Quality Assessment of Despeckling Filters Based on the Analysis of Ratio Images

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Overview:  Characterizing and filtering speckle variation is crucial in SAR image processing and analysis. Evaluating speckle filter performance would require a universal metric, which is still practically lacking. This paper presents an interesting effort in quantifying SAR speckle filter qualities. 

(1) Using ratio images is the basis of the proposed metrics. However, the authors should clarify all the assumptions as well as the limitations surrounding the relevant discussion. Formulating the ratio images with Gamma distribution is valid only for fully developed speckle. Do you expect the method applicable at higher resolutions? The resolution is quite low for the test cases (BTW, what is the resolution? It matters!). Even on these low-resolution images, the scattering response from urban structures does not qualify for fully developed speckle - indeed, one can see the silhouette of those structures in the ratio images.

(2) For a recommendation, I also suggest that the authors delineate the rationales for adopting the visibility graph in the introduction. The majority of the introduction was devoted to the development of different speckle filters and evaluation metrics. However, when it came to the proposed method, the description suddenly became very much condensed, e.g., to lines 90-92.

(3) Typo in line 112 - it should be “XN/X”.  The experiments in Section (3) are performed on 5 “SLC” SAR images, but the ENL of the ratio images in a homogeneous area is evaluated to be 2.95 (lines 314-325). Isn’t it supposed to be 1, referring to that ratio image definition “XN/X”?

(4) Maybe the scope of this paper is meant to be restricted to machine learning based speckle filtering approaches. Nevertheless, I would caution that the proposed metrics, as built on statistical analysis, might be more genuine to the common statistical-based speckle filtering approaches. I believe the authors have performed similar experiments, for example, on the Lee filter (and/or the non-local filter). If so, the authors may at least comment on those for a useful comparison.  

Author Response

Please see the attachment

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This article proposes a quantitative and qualitative evaluation framework for SAR despeckling filters based on the analysis of ratio images. The authors introduce a new metric called Texture-Divergence Measurement (TDM), which integrates Haralick texture features (Contrast and Variance) and the Jensen-Shannon Divergence (JSD) to assess how closely the ratio images resemble ideal pure speckle noise. Additionally, the study utilizes Image Horizontal Visibility Graphs (IHVG) to qualitatively visualize residual structures in the ratio images. The method is applied to five SAR images using four despeckling algorithms (AE, FANS, Monet, SCUNet), and the results suggest that FANS and Monet provide a better balance between noise suppression and structure preservation compared to SCUNet. While the paper presents a novel approach, there are several issues regarding the methodology description, experimental validation, and data presentation that need to be addressed.

1. The abstract states that normalization was performed "through a sigmoid-based transformation". However, in Section 3.5, the text describes a normalization process based on dividing by reference values and applying a saturation threshold where(D_JS>=5.0)is constrained . This contradiction must be resolved to clarify the actual mathematical method used.
2. The experimental dataset is currently limited to five SAR images from a single region (Toronto). To verify the robustness of the TDM metric, the authors should include images from diverse environments (e.g., agricultural fields, ocean, mountainous terrain) and different sensors or resolutions.
3. In Equation (18), the weight parameter α is fixed at 0.5 . There is no justification or sensitivity analysis provided for this choice. The authors should discuss how varying α  affects the ranking of the filters or provide a rationale for using equal weighting.
4. In Table 1, the M-estimator value for SCUNet sample 3 is extremely high (698.89) compared to other samples. The authors should check if this is an outlier or a calculation error and briefly explain why SCUNet performs so differently on this specific sample.
5. In Table 2, the column  D_JS  shows SCUNet values saturated at 1.00 (or higher, capped by normalization). Since the raw JSD values in Table 1 are very high (up to 19.88), the normalization step in Table 2 seems to compress the worst performer too aggressively. The authors should comment on whether this saturation hides the magnitude of SCUNet's deviation.
6. There are spelling errors in the text that need correction, such as "despeclking" in the "Highlights" section and "feautures" in Section 3.3. Please proofread the manuscript carefully.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf