C-RISE: A Post-Hoc Interpretation Method of Black-Box Models for SAR ATR

Round 1

Reviewer 1 Report

This paper proposed a post-hoc interpretation method for black-box models named C-RISE. The ideal is interesting, and the experiments are broadly convincing. Here are several concerns that need to be considered:

(1)    There are several parameters that need to be set in the proposed C-RISE algorithm. The authors simply set k=4, N=2000, s=8, p=0.5, but give no explanation. Especially, the authors also said experimental results were sensitive to the number of clusters, therefore, it is necessary and also meaningful to elaborate how to set a reasonable the value of k.

(2)    The authors did experiments on MSTAR database but only under SOC. I strongly suggest they also do experiments under extended operating conditions (EOC).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposed an approach named C-RISE, which builds upon the RISE algorithm to provide a post-hoc interpretation technique for black-box models used in SAR image target recognition. The following points are for the authors’ reference:

Q1: The transferable mentioned in line 15 of the abstract was not reflected in the method design and experiment.

Q2: From lines 173 to 192, the RISE algorithm is explained twice.

Q3: What do $z$ and ${z_i}$ represent in Equation 4?

Q4: In Equation 5, whether $\lambda $ or $M\left(\lambda\right)$ denote the pixel with a value of 1 in the mask, and what the mean of ${E_M}\left[\cdot\right]$?

Q5: At the beginning of section 3, it was mentioned that in order to solve the problems of noise, energy dispersion, and inaccurate positioning, a C-RISE algorithm based on RISE was proposed. However, in the subsequent detailed introduction of the algorithms, there was no introduction to the algorithms designed for these problems, nor was it reflected how they were solved.

Q6: At the beginning of section 3.1, it is mentioned that the RISE algorithm has high complexity, but it seems that the proposed Mask Generation method is more complex than RISE. In addition, what is the role of smoothness and consistency in the target spatial structure mentioned in line 234, and how is it achieved?

Q7: What the mean of $$a_i^j\left[ l \right]$$ and $l$ in Equation 18?

Q8: The experiment is only conducted based on Alex network. Is it effective on other typical SAR target recognition networks?

Q9: Line 308, error in fig1, expecting fig6.

Q10: What’s the mean of Conservation and Occlusion Test in section 4.3?

Moderate editing of English language.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper introduces a novel post-hoc interpretation method for black-box models in SAR ATR, which builds on the RISE algorithm. The proposed C-RISE algorithm offers several advantages over other methods, including its ability to group mask images that capture similar fusion features using a clustering strategy. The paper is well written and I would suggest revisions as listed below

1.      Can you explain how the clustering strategy used in C-RISE helps to concentrate more energy in the heatmap on the target area?

2.      What are some potential applications of C-RISE beyond SAR ATR?

3.      Can you provide more details on the qualitative analysis and quantitative calculation used to evaluate the interpretation effects of different methods and C-RISE algorithm?

4.      How was the clustering strategy used in C-RISE implemented, and what parameters were used to determine the number of clusters?

5.      The literature review of this paper is good. However, the authors missed out an important family of data-driven methods for field interpolation/reconstruction, namely the data assimilation methods. Some references to build the related work:

https://wires.onlinelibrary.wiley.com/doi/full/10.1002/wcc.535

https://www.mdpi.com/2072-4292/14/13/3228

 

https://www.sciencedirect.com/science/article/pii/S016727890600354X?casa_token=xSHIW_Un6aMAAAAA:orcUgYlNkOi8abIi-L91Mn0He2wnSsoWPrYJDk7EyzLHV79eJn2HMb4TF8cmNMOSuc5QESM1Zw

the english quality is acceptable

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have answered all my concerns, I think the manuscript now can be acceptable for publication.