JointNet: Multitask Learning Framework for Denoising and Detecting Anomalies in Hyperspectral Remote Sensing
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a multitask learning for HSI anomaly detection and denoising, Experimental results demonstrate the effectiveness of the proposed method. The following problems should be addressed.
1. The introduction reviews the denoising methods, and the related work discusses the detection work, which seems somewhat unreasonable. The authors should also discuss methods that integrate these two tasks.
2. The images in the article exceed the page margins.
3. Figure 2 should illustrate which part of the AE module serves as the minimum noise fraction rotation (MNF) algorithm.
4. The reviewer suggests discussing some deep learning methods related to HSI processing.
5. The authors should conduct separate comparative analyses of hyperspectral denoising methods and anomaly detection methods to demonstrate the efficacy of handling both tasks simultaneously.
6. The authors also need to compare the joint denoising and detection methods to validate the effectiveness of the proposed method.
7. The author need to conduct ablation experiments on denoising and anomaly detection to demonstrate the effectiveness of the proposed network.
8. Real data experiments are needed to verify the robustness of the proposed method (if any).
Comments on the Quality of English LanguageN/A
Author Response
Comments 1: The introduction reviews the denoising methods, and the related work discusses the detection work, which seems somewhat unreasonable. The authors should also discuss methods that integrate these two tasks.
Response 1: Thank you for pointing this out, and we agree with this comment. We have included a discussion of the integration of the two tasks in the introduction, and have shown that the main body of the method is anomaly detection. The revised content is highlighted in red color in the manuscript.
Comments 2: The images in the article exceed the page margins.
Response 2: Thank you for pointing this out, and we agree with this comment. We have adjusted the formatting of the images in the article.
Comments 3: Figure 2 should illustrate which part of the AE module serves as the minimum noise fraction rotation (MNF) algorithm.
Response 3: Thank you for pointing this out, and we agree with this comment. We have modified figure 2, and further detailed the correspondence between the encoder of the AE and the MNF algorithm in figure 3.
Comments 4: The reviewer suggests discussing some deep learning methods related to HSI processing.
Response 4: Thank you for pointing this out, and we agree with this comment. We have discussed some DL methods related to HSI processing in the introduction and related work. The revised content is highlighted in red color in the manuscript.
Comments 5: The authors should conduct separate comparative analyses of hyperspectral denoising methods and anomaly detection methods to demonstrate the efficacy of handling both tasks simultaneously.
Response 5: Thank you for pointing this out, and we agree with this comment. We have restructured and improved the experimental section, where subsection 4.5.1 compares the denoising performance, and subsection 4.5.2 compares the anomaly detection performance.
Comments 6: The authors also need to compare the joint denoising and detection methods to validate the effectiveness of the proposed method.
Response 6: Thank you for pointing this out, and we agree with this comment. Since there are fewer existing studies on joint denoising and anomaly detection, and the subject of the proposed method in this article is anomaly detection, we have mainly selected anomaly detection models with the same role for comparison in the comparison experiments section, and added a comparison algorithm to validate the effectiveness of the proposed method.
Comments 7: The author need to conduct ablation experiments on denoising and anomaly detection to demonstrate the effectiveness of the proposed network.
Response 7: Thank you for pointing this out, and we agree with this comment. We have included relevant ablation experiments in the experimental section, and tested the effects of performing the two tasks independently and jointly. The results validate the effectiveness of the proposed method.
Comments 8: Real data experiments are needed to verify the robustness of the proposed method (if any).
Response 8: Thank you for pointing this out, and we agree with this comment. We currently have no way to obtain real noisy hyperspectral images. We have selected shared datasets common to existing denoising and anomaly detection studies to test the experimental, and do the same type of preprocessing sessions. The results can confirm the effectiveness of the model. Further attempts will be made to use real datasets, and the aspiration will be added to the outlook section. The revised content is highlighted in red color in the manuscript.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper introduces a multitask learning-based anomaly detection method for noisy hyperspectral images, which is innovative as it combines denoising and anomaly detection tasks, sharing feature information to enhance detection accuracy. However, this manuscript should be further improved before it can be considered for publication.
1. There are some minor grammatical errors and inappropriate phrasing throughout the manuscript. It needs thorough proofreading to improve the grammar.
2. The methodology section is sparse on details regarding the JointNet model. Please provide a comprehensive description of the network layers, activation functions, loss functions, and hyperparameters used. A detailed architectural diagram of JointNet should be included to aid understanding.
3. The discussions of the results are somewhat inconsiderate. Please provide a deeper analysis of the performance metrics. Please explain why JointNet outperforms other models and include potential reasons for any anomalies or unexpected results observed in the data.
4. The experimental datasets used in this paper are a little insufficient so the generalized effect may be unavailable. Additionally, more information, including the source, number of samples, dimensions, and any preprocessing steps such as normalization or augmentation, should be included.
5. Most importantly, the proposed methods adopt the groundtruth in the cross-entropy loss for binary classification, which is unscientific in conventional anomaly detection. Therefore, the authors should explain the reason and how to apply it to real detection scenarios where groundtruth is unavailable.
6. The critical formulations like “A” and “q(X)” should be marked in the overall architecture for better understanding.
7. The auto-encoder network is too simple to model the background distribution. Therefore, it is recommended to use advanced variants such as variational auto-encoder (VAE) and adversarial auto-encoder (AAE).
Comments on the Quality of English LanguageThere are some minor grammatical errors and inappropriate phrasing throughout the manuscript. It needs thorough proofreading to improve the grammar.
Author Response
Comments 1: There are some minor grammatical errors and inappropriate phrasing throughout the manuscript. It needs thorough proofreading to improve the grammar.
Response 1: Thank you for pointing this out, and we agree with this comment. This article has been processed using the language editing.
Comments 2: The methodology section is sparse on details regarding the JointNet model. Please provide a comprehensive description of the network layers, activation functions, loss functions, and hyperparameters used. A detailed architectural diagram of JointNet should be included to aid understanding.
Response 2: Thank you for pointing this out, and we agree with this comment. We have enhanced the detailed description of the JointNet model in the Methodolgy, and labeled the correspondence between the encoder’s components and the MNF algorithm in Figure 3. The revised content is highlighted in red color in the manuscript.
Comments 3: The discussions of the results are somewhat inconsiderate. Please provide a deeper analysis of the performance metrics. Please explain why JointNet outperforms other models and include potential reasons for any anomalies or unexpected results observed in the data.
Response 3: Thank you for pointing this out, and we agree with this comment. We have included further analysis of the results and potential anomalies or unexpected results in the thematic experiments of the Experiment.
Comments 4: The experimental datasets used in this paper are a little insufficient so the generalized effect may be unavailable. Additionally, more information, including the source, number of samples, dimensions, and any preprocessing steps such as normalization or augmentation, should be included.
Response 4: Thank you for pointing this out, and we agree with this comment. We have selected shared datasets common to existing denoising and anomaly detection studies to test the experimental, and do the same type of preprocessing sessions. A detailed description of the dataset is discussed in the 4.1 section. The results can confirm the effectiveness of the model. Further attempts will be made to use real datasets, and the aspiration will be added to the outlook section. The revised content is highlighted in red color in the manuscript.
Comments 5: Most importantly, the proposed methods adopt the groundtruth in the cross-entropy loss for binary classification, which is unscientific in conventional anomaly detection. Therefore, the authors should explain the reason and how to apply it to real detection scenarios where groundtruth is unavailable.
Response 5: Thank you for pointing this out, and we agree with this comment. The cross-entropy loss here is designed for subtask II of the proposed method, which is critical to bicategorize the pixels to be tested to noise and anomaly targets based on the reconstruction error and the noise component obtained through subtask I. The physical meaning of ground truth here is whether the pixel to be tested is noise, and the value can be obtained indirectly from the noise component. The detailed description of the process has been further refined in the article. The revised content is highlighted in red color in the manuscript.
Comments 6: The critical formulations like “A” and “q(X)” should be marked in the overall architecture for better understanding.
Response 6: Thank you for pointing this out, and we agree with this comment. We have described the local loss function in more detail in the 3.4 subsection, where “p(X)” denotes the ground truth, numerically obtained indirectly from the noise component, and “q(X)” denotes the prediction of the label of the pixel to be tested, numerically equal to the noise score “A”. The revised content is highlighted in red color in the manuscript.
Comments 7: The auto-encoder network is too simple to model the background distribution. Therefore, it is recommended to use advanced variants such as variational auto-encoder (VAE) and adversarial auto-encoder (AAE).
Response 7: Thank you for pointing this out, and we agree with this comment. We have added this proposal to the outlook, then refined the proposed model and further research.
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper presents a multitask learning-based anomaly detection method for noisy HSI. A JointNet model is developed to avoid the loss of image details. In addition, the authors propose a new loss function to capture the shared information across different tasks. Experimental results demonstrate the effectiveness of the design. My comments are as follows:
1. In the introduction, the literature review and problem statements should be improved. The authors provide the literature review for both denoising and anomaly detection. It is unclear which one is the main goal of this paper. The problem statement should be improved. Moreover, the authors should improve the literature review on anomaly detection.
2. There exist a number of typos. For instance, “by separate, denoising, and anomaly detection…” in line 102. “uses plug-and-play a prior ….” In line 163. The authors need to proofread the paper.
3. The authors can provide the results with varying parameters to investigate their influence on the performance.
4. The best results in the tables can be annotated in bold.
Comments on the Quality of English LanguageThere exist a number of typos. For instance, “by separate, denoising, and anomaly detection…” in line 102. “uses plug-and-play a prior ….” In line 163. The authors need to proofread the paper.
Author Response
Comments 1: In the introduction, the literature review and problem statements should be improved. The authors provide the literature review for both denoising and anomaly detection. It is unclear which one is the main goal of this paper. The problem statement should be improved. Moreover, the authors should improve the literature review on anomaly detection.
Response 1: Thank you for pointing this out, and we agree with this comment. We have refined the problem statements in the Introduction, emphasized out that the main goal is anomaly detection, and enhanced the literature review on anomaly detection in Related work. The revised content is highlighted in red color in the manuscript.
Comments 2: There exist a number of typos. For instance, “by separate, denoising, and anomaly detection…” in line 102. “uses plug-and-play a prior ….” In line 163. The authors need to proofread the paper.
Response 2: Thank you for pointing this out, and we agree with this comment. We have processed this article using language editing.
Comments 3: The authors can provide the results with varying parameters to investigate their influence on the performance.
Response 3: Thank you for pointing this out, and we agree with this comment. For space reasons, we pick the main parameters and give in subsection 4.2 the effect of different parameters on the results, as shown by the graphs and tables. The revised content is highlighted in red color in the manuscript.
Comments 4: The best results in the tables can be annotated in bold.
Response 4: Thank you for pointing this out, and we agree with this comment. We have annotated the best results in bold in all the tables.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsMy concerns are not well addressed by the authors. The reply is not satisfactory as some of my suggestions are not taken into consideration or dealt with perfunctorily, for example, adding compared methods, or adding more experimental results. I'm afraid it's hard to improve the quality of this article in this situation. Considering that Remote Sensing is a top-class journal, I recommend rejecting this manuscript.
Comments on the Quality of English LanguageN/A
Reviewer 2 Report
Comments and Suggestions for AuthorsAll my comments are addressed. The revisions have significantly improved the quality of the manuscript. However, there are still two minor issues that need to be addressed:
1. The paper lacks a detailed explanation of the overall framework, i.e., a detailed explanation of Figure 2.
2. The letters in Equations (1), (2), and (3) have not been explicitly explained in the paper.
Reviewer 3 Report
Comments and Suggestions for AuthorsThanks for the revision. I have no more comments.
