Scalable Hyperspectral Enhancement via Patch-Wise Sparse Residual Learning: Insights from Super-Resolved EnMAP Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

A novel parallel patch-wise sparse residual learning (P2SR) algorithm is proposed in the manuscript for resolution enhancement based on fusion of HSI and MSI. However, the motivation for the proposed method is unclear, the novelty and key contributions are not explicitly stated, and the introduction of the method lacks depth and detail.

 

1. Please articulate the motivation behind the proposed method in the abstract, and provide a clearer overview of the overall approach.
2. Some latest references can be cited.
3. Please highlight the challenges to be solved and innovations in the introduction.
4. The description of the method is insufficient and lacks coherence. It is suggested that the methodology section be divided into several sub-sections for a clear and organized presentation.
5. The methodology section would benefit from including key equations, such as those for sparse coding, FISTA optimization, and guided filtering, to strengthen the theoretical rigor.
6. Phrases like “multi-decomposition techniques” should specify which methods are used.
7. Please consider adding comparisons with more recent methods.

Author Response

Author’s Reply to Reviewer 1

We thank the reviewer for the constructive critical reviews. A point-by-point response is provided for all comments by the reviewer.

Comment 1: Please articulate the motivation behind the proposed method in the abstract, and provide a clearer overview of the overall approach.

Response 1: The motivation with regard to solving existing challenges is now added to the abstract.

 

Comment 2: Some latest references can be cited.

Response 2: The review of the methods is now extended in the introduction with some more relevant and recent studies. See references [42-50].

 

Comment 3: Please highlight the challenges to be solved and innovations in the introduction.

Response 3: The challenges with regard to existing methods and the proposed innovation are now added to the introduction. A subsection of “significant contributions” with regard to solving these challenges is also added to the discussion.

 

Comment 4: The description of the method is insufficient and lacks coherence. It is suggested that the methodology section be divided into several sub-sections for a clear and organized presentation.

Response 4: For a more detailed description of the method, an algorithm table is added to the manuscript in Appendix A. This description is in addition to the flowchart and detailed explanation of the algorithm which already exists in the proposed method section. We also provide a pseudo code of the source-implementation for clarifying the actual working of the algorithm in the same appendix section.

 

Comment 5: The methodology section would benefit from including key equations, such as those for sparse coding, FISTA optimization, and guided filtering, to strengthen the theoretical rigor.

Response 5: As a preliminary description of the key techniques already exists, we have added key equations of techniques in the algorithm table as mentioned in Response 4. 

 

Comment 6: Phrases like “multi-decomposition techniques” should specify which methods are used.

Response 6: The phrase “multi-decomposition” is now clarified at an initial point in the abstract - which states use of ICA, NMF and 3DWT decomposition methods.

 

Comment 7: Please consider adding comparisons with more recent methods.

Response 7: We appreciate the suggestion to include comparisons with more recent methods. Our literature review identified several cutting-edge approaches, many leveraging deep learning and promising future source-code releases. To ensure robust and reproducible comparisons, we included two recent, widely-cited deep learning methods—SSRNet and ResTFNet—both of which offer open-weight pre-trained models. These models are frequently used as benchmarks, enabling seamless comparison with other studies. Despite an extensive search, we found that many newer methods lack publicly available source code and more importantly, pre-trained-open weight model files. For recent studies where the source code is available, most of them lack pre-trained model weights (and only provide model architecture). The open-weight models files are critical for inclusion in this study as we rely on limited data and they also ensure a fair and meaningful comparison.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors introduce P2SR, a novel parallel algorithm for resolution enhancement by fusing hyperspectral (HSI) and multispectral (MSI) data. P2SR employs multi-decomposition to extract spatial-spectral features, forming a sparse dictionary for efficient reconstruction via first-order optimization. The enhanced HSI is reconstructed using sparse-dictionary features from low-res HSI, refined by an MSI-guided filter to boost spatial fidelity and reduce artifacts. Overall, the paper is of scientific sounds. 
Key Advantages:

1. Parallel HPC deployment ensures scalability for large datasets.
2. Outperforms traditional and state-of-the-art (SOA) methods in quantitative metrics and spatial quality.
3. Enhances geological features, reduces mixed pixels, and sharpens spatial details.
4. Preserves critical spectral signatures (e.g., Fe²⁺ absorption) for precise mineral mapping, land-use analysis, and environmental monitoring.

Main concerns:

1. What are the advantages of the proposed P2SR method compared to other resolution enhancement algorithms.
2. Generally, quantitative results can be added to Abstract.
3. Introduction: What is the motivation of this paper? In other words, what problems exist and what the authors want to solve. In addition, the main idea of the proposed method can be expressed more clearly.
4. The main contributions of this work can be concluded point by point.
5. The resolution of Figure 1 can not meet the publication requirements.
6. It is suggested to write an algorithm table to describe the proposed method.
7. In addition to the ResTFNet and SSRNet methods, more advanced methods based on deep learning should also be added for comparison.
8.  In order to demonstrate the capabilities of each part of the method, ablation experiments are very necessary.
9. Discussion: It is suggested that the author discuss and analyze the proposed method in greater depth based on the experimental results. 
10. The limitations of this work should also be concluded in conclusion part.

Author Response

Author’s Reply to Reviewer 2

 

We thank the reviewer for the constructive critical reviews. A point-by-point response is provided for all comments by the reviewer.

 

Comment 1: What are the advantages of the proposed P2SR method compared to other resolution enhancement algorithms.

Response 1: The proposed P2SR method offers several key advantages over existing hyperspectral resolution enhancement algorithms:

 

Local Feature Awareness: Unlike methods such as CNMF and HySure that rely on strong global assumptions, P2SR adopts a patch-based approach that leverages local data characteristics. This leads to more precise and adaptive enhancement across diverse spatial regions within a scene.

 

Low Data and Computational Demands: In contrast to deep learning-based techniques, P2SR requires minimal training data and has significantly lower computational overhead. This makes it especially well-suited for scenarios with limited labeled data or for near-real-time applications where fast processing is critical.

 

Preservation of Detail and Noise Suppression: By integrating sparse coding with iterative thresholding, P2SR effectively suppresses noise while avoiding the over-smoothing effects common in generative models. This results in sharper and more reliable reconstructions.

 

Guided Fusion Strategy: P2SR benefits from a guided fusion mechanism that incorporates high-resolution spatial reference data. This leads to superior performance compared to single image super-resolution approaches that typically operate on unimodal inputs.

 

These characteristics collectively make P2SR a robust and efficient choice for hyperspectral resolution enhancement across a wide range of applications. This is now clearly stated in the Introduction section of the paper.

 

Comment 2: Generally, quantitative results can be added to Abstract.

Response 2: A few key quantitative results are now added to the Abstract

 

Comment 3: Introduction: What is the motivation of this paper? In other words, what problems exist and what the authors want to solve. In addition, the main idea of the proposed method can be expressed more clearly.

Response 3: The motivation of the paper is clarified further in the introduction section with specific problems in the existing methods and solutions suggested by the proposed P2SR method.

 

Comment 4: The main contributions of this work can be concluded point by point.

Response 4: The main contributions are discussed point-by-point in the newly added subsection 5.1.

 

Comment 5: The resolution of Figure 1 can not meet the publication requirements.

Response 5: A high-defination figure for Figure 1 will be provided at the production stage of the paper.

 

Comment 6: It is suggested to write an algorithm table to describe the proposed method.

Response 6: An algorithm table is added to the manuscript in Appendix A. This description is in addition to the flowchart and detailed explanation of the algorithm which already exists in the proposed method section. We also provide a pseudo code of the implementation for clarifying the actual working of the algorithm in the appendix.

 

Comment 7: In addition to the ResTFNet and SSRNet methods, more advanced methods based on deep learning should also be added for comparison.

Response 7: We appreciate the suggestion to include more deep learning comparisons. Our literature review identified several cutting-edge approaches, many leveraging deep learning and promising future source-code releases. To ensure robust and reproducible comparisons, we included two recent, widely-cited deep learning methods—SSRNet and ResTFNet—both of which offer open-weight pre-trained models. These models are frequently used as benchmarks, enabling seamless comparison with other studies. Despite an extensive search, we found that many newer methods lack publicly available source code and more importantly, pre-trained-open weight model files. For recent studies where the source code is available, most of them lack pre-trained model weights (and only provide model architecture). The open-weight models files are critical for inclusion in this study as we rely on limited data and they also ensure a fair and meaningful comparison.

 

Comment 8: In order to demonstrate the capabilities of each part of the method, ablation experiments are very necessary.

Response 8: The authors agree that ablation experiments are necessary in most state-of-the-art methods (although most oftenly not performed as evident in literature!), and are more common to deep learning methods that use neural networks where contribution of each component can be evaluated by dropping particular components (e.g., dropout, residual connection or attention modules). We proposed a non-neural network based method that cannot be executed with ablations as each constituent component is executed sequentially and any kind of ablation will lead to a broken pipeline. This is evident from the described algorithm given in Appendix A. Hence ablation experiments are not applicable and relevant for the proposed method.

 

Comment 9: Discussion: It is suggested that the author discuss and analyze the proposed method in greater depth based on the experimental results.

Response 9: The discussion section is now provided with greater depth based on the experimental results with content organized into explicit sub-sections with more details. 

 

Comment 10: The limitations of this work should also be concluded in conclusion part.

Response 10: The limitations of the experiments are now added to the conclusion part and also explicitly discussed in the discussion section.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposes a fusion method of hyperspectral images and multispectral images based on parallel block sparse residual learning, aiming to improve the spatial resolution of HSI. The method proposed in this paper is innovative. However, there are still many problems that need to be improved before a possible publication. Below are detailed suggestions.

1. The abstract part lacks the description of the background and challenges of the research field.
2. The motivation of the proposed method is not clear.
3. The contribution of the introductionneeds to be described in points.
4. The review of the existing technology is insufficient. It is suggested to add therelated work section. 
5. The description of the proposed method is not sufficient.
6. The comparison algorithm in this paper is insufficient and should include the latest deep learning-based methods.
7. Dataset description, experimental setup and evaluation strategies should be included in the section of experiments and analysis.
8. The qualitative and quantitative analyses lack depth.
9. The conclusion lacks limitations and directions for future improvements.

Author Response

Author’s Reply to Reviewer 3

 

We thank the reviewer for the constructive critical reviews. A point-by-point response is provided for all comments by the reviewer.

 

Comment 1: The abstract part lacks the description of the background and challenges of the research field.

Response 1: The background and overarching challenges relevant to the work are added to the abstract

 

Comment 2: The motivation of the proposed method is not clear.

Response 2: The motivation with regards to solving existing challenges in hyperspectral super resolution domain are added to the introduction.

 

Comment 3: The contribution of the introduction needs to be described in points.

Response 3: The contribution of the proposed method as compared to methods described in the introduction are now added to the Discussion in sub-section 5.1.

 

Comment 4: The review of the existing technology is insufficient. It is suggested to add the related work section.

Response 4: The review of the existing technology is now extended in the introduction with some more relevant and recent studies. See references [42-50].

 

Comment 5: The description of the proposed method is not sufficient.

Response 5: A table describing the algorithm of the method is added to the manuscript in Appendix A. This description is in addition to the flowchart and detailed explanation of the algorithm which already exists in the proposed method section. We also provide a pseudo code of the implementation for clarifying the actual working of the algorithm in the appendix.

 

Comment 6: The comparison algorithm in this paper is insufficient and should include the latest deep learning-based methods.

Response 6: We appreciate the suggestion to include comparisons with more recent deep learning methods. Our literature review identified several cutting-edge approaches, many leveraging deep learning and promising future source-code releases. To ensure robust and reproducible comparisons, we included two recent, widely-cited deep learning methods—SSRNet and ResTFNet—both of which offer open-weight pre-trained models. These models are frequently used as benchmarks, enabling seamless comparison with other studies. Despite an extensive search, we found that many newer methods lack publicly available source code and more importantly, pre-trained-open weight model files. For recent studies where the source code is available, most of them lack pre-trained model weights (and only provide model architecture). The open-weight models files are critical for inclusion in this study as we rely on limited data and they also ensure a fair and meaningful comparison.

 

Comment 7: Dataset description, experimental setup and evaluation strategies should be included in the section of experiments and analysis.

Response 7: Thank you for the suggestion to consolidate the dataset description, experimental setup, and evaluation strategies into the experiments and analysis section. We agree that clarity and coherence in presenting these elements are crucial. However, we believe maintaining a separate “Dataset and Study Sites” section is justified, given the method’s application across multiple datasets and three distinct study sites. These sites are central to the study, as their unique characteristics are as critical as the proposed technique itself, particularly in the context of the “Application-Oriented Assessment” subsection. Combining the dataset description with the experimental setup could obscure the significance of the study sites and dilute the focus on their application. Moreover, the current organization also fits MDPI Remote Sensing’s recommended structure.

Comment 8: The qualitative and quantitative analyses lack depth.

Response 8: The qualitative and quantitative analysis is made more extensive and detailed.

 

Comment 9: The conclusion lacks limitations and directions for future improvements. 

Response 9: The conclusion now consists of limitations and future improvement scope.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

1. Please add more technical details.
2. It is recommended to provide a computational complexity analysis to prove that the method  reduces computational overhead.

Author Response

Comment 1: Please add more technical details.

Response 1: More technical details are now added, especially in-relation to computations.

 

Comment 2: It is recommended to provide a computational complexity analysis to prove that the method  reduces computational overhead.

Response 2: There are several ways undertaken to reduce computational complexity. A few of them are listed here: i. Processing patches in parallel for scaling across multiple cores. ii. Use of Numba JIT compiler for faster matrix operations within the iterative loops. iii. Early stopping in iterative techniques like FISTA for faster convergence. iv. Giving preference to vectorised packages for implementation.  v. Using log-based debugging for allowing the overall process to continue uninterrupted instead of halting execution. 

For providing proof of reducing computational complexity, the authors have now provided  specialised demo files (https://github.com/naikp13/hsi_enhancement/tree/main/demo) on github where the process bar demonstrates reduction in the computational overhead. 

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all concerns.

Author Response

Comment: The authors have addressed all concerns.
Response: Thank you for all the constructive comments.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed all my comments, and  I have no comments now.

Author Response

Comment: The authors have addressed all my comments, and  I have no comments now.

Response: Thank you for all the constructive comments.