A Phase Filtering Method with Scale Recurrent Networks for InSAR
Round 1
Reviewer 1 Report
In the newly inserted parts I found two things that should be corrected:
line 289: "R is near 1 because s1 and s2 is very large" -> "R is near 1 because s1 is much larger than s2" (linear patch in [47])
line 397: it should be (250-150)/10+1
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Thank you for the opportunity to evaluate the paper after re-editing. The article is much better written than the 1st version, it is more understandable for the reader. However, the authors did not avoid a few methodological errors, details of which I will describe below. Most of the comments concern the revised part of the article.
Line 385 and 397. In line 385 the authors write about 25600 pairs of real SAR photos, in the next line about 900 photos, and in line 397 only about 396. Please explain where, from the number of almost 400 pairs of interferometric photos, such a large number of pairs for analysis were obtained.
Line 405: The authors used only 2 days of satellite flights on the sample surface and as a reviewer I do not understand how it was possible to develop 26.5 thousand. pairs of interferograms, please provide a precise explanation.
Line 438: Line 438: Were TerraSAR-X data also split into smaller sections? Two days of the exhibition on February 12 and 23 is only one pair!
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
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
The reviewed article concerns the use of artificial neural networks in the filtration of SAR images. This is an interesting issue, but use the "black box" of neural networks in the analysis of precise SAR data may be questionable. The authors of the study did not fully explain the described research problem. The introduction to the input data filtering in image processing has been very well developed
Below are the most important detailed comments on the study.
Line 105-106 The main assumption of the SAR methods is deformation analysis, the determination of the height is only additional information - a relative value.
Figure 1, it would be good for the authors to add all the markings used in formulas 1-3, or to propose a new, more detailed drawing.
The accuracy of the method should be analyzed in the "Problem description" section, as the accuracy of the study results from the input values and not from the parameters of the neural network. This approach only answers the quality of the network, not the accuracy of the method.
Figure 6. Scaling an image causes to loss of information and need to generalize the source data. The algorithm used increases the loss of information, scaling "down" and then "up". Are the data not mutually exclusive with subsequent scaling steps?
lines 250-251: The authors used a very large number of pairs of photos to develop, both theoretical and real. What will happen if we use only a few dozen pairs of interferometric images in the study? A large number of pairs of images gives great opportunities for data repeatability, a small number of pairs does not give such results. Is the use of self-learning neural networks only applicable to large amounts of data?
Figure 11. What information is on the vertical axis and what is on the horizontal axis? Please describe the drawing.
Line 268: The authors tested different filters, or for the same number of pairs of images?
Line 290: Please explain the SNR abbreviation that appears many times in the study.
Line 315: Please describe in more detail the data used: time range, number of views, type of path, etc.
Lines 321-322. Comparing the filtering results in Figures 15 and 16, there are no significant differences (apart from c) in the results of the methods used and the proposed by the authors. Is it therefore useful to use a new method that requires a very large amount of source data to be valid?
Conclusions
Please improve the conclusions as it contains the information already described in the summary, please describe the values received in more detail with a critique of the results obtained.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The submitted paper is a valuable contribution to the field and one of the first to use deep learning for InSAR phase filtering. Results presented display highest quality. In particular I was enthralled by the capacity to recover high frequency fringe patterns (figure 15 d)). Unfortunately, the method is only insufficiently documented, in particular with regard to the elements of the network and the training process. In order to see if I could find the missing details in the cited papers or in papers referenced there I selectively read 14 other papers on deep learning. Still many questions remained unresolved after doing so, which are addressed below. Please answer them detailed keeping in mind reproducibility and that not everyone in the remote sensing community has yet a good background knowledge of deep learning. The importance of having a good documentation is the reason why I ask a major revision of this interesting paper.
"->" indicates a suggestion to improve the formulations
lines 5+6: "we propose a phase filtering method based on deep learning to efficiently filter out the noise in the interferometric phase in this paper"
-> "we propose in this paper a phase filtering method based on deep learning to efficiently filter out the noise in the interferometric phase"
line 25: "becomes" -> "has become"
line 30: "the phase unwrapping"->"phase unwrapping"
line 81: "detail preservation" instead of "detail retention" would be more familiar to me. Maybe you should change this throughout the text.
three lines below formula (3): What is meant with "principal value"? Do you mean expected value?
two lines above formula (4): You have introduced interferometric phase as being real. The imaginary part of a real number is zero. Please avoid unorthodox use of terminology. Furthermore, I would recommend to make
your different notions of phase visible in formulas by a consistent use of symbols across the paper. E.g. phi for phases that are wrapped to (-pi,pi], varphi for phases that can assume arbitrary real values and still another character for real and imaginary part of exp(i*varphi). This would help to understand what notion of phase is meant at one glance, in particular in formulas (6)-(8).
some lines above line 125: As in the first of the formulas (2) it is always possible to obtain a phase in the interval (-pi,pi]. Why is this an argument for filtering real and imaginary part separately? Do you want to avoid processing an image with 2*pi phase jumps?
line 145-147: Please explain more precise what has been done here. How do you get the values?
Figure 2 (a)+(d): It would look nice if the colors of the blocks were chosen depending on their heights/values. That way the transition to the 256x256 matrix would be visualized in a more obvious way.
lines 168-170: What optimization algorithm do you use?
Figure 5: Please explain in the text how you do up-sampling and down-sampling.
line 187: "long-term term memory" -> "long short-term memory"
lines 191-202: Please explain what are the different channels and what happens when number of channels is doubled or halfed. Same for feature maps. Furthermore, please give for all layers (e.g. in form of a table) sizes and for all convolutions/deconvolutions the sizes of their kernels and other relevant information (stride, padding).
line 210: "add" -> "added"
Figure 6: Your ResBlock does not to coincide with that of [27] (it adds an ReLU). Is that intended? Please comment on this.
Formula (8): I would suggest to give the terms in such an order that it is easier to see that one runs around a square.
Formula (9): How do you define C1 and C2?
line 247: "base" -> "based"
lines 249-251: What is your "curriculum"? Have you used in every iteration the full set of training data? Or have you followed some particular strategy for training? You should also describe your training and test data in some detail. E.g. how were parameters (cp. lines 155-163) distributed? Do you use some data augmentation technique? What do you do to avoid overfitting and to keep the generalization error small? There can be found many tricks in literature. Do you use any of them? What is your optimization algorithm? How many iterations? How do you initialize the parameters?
line 252: What is meant by "out-of-order manner"? Please explain this.
Figure 7: I would suggest that tick labels on the colorbar include -1 and 1.
Figure 8: It would be good if labeling of colorbars were consistent across figures.
Figure 9: How do you define/determine your histogram curve?
Figure 10(b): "Glodstein" -> "Goldstein"
line 279: "detail" -> "detailed"
Table 1: "SSIM" -> "Mean SSIM"
line 296: "is closer zero" -> "is closer to zero"
line 294: "Figure 13" -> "Figure 12"
Figure 12: The colorbar is missing.
Figure 13: "SSIM" -> "Mean SSIM"
line 352: "Due to" -> "Because"
line 361: There are already other publications! See literature list below.
lines 362-362: "with a stronger balance capacity in" -> "with the capacity for better"
lines 363-364: Suggestion: "A second topic of future work will be to apply the proposed method to further real data sets."
Literature:
A. Hirose, Complex-Valued Neural Networks, ser. Studies in Computational
Intelligence. Springer Berlin Heidelberg, 2012, vol. 400.
S. Mukherjee, A. Zimmer, N. K. Kottayil, X. Sun, P. Ghuman and I. Cheng,
"CNN-Based InSAR Denoising and Coherence Metric,"
2018 IEEE SENSORS, New Delhi, 2018, pp. 1-4, doi: 10.1109/ICSENS.2018.8589920.
Sun, X.; Zimmer, A.; Mukherjee, S.; Kottayil, N.K.; Ghuman, P.; Cheng, I.
DeepInSAR—A Deep Learning Framework for SAR Interferometric Phase Restoration and Coherence Estimation.
Remote Sens. 2020, 12, 2340.
Deep Learning Meets SAR
Xiao Xiang Zhu, Sina Montazeri, Mohsin Ali, Yuansheng Hua, Yuanyuan Wang, Lichao Mou, Yilei Shi, Feng Xu, Richard Bamler
SUBMITTED TO IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE, 2020
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Dear authors, thank you for considering most of the comments in the review. I have a few more that specify the previous ones, and in some cases result from the explanations and answers given to me.
Note 1 Proposal method: The neural network with subnetworks has been described in great detail, but in my doubts are raised by the architecture of the entire network, is it variable for different tasks, or is it constant? Have other neural networks architectures has been tested and what were the test results?
Note 2 lines 279-283: Please explain in more detail the method of adopting the Q matrix in the calculation of real data, which is very important in assessing the suitability of the method for analysis.
Note 3 Lines 294-296 Lines 294-296 The number of test and training pairs is very large, very often impossible to obtain from real data. The authors in line 377 report that they used 2 days of exposure. I can't understand how they got 10,000 pairs of test and 2,500 training interferograms over 2 days. The exposure had to be made approximately every 15 seconds. In fact, SAR images for the same area are taken every 6 days, and in the 1990s they were taken every 12 days, or even less frequently.
Note 3 Figure 10 and 12. Please add descriptions in the figure, not just in the text only.
Note 4 Line 377: The use of 2 days for an area convenient for observation does not reflect the validity of the proposed method. Please test more complex areas (e.g. the Alps in Italy, the Himalayas in Tibet, Greenland, or the Amazon rainforest in Brazil) to confirm the validity of the proposed calculation method. In addition, please use real data available for free for analysis (www.esa.com).
Summing up, the method is interesting, but if it is not verified on real data, it does not bring anything new to the solution of the research task. Unfortunately, the test data can be checked many times and in various ways, but without confirmation in the calculations based on real data, the method is inappropriate.
Reviewer 2 Report
Dear authors,
Thank you very much for the extensive editing of the first version. I recommend the publication of this very nice paper. Please check the minor issues I found in the revision for the final version.
line 297: Xavier is the first name. "Xavier" -> "Glorot's"
line 330: "SSIM" -> "MSSIM"
line 401: "SSIM" or "MSSIM"? Should this be contained in tabel 4?
line 431: "in detail" -> "detail"
