Improved General Polarimetric Model-Based Decomposition for Coherency Matrix

Round 1

Reviewer 1 Report

The work developed in this paper is sufficiently interesting and the results are also encouraging. However, before publication the paper need to be edited in some parts. First of all, since this paper is an improved version of the reference [33], I suggest to better emphasize this aspect and also to cite reference [33] each time the method of Chen is indicated in the body of text. 

In Section 3, I suggest to directly illustrate the proposed method, and move details on the CSM and GVSM in Appendix. Conversely Section 3.2 that is the core of this paper should be improved in its presentation, with all details provided within it.

There are some typos, e.g., "C-band polari-metric" in Abstract. In some parts you refer to figures as "Fig." otherwise as "Figure". Please, check the entire paper.

In the first paragraph of Introduction some recent publications about PolSAR image classification could also be cited to enrich the Literature overview:
[1] "Screening Polarimetric SAR Data via Geometric Barycenters for Covariance Symmetry Classification." IEEE GRSL 2023.
[2] "Physics-Aware Training Data to Improve Machine Learning for Sea Ice Classification from Sentinel-1 SAR Scenes." IGARSS 2022.
[3] "Multi-channel synthetic aperture radar based classification of maritime scenes." IEEE Access 2020.

In the opinion of this Reviewer, it would be useful to add a Notation section just after the end of Introduction.

The quality of English is sufficiently good even if the text should be more fluid. I suggest another carefully proofreading of the manuscript by the authors before next submission.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

 

In this paper, the author proposed an improved generalized polarimetric model-based decomposition method to reduce reliance on the initial input values, and improve the computational efficiency.

Detailed comments are:

 1）          In page 13, The number of Table 2 in should be three but not two. Please modify it.

2）          In page 13, the authors write “The average time consumption of each pixel of the proposed method is the least among these methods.”. But in the table below, the time consumption per pixel of the GMD-GVSM method is lower than that of the Proposed method. Please modify it.

3）       In section 3.2, the authors introduce a criterion area descriptor Dn. Could you provide some explanation of the physical meaning of this formula? Why it can be a criterion between urban and non-urban areas?

 

 Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The model-based decomposition presented aims to avoid misrepresenting oriented building scatter as volumetric scatter while simultaneously reducing computational overhead. 

The first goal obtains via an urban vs. non-urban discriminator, eqs.  (29)-(31). A threshold is set to choose the appropriate scattering model, with the assumption that urban areas contain no volume scattering and non-urban areas no "cross-scattering". How sensitive are the results to the threshold value? What does the urban / non-urban mask look like in practice? The critical regions are where equality (or near equality) holds in eq. (31). How much do the decompositions change going from just below to just above the urban threshold value?

A second threshold determines whether the decomposition incorporates the 'Generalized Volume Scattering Model' or the Yamaguchi model from 2015. Again how is the threshold set? How sensitive are the results to the threshold value? Where and when is the standard Yamaguchi model employed? 

For any threshold, a sensitivity analysis is necessary. Are these values fixed across all datasets, radar frequencies, etc., or is there a heuristic that is used to determine their values? Since the canonical scattering models used in your decomposition critically depend on the threshold values, questions regarding the thresholding must be fully addressed in the text of the manuscript.

The second goal is met by judiciously separating gamma and the 2 orientation angles from the remaining variables. The solution for these 3 variables is a non-linear problem, which you solve by brute force. As long as the non-linearities are well behaved, and the residual remains a smooth function of these variables, a brute-force technique may work. However, to guarantee the optimum solution the entire 3-D space must be tested explicitly. Is this done in practice? (See Point 2 below.) Does the residual appear to be a smooth function? A few comments regarding the solution of the non-linear part of the problem would be welcome. 

Generally speaking, you have designed a hierarchical decision tree algorithm that chooses one of three main paths depending upon the desired 'volume' scattering model. Are there approaches other than decision trees to select among a set of canonical scattering models? 

Minor notes & typos:

1. Eq. (8) the first matrix component is wrong.

2. Line 257, 'according to (39), since Re[T_23] = 0  [29], and ...' seems in error. There is no reason for any component, real or imaginary, of the observed coherency matrix to be zero. Please recheck the ranges of the theta angles. 

3. Lines 447-449, 'By simplifying the nonlinear equations ... local minimum and initial values selection problems are avoided.' is over-stated, especially if the full 3-D variable space is not covered, see above and Point 2. You 'avoid' the problem by brute force testing of every possible solution before choosing the one that minimizes the residual. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf