SMD-Net: Siamese Multi-Scale Difference-Enhancement Network for Change Detection in Remote Sensing
Round 1
Reviewer 1 Report
This paper proposed a siamese change detection method, named SMD-Net, for bi-temporal remote sensing change detection. In the SMD-Net, a siamese residual multi-kernel pooling module (SRMP)and a feature difference module (FDM) is used to enhance change information. Experiments show it is effective. There are some questions:
1. Description of the SDM-Net strcture should be reconsidered carefully. The figures(Fig.1-3) are a little confused, for example, in figure 1(a) ,the module FDM and SRMP have only 1 input, it is the concatenation of the 2 feature maps from image 1 and image 2 respectively,
but in Figure 2 and 3, both of the module have 2 inputs(2 feature maps). And in Fig. 1(a), there are deep green blocks with light
green babkground. Is the deep green blocks share the same meaning as in Fig. 1(b)? And what is the light green background mean?
2. At P6 paragraphy 2 line 11, "Figure 1(b)" should be Figure 1(c).
3. In my opinion, the description of dataset and the evaluation metrics (in the subsections 3.1.1, 3.1.2 and the section 3.2 respectively) are very detailed. And most of these sentences are from reference [38]. those are not important parts of your paper, could these be simplified?
4. Quality of training data is a general problem to all deep learning models. in most cases,we work hard on improving the data quality, to weaken their effects to the model. And in some special case, when the robustness of the model is our goal, some noise data will be added consciously to improve the robustness of the model. Obviously, it is is not research problem of the paper. Therefore, subsection 4.1, I think, is not necessary.
5. In most tables, the evaluation metrics are listed in order P, R, F1 and OA ,but in the text, they are listed in reverse order. It is inconvenient to the readers.
6. In this paper, P, R, F1 and OA is selected as the evaluation metrics. Maybe the intersection over union( IOU) , Kappa are more accurate indicators to evaluate the experiment results, therefore those 2 metrics are adviced.
Author Response
Dear reviewer, thank you very much for your invaluable comments which helps improve our manuscript significantly. Please read our responses and new manuscripts.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Please find the review in the attached file.
Comments for author File: Comments.pdf
Author Response
Dear reviewer, thank you very much for your invaluable comments which helps improve our manuscript significantly. Please read our responses and new manuscript.
Author Response File: Author Response.pdf
Reviewer 3 Report
The article presents a change detection method that aims to improve the results of the following parameters: on object completeness, detection of small objects and edge of objects. The literature review is well done and provides a good context for the method presented. The section that details the method proposition is also well structured. However, the analysis of results and the conclusion are not at the same level of quality as the other sections of the manuscript. In the results analysis section, the authors make an excellent comparison between the general results of the proposed method using measures such Overall Accuracy (OA), Precision (P), Recall (R) and F1-score. They also perform an interesting analysis of the quality of the bases used, identifying overlaps between the training, validation, and test samples. However, the greatest contribution proposed by the method would be the improvement of three types of common problems in change detection algorithms: object completeness, small object detection, and object edge.
This contribution is not robustly proven in the presented analysis of results. The authors only indicate the improvement of these three problems in a qualitative way with the help of some figures selected as an example. There is no quantitative analysis of how the method differs from others in terms of in object completeness, small object detection, and object edge.
For the publication of the manuscript in this journal, I recommend that the authors complement the analysis of results by quantitatively evaluating the performance of the method in terms of in object completeness, small object detection, and object edge. Another point that I would like to see complementing the analysis of results would be the quantitative assessment of how much the proposed method was able to identify changes that were not identified in the databases used. Another recommendation is that the authors also redo their conclusions so that they not only present the positive points of the experiment, but also make a critical analysis of the limitations of the experiment, proposing future works to overcome these limitations.
Author Response
Dear reviewer, thank you very much for your invaluable comments which helps improve our manuscript significantly. Please read our responses and new manuscript.
Author Response File: Author Response.pdf
Reviewer 4 Report
Please see the attached file.
Comments for author File: Comments.pdf
Author Response
Dear reviewer, thank you very much for your invaluable comments which helps improve our manuscript significantly. Please read our responses and new manuscript.
Round 2
Reviewer 1 Report
There are 2 questions:
- Question about the SDM_net structure in Fig. 1:
In ResNet-34, there are 16 res-blocks and a convolutional layer (with 7×7 filter) and a pooling layer at beginning and a global average pooling layer at the end. In Fig. 1, it seems your model has 5 modules, then how are the modules divided?
And in Fig. 1, the encoder is a trapezoid (ladder-sharped), why? The output feature maps are reduced in shape, but the blocks in network is not. And also in Fig. 1, are you sure the decoder is a trapezoid？
- It is nice the IOU and kappa are added to evaluate the experiment results. But it is added only in Tables. Could they be analyzed in the text? It is the same to evaluation index P and R.
Author Response
Dear reviewer, thank you very much for your invaluable comments which helps improve our manuscript significantly. Please read our responses and new manuscripts.
Author Response File: Author Response.pdf
Reviewer 2 Report
Dear authors, dear editors,
following the adjustments to the submission and the authors' reply to the review, I recommend the submission for acceptance.
Besides including additional information into the text to clarify my minor objections, the authors did address my major concern: To assess the impact of pre-training in the considered networks, the authors provided additional experiments and clarified the role of pre-training in the text. That given, I assume the readers can clearly evaluate the authors' results and experimental design. No further concerns remain.
Author Response
following the adjustments to the submission and the authors' reply to the review, I recommend the submission for acceptance. Besides including additional information into the text to clarify my minor objections, the authors did address my major concern: To assess the impact of pre-training in the considered networks, the authors provided additional experiments and clarified the role of pre-training in the text. That given, I assume the readers can clearly evaluate the authors' results and experimental design. No further concerns remain.
Reply: Thank you very much for your comments which helps improve our manuscript.
Reviewer 4 Report
All my concerns have been answered.
Author Response
All my concerns have been answered.
Reply: Thank you very much for your comments which helps improve our manuscript.
