A Multi-Dimensional Deep Siamese Network for Land Cover Change Detection in Bi-Temporal Hyperspectral Imagery
Round 1
Reviewer 1 Report
Thanks for the possibility to review your work. I have some revisions I would kindly ask authors to address.
1. In the first section of the article, the authors listed many examples of using DL methods for HCD. But the authors only lists the methods, and does not explain the experimental results parameters of the listed methods. Such a presentation would not reflect the strengths of the authors' method.
2. In the Second section of the article, both two figures are numbered "Figure 2", which is an obvious mistake. The author is advised to check the details of the whole article. They should avoid similar mistakes.
3. In the Second section of the article, the datasets used by the authors are relatively old. Compared with recent datasets, considering sharpness, changes in the shape and scale of farmland in recent years, and other conditions, will these relatively old images have an impact on the experimental results.
4. In the section 3.2.3 of the article, the authors use the Otsu thresholding method to segment objects from the background, has the author compared other image thresholding methods, such as optimal thresholding and adaptive thresholding. Will the segmentation results of different methods affect the experimental results?
5. In the section 3.3 of the article, the author says ‘The proposed architecture has three 3D dilated convolution layers, a 2D convolution layer, and a 1D convolution layer.’ Can the combination of different layers bring better results to the experiment? The author can add comparative experiments to verify.
6. In the Fourth section of the article, in Table 2, the authors compare the comparison of OA parameters using the original Otsu method and the hierarchical Otsu method for the farmland dataset. Are these results test results or validation results？
7. In the Fourth section of the article, the author says ‘our method had lower MD and FA rates for both datasets, indicating its higher performance in HCD’, but the FA rate indicators are listed in Table 4 and Table 5. The SU method and the MSU method have more advantages in the two datasets. The authors are advised not to compare FA rates.
8. In the references, the author cited articles from 2013 and 2016. Deep learning technology is developing rapidly. The methods used in these older articles should not represent the level of the field in recent years. To highlight the advantages, it is recommended that authors cite the state of the art articles with relatively recent years for comparison.
Author Response
Title: A Multi-Dimensional Deep Siamese Network for Land Cover Change Detection in Bi-Temporal Hyperspectral Imagery
Dear Editor, Reviewers,
The authors thank you and the two reviewers for valuable comments, by which the current manuscript improved considerably. We answered all the questions below, and the comments were implemented in the Revised manuscript (please see the yellow highlighted areas in the revised manuscript). The main corrections are listed below:
(1) The grammar and structure of the manuscript were improved.
(2) To better evaluation of the proposed method's performance, two state-of-the-art methods were implemented and the results were compared with those of the proposed method.
(3) More HCD methods were reviewed in the introduction section.
(4) The limitation of the proposed method was discussed in more detail.
(5) The ablation analysis was added.
Response to Reviewer 1:
- In the first section of the article, the authors listed many examples of using DL methods for HCD. But the authors only list the methods and do not explain the experimental results parameters of the listed methods. Such a presentation would not reflect the strengths of the authors' method.
Answer 1: As suggested, the authors added the details of the experimental results of the listed methods in the revised manuscript (Section 1, Paragraph 3, Lines 60, 66, 71, 82, 87).
- In the Second section of the article, both two figures are numbered "Figure 2", which is an obvious mistake. The author is advised to check the details of the whole article. They should avoid similar mistakes.
Answer 2: The number of figures is revised. (Section 3, Figure 3).
- In the Second section of the article, the datasets used by the authors are relatively old. Compared with recent datasets, considering sharpness, changes in the shape and scale of farmland in recent years, and other conditions, will these relatively old images have an impact on the experimental results.
Answer 3: As suggested, the authors explored the recent hyperspectral change detection datasets. However, there are many change detection benchmark datasets (Below Table) but we did not find a recent benchmark dataset for hyperspectral images. It is worth noting that the generation of the new dataset with reliable ground truth needs more time (more than 30 days as). Furthermore, the datasets used in this study are benchmark HCD datasets that have been used in many studies. The utilizing new dataset need reliable ground truth which is not available to us.
- In section 3.2.3 of the article, the authors use the Otsu thresholding method to segment objects from the background, as the author compared other imaging thresholding methods, such as optimal thresholding and adaptive thresholding. Will the segmentation results of different methods affect the experimental results?
Answer 4: We agree that utilizing adanaved thresholding methods might slightly affect the experimental results. However, we should note that although the Otsu thresholding is a simple thresholding method, other thresholding methods have a high computational cost and more complexity. This, the authors employed the Otsu thresholding in a hierarchical manner to improve the quality of sample dataset generation.
- In section 3.3 of the article, the author says ‘The proposed architecture has three 3D dilated convolution layers, a 2D convolution layer, and a 1D convolution layer.’ Can the combination of different layers bring better results to the experiment? The author can add comparative experiments to verify.
Answer 5: As suggested, the authors added the ablation analysis in the revised manuscript (Section 4, Last paragraph, Table 6, Lines 374-377)
- In the Fourth section of the article, in Table 2, the authors compare the comparison of OA parameters using the original Otsu method and the hierarchical Otsu method for the farmland dataset. Are these results test results or validation results？
Answer 6: The presented results are for the selected sample regions in Figure 6-a and Figure 6-d. Moreover, the authors added the ablation analysis in the revised manuscript (Subsection 4, paragraph 2, Line 255).
- In the Fourth section of the article, the author says ‘our method had lower MD and FA rates for both datasets, indicating its higher performance in HCD’, but the FA rate indicators are listed in Table 4 and Table 5. The SU method and the MSU method have more advantages in the two datasets. The authors are advised not to compare FA rates.
Answer 7: As suggested, the authors revised this issue (Section 2.2, Line 175-179).
- In the references, the author cited articles from 2013 and 2016. Deep learning technology is developing rapidly. The methods used in these older articles should not represent the level of the field in recent years. To highlight the advantages, it is recommended that authors cite state-of-the-art articles from relatively recent years for comparison.
Answer 8: Based on your comment, the authors discussed more articles in the revised manuscript. Furthermore, two new state-of-the-art deep learning-based methods were implemented in the new version of the paper (See, Section 1, Paragraph 3, Line 78-82, Subsection 3.4, Line 230-233, Section 4, Figures 7, 8 ,9, 10, Tables 4, 5).
Reviewer 2 Report
Thank you for giving me this opportunity to read the manuscript entitled "A Multi-Dimensional Deep Siamese Network for Land Cover Change Detection in Bi-Temporal Hyperspectral Imagery". The topic of this manuscript is interesting. However, some issues still need to be addressed before it could be considered for publication in Sustainability.
- Please replace the keywords that already appear in the manuscript's title with close synonyms or other keywords, which will also facilitate your paper to be searched by potential readers.
- The scale, compass, and legend should be added to the maps in the Figures.
- More study regions with larger areas are suggested to be used as experimental data to demonstrate the robustness of the method.
- Lines 25-26: “The Earth's surface constantly changes due to different factors, such as climate change and anthropogenic activities[1,2].” Some newly published papers could be cited to support the statement here, for example, the paper titled “How does urban expansion impact people's exposure to green environments? A comparative study of 290 Chinese cities.”
- Limitations should be added as a sub-section of the Discussion section.
- Some grammatical errors exist in the manuscript. Therefore, a critical review of the manuscript language will improve readability.
Author Response
Title: A Multi-Dimensional Deep Siamese Network for Land Cover Change Detection in Bi-Temporal Hyperspectral Imagery
Dear Editor, Reviewers,
The authors thank you and the two reviewers for valuable comments, by which the current manuscript improved considerably. We answered all the questions below, and the comments were implemented in the Revised manuscript (please see the yellow highlighted areas in the revised manuscript). The main corrections are listed below:
(1) The grammar and structure of the manuscript were improved.
(2) To better evaluation of the proposed method's performance, two state-of-the-art methods were implemented and the results were compared with those of the proposed method.
(3) More HCD methods were reviewed in the introduction section.
(4) The limitation of the proposed method was discussed in more detail.
(5) The ablation analysis was added.
Response to Reviewer 2:
- Please replace the keywords that already appear in the manuscript's title with close synonyms or other keywords, which will also facilitate your paper to be searched by potential readers.
Answer 1: As suggested, the authors changed the keywords (Section Abstract, Keywords, Line 21).
- The scale, compass, and legend should be added to the maps in the Figures.
Answer 2: These datasets are benchmark data and were in mat format. Thus, we couldn’t add scale and compass to the figures. However, the authors added more details about the label of images in the caption of the figures (Figure 9, Figure 10).
- More study regions with larger areas are suggested to be used as experimental data to demonstrate the robustness of the method.
Answer 3: The datasets used in this study are benchmark HCD datasets that have been used in many studies. The utilizing new dataset needs reliable ground truth which is not available to us. The used datasets have highly complex change classes. Moreover, It is worth noting that the generation of the new dataset with reliable ground truth needs more time (more than 30 days).
- Lines 25-26: “The Earth's surface constantly changes due to different factors, such as climate change and anthropogenic activities [1,2].” Some newly published papers could be cited to support the statement here, for example, the paper titled “How does urban expansion impact people's exposure to green environments? A comparative study of 290 Chinese cities.”
Answer 4: Based on your comment, the authors discussed more articles in the revised manuscript (Section 1, Paragraph 3, Line 25).
- Limitations should be added as a sub-section of the Discussion section.
Answer 5: As suggested, the limitation of this research was discussed in more detail (Section 5, Last paragraph, Lines 324-377)
- Some grammatical errors exist in the manuscript. Therefore, a critical review of the manuscript language will improve readability.
Answer 7: Based on your comment, we reviewed and revised the manuscript for grammatical errors.
Round 2
Reviewer 1 Report
All my questions are well sovled. The manuscript is greatly improved. I think that the manuscript could be accepted by the journal.
