Sea Ice Detection from GNSS-R Data Based on Residual Network

Round 1

Reviewer 1 Report

The authors applied the ResNet model to detect the sea-ice from TDS-1 measurements. The paper is well organized, and can be accepted in Remote Sensing with minor revisions. The detailed comments are presented as follows:

Ⅰ. Major problems

1. In the paper, the ResNet model is compared to the AlexNet, LeNet, and Decision tree, and got the best results. However, the detail structure info of the AlexNet and LeNet are not present in the paper, since the different struct of the CNN model may result in different result. It’s better to present the details of the AlexNet and LeNet, which makes your conclusions more convincing.

2. In Line 218, the original size of the TDS-1 DDM is 128*20. After the stretching, the DDM size changes to 32*32? Please describe this part in detail.

3. In Section 3.2.3, it’s better to list a table or draw a statistics map to show the error of sea ice detection compared to the ground truth for each method.

4. There are many inappropriate English expressions in this paper. It needs to be polished.

Ⅱ. Minor problems

1.      In Section 2.3, Line 202, the serial number used here is inappropriate.

2.      Line 206, the expression “Literature[25]….” is inappropriate. We prefer to use “Author[25]…”.

3.      Line 277, “consequently leading to reduced sea ice detection accuracy”. Or leading to the reduced seawater detection accuracy.

4.      The index “F1-score” needs to be explained in the paper.

5.      Line 335, the “it can be observed that the SNR of LeNet decreases with increasing noise…”, what’s the meaning of “the SNR of LeNet”?

6. The titles of section used in the paper needs to be improved?

There are many inappropriate English expressions in this paper. It needs to be polished.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The paper “Sea Ice Detection From GNSS-R Data Based on Residucal Network” described the methodology of applying resnet based CNN on specular reflection information observed by satellite gnss-r receiver with TDS-1.

Application of CNN over DDM is interesting idea. However, Still need further information needs to be described with revision data presentation.  Could you explain why you employed TDS-1 but not CyGNSS data for your work? What is the advantage of TDS-1 over the CyGNSS, newer observation system?

Fundamentally, It made me difficult to evaluate the paper due to the lack of information about how did you treat footprint size and spatial resolution of each specular point (As you know, each specular point has greatly different footprint size). This information is very important to set the window size of convolutional window size over an image (and how to treat ground truth data as well). DDM information also contains shape and spatial contribution weight of target signature. Please also describe how your CNN treat such spatial information.

Specific comments are as follows. Another difficulty is that there was no description about the ground truth data in methodology section. Please provide the information and also present a figure of scatter plot between gnss-r reflectivity and ground truth based physical values.

1)     L56-L70. As described here, each scientist conducts data quality control or removal of low quality specular point data. But in your paper, I could not find the description how did you do it in this study. Please provide it in section 2-1.

2)

L86-L94 can be skipped. Rather than that, please describe about the receiver/satellite.

If you used satellite data such TDS-1, please describe the quantity/quality of data you used. (e.g., duration of observation period you used, number of specular points,  range of incidence angle e.t.c.,). If you conducted some quality control or data removal, please also describe the criteria of data removal as well.

Please collect English of the legend of figure 2. “ddms prone to detection errors” is weird.

Figure 2  please describe the value/unit of delay/doppler on the horizontal/vertical axis in the figure. Please also collect English of the legend of figure 2. “ddms prone to detection errors” is weird.

Please remove subsection 2.2 L131-137

It is too short and contained information is too little.

L138 3.1.1.1 subsubsection.

This is not appropriate to start suddenly 3.1.1. and title “subsubsection” is not appropriate.

Did you apply CNN over DDM image?

Did you apply CNN over DDM or image? Please clearly state at the beginning.

Describing sattlite informati from section 3/result is not appropriate. Please move this to section 2

L211-L222. The satellite information appeared suddenly here. It is weird since satellite info. Is describe in section-3 later...

Insection 2, please describe your ground truth data as well.

some times appropriate literature are found like a legend of figure 2-2. more importantly, logic is sometimes corrupted. please have an extinsive edition. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript uses deep learning to detect sea ice through remote sensing images with very important scientific significance. The structure of the paper is rigorous and the discussion is complete, but a brief introduction to the sea ice situation in the study area is required. Suggest minor revisions for publication.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

This article introduces a sea ice detection method using Residual Neural Network (ResNet) based on Delay Doppler maps (DDMs) collected from TechDemoSat-1 (TDS-1). The proposed method demonstrates a high accuracy of 98.61% in sea ice detection even during the melting period. ResNet's utilization results in improved accuracy, robustness to noise, and consistent stability during the melting phase compared to other sea ice detection algorithms.

1. In page 7, line 202, the authors discuss the DDM preprocess in terms of invalid data and noise removal. The authors has mentioned the four rows as the area for the nise calculation. Lines 207 to 209 are not clear.

2. In noise removal process, although Yan et al. (2017 and 2019; references 23 and 25) have given a strong methodology for DDM noise assessment, recent studies are recommended to be cited and followed in terms of noise removal, and in the next step, coherent detection, which is a modern topic in recent GNSS-R missions. The authors may refer to https://doi.org/10.1109/TGRS.2020.3009784, as an example.

3. The elevation angle range is recommended to be included among the data description as it can be decisive in terms of SNR, which is discussed in section 3.2.2 pf the manuscript.

4. Since the paper focuses on the application of GNSS-R in the remote sensing of cryosphere, the authors are recommended to refer to https://doi.org/10.3390/rs12172721 as one of the most recent applications of GNSS-R for lake ice remote sensing.

5. In section 3.2.3, the authors have used NOAA data as a source of validation. The authors are recommended to mention the accuracy of NOAA data. This article https://doi.org/10.3390/rs11020155 can be a good reference.

6. The authors are recommended to include the sea ice types in their manuscript, i.e., FYI, MYI, etc., as they may have different signature on GNSS-R signals. This paper can be a good refernce https://doi.org/10.1016/j.rse.2019.05.021

7. Minor language edition is required.

Minor edition is required, e.g., see the line 159 and 160. 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The paper has been improved. but I suggest to merge figure 6and 7. and reconstruct figure 2 whose subfigure is spacerly allocated.

it was improved

Author Response

Please see the attachment.

Author Response File: Author Response.docx