A U-Net Approach for InSAR Phase Unwrapping and Denoising

Round 1

Reviewer 1 Report (Previous Reviewer 2)

This manuscript proposes a multi-stage phase unwrapping deep neural network framework based on U-Net, to jointly unwrap and denoise InSAR data. Experimental results demonstrate that proposed approach outperforms related work in the presence of phase noise and discontinuities. In summary, the research is interesting and provides valuable results. However, I have the following concerns regarding this paper：

(1) The Classical methods to solve SBPU and Deep learning for Phase Unwrapping section should include a summary table containing the following information for each related work: a) reference number, b) brief methodology, c) highlights, and d) limitations. 

(2) Vision technology integrated with deep learning is emerging these years in various engineering fields. The authors may add more state-of-art computer vision application articles for the integrity of the introduction (Novel visual crack width measurement based on backbone double-scale features for improved detection automation; Engineering Structures.).

(3) A comprehensive discussion of the results is necessary.

(4) It is advisable to review the language used in the paper, particularly the utilization of active and passive sentence structures. 

(5) In the introduction, it was mentioned that the unavailability of large publicly available datasets and accurate simulators are another challenge. However, in the experimental section, the simulator was directly used to generate the dataset. So what is the difference between datasets and emulator? ”The unavailability of large publicly available datasets, huge computational resources, and accurate simulators is another challenge that especially affects deep learning-based solvers.””The training, validation, and test interferogram datasets were generated using the simulator proposed by Xu et al [43]. The simulator provides capabilities of creating phase maps of many complex mathematical patterns and also allows adding noise to the interferograms while generating the dataset.”

(6) For visual measurement applications, please refer to Seismic performance evaluation of recycled aggregate concrete-filled steel tubular columns with field strain detected via a novel mark-free vision method; Structures. Visual measurement of dam concrete cracks based on U-net and improved thinning algorithm;  Journal of Experimental Mechanics.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

This study develops a new method to unwrap SAR interferograms using a deep-learning architecture called U-net. The manuscript is well-organized, and this study is well-motivated. Also, the proposed methodology looks plausible. Therefore, I do not have many fundamental comments, but I only have rather technical comments. I want the authors to address the following points to improve the manuscript. 

 

1. The author claims that the method developed in this study is robust in the presence of noise. However, Figures 8, 9, and 10 indicate that the authors worked on synthetic images with a high signal-to-noise ratio. Unwrapping these images is easy after spatial smoothing to eliminate short-wavelength noises. I suggest the authors work on synthetic data with more noise so that it is challenging to unwrap from visual inspection. 

 

2. Related to the comment above, what noises are added in the synthetic interferogram? The authors argue that they generated synthetic interferograms using a Xu et al. method (Line 305). However, more detailed explanations about the added synthetic noise, as well as signals, are necessary. 

 

3. Most of the figures are too small to be visible. I want the authors to enlarge figures so that at least the characters in each figure are visible. Also, some figures need to be of higher resolution. 

 

4. Figures 6 and 7 should be in the main text as equations. 

 

5. Line 19: I understand that RMSE is unitless (equation 15), but at this point, readers do not know the definition of RMSE. Therefore, I suggest the author 1) unabbreviate RMSE and 2) describe that the error is low. 

 

6. Line 43: w.r.t: with respect to 

 

7. Line 44 Remove the comma after "the presence of noise." 

 

8. Line 149: "SNAPU" should be "SNAPHU." 

 

9. Line 201: The authors claim N=3 here, but I wonder if N should be 2 because this study attempts to unwrap a single interferogram in which the data is distributed in two dimensions. 

 

10. Line 258: N^2XN^2: I did not understand what X stands for. Is it the multiplier? Then, the authors should represent it as "*." 

 

11. Line 315: I am afraid that 5000 interferograms are not enough as the training dataset. I wonder if thousands or hundreds of thousands of images are necessary for the training dataset. How is this number of images justified? 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report (New Reviewer)

The author describes a new deep-learning-based phase unwrapping algorithm. Overall, the article is well-structured and comprehensive, detailing the entire process of the algorithm, including dataset generation, model training and validation, and performance evaluation. The proposed algorithm is compared with existing algorithms, and its advantages in different performance indicators are demonstrated using RMSE and SSIM. However, I believe the article needs a more specific and detailed description of the methods; the current version lacks technical details and explanations, which are not conducive to the understanding and reproduction of the methods. Specific suggestions are as follows:

1. Page 1 Line 23: The introduction to the background is quite brief. Although there is some introduction to the phase unwrapping problem, more detailed background information can be added, such as the importance of phase unwrapping in practical applications, existing unwrapping algorithms, and their limitations.

2. Page 6 Line 232: It is better to give the corresponding references where deep learning segmentation-based phase estimation methods suffer from poor noise immunity.

3. Page 6 Line 237: The author claims to propose a "two-stage" method, but it is difficult to clearly see the two stages in Figure 5.

4. Page 9 Line 304: When describing the experimental dataset, the author mentions the use of a simulator proposed by Xu et al. to generate the dataset but does not mention the specific details of the simulator or the reasons for using it. I suggest adding details about the simulator.

5. Page 9 Line 308: The author claims that the simulator supports the generation of interferograms with more than 2π jumps, changes were made to restrict the phase jumps to less than 2π to satisfy Itoh’s continuity assumption, I suggest explaining this in more detail. Additionally, the author needs to provide more information on network training and construction, such as specific loss functions, network initialization methods, and whether data augmentation techniques were used.

6. Page 9 Line 303: Although the author provides quantitative results, comparison images (such as error maps) might help visually demonstrate the algorithm's performance and differences from other algorithms.

7. Page 9 Line 303: In the comparative experiments, although the RMSE and SSIM values of different algorithms are provided, there is no in-depth analysis or discussion of these results. It is recommended to add some analysis or discussion about the experimental results, such as why the proposed algorithm can achieve better performance, and what makes it perform better in noise handling, etc.

8. An important part of noise reduction methods is the verification of results, and the author should add some new references, such as, doi: 10.1016/j.fuel.2023.129584

Fair.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report (New Reviewer)

Can be accepted.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

This manuscript is not a typical research paper and need big improvement before the re-submission.

1. The literature reviews are not included in the Introduction. Too much basic knowledge was presented. 

2. The background, novelty, results was not clearly shown in the Abstract. 

3. The simulation case for the unwrapping test is too simple, which could be easily unwrapped by the tranditional methods. It is suggested to simulated some difficult case or real data to emphasize the advantages of the proposed method. 

Extensive editing of English language required

Reviewer 2 Report

This manuscript presents a novel multi-stage phase unwrapping deep neural network framework based on U-Net for joint unwrapping and denoising of InSAR data. The experimental results demonstrate that the proposed approach outperforms existing methods in terms of handling phase noise and discontinuities. Overall, the research is interesting and provides valuable results. However, I have the following concerns regarding this paper:

(1) Inconsistency in the description: The abstract states, "In this paper, we present a multi-stage phase unwrapping deep neural network framework based on U-Net, to jointly unwrap and denoise InSAR data," while in the "3. Materials and Methods" section, it is mentioned, "Our architecture is a two-stage Deep Neural Network (DNN)." The description should be consistent throughout the manuscript.

(2) Lack of quantified data in the abstract: The abstract does not provide quantified data to illustrate the superiority of the proposed method over other existing methods.

(3) Poor image clarity: The figures, especially Figure 9, have unclear axes and titles, which should be addressed to improve their clarity.

(4) Irrelevant references in the "2. Related Works" section: The paper cites numerous works that are unrelated to the experimental comparison, such as "Herraez et al. [50], Chartrand et al. [51], Perera et al. [52]." Additionally, the methods used for comparison are not adequately described.

(5) Lack of experimental setup details: The manuscript does not specify the hardware devices or development environment used for the experiments. Additionally, there is a lack of information regarding experimental parameters such as learning rate, number of iterations, batch size, etc.

 

(6) Insufficient evaluation metrics: The paper only includes "average mean square error (MSE)" and "worst-case mean square error" as evaluation metrics, which may not sufficiently demonstrate the advancement of the proposed method. Consider including additional evaluation metrics to provide a comprehensive assessment.

Overall, addressing these concerns would improve the clarity, consistency, and comprehensiveness of the manuscript.