Transfer-Aware Graph U-Net with Cross-Level Interactions for PolSAR Image Semantic Segmentation

Round 1

Reviewer 1 Report

The goal of this paper is to propose a novel graph convolutional network, CLIGUNet, for robust PolSAR image classification, overcoming the limitations of existing approaches by introducing weighted max-relative spatial convolution and multi-scale dynamic graphs for enhanced feature learning and generalization across datasets.

The authors may highlight the significance of integrating advanced statistical and machine learning approaches in remote sensing applications to provide a broader context for their work on PolSAR image classification with graph convolutional networks. To this aim, the authors may refer to  https://doi.org/10.1007/s11069-023-06240-2 and https://doi.org/10.1016/j.rse.2023.113891, as two most recent examples.

Author Response

Dear Reviewer:

We sincerely thank you for your constructive advice and valuable comments, which are of great help in improving this manuscript. Accordingly, the revised manuscript has been systematically improved. Moreover, the originality and innovations of our proposed CLIGUNet have also been emphasized, which have been stressed in the abstract, introduction, methodologies, experiments, and conclusion.

We have carefully read the two journal papers and decided to use them as reference papers, which have been marked out in red color, to provide a broader context for our work on PolSAR image classification with graph convolutional networks. 

Yours Sincerely,

Shijie Ren

Feng Zhou

Lorenzo Bruzzone

26^th March 2024

Reviewer 2 Report

I have carefully reviewed your paper titled "Transfer-Aware Graph U-Net with Cross-Level Interactions for PolSAR Image Semantic Segmentation" and found it to be a significant contribution to the field of remote sensing and semantic segmentation. Below are comments and suggestions aimed at further improving the quality and impact of your work:

1.     The paper is generally well-written and organized. However, providing a clearer motivation at the beginning of the paper, explaining the importance and challenges of PolSAR image segmentation, would better engage readers and set the context for your work.

2.     The incorporation of graph-based structures into the U-Net architecture is intriguing and novel. However, more detailed explanations of how these structures are integrated and their advantages over traditional CNN architectures would enhance the understanding of your approach.

3.     While you briefly mentioned the transfer learning strategy, providing more insights into the choice of source domain, the process of domain adaptation, and how it benefits PolSAR image segmentation would enrich the methodological description.

4.     The experimental setup section lacks some details, such as the specific datasets used, the size and characteristics of the datasets, and the evaluation metrics employed. Including this information would make the experimental results more reproducible and meaningful.

5.     While the experimental results demonstrate the effectiveness of your approach, it would be beneficial to include comparisons with a wider range of state-of-the-art methods, including both traditional and deep learning-based approaches, to provide a more comprehensive assessment of your method's performance.

6.     Conducting ablation studies to analyze the contribution of different components of your proposed model would help understand the effectiveness of the individual design choices and provide insights into areas for potential improvement.

7.     Including visualizations of segmentation results, such as segmentation maps or qualitative comparisons with ground truth annotations, would aid in understanding the performance of your model and provide intuitive insights into its strengths and weaknesses.

8.     Acknowledging the limitations of your approach, such as its sensitivity to certain types of terrain or weather conditions, and discussing potential strategies for mitigating these limitations would add depth to your paper and guide future research directions.

9.     Providing access to the code implementation of your proposed method would facilitate reproducibility and enable other researchers to build upon your work. Consider sharing your code on platforms like GitHub along with clear documentation and instructions.

10. Strengthening the conclusion section by summarizing the key findings, reiterating the significance of your contributions, and outlining specific avenues for future research would provide a satisfying closure to your paper and inspire further exploration in the field.

Moderate editing of English language required

Author Response

Comment 1. The paper is generally well-written and organized. However, providing a clearer motivation at the beginning of the paper, explaining the importance and challenges of PolSAR image segmentation, would better engage readers and set the context for your work.

Author response:

Thank you very much for the comment. In our revised manuscript, we have rewritten the introduction to better describe and emphasize the novelty of our method, together with more citations to better engage readers and explain the importance of PolASR image segmentation.

Comment 2. The incorporation of graph-based structures into the U-Net architecture is intriguing and novel. However, more detailed explanations of how these structures are integrated and their advantages over traditional CNN architectures would enhance the understanding of your approach.

Author response:

Thank you very much for the comment. In our revised manuscript, we have revised the introduction and conclusion to better describe how these structures are integrated into our method, and the ablation study of each component with time cost also analyzes the advantages of our method over traditional CNN architectures.

Comment 3. While you briefly mentioned the transfer learning strategy, providing more insights into the choice of source domain, the process of domain adaptation, and how it benefits PolSAR image segmentation would enrich the methodological description.

Author response:

Thank you very much for the comment. In our revised manuscript, we have revised Section 3.5 to provide more details about the choice of source domain, target domain, and how these domains are processed before transfer. It is worth noting that Section 3.5 aims to test the generalization performance of comparative methods without using an additional domain alignment module or loss function to minimize the distribution discrepancy between the source and target domains, this is because the target PolSAR datasets are usually not available in real applications. We will try to propose more robust PolSAR image segmentation methods with domain shifts in our future works.

Comment 4. The experimental setup section lacks some details, such as the specific datasets used, the size and characteristics of the datasets, and the evaluation metrics employed. Including this information would make the experimental results more reproducible and meaningful.

Author response:

Thank you very much for the comment. In our revised manuscript, we have added Table 3 in Section 3.1 to show these details, including the dataset name, radar platform, band, year, resolution, polarimetric type, image size, and land cover types.

Comment 5. While the experimental results demonstrate the effectiveness of your approach, it would be beneficial to include comparisons with a wider range of state-of-the-art methods, including both traditional and deep learning-based approaches, to provide a more comprehensive assessment of your method's performance.

Author response:

Thank you very much for the comment. In our revised manuscript, we have added SVM as the representative of the traditional method, together with the segmentation results of the PolSAR datasets, which can be found in Figure 6, Figure 7, and Table 6, Table 7.

Comment 6. Conducting ablation studies to analyze the contribution of different components of your proposed model would help understand the effectiveness of the individual design choices and provide insights into areas for potential improvement.

Author response:

Thank you very much for the comment. In our revised manuscript, we have analyzed the effectiveness of each component in the ablation studies to better describe and emphasize the novelty of our method. Besides, the time cost and improvements listed in the tables show that our proposed components can help to improve the segmentation performance of our CLIGUNet.

Comment 7. Including visualizations of segmentation results, such as segmentation maps or qualitative comparisons with ground truth annotations, would aid in understanding the performance of your model and provide intuitive insights into its strengths and weaknesses.

Author response:

Thank you very much for the comment. In our revised manuscript, we have added the ground truth of each dataset in Section 3.3, Section 3.4, and Section 3.5, please find them in Figure 6, Figure 7, Figure 8, and Figure 9.

Comment 8. Acknowledging the limitations of your approach, such as its sensitivity to certain types of terrain or weather conditions, and discussing potential strategies for mitigating these limitations would add depth to your paper and guide future research directions.

Author response:

Thank you very much for the comment. In our revised manuscript, we have acknowledged the limitations of our approach in Section 4. In our future work, we will consider analyzing the sensitivity of existing deep-learning approaches to certain types of terrain, land covers, weather conditions, frequency bands, resolution, and radar incidence angles.

Comment 9. Providing access to the code implementation of your proposed method would facilitate reproducibility and enable other researchers to build upon your work. Consider sharing your code on platforms like GitHub along with clear documentation and instructions.

Author response:

Thank you very much for the comment. We have been working on the code to make it easier to understand and adaptable to other datasets. Our code is expected to be released on GitHub after our paper is accepted.

Comment 10. Strengthening the conclusion section by summarizing the key findings, reiterating the significance of your contributions, and outlining specific avenues for future research would provide a satisfying closure to your paper and inspire further exploration in the field.

Author response:

Thank you very much for the comment. In our revised manuscript, we have summarized the main innovations and outlined specific avenues for future work in Section 4.

Yours Sincerely,

Shijie Ren

Feng Zhou

Lorenzo Bruzzone

31^st March 2024

Reviewer 3 Report

This manuscript introduces a transfer-aware graph U-Net with cross-level interactions to achieve robust semantic segmentation for polarimetric synthetic aperture radar (PolSAR), which could produce better classification results than SOA methods and handle unknown imagery with similar land covers. The methodology is described in detail and clearly. The organization of the manuscript is fair, and the mathematical derivations are reasonable. The main innovations of this method lie in that it can generalize effective feature representations across various resolutions with trainable residual transformers, which also enables simultaneous learning of latent features from batch input with multi-scale dynamic graphs, and that is the main contribution of the paper as well. Besides, the weighted max-relative spatial convolution derived from symmetric revised Wishart distance enables trainable adjacency matrices with refined graph topology, and that is also a spotlight of the paper. Hence，I recommend this manuscript can be accepted after minor revision.

 There are a few details that need to be refined：

(1) The organization of the experimental settings can be improved by addressing the superiority over the SOA methods at the end of section 3.1.4. Besides, the training and validation data used for the different methods in the section 3.5 (generalization studies across datasets) shall be indicated in section 3.1.4.

(2) Some concepts are repeated more than once, e.g. the abbreviations (CLIGUNet) in line 153 and line 158.

(3) There are problems with the notation. In the paper, some symbols are used with different meanings. This creates confusion and should be fixed. For example, in line 250, N denotes the batch size of CLIGUNet. In line 280, N denotes the dimension of the second power of A, which should be the number of pixels in each image patch.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments follow in the review letter

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

The reviewer highlights the authors' great flexibility and patience with criticism realized during the review process. The answers and updates to the texts were enlightening and helped clarify any doubts during the second revision process.

However, I leave two suggestions which, in the opinion of the reviewer, could be added to the conclusion of the work to goal to improve it. They follow:

1) The authors worked hard to insert the processing times in the table. Because of this, they could insert comments on the methods' processing times in the article's conclusion, as they did in the response letter.

2) In the conclusion, like future works, the authors could mention the intention not to use pre-processing and to adapt the proposed method to non-Gaussian distributions, as mentioned in the response letter.

 Yours sincerely.

The reviewer

Comments for author File: Comments.pdf

Author Response

Comment 1. The authors worked hard to insert the processing times in the table. Because of this, they could insert comments on the methods' processing times in the article's conclusion, as they did in the response letter.

Author response:

Thank you very much for the comment. In our revised manuscript, we have made comments on the processing time of our CLIGUNet and comparative methods in the conclusion, which can be found in Section 4.

Comment 2. In the conclusion, like future works, the authors could mention the intention not to use pre-processing and to adapt the proposed method to non-Gaussian distributions, as mentioned in the response letter.

Author response:

Thank you very much for the comment. In our revised manuscript, we have mentioned the intention not to use prepossessing and possible research directions with non-Gaussian distributions in future works, which can be found in Section 4.

Yours Sincerely,

Shijie Ren

Feng Zhou

Lorenzo Bruzzone

9^th April 2024