SMNet: Symmetric Multi-Task Network for Semantic Change Detection in Remote Sensing Images Based on CNN and Transformer

Round 1

Reviewer 1 Report

Line13-20: The abstract part can describe the advantages of the model and how to improve it without describing the experimental process.

Line80-88: The main contribution can be more specific, such as what results are achieved.

Line 168-169: Comparing ResNet and PR is to briefly describe the differences between the two, which can better reflect the advantages of PR.

Line204:Explain what each element in the formula means. (C,H,W).

Line224-225 You say that your rough operation will lose details, so is there any way to optimize this operation? I thought it might take your experiment a step further.

Line 296 What accounts for the apparent imbalance between the positive and negative samples? On the other hand, you say that oa is not guaranteed why choose it as an outcome indicator?

Line414 How do you choose your five N-level Settings? What is the effect of the other level Settings?

Line434-438 Your conclusion may seem too brief and lacking in depth, such as the theoretical and practical significance and value of the paper, and suggestions for further research on the topic

In summary, I was happy to review your manuscript:” Multi-task Semantic Change Detection for Remote Sensing Images Based on CNN and Transformer”. Overall, it is clear that a lot of work has gone into this research. But This innovation of  manuscript will be constantly excavated and verified by supplement experiments. Overall Recommendation is Reconsider after major revision (control missing in some experiments).

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The study offers a multi-task semantic change detection method for remote sensing images based on CNN and transformer. The following is the specific suggestions:

1. The study propose a novel symmetric multi-task network (SMNet), so SMNet should be shown in the title.

2. Section 2 should not only list the related work. Section 2 may add some analysis content about the related work.

3. Figure 6, GT is the ground truth used for validating the results of SMNet method. However, GT images are not complete.

4. Line 370, water is effectively detected by SMNet. However, from Figure 6, water is also effectively detected by BIT.

5. Figure 8 is not distinct.

6. Table 1, the gap between BIT and SMNet is very small.

7. Line 396, the addition of the multi-task loss achieves a gain of 0.3% and 0.73% for mIoU and Sek, respectively. The values 0.3% and 0.73% is too small, which is meaningless. These values may be accidental errors.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Summary: Dense change detection with semantic category information is a challenging computer vision task. This paper proposes a novel method for semantic change detection by gathering local and global semantic information using residual block and transformer block and utilizing a multi-content fusion module to separate the foreground and background information. Finally, the proposed method shows good performance on the SECOND dataset.

My comments are provided below.

1. Novelty: The proposed modules such as PR, TB, and MCFM are not novel which makes the novelty of the paper weak. However, the application of these approaches to SCD is novel. The authors are encouraged to explain how their contribution is novel.

2. In Sec. 3.2, what does the input feature X refer to?

3. Benchmarks: Evaluating results on only one dataset is not sufficient and may not fully demonstrate the generalizability of a model. Experiments should be performed on other datasets such as Landsat-SCD. Also, the authors only compared the performance of the proposed method with a few existing methods. They are encouraged to compare the proposed method with other existing methods as mentioned in [1].

[1] https://arxiv.org/pdf/2212.05245v3.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

In summary, I was happy to review your manuscript:” SMNet: Symmetric Multi-task Network for Semantic Change 2 Detection in Remote Sensing Images Based on CNN and 3 Transformer”.  Overall Recommendation is Accept after minor revision (corrections to minor methodological errors, text editing and Moderate English changes required). 

Reviewer 3 Report

My comments are addressed. I recommend accepting the paper.