A Neural Network for Hyperspectral Image Denoising by Combining Spatial–Spectral Information
, Boyu Pang 1,2 and Dexin Sun 1,2,*Round 1
Reviewer 1 Report (Previous Reviewer 3)
Comments and Suggestions for AuthorsThis paper introduces a deep learning-based denoising approach for hyperspectral images (HSI). The method takes into account the spatial-spectral correlation inherent in HSI data. It employs multi-resolution convolutional blocks to capture both spatial and spectral features, adaptively combines contextual information, and utilizes a residual connection to learn mapping characteristics at various levels, thereby enhancing feature fusion and representation. Ultimately, it approximates the residual mixed noise. The language and format of the article have undergone significant revisions, resulting in an overall clarity and improved readability. The following questions may interest me.
1. The network employs a “residual-like” denoising approach, which has shown promising results. What are the advantages of this denoising network compared to other methods that use attention mechanisms for denoising?
2. The model incorporates a substantial number of convolutional operations, resulting in a training time of up to 11 hours. I would like to inquire how the author plans to strike a balance between performance and time consumption in future research?
Comments on the Quality of English LanguageNone.
Author Response
General Comments: This paper introduces a deep learning-based denoising approach for hyperspectral images (HSI). The method takes into account the spatial-spectral correlation inherent in HSI data. It employs multi-resolution convolutional blocks to capture both spatial and spectral features, adaptively combines contextual information, and utilizes a residual connection to learn mapping characteristics at various levels, thereby enhancing feature fusion and representation. Ultimately, it approximates the residual mixed noise. The language and format of the article have undergone significant revisions, resulting in an overall clarity and improved readability. The following questions may interest me.
Response: Thank you very much for reading our manuscript carefully and giving a positive comment, and thank you for recognizing our work, we have carefully considered and responded to your valuable suggestions.
Point 1: The network employs a “residual-like” denoising approach, which has shown promising results. What are the advantages of this denoising network compared to other methods that use attention mechanisms for denoising?
Response 1: Thank you for your insightful questions. Residual networks are effective in addressing the problem of network degradation that arises in deep networks. By using residual connections, we can add more trainable layers to the network, thereby improving its performance. This allows us to leverage the benefits of deeper architectures without suffering from vanishing gradients or gradient explosions.
On the other hand, attention mechanisms are designed to focus limited computational resources on key information. They enhance the importance of relevant features while suppressing irrelevant or noisy information. By doing so, attention mechanisms can improve the efficiency and effectiveness of neural networks by allocating resources where they are most needed.
Both are designed for neural networks to do their job better, and can be combined to make the network perform better without comparing the better and the worse. In the manuscript's concluding section, our future research direction states: " Additionally, the inclusion of attention mechanisms in denoising networks could enhance the network's feature extraction capabilities."
Point 2: The model incorporates a substantial number of convolutional operations, resulting in a training time of up to 11 hours. I would like to inquire how the author plans to strike a balance between performance and time consumption in future research?
Response 2: Thank you for your careful review of our manuscript and for raising valuable questions. Our current optimization approach involves incorporating the Batch Normalization (BN) layer into the network training process. BN is a technique used to normalize the mean and variance of a single batch of training data, which helps address the issue of vanishing gradients in shallow networks and allows for an increase in the learning rate of the network.
In support of our approach, we consulted and studied a paper titled "How does batch normalization help optimization?" by Santurkar et al. (2018). The authors of that paper primarily focused on the impact of BatchNorm on training, but they also mentioned that their findings suggested BatchNorm's potential to improve generalization. This indicates that BN can enhance the learning rate of the network, facilitate faster convergence, and increase the network's ability to generalize to unseen data.
Reviewer 2 Report (Previous Reviewer 5)
Comments and Suggestions for AuthorsThe authors have revised this paper according to my suggestions.
Author Response
Thank you to the reviewer for your positive comments about our manuscript. Thank you very much for your time and effort in the review process.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsOverall the manuscript is well structured and presented, while the topic is of interest to several domains of applications. I have nothing to add or ask by the authors, but one question (food for thought, maybe for upcoming research) whether the proposed architecture can also be applied and be effective in cases of spatial distortion of the HS images. The scientific content is solid and the tests performed support the results.
Author Response
Please see the file attached:
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsHyperspectral cameras are disturbed by various types of noise when acquiring images such as the non-uniformity of the sensor response and changes in the transmission path through the atmosphere. Achieving high-fidelity removal of multiple types of complex noise in hyperspectral
images poses a technical difficulty in the field of hyperspectral data processing. Most existing methods are based on specific types of noise with isolated research, the scope of application of the associated denoising models is limited, the ability to process complex noise situations is lacking, and their versatility and robustness are insufficient. This paper proposes a denoising method based on the network that fully considers the spatial structure and spectral differences of noise in an image data cube. The proposed network takes the DN value of the current band and horizontal, vertical and spectral gradients as inputs. A multi-resolution convolutional module is utilized to accurately extract spatial and spectral noise features, and features at different levels are aggregated through residual connections. Finally, the residual mixed noise is approximated. Both simulated and real cases verify the effectiveness of the proposed denoising method. Compared with six commonly used HSI denoising algorithms, the proposed method presents superior results in terms of its visual effect and evaluation indicators.
1) At the end of the abstract, it will be more intuitive and convincing to illustrate the qualitative results of a large number of experiments for verifying the superiority and effectiveness.
2) Some of the related work is ignored. The authors ignore some relevant papers. For example, “Image Dehazing by an Artificial Image Fusion Method Based on Adaptive Structure Decomposition” in IEEE Sensors Journal, and “A Novel Fast Single Image Dehazing Algorithm Based on Artificial Multi exposure Image Fusion ”, in IEEE Transactions on Instrumentation and Measurement. The authors should compare their method with it carefully.
3) In the Section 2. Spatial–Spectral Denoising Network, the thesis mentions that “The features are extracted through convolution at different resolutions, and the fully cascaded network is used to learn the features at different receptive field sizes. Residual connections use multiple feature mappings obtained from the intermediate connection layers.”. What adjustments and adaptations are made for convolution operations at different resolutions?
4) In the Section 2. Spatial–Spectral Denoising Network, the thesis mentions that “Due to its unique directional structure, spatial gradient information can effectively highlight sparse noise to a certain extent. Hyperspectral images contain rich spectral information, and each band's noise level and noise type usually differ, providing additional complementary information. Thus, it is significant to fully use both spatial and spectral gradient information when performing HSI denoising.”. What are the operational contents and processing procedures for highlighting sparse noise in directional structures?
5) The explanation of equation (6) is not enough, What the symbol on the right side of the equation represents puzzles the readers.
6) Make sure your conclusions appropriately reflect on the strengths and weaknesses of your work, how others in the field can benefit from it, and thoroughly discuss future work.
7) In the reference section, it will be better to search and cite more latest research, which can better reflect the innovation of this thesis.
Author Response
Please see the file attached:
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper proposes an HSI denoising method based on deep learning. The proposed method fully considers the spatial-spectral correlation of hyperspectral images. It captures spatial and spectral features with multi-resolution convolutional blocks, adaptively fuses the surrounding information, and then uses a residual connection to learn the mapping characteristics at different levels, thus enhancing feature fusion and representation. Finally, the residual mixed noise is approximated. The results of simulations and experiments with real data demonstrated the superiority of the proposed method over other state-of-the-art methods in processing different noise types, including Gaussian noise, stripe noise, deadline noise, and mixtures thereof. However, it is difficult to follow the paper, and there are no highlights in the paper to engage the reader：
1. Insufficient innovativeness. Specifically, the paper only makes simple combinations and overlays on existing modules (e.g., MCB, residual structures).
2. Contributions are not clear and the algorithms corresponding to the main contributions are not elaborated.
3. The formulas are unclear, e.g., formulas 8 and 9. Furthermore, the font of the formulas does not match the font of the text.
4. It is difficult to follow the experimental sections. For example, in subsections 3.2 and 3.3 of the experiment, the authors used multiple graphs to show the denoising effect of different methods, and again used different colors and symbols to represent different methods and data, which makes the representation confusing and confounding.
5. The experimental section lacks settings for specific parameters such as filter size, threshold, etc.
6. There are some confusions, such as line 403.
Comments on the Quality of English LanguageEnglish very difficult to understand.
Author Response
Please see the file attached:
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsComments for author File: Comments.pdf
English usage is generally good, but there are peculiarities and questionable statements scattered through the manuscript. Several are noted in the attachment.
Author Response
Please see the file attached:
Author Response File: Author Response.pdf
Reviewer 5 Report
Comments and Suggestions for AuthorsThis paper proposed a new HSI denoising algorithm. The paper is well written. I have some comments about it.
Lines 85-90: These descriptions should be moved to the result and conclusion sections instead of the introduction section.
The font and format of the equations should be checked.
Eqs. 8, 9, explain what the symbol of a circle means.
The training time of the deep learning methods should both be listed.
In the real image data experiment, the improvement of the proposed method seems obscure from visual comparison. Providing more quantitative metrics to show the proposed method outperformed others is necessary.
It is expected to give a link to the code so that other researchers can use the proposed method for comparison.
Comments on the Quality of English LanguageThe languages can be further polished.
Author Response
Please see the file attached:
Author Response File: Author Response.pdf
