprint?
Fig. 1 – I’m not sure if there is a benefit to having the sea ice freeboard grid underneath the points, especially as it isn’t referenced in the text or caption. I think it makes the points harder to see. In the legend boxes, please also make the point symbols larger.
Lines 149-151 – These sentences would benefit from rephrasing; it states that there are a lack of summer snow depth datasets but then introduces the SnowModel-LG without context of how this modelled data can therefore be produced.
Lines 171-172 – What is meant by ‘misaligned tracks’? How do you determine if a track is misaligned? I’m concerned this would mean only tracks which seem to agree with the underlying image are retained.
Fig. 3 – Some of the figure labels are overlapping, which I appreciate is hard given the number of boxes. Do all of the axes need labelling, especially as for each row they are the same?
Table 1 – Is it possible to include monthly performance statistics for summer? I would be interested in how the lead classification performs in July/August based on my comment above.
Lines 399 – 401 – Yes, there is a substantial improvement in coverage, well done!
Section 4.6. – I think this section should come higher up in the text as it doesn’t feel as relevant at the end of the sea ice thickness/volume validation.
Grammatical Comments
Line 16 – Incomplete reference. This reference is also linked to a dataset, rather than a paper in the full reference list.
Line 59 – Reference needs brackets.
Line 89 – Reference needs brackets.
Line 109 – Reference needs brackets.
Lines 115-119 – There is some repeated text here from the paragraph above.
Line 142 – PIOMAS acronym has already been defined.
References – There are several incomplete references/DOIs. Please check.
Dawson, G. et al. (2022). A 10-year record of Arctic summer sea ice freeboard from CryoSat-2. Remote Sensing of Environment. 268, 112744. doi.10.1016/j.rse.2021.112744
Rinne, E. and Similä, M. (2016). Utilisation of CryoSat-2 SAR altimeter in operational ice charting. The Cryosphere, 10, 121-131. doi:10.5194/tc-10-121-2016
Laxon, S. (1994). Sea ice altimeter processing scheme at the EODC. International Journal of Remote Sensing, 15:4, 915-924. doi.10.1080/014311694408954124
Zygmuntowska, M. et al. (2013). Waveform classification of airborne synthetic aperture radar altimeter over Arctic sea ice. The Cryosphere, 7, 1315-1324. doi:10.5194/tc-7-1315-2013.
Citation: https://doi.org/10.5194/egusphere-2025-2789-RC2 - AC2: 'Reply on RC2', Anne Braakmann-Folgmann, 22 Sep 2025
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2025-2789', Sanggyun Lee, 24 Aug 2025
This study follows Dawson et al. (2022) and Landy et al. (2022), extending their work by adding more training samples and improving the CNN architecture. While the contributions are clear, there remain parts of the Methods section that require further elaboration to improve readability and reproducibility. I therefore recommend publication after a major revision.
General comments:
P4, L97: Since the addition of training samples is a key contribution of this study, I recommend providing a table summarizing the newly added samples, including their geographic region, month, and satellite source.
P4, L117: Could the authors clarify why a 7 km window was chosen?
P8, L172: Apparently, the lead and ice samples were manually extracted using visual inspection. However, as noted in Section 4.2, distinguishing between thinned floes and good floes is not straightforward in Sentinel-1 imagery. Could the authors comment on the potential impact of human error during the manual extraction of training samples, and how such uncertainty was mitigated?
P8, L175: If I understand correctly, Istomina et al. (2016) does not explicitly demonstrate that melt ponds and leads can be spectrally distinguished. It may be helpful for the authors to clarify how this reference supports their statement on spectral separation.
L8, 187: Since distinguishing leads from melt ponds is the most critical challenge in summer, could the authors clarify whether melt ponds are assumed to be entirely included in the noisy floe class? In addition, are refreezing ponds in August and September also included in this class? From a sea surface height (SSH) perspective, it may be difficult to separate thinned floes from refreezing ponds. It would be helpful if the authors could explain how refreezing ponds are treated in their classification.
Figure 3: It would strengthen the manuscript if the authors could provide quantitative evidence that the addition of the thinned floe class improves the classification performance.
Figure 6: Since surfaces with diverse geophysical conditions also occur in July, August, and September, it would be helpful to include additional examples of CryoSat-2 tracks with coincident imagery from these months. This would further support the robustness of the classification.
Minor comments:
P1, L16: One reference is missing the publication year. Please correct this.
Figure 1: The caption of Figure 1 is unclear as currently written. Please rephrase it to improve clarity.
P8, L163-164: The text refers the reader to Dawson et al. (2022), but without further explanation it may be difficult to understand Figure 3 and the use of the 11-point window. A short description would be helpful.
P10, L209: Please specify whether pooling refers to max pooling or mean pooling.
P10, L221: Did the authors also test the Adam optimizer? If so, was there any improvement in performance compared to RMSProp?
Table 1 and 2: Please place the captions above the tables rather than below.
Table 1: For the ice user/producer accuracies, does this metric include both good and noisy floes together? Please clarify.
Citation: https://doi.org/10.5194/egusphere-2025-2789-RC1 - AC1: 'Reply on RC1', Anne Braakmann-Folgmann, 22 Sep 2025
-
RC2: 'Comment on egusphere-2025-2789', Anonymous Referee #2, 25 Aug 2025
General Comments
In this manuscript, the authors build on the summer CryoSat-2 lead/floe classification convolution neural network (CNN) of Dawson et al. (2022) by expanding and improving the training dataset and introducing a new thin-ice classification. They then evaluate Arctic summer sea ice thickness and volume using this improved lead/floe classification. This is an important challenge for the sea ice community as melt ponds on summer sea ice complicate classification of the sea surface in radar altimetry, and historically, sea ice thickness estimates excluded summer for this reason. The authors find an increase in accuracy when using the new dataset, which is promising for improving current summer sea ice thickness retrievals.
Overall, this manuscript is well-written and relevant to the wider science community, where desire for year-round Arctic sea ice observations is high. However, I do have some general concerns, particularly regarding the thin ice classification.
- My biggest concern arises from Fig 3., where different waveform parameters are plotted for leads, good floes, noisy floes, and the new thinned floes classification. At present, I am not convinced by the characteristics used to determine the thinned floe classification. I would not expect a decrease in pulse peakiness (PP) over thinned floes, but rather an increase compared to good floes. We tend to see a reduction in peak power over thick floes and thin floes are likely to return a more specular waveform and thus have a higher PP (e.g., Rinne and Similä, 2016; Laxon, 1994). Zygmuntowksa et al. (2013) find that PP values from thin ice and leads are similar. I think further evaluation and clarification of this new class is required.
- The lack of leads in the central Arctic in July and August (Fig 7.) is noted as likely an artefact caused by the conservative treatment of leads in the classifier design, which I understand. However, this does make me doubt its performance in the peak of summer, as there are areas of almost 0% leads during these months, but almost 100% good floes. Does this mean sea ice freeboard/thickness cannot be calculated at all in these grid cells? What happens if the classifier is relaxed for leads? Have you assessed against imagery in these cases, for example? What are the classifier performance statistics during these months specifically?
- I think an opportunity has been missed in Fig. 6 to provide a more in-depth visual comparison to imagery. You have included one SAR image and a qualitative assessment in Section 4.2., but ~3 examples would be useful here so we can get a better understanding of how the new classification performs particularly if the ice is more complex or during July/August given the poorer lead detection in those months, rather than fairly consolidated ice as it is in this example.
Specific Comments
Figures – I note specific cases below, but generally the text and symbols on the figures are too small, and at times the labels are overlapping/not visible. Please fix.
Line 90 – Is this limit based on the CryoSat-2 footprint?
Fig. 1 – I’m not sure if there is a benefit to having the sea ice freeboard grid underneath the points, especially as it isn’t referenced in the text or caption. I think it makes the points harder to see. In the legend boxes, please also make the point symbols larger.
Lines 149-151 – These sentences would benefit from rephrasing; it states that there are a lack of summer snow depth datasets but then introduces the SnowModel-LG without context of how this modelled data can therefore be produced.
Lines 171-172 – What is meant by ‘misaligned tracks’? How do you determine if a track is misaligned? I’m concerned this would mean only tracks which seem to agree with the underlying image are retained.
Fig. 3 – Some of the figure labels are overlapping, which I appreciate is hard given the number of boxes. Do all of the axes need labelling, especially as for each row they are the same?
Table 1 – Is it possible to include monthly performance statistics for summer? I would be interested in how the lead classification performs in July/August based on my comment above.
Lines 399 – 401 – Yes, there is a substantial improvement in coverage, well done!
Section 4.6. – I think this section should come higher up in the text as it doesn’t feel as relevant at the end of the sea ice thickness/volume validation.
Grammatical Comments
Line 16 – Incomplete reference. This reference is also linked to a dataset, rather than a paper in the full reference list.
Line 59 – Reference needs brackets.
Line 89 – Reference needs brackets.
Line 109 – Reference needs brackets.
Lines 115-119 – There is some repeated text here from the paragraph above.
Line 142 – PIOMAS acronym has already been defined.
References – There are several incomplete references/DOIs. Please check.
Dawson, G. et al. (2022). A 10-year record of Arctic summer sea ice freeboard from CryoSat-2. Remote Sensing of Environment. 268, 112744. doi.10.1016/j.rse.2021.112744
Rinne, E. and Similä, M. (2016). Utilisation of CryoSat-2 SAR altimeter in operational ice charting. The Cryosphere, 10, 121-131. doi:10.5194/tc-10-121-2016
Laxon, S. (1994). Sea ice altimeter processing scheme at the EODC. International Journal of Remote Sensing, 15:4, 915-924. doi.10.1080/014311694408954124
Zygmuntowska, M. et al. (2013). Waveform classification of airborne synthetic aperture radar altimeter over Arctic sea ice. The Cryosphere, 7, 1315-1324. doi:10.5194/tc-7-1315-2013.
Citation: https://doi.org/10.5194/egusphere-2025-2789-RC2 - AC2: 'Reply on RC2', Anne Braakmann-Folgmann, 22 Sep 2025
Peer review completion
Journal article(s) based on this preprint
Data sets
Training data Anne Braakmann-Folgmann, Jack C. Landy and Geoffrey Dawson https://doi.org/10.5281/zenodo.15645704
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,398 | 190 | 40 | 1,628 | 73 | 74 |
- HTML: 1,398
- PDF: 190
- XML: 40
- Total: 1,628
- BibTeX: 73
- EndNote: 74
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Anne Braakmann-Folgmann
Jack C. Landy
Geoffrey Dawson
Robert Ricker
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(15650 KB) - Metadata XML
This study follows Dawson et al. (2022) and Landy et al. (2022), extending their work by adding more training samples and improving the CNN architecture. While the contributions are clear, there remain parts of the Methods section that require further elaboration to improve readability and reproducibility. I therefore recommend publication after a major revision.
General comments:
P4, L97: Since the addition of training samples is a key contribution of this study, I recommend providing a table summarizing the newly added samples, including their geographic region, month, and satellite source.
P4, L117: Could the authors clarify why a 7 km window was chosen?
P8, L172: Apparently, the lead and ice samples were manually extracted using visual inspection. However, as noted in Section 4.2, distinguishing between thinned floes and good floes is not straightforward in Sentinel-1 imagery. Could the authors comment on the potential impact of human error during the manual extraction of training samples, and how such uncertainty was mitigated?
P8, L175: If I understand correctly, Istomina et al. (2016) does not explicitly demonstrate that melt ponds and leads can be spectrally distinguished. It may be helpful for the authors to clarify how this reference supports their statement on spectral separation.
L8, 187: Since distinguishing leads from melt ponds is the most critical challenge in summer, could the authors clarify whether melt ponds are assumed to be entirely included in the noisy floe class? In addition, are refreezing ponds in August and September also included in this class? From a sea surface height (SSH) perspective, it may be difficult to separate thinned floes from refreezing ponds. It would be helpful if the authors could explain how refreezing ponds are treated in their classification.
Figure 3: It would strengthen the manuscript if the authors could provide quantitative evidence that the addition of the thinned floe class improves the classification performance.
Figure 6: Since surfaces with diverse geophysical conditions also occur in July, August, and September, it would be helpful to include additional examples of CryoSat-2 tracks with coincident imagery from these months. This would further support the robustness of the classification.
Minor comments:
P1, L16: One reference is missing the publication year. Please correct this.
Figure 1: The caption of Figure 1 is unclear as currently written. Please rephrase it to improve clarity.
P8, L163-164: The text refers the reader to Dawson et al. (2022), but without further explanation it may be difficult to understand Figure 3 and the use of the 11-point window. A short description would be helpful.
P10, L209: Please specify whether pooling refers to max pooling or mean pooling.
P10, L221: Did the authors also test the Adam optimizer? If so, was there any improvement in performance compared to RMSProp?
Table 1 and 2: Please place the captions above the tables rather than below.
Table 1: For the ice user/producer accuracies, does this metric include both good and noisy floes together? Please clarify.