Hyperspectral Image Denoising Based on Principal-Third-Order-Moment Analysis

Round 1

Reviewer 1 Report

This manuscript proposes a low-rank-based denoising method, aiming to preserve the spectral characteristic of small targets while reducing noise. Denoising is an important topic for the application of hyperspectral images. Generally, this paper is technically correct. However, I have some major concerns about the methodology and experiments that should be explained and solved. The major concerns for the manuscript are given in detail as follows.

1. Since the noise in real hyperspectral images is complex, I suggest the authors simulate some complex mixed noise including impulse noise, gaussian noise, and stripes/dead lines. Also, the SOTA mixed noise removal methods should be incorporated into the experiments to fully evaluate the effectiveness of the proposed method.

2. A simulated dataset is not enough, please provide at least one more simulated experiment.

3. In the Introduction, some related low-rank-based mixed noised removal methods are missing, for example, FastHyMix, RoSEGS, etc.

4. In the experiments, which target detection method is employed to detect the small targets? Please include some details about the methods and the parameters used.

5. Since there are already some anomaly detection methods taking into account the impact of noise, I am very curious to see a comparison of the proposed method to some of these methods.

6. Please include the computation time for all the methods to fully evaluate the proposed methods.

Please have the manuscript thoroughly and carefully checked again before submitting the manuscript. I have found several mistakes in this version.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

 - On the model, it is not clear why we should focus on the co-third-order moment tensor M, and why maximize M ×1 u ×2 u ×3 u leads to good denoising results. Does noise reduce this value of objective function? The authors must clearly highlight the motivation and the physical meaning of the proposed algorithm. 

 - There are many notations, the authors should summarize them in a Table to make it reader-friendly. 

 - When reviewing related works, the authors ignore two important directions: noise modeling based methods [R1-R2] and total variation based methods [R3-R6].

[R1] Denoising Hyperspectral Image With Non-i.i.d. Noise Structure. Yang Chen;Xiangyong Cao;Qian Zhao;Deyu Meng;Zongben Xu. IEEE Transactions on CyberneticsYear: 2018 | Volume: 48, Issue: 3

[R2] Hyperspectral Image Denoising by Asymmetric Noise Modeling. Shuang Xu;Xiangyong Cao;Jiangjun Peng;Qiao Ke;Cong Ma;Deyu Meng IEEE Transactions on Geoscience and Remote SensingYear: 2022 | Volume: 60

[R3] Graph Spatio-Spectral Total Variation Model for Hyperspectral Image Denoising. Shingo Takemoto;Kazuki Naganuma;Shunsuke Ono. IEEE Geoscience and Remote Sensing LettersYear: 2022 | Volume: 19

[R4] Hyperspectral Image Denoising Using Spatio-Spectral Total Variation. Hemant Kumar Aggarwal;Angshul Majumdar. IEEE Geoscience and Remote Sensing Letters

[R5] S. Xu, J. Zhang, J. Wang, C. Zhang, “Hyperspectral Image Denoising by Low-Rank Models with Hyper-Laplacian Total Variation Prior,” Signal Processing

[R6] Hyperspectral Image Mixed Noise Removal Based on Multidirectional Low-Rank Modeling and Spatial–Spectral Total Variation. Minghua Wang;Qiang Wang;Jocelyn Chanussot;Dan Li. IEEE Transactions on Geoscience and Remote Sensing. 

 - The method proposed in this paper shows promising results in terms of noise suppression and target preservation. However, it would be beneficial to provide a more detailed explanation of the underlying theoretical framework to better understand the advantages of this approach over existing methods.

 - In the real-world experiments, GF-5 dataset seems not to be the widely used imagery. The authors should provide the download link for it. Also, the authors should make the proposed algorithm open access (for example, upload it on GitHub). Otherwise, the readers cannot check the reproducibility of this algorithm on GF-5 real-world dataset. 

 - The simulated experiments are not convincing. Firstly, only i.i.d. Gaussian noise is considered. The authors should refer to the paper (https://ieeexplore.ieee.org/document/9975834), and conduct non-i.i.d. Gaussian noise and mixed noise experiments. Secondly, instead of MPSNR and MSSIM, actually ERGAS and SAM are the better metrics to evaluate HSI quality, so the authors should report their values. 

 - Although the computational time is reported in Table 2, the paper does not discuss the computational complexity of the proposed method.

 - A typo in line 195:  mathbf Xb -> \mathbf{Xb}

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors,

you present a new approach to denoise HSIs based on the 3rd order moment analysis.

Although the study is interesting , there is space for improvement with regards to the presentation of your research.

Some suggestions:

 line 17:  I disagree that HSIs have low spatial resolution, that depends on the separation distance between target and sensor. So please rephrase here to convey your message correctly.

line 23: what do you mean by "prior knowledge of the HSI signal"? Please elaborate

lines 127-128: you have proven equation (5), so please rephrase.

line 130: "it has been observed ..." how and based on what data?

line 165: you are using 2 HSIs images for your separate experiments, please elaborate more with appropriate references on the sensors' and platforms' details. This should include details on the bands you have used for your study.

line 177: how do you prove the high image quality  here?

General comments:

Introduction

you could refer to some numerical results from your study to state the improvement range

Figures

captions should be more detailed and even state the model names instead of having a, b, c etc

please place figures appropriately in the text as having figures before having those discussed in the main body confuses the readers

Conclusions

this section is poorly developed and needs to be conclusive but yet informative.

Thank you

Please make sure you have clear statements and revise of the text

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I am satisfied with the current version.

The quality of the English language is OK for me.

Reviewer 2 Report

Minor editing of language is required before publication. 

Minor editing of language is required before publication.