Change Detection for Forest Ecosystems Using Remote Sensing Images with Siamese Attention U-Net

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Good work to study the topic of forest change detection by remote sensing images and deep learning methods, but some issues should be addressed. My comments are as follows:

1) First of all, the title of this paper is not acceptable. What this paper studies is simply about the change detection of forest, but not to accurately identify "High-Risk Areas". For forest, "High-Risk" is a wide concept, the deforestion, the loss of forest caused by wildfire, the tree diseases can be all included as "High-Risks".  Besides, the remote sensing technique should be also listed in the title. I suggest to change the title as "Change Detection for Forest Ecosystems Using Remote Sensing Images  with Siamese Attention U-Net". Correspondingly, some presentations in this paper also need to be made changes.

2) The abstract part is not well organized. Too much contents about the accuracies of the proposed method are provides, but too less contents about the proposed method itself are introduced.

3) The research situation is generally well introduced in the Introduction Section, but, to better highlight the innovations of this study, the drawbacks of the current studies need to be comprehensively introduced.

4) Section 3.2: The annotation process, or to say, the training sample generation step is very important the proposed method. More details about this process should be given out, for example, list more sample images acquired in different places, introduce about the ratio between the training samples and the testing samples.

5) How about the training time and the testing time of the proposed method and other methods in Table 2?

Author Response

Reviewer#1, Concern #1: First of all, the title of this paper is not acceptable. What this paper studies is simply about the change detection of forest, but not to accurately identify "High-Risk Areas". For forest, "High-Risk" is a wide concept, the deforestation, the loss of forest caused by wildfire, the tree diseases can be all included as "High-Risks".  Besides, the remote sensing technique should be also listed in the title. I suggest to change the title as "Change Detection for Forest Ecosystems Using Remote Sensing Images with Siamese Attention U-Net". Correspondingly, some presentations in this paper also need to be made changes.

 

Author response: Thank you very much for highlighting this issue. We understand that the necessity about of changing the title. Therefore we have changed the title as “Change Detection for Forest Ecosystems Using Remote Sensing Images with Siamese Attention U-Net”.

 

 

Reviewer#1, Concern #2: The abstract part is not well organized. Too much contents about the accuracies of the proposed method are provides, but too less contents about the proposed method itself are introduced.

 

Author response: Thanks a lot for pointing out concerns regarding the abstract. We have modified the abstract as mentioned below. (Line #4-11)

 

The updated abstract is:

Forest ecosystems are critical components of Earth's biodiversity and play vital roles in climate regulation and carbon sequestration. They face increasing threats from deforestation, wildfires, and other anthropogenic activities. Timely detection and monitoring of changes in forest landscapes pose significant challenges for government agencies. To address this challenges, we propose a novel pipeline by refining U-Net design with employing two different schema of early fusion networks and Siam network architecture capable of processing RGB images specifically designed to  identifying high-risk areas in forest ecosystems through change detection across different time frames in the same location. It annotates ground truth change maps in such time frames using encoder-decoder approach with the help of an enhanced feature learning and attention mechanism. Our proposed pipeline, integrated with ResNeSt blocks and SE attention techniques, achieved impressive results in our newly created forest cover change dataset. The evaluation metrics reveal a Dice score of 39.03%, a Kappa score of 35.13%, an F1-score of 42.84%, and an overall accuracy of 94.37%. Notably, our approach significantly outperformed multitasking model approaches in the ONERA dataset, boasting a precision of 53.32%, a Dice score of 59.97%, and an overall accuracy of 97.82%. Furthermore, it surpassed multitasking models in the HRSCD dataset, even without utilizing land cover maps, achieving a Dice score of 44.62%, a Kappa score of 11.97%, and an overall accuracy of 98.44%. Although the proposed model had a lower F1-score than other methods, other performance metrics highlight its effectiveness in timely detection and forest landscape monitoring, advancing deep learning techniques in this field..

 

 

Reviewer#1, Concern #3: The research situation is generally well introduced in the Introduction Section, but, to better highlight the innovations of this study, the drawbacks of the current studies need to be comprehensively introduced.

 

Author response: Thanks a lot for this comment. We have added one more objective which can highlight overall innovation of the study and we have modified the paragraph in line number 83 and 124 and to mention drawbacks of the current studies.

 

The updated texts are:

 

1. Propose a novel attention mechanism for sequential satellite images to minimize the additional parameters to train detect changes and need of data for auxiliary tasks.
2. However, learning from small sample sets is helpful for developing algorithms when there is a lack of large amounts of labeled data, especially since deep architectures have shown higher potential in such practices. Although, in order to impose effective prevention mechanisms against deforestation, the accuracy obtained from such algorithms might be invaluable due to inadequate numbers of existing datasets.

 

Reviewer#1, Concern #4: The annotation process, or to say, the training sample generation step is very important to the proposed method. More details about this process should be given out, for example, list more sample images acquired in different places, and introduce the ratio between the training samples and the testing samples.

 

Author response: Thanks a lot for pointing out this. We have incorporated the suggested annotation process as most specifically mentioning the annotation mechanism along with an image (Figure 7). (line #272-285)

 

Reviewer#1, Concern #5: How about the training time and the testing time of the proposed method and other methods in Table 2?

 

Author response: Thank you for this valuable comment.  We have included training and testing time along with accuracy in a new table Table 3 and added a descriptive paragraph for that. (line #427-437)

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors,

Your work is interesting, innovative and generally well written. Some suggestions:

- Please revise the abstract to better focus the framework of the work and its importance in the forestry sector.

- The introduction also needs a revision (especially in the first part) to focus the framework of your research. Forestry is a research topic where ML and DL are intensively used to predict and monitor phenomena; therefore you need to be precise in identifying, criticising, boundaries of innovation and state of the art.

- Why RGB images? Currently we have a high availability of remote sensing data, acquired by satellite (sometimes free), another remote and proximal sensor. You have to argue better this chosen.

- Figure 12 is too small and has a poor resolution.

- In conclusion, you can better describe the next steps of the research and emphasise the practical issues of your research.

Author Response

Reviewer#2, Concern #1: Your work is interesting, innovative and generally well written. Some suggestions: Please revise the abstract to better focus the framework of the work and its importance in the forestry sector.

Author response: Thanks a lot for this comment. We have revised the abstract by more specifically elaborating the framework.

Reviewer#2, Concern #2: Your work is interesting, innovative and generally well written. Some suggestions: Please revise the abstract to better focus the framework of the work and its importance in the forestry sector The introduction also needs a revision (especially in the first part) to focus the framework of your research. Forestry is a research topic where ML and DL are intensively used to predict and monitor phenomena; therefore you need to be precise in identifying, criticising, boundaries of innovation and state of the art.

Author response: Thanks a lot for this comment. We have added one more objective which can highlight overall innovation of the study.

The updated texts are:

1. Propose a novel attention mechanism for sequential satellite images to minimize the additional parameters to train detect changes and need of data for auxiliary tasks. (line 124)

Reviewer#2, Concern #3: Why RGB images? Currently we have a high availability of remote sensing data, acquired by satellite (sometimes free), another remote and proximal sensor. You have to argue better this chosen.

                                                                                                                                            

Author response: Thank you for this valuable comment.  We have incorporate following justifications in line number 246-250.

The added justifications are:

1. Accessibility: RGB images are widely available, often free, and easier to acquire than multi-spectral or hyper-spectral images, which might be expensive or restricted.
2. Processing Simplicity: RGB images require less computational power and simpler processing techniques compared to multi-spectral or hyper-spectral data, which can be more complex and resource-intensive to analyze.

 

Reviewer#2, Concern #4: Figure 12 is too small and has a poor resolution.

 

Author response: Thanks a lot for pointing out this. We have changed the Figure 12 with high resolution images.

 

Reviewer#2, Concern #5: In conclusion, you can better describe the next steps of the research and emphasise the practical issues of your research.

 

Author response: Thank you very much for highlighting this concern. We understood the significance of emphasizing the future step of this work and practical complications of this study. Therefore, we have addressed the future work of this study and the practical issue of this study in the conclusion part. (line #511-518)

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript is clear and well-organized, and the authors' efforts are greatly appreciated. However, a significant concern is the low F1-score combined with high accuracy in the results. This discrepancy suggests that while the model successfully classifies the majority class (non-deforested areas), it struggles to effectively distinguish deforested areas, leading to poor performance in identifying the minority class. As a result, the proposed model may not be effectively capturing deforestation areas, which raises questions about the overall contribution and broader impact of this study.

1. In the introduction (line 112), the authors claim to propose a 'novel deep-learning model' for identifying deforestation areas; however, the specific innovations or novel elements introduced by the model are not clearly articulated. It seems the study may be applying existing deep learning models to the deforestation context without introducing significant innovation. Greater clarity and detail are needed to clearly outline what sets this model apart. Additionally, a comparison with existing approaches should be provided to highlight both similarities and differences, thereby underscoring the model's unique contributions.

2. In the abstract, is there a specific reason the F1-score is not mentioned in lines 11 and 13? Additionally, in Section 4.1 and Table 1, there is a lack of description regarding the F1-score, and it is also absent in the results in Table 4. The F1-score is a crucial metric, particularly in evaluating model performance on imbalanced datasets, as it provides a more comprehensive assessment than accuracy alone. Please address these issues.

3. In line 104, I recommend removing the term 'accurately' given the low F1-score results.

 

4. I suggest comparing your results, particularly the F1-score, with those in existing studies or literature. This comparison could help justify the observed low F1-score. Additionally, this significant limitation should be acknowledged in the abstract, introduction, discussion of the relevant results, and conclusion. I recommend including a thorough discussion of this limitation in a dedicated paragraph within the conclusion.

Author Response

Reviewer#3, Concern #2: In the introduction (line 112), the authors claim to propose a 'novel deep-learning model' for identifying deforestation areas; however, the specific innovations or novel elements introduced by the model are not clearly articulated. It seems the study may be applying existing deep learning models to the deforestation context without introducing significant innovation. Greater clarity and detail are needed to clearly outline what sets this model apart. Additionally, a comparison with existing approaches should be provided to highlight both similarities and differences, thereby underscoring the model's unique contributions.

Author response: Thanks a lot for pointing out this issue. We have changed the novel deep learning mode to enhanced deep learning model, because we modified the existing architecture of the deep learning models.

Please refer Table 4 and 5 regarding the comment of comparison with existing approaches.

Reviewer#3, Concern #3: In the abstract, is there a specific reason the F1-score is not mentioned in lines 11 and 13? Additionally, in Section 4.1 and Table 1, there is a lack of description regarding the F1-score, and it is also absent in the results in Table 4. The F1-score is a crucial metric, particularly in evaluating model performance on imbalanced datasets, as it provides a more comprehensive assessment than accuracy alone. Please address these issues.

Author response: Thanks a lot for pointing out this issue. Since all of the models’ F1-Score comparatively low, we have considered mainly the accuracy dice and precision score. However, we have mentioned it in the abstract as well. (line #17-19)

Reviewer#3, Concern #4: In line 104, I recommend removing the term 'accurately' given the low F1-score results.

Author response: Thanks a lot for pointing out this issue. We removed the term since our F1-Score is low. (line #109)

Reviewer#3, Concern #5: In line 104, I recommend removing the term 'accurately' given the low F1-score results I suggest comparing your results, particularly the F1-score, with those in existing studies or literature. This comparison could help justify the observed low F1-score. Additionally, this significant limitation should be acknowledged in the abstract, introduction, discussion of the relevant results, and conclusion. I recommend including a thorough discussion of this limitation in a dedicated paragraph within the conclusion.

Author response: Thanks a lot for the suggestion. We have acknowledged the experimental issue of low F1-Score in the abstract, results and discussion and conclusion.

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have no more questions.

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for the diligent revisions. The manuscript has improved in this version. I recommend accepting this revised submission.