Multi-Step Unsupervised Domain Adaptation in Image and Feature Space for Synthetic Aperture Radar Image Terrain Classification
Round 1
Reviewer 1 Report (Previous Reviewer 2)
Comments and Suggestions for AuthorsSince the authors of the manuscript correctly took into account my comments, I propose that it be accepted for publication in the Remote Sensing Journal.
Reviewer 2 Report (Previous Reviewer 3)
Comments and Suggestions for AuthorsThe authors have made a commendable effort to address the previous reviewers' comments, and the manuscript is well-organized.
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
Comments and Suggestions for AuthorsAs for pointing out that existing domain adaptation methods, these lack the joint consideration of domain change in image space and feature space, leading to capture an inadequate perception of terrain features, for this one must consider that this is due to multiple factors. However, what you consider in specific to evaluate the Style Transformation and Domain Metrics (STDM) with your proposal to adapt an unsupervised STDM-UDA domain framework for terrain feature classification seems a very interesting challenge.
What is striking, is that in the process of integrating or building a domain-independent adaptive network, and that this consists of two stages to reduce the domain differences in the image space and semantic space in order to migrate the source annotation information and thus achieve terrain classification in the target domain.
To this end, I question the following:
Since SAR images in their raw or native state have a large dynamic range due to their spatial resolution and indeed standard CNNs do not process very wide dynamic ranges, then the challenge is how to effect a dynamic range without affecting the alteration of the pixel resolution type of the image.
What do you think might change in the generalization capability of the SAR imaging model if another neural network strategy is changed or experimented with to adapt a dynamic range without affecting the alteration of the pixels?
Did you experiment with various neural network methods, and which ones were these?
In evaluating the five high-resolution single-channel SAR images and Considering only the consistency of the bands, how do you explain that image style transfer and adaptive domain segmentation occurred, given that the spatial resolutions of each image vary and the coverage sizes were different from each other?
Detail about this questioning.
Finally, it strikes me that in your proposal you copied three times the data value of a single image channel to form a three-channel image and the experiment was performed without any data enhancement operation.
Could you please detail if the 95% truncation threshold was maintained?
Comments on the Quality of English LanguageIn the case of English writing, I only recommend that you give a final check to possible grammar and spelling details that could be improved.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe original achievement of the authors of the paper is the synthesis of the synthetic aperture radar image terrain classification method using a multi-stage adaptive structure of the unsupervised domain.
Comments:
1.The manuscript in its current form is a research report rather than a scientific article.
2. The abstract is not a summary of the work, its first half should be moved to the Introduction, as a justification of the thesis of the work.
3. In the Introduction, you should include the thesis of the thesis, and then the objectives of the work, which become its chapters.
4. For equations (1-16), indicate which ones are original and which are from the literature and must be cited.
5. Conclusions are not intended to constitute a summary of the work, but to contain specific quantitative and qualitative conclusions from the research.
6. In Conclusions, include a detailed plan for further research on the topic of the paper.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe paper discussing the Multi-Step Unsupervised Domain Adaptation in Image and Feature Space for SAR Image Terrain Classification seems technically sound. However, a few potential areas might require clarification and improvement.
In the abstract section
The text claims "Experiments conducted on several image pairs with varying degrees of domain gap show that the proposed approach achieves competitive classification accuracy over high-resolution broad scenes without labels." This statement lacks specificity regarding the datasets used.
In the introduction section
"all-day, all-weather capability" - This phrase may imply that SAR imaging works equally well in all lighting and weather conditions, which is not entirely accurate. SAR imaging can work in various weather conditions, but extreme conditions like heavy rain or dense fog may still affect its performance.
"On the one hand, in contrast to optical images, SAR image labeling is arduous, infrequent, and costly to acquire." - This sentence is accurate, but it could be improved by specifying why
SAR image labeling is arduous, infrequent, and costly compared to optical image labeling.
3. Method section
"STDM-UDA consists of two independent steps based on adversarial DA: image style transfer network and adaptive segmentation network." - This sentence is misleading. While it states that STDM-UDA consists of two steps, it doesn't clarify what "adversarial DA" refers to, and it's unclear how the steps are independent.
3.2. Image Style Transfer Network section
The statement lacks specificity on how CycleGAN is used for unpaired SAR image style transfer.
4.1. Experimental Dataset section
The text mentions that certain terrain categories will be ignored during training and testing but does not explain why or how this decision was made.