Review Reports

Round 1

Reviewer 1 Report

This paper aims at tackling the image change detection problems, such as missed detection of features at different scales and incomplete region detection. This paper jointly uses the Siamese network and multi-scale attention mechanism. Moreover, it improves the contrastive loss function to fix the imbalance problem. The proposed method is evaluated on two datasets and outperforms a few existing methods. However, this paper is too premature to be accepted, due to the following reasons:

1.      The writing of this paper can be improved, including typos and grammatical errors.

2.      The quality of figure 3 is bad. Some lines do not correctly connect to the correct elements, which makes it hard to understand the architecture.

3.      In line 209, the thresholds m_u, m_c are set to 0.2 and 2.0, respectively. Please conduct comparative experiments to verify these values have the best performance.

 

4.     Only 4 algorithms are employed as the comparative algorithms. Please add more algorithms in the comparative experiments to verify the proposed algorithm can effectively improve the change detection performance of remote sensing image. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors presented a neural network architecture for change detection in remote sensing images. I have the following concerns:

1. In Section 3, the description of which dataset is used for training, validation and testing is not clear. The reader would feel both are used, and at the same time line 261 in section 4.1 mentions YNCD is used for evaluation.  That also makes interpreting section 4.2.2 harder.

2. Dividing the dataset into training, validation and testing after augmenting the data by cropping and rotating is tricky, because for validation and testing to be efficient, the images in these groups should be completely independent of the training, not derived from it by rotation and resizing for example. Hence, if the authors can comment on how datasets are created then that would be informative 

 

3. The datasets used are based on very few images. Cropping and rotating images is a good way to enrich the dataset, but still the original dataset needs to include more images.

4. Though it is not very clear, but I assume in section 4.2.2 the authors compared including and excluding part of YNCD in training. If testing is done exclusively on YNCD as mentioned in line 261, then it makes sense that adding part of it to the training will enhance the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The work is promising. However, there are some concerns:

1. One of the main basic aspects and essential subjects discussed in the work is SEGMENTATION. Hence, there should be a thorough updated discussion about the appropriate methods of segmentation tasks and/or segmentation parameters. The references on segmentation are all right but still need to be extended to cover more up-to-date details about scene and semantic segmentation.  

2. To support the segmentation results I suggest you added another figure similar to Fig. 5 but containing other details and showing exactly what changes had been done.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The layout of the paper is bad. Many tables locate in two pages

Reviewer 2 Report

The authors have addressed my concerns.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.