Classification of sea ice types in Sentinel-1 synthetic aperture radar images

Park, Jeong-Won; Korosov, Anton Andreevich; Babiker, Mohamed; Won, Joong-Sun; Hansen, Morten Wergeland; Kim, Hyun-Cheol

doi:https://doi.org/10.5194/tc-14-2629-2020

Articles | Volume 14, issue 8

https://doi.org/10.5194/tc-14-2629-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/tc-14-2629-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 14, issue 8

Research article

|

20 Aug 2020

Research article |

| 20 Aug 2020

Classification of sea ice types in Sentinel-1 synthetic aperture radar images

Jeong-Won Park, Anton Andreevich Korosov, Mohamed Babiker, Joong-Sun Won, Morten Wergeland Hansen, and Hyun-Cheol Kim

Download

Final revised paper (published on 20 Aug 2020)
Preprint (discussion started on 11 Jun 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of Park et al', Anonymous Referee #1, 28 Jun 2019
- AC1: 'Response to referee#1', Jeong-Won Park, 02 Dec 2019
RC2: 'Classification of ice types in Sentinel-1 SAR - review', Anonymous Referee #2, 16 Jul 2019
- AC2: 'Response to referee#2', Jeong-Won Park, 02 Dec 2019
RC3: 'Review of “Classification of Sea Ice Types in Sentinel-1 SAR images” by Jeong-Won Park et al. 2019 TCD', Anonymous Referee #3, 17 Jul 2019
- AC3: 'Response to referee#3', Jeong-Won Park, 02 Dec 2019

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (further review by editor and referees) (12 Dec 2019) by John Yackel

AR by Jeong-Won Park on behalf of the Authors (22 Jan 2020) Author's response Manuscript

ED: Publish subject to revisions (further review by editor and referees) (04 Feb 2020) by John Yackel

ED: Referee Nomination & Report Request started (03 Mar 2020) by John Yackel

RR by Anonymous Referee #1 (15 Mar 2020)

Suggestions for revision or reasons for rejection

I would like to thank the authors for addressing the comments given on the first round of revisions. There are a few more comments that I would like the authors to address before publication. Below are the original comments, authors responses start with * and new comments with _.

Specific comments
Mahmud et al 2018 showed that different sea ice types have different incidence angle dependent slopes. Have you considered using a sea ice type dependent slope factor? Moreover, how does the work by Mahmud et al 2018 fit in with the incidence angle dependencies presented here?
*Typically, the slope for open water is higher than that for sea ice, thus the correction works in a way reducing the difference in sigma naught for open water as well. As the review pointed, different sea ice types have different incidence angle dependent slopes; however again, ice type-specific correction prior to ice type classification is controversial. Although estimating ice type dependent slope is not a part of this manuscript, we provide the values derived from the training/validation dataset for the review purpose only.
_Studying the figures provided in the review, the HV sigma naught values for the different ice types and open water used in the training/validation data, it is striking that there is hardly any variation across incidence angle, but also that there is limited difference in backscatter values between the different types of surface. The values in the HV- data presented appear to be very low, especially considering that there is a difference in the backscatter values, e.g. for the open water, presented in Figure 7 and 8. As a side note would a scalebar in dB for Figure 7 and 8 be beneficial. Moreover, the values are close to the nominal noise floor provided with the meta data in the Sentinel-1 images. Is the lack of trend an effect of the thermal noise correction applied to the images?

_Work by e.g. Isleifson et al 2010 found slopes for new ice and all 4 channels given a range of incidence angles and sufficient SNR. Work by Gill et al 2015 also found slopes of -0.25 for the HV-channel and Scharien et al, 2014 also found slopes for the HV-channel. How does this work relate to those studies?

How do you define “good match” (P7 R3)? Temporal overlap? Spatial overlap? How was this manual selection of images performed? How were the open water vs sea ice charts, that are used as initial input into the classifier, derived?
*Both temporal and spatial overlaps are important. Since the SAR image itself is potentially one of the sources for ice charting at the ice services, some images spatially match well with the shape of the polygons in ice chart. Temporal window length of 3 days from the publication date of ice chart was used for squeezing the number of images to make decisions of use/discard for training.
_What is a good match? Please state this clearly in the manuscript. The sentence now simply states that spatially and temporally good matches were used. How large is the spatial overlap? What is meant with squeezing the number of images?

Figure 8. In Figure 3 and 7 the ice chart from 3 days later is used yet in Figure 8 the image from the same day is used. Please be consistent in which time interval is used for these weekly ice charts.
*Since the reference ice chart is published weekly, the same NIC ice chart both in Figure 7 and 8 is supposed to be valid for both dates of Figure 7 and 8, but it is not as shown. The common time interval of 3 days in Figure 3 and 7 is just a coincidence. It can be one day or two days depending on the date of SAR image that used as source materials for ice charting at ice services. If the satellite images acquired two days prior to the publication of weekly ice chart, then the overall distribution of ice would represent the status of that time. Note that among the 57 images used for training, 42 and 35 percent of the images were acquired 2 and 3 days prior to the publication date of the corresponding weekly ice chart, respectively.
_For the validation data in 2018 and 2019 how many of those images were acquired at 2 and 3 days prior to the ice chart used to validate them?

_The new ice areas likely change at short time scales. What is the impact of having 42% (2-days) and 35% (3-days) time separation between the image acquisition and the sea ice chart? Is there a consistent time difference between the 2018 and 2019 data with respect to the Sentinel-1 images and the sea ice charts? Comparing the accuracy of the sea ice classifications with the temporal difference might provide some answers here.

Technical comments
In many places references to the appropriate work is missing, e.g. P1, R27, P4 R15-16 and P6 R9-10. Please carefully revise the manuscript to include references to earlier work.
*Corrected.
_The P6 R9-10 still has no reference to previous work.

P6 R10-12 you argue that when training dataset are prepared manually the sample size is usually less then 20 images. Please provide several references to support this statement.
_In Liu et al 2015 only one image is used for the training data please revise the manuscript accordingly (P6 R29). Another image is used for the validation of their study but only one for the training data.

_The reference Leigh et al 2014 does not provide slopes for either the HH- or the HV-channel in their study. Please revise the manuscript accordingly.

In Figure 9 the sea ice type FYI thin is include in the sea ice classification results. Please clarify what thin FYI means. Why is this class not used throughout?
*The label was wrong. It was just FYI without “thin”. As the bottom panels (SAR results) are irrelevant in the context of the corresponding description, they were removed in the revised manuscript.
_In Figure 9 the label thin FYI is still present.

_Throughout the manuscript first year ice is not consistently named: e.g. Figure 3 FYI (top left) FY ice (bottom left), Figure 7 and 8 First-year ice. Page 3, Row 9. First year. Moreover, FYI is not defined within the manuscript.

_P1 R11. Number -> amount
_P1. R20 What is operational manner?
_P2. R1-3. Reference to “HH and HV widely used for ice edge detection etc” is missing.
_P2. R29-30. Please specify type of bias.
_P4. R.23 “…noise, some which originate...”
_P8 R13. Clarify easily or remove it. “…borders of the open water …”
_P8 R13-15. Unclear sentence please revise.
_P10 R18. For FC1 and FC2 there is an improvement, however for FC3 there is a -0.05 reduction in the Kappa values. Please revise the sentence to account for this.
_P11 R20. “different marginal sides”. What does this mean?
_P11 R21. Which figure is this image boundary visible in?
_P11 R22. What does extreme marine conditions mean? Please clarify.
_P11 R23-25. Are high wind conditions not accounted for/included in your training data?

References
J. P. S. Gill, J. J. Yackel, T. Geldsetzer, and M. C. Fuller, “Sensitivity of C-band synthetic aperture radar polarimetric parameters to snow thickness over landfast smooth first-year sea ice,” Remote Sens. Environ., vol. 166, pp. 34–49, Sep. 2015.

Scharien R. K., Landy J., and Barber D. G. (2014), First-year sea ice melt pond fraction estimation from dual-polarisation C-band SAR – Part 1: In situ observations, The Cryosphere, 8, 2147–2162. doi:10.5194/tc-8-2147-2014

Hide

RR by Anonymous Referee #2 (16 Mar 2020)

RR by Anonymous Referee #3 (20 Mar 2020)

ED: Reconsider after major revisions (further review by editor and referees) (24 Mar 2020) by John Yackel

AR by Jeong-Won Park on behalf of the Authors (11 May 2020) Author's response

ED: Referee Nomination & Report Request started (27 May 2020) by John Yackel

RR by Anonymous Referee #3 (09 Jun 2020)

RR by Anonymous Referee #1 (11 Jun 2020)

Suggestions for revision or reasons for rejection

The manuscript has now been improved since the last round of revisions, however some changes still needs to be made before it can be accepted for publication.

Specific comments
How is the overall accuracy calculated? If only 12% of the potentially dominant (?) class mFYI is correct how does this affect the overall accuracy of the algorithm? And if the algorithm now largely for summer is an open water vs sea ice classification, should it be clarified in the abstract that the method is not really a sea ice type classifier.

As the comparison with OSI SAF data now is quite a large section of the manuscript, consider moving some of the explanations on page 12 to the methods section, e.g. a validation section. Moreover, I suggest to convert old ice into MYI or vice versa when comparing the two results. Just clearly state what has been done. You claim that the reason for low accuracy for FYI in your method compared to the OSI-403-c is that “A large extent of old ice in NIC ice chart is annotated as multi-year ice in the OSI-403-c. This does not clarify why your algorithm has a large error for FYI classification. Please clarify.

A recently accepted study by Lohse et al. 2020 investigated slopes for Sentinel-1 images and found slopes of: -0.27 (HH) and -0.26 (HV) for level FYI, -0.23 (HH and HV) for MYI and for open water -0.72 (HH) and -0.33 (HV). Consider including these as well in your comparison given that they are covering the same area (though they have a Pan-Arctic coverage) and are also using Sentinel-1 EW images. Within their work they are using 80 Sentinel-1 EW images to establish these slopes and are using overlapping optical images and expert opinions from the Norwegian Ice service to establish the training and validation data. Consider also including this among other recent sea ice type studies as it is pointed out in the paper that such studies using Sentinel-1 images are missing.

Technical comments
Within the text when others use sea ice experts to help establish training and validation data they are referred to as: e.g. informal sources (P13, R3). Yet when you are using potentially the same experts they are referred to as: credible sources (P4, R1). In my opinion the experts are equally credible, please revise accordingly.

P1R.12 works -> work
P1R19. Two?
P1R22-23. Unclear sentence please revise.
P5 R1. Consider rewriting this sentence. HV generally has lower backscatter values but the noise subtraction does not especially yield negative values in HV, low values in both channels will be equally affected.
P5 R11. Consider adding also Lohse et al., 2020 here.
P10 R25. The FYI also has a small accuracy decrease here, please revise.
P12 R22-23. Unclear sentence please revise.

References
Lohse J., Doulgeris A.P and Dierking W., Mapping Sea Ice Types from Sentinel-1 Considering the Surface-Type Dependent Effect of Incidence Angle, Annals of Glaciology, 2020, doi: 10.1017/aog.2020.45

Hide

RR by Anonymous Referee #2 (12 Jun 2020)

ED: Publish subject to revisions (further review by editor and referees) (14 Jun 2020) by John Yackel

AR by Jeong-Won Park on behalf of the Authors (22 Jun 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (25 Jun 2020) by John Yackel

RR by Anonymous Referee #1 (26 Jun 2020)

RR by Anonymous Referee #2 (11 Jul 2020)

ED: Publish as is (14 Jul 2020) by John Yackel

AR by Jeong-Won Park on behalf of the Authors (17 Jul 2020) Manuscript

Short summary

A new Sentinel-1 radar-based sea ice classification algorithm is proposed. We show that the readily available ice charts from operational ice services can reduce the amount of manual work in preparation of large amounts of training/testing data and feed highly reliable data to the trainer in an efficient way. Test results showed that the classifier is capable of retrieving three generalized cover types with overall accuracy of 87 % and 67 % in the winter and summer seasons, respectively.