Comparative Analysis of Despeckling Filters Based on Generative Artificial Intelligence Trained with Actual Synthetic Aperture Radar Imagery

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a generative artificial intelligence (GenAI) method for Synthetic Aperture Radar (SAR) imagery which is a meaningful endeavor. However, the current experimental setup and analysis lack sufficient rigor and require further refinement. Further comments are outlined below:

1.    Ensure that the introduction clearly states the research problem and contributions of your work to make the paper more engaging and focused.

2.    The use of actual SAR imagery is a strength. However, detailing the diversity and representativeness of your dataset could strengthen the validity of your findings.

3.    More detailed information on the training process, such as hyperparameter settings, model architectures, and training duration, would enhance reproducibility.

4.    On page 16, the authors propose that a higher ENL value corresponds to a reduced level of noise in the image, but ENL reflects the strength of the SAR image containing multiplicative noise. The smaller the ENL, the less multiplicative noise contained in the image and the higher the image quality.

5.    The equations lack an analysis of the meaning of the variables, such as in Equation 4 and Equation 5.

6.    In the Experimental Results section, the results lack a detailed and in-depth analysis. It is recommended to include a more comprehensive evaluation with different scenarios.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

• Highlight how the proposed generative AI-based despeckling filters can be applied to real-world SAR image analysis and practical use cases.
• Simplify some of the technical details to make the paper more accessible to a broader audience without compromising scientific depth.
• Provide a more detailed discussion of the advantages and limitations of the filters compared to existing methods.
• Suggest potential future research, such as expanding the dataset or testing the filters under different environmental conditions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript aims to evaluate the performance of various despeckling techniques, including traditional, deep learning, and transformer-based models, for SAR images.

The work is conceived and written very well, however I found two main weaknesses 1) in the introduction the most modern despeckling algorithms using
artificial intelligence are adequately described, but a section that
talks about the classic despeckling techniques that do NOT use artificial
intelligence is completely missing (i.e. the various types of geometric
or spectral filters and multilookong), which were used in the work anyway
(like Lee filters and FAN). 2) At the end the problem of the computational cost of the various filtering
systems is only mentioned (lines 463-465) and it is not explained that
modern systems based on artificial intelligence and deep learning need
a time expensive training to work. Furthermore, filters like Lee are
immediately usable and are already implemented in most suites for the
processing of satellite data. Please try to find a quantitative metrics that
describe the computational cost or labour effort intensity of each filter and
include it in the discussion. In addition to this, please correct these parts: - Caption of figure 2: please explain what Ground Truth coincides with,
- The part that is found between the lines 286-288 goes in the conclusions  

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors the authors comprehensively answered almost all the questions they asked. My opinion is positive for the publication of the manuscript