SEL-Net: A Self-Supervised Learning-Based Network for PolSAR Image Runway Region Detection
Round 1
Reviewer 1 Report
The reviewer would like to thank the authors for this thoughtful manuscript. This work has good potential. The authors are requested to put in some additional efforts to improve the quality of this manuscript.
Introduction
The authors are requested to cite the following Earth observation article reporting the influence of atmospheric warming on cryospheric entities. Please elaborate on how the proposed methodology can assist such investigations for the detection of linear debris free glacial sections which holds similarity to runway area detection.
-Shugar et al, A massive rock and ice avalanche caused the 2021 disaster at Chamoli, Indian Himalaya, Science, 2021.
Quaternion Neural Network (QNN)
The authors are requested to discuss the strength of other supervised machine learning methods like QNN which have been used with PolSAR data. Please cite the following articles and highlight their contributions and utility with regards to the method proposed in the present investigation.
-Usami, N., et al., 2016. “Proposal of Wet Snow Mapping with Focus on Incident Angle Influential to Depolarization of Surface Scattering.” In IEEE International Geoscience and Remote Sensing Symposium 1544–1547.
-Fang Shang and Akira Hirose, "Quaternion neural-network-based PolSAR land classification in poincare-sphere-parameter space," IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 9, pp. 5693-5703 , 2014.
Crucial Polarimetric Feature: Variable Importance
Please highlight the PolSAR variables impacting the final performance and improvement.
Mathematical Expressions
Please provide only the relevant mathematical expressions.
Figures
The authors are requested to provide high-resolution figures. The metrics are not clear like in Fig. 9 and 18. The legend is almost invisible. Please remove the percentages from Fig. 26.
Tables
Please present the results in the barplot format for a better understanding.
Discussion and Conclusion
The authors are requested to list the key contributions in this section. At the moment the section is not detailed enough. Both these sections should be separate.
A grammatical check is necessary before the final publication.
Author Response
Dear Reviewer,
We want to thank you for the careful reviews, thoughtful and constructive comments, and valuable suggestions. We have made all the necessary amendments in our revised manuscript according to the comments and suggestions. All of the modified parts are highlighted in the revised manuscript. Please see the attachment for our reply to your concerns and comments.
Thank you for your time and consideration.
Best wishes,
Ping Han, Yanwen Peng, Zheng Cheng, Dayu Liao，Binbin Han
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors proposed a runway detection method from PolSAR images using self-supervised learning. This paper is technically sound, with moderate novelty by introducing additionally modules tailored for the tasks of interest. This paper is subject to changes listed below before publication.
1. I suggest the authors to simplify the abstract. There is not need to include the background/introduction part. Start straight with what you have proposed and what the results are.
2. Page 2, Line 86, what is D-Unet?
3. Page 6, Line 211, the purpose of the coherency matrix is not to “reduce the speckle”. Mathematically speaking, C = E(s * s^T) and should be estimated by multi-looking or spacial filtering.
4. Page 6, Line 236, the PauliRGB gives you 3 images, then you have 4 more feature images, wouldn’t that give you 7 images instead of 5?
5. The “Dynamic Dictionary” section needs clarification. I think what you meant to say is that the dictionary is kept at a fixed size, and when the # of items exceeds the size, the item is evicted in a FIFO way?
6. A general question: are you using single-look PolSAR image or multi-look PolSAR image? If it is the former, what is the spacial filtering used to get the T matrix?
7. Have you tried speckle filtering before detection? It seems a useful pre-processing given the large speckle noise of the PolSAR image and I’d love to see how or whether it can improve the performance.
Author Response
Dear Reviewer,
We want to thank you for the careful reviews, thoughtful and constructive comments, and valuable suggestions. We have made all the necessary amendments in our revised manuscript according to the comments and suggestions. All of the modified parts are highlighted in the revised manuscript. Please see the attachment for our reply to your concerns and comments.
Thank you for your time and consideration.
Best wishes,
Ping Han, Yanwen Peng, Zheng Cheng, Dayu Liao，Binbin Han
Author Response File: Author Response.docx
Reviewer 3 Report
In this paper, the authors construct a self-supervised learning detection network SEL-Net for airport runway area in PolSAR images. By improving the network structure, the authors reduce the semantic information loss in the network propagation process and enhance network’s ability to extract deep features and edge information. The analysis is convincing and the conclusions sounds reliable. Generally, this paper is well organized and the authors presented solid and comprehensive content.
Detailed comments are:
1) Since the authors concatenate multiple submodules in the model based on MOCO and Unet, such as EFM UIM, it is necessary to list the complexity of the entire model.
2) In Section 3.1, the authors take T11 and other feature images as additional input. Are the feature images filled with 3-channel images used? This will reduce the efficiency of the model. It would be easier to understand if the specific dimensions of the network structure were indicated in Figure 4.
3) In Section 3.2.5, the authors use a linear combination of BCE and DICE as the final loss of the detection network. How to optimize the weight coefficient λi of the loss function?
4) In Section 4.1.1, the authors have mentioned that the training results are affected by noise. What is the noise level of the dataset? Any despeckling method applied before training? Please explain it more.
5) The format of variables should be standardized and consistent.
Author Response
Dear Reviewer,
We want to thank you for the careful reviews, thoughtful and constructive comments, and valuable suggestions. We have made all the necessary amendments in our revised manuscript according to the comments and suggestions. All of the modified parts are highlighted in the revised manuscript. Please see the attachment for our reply to your concerns and comments.
Thank you for your time and consideration.
Best wishes,
Ping Han, Yanwen Peng, Zheng Cheng, Dayu Liao，Binbin Han
Author Response File: Author Response.docx
Reviewer 4 Report
This paper proposes a self-supervised learning-based network for runway region detection. The writing is poor. Please revise the language and logic of the article to highlight contributions. The issues are as follows:
1. In Introduction, Para 2, the authors state “one significant drawback of conventional approaches is that they are less suitable for real-time detection applications”, is your method suitable for real-time detection applications?
2. The shortcomings of the existing methods need to be revised. One of shortcomings is the scarcity of annotated training data. However, many methods have been proposed to detect runway with less labeled training data.
3. What is the shortcoming of the PolSAR runway detection algorithm?
4. Figure 2, “segmentation”, there is a space between “m” and “e”.
5. Figure 5, pay attention to “images”. It is inappropriate for y1 and y2 to be represented by dashed boxes.
6. The difference between the two yi in Figure 8?
7. The proposed method is usually placed in the last.
Minor modification is required.
Author Response
Dear Reviewer,
We want to thank you for the careful reviews, thoughtful and constructive comments, and valuable suggestions. We have made all the necessary amendments in our revised manuscript according to the comments and suggestions. All of the modified parts are highlighted in the revised manuscript. Please see the attachment for our reply to your concerns and comments.
Thank you for your time and consideration.
Best wishes,
Ping Han, Yanwen Peng, Zheng Cheng, Dayu Liao，Binbin Han
Author Response File: Author Response.docx
