Deep Siamese Networks Based Change Detection with Remote Sensing Images
Round 1
Reviewer 1 Report
In the article, the authors collected a change detection dataset with 862 labelled image pairs and they proposed a supervised change detection method based on a deep siamese semantic segmentation network to handle the proposed data effectively. However, I have some comments on the article and suggested some improvements before it is considered to be published. (see comments).
- Many acronyms were used in the manuscript; a list of acronyms should be provided before the ‘references’ section for easy readership.
- In the dataset section (page 4), there is no mention of the Ground Sampling Distance (GSD) of the satellite images. GSD is necessary to understand the quality of the dataset. Authors should include the information.
- Authors may consider making the code open to the public to strongly support the paper’s overall claim (if possible).
Overall, the paper formulated a well-designed experiment to justify the claim.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This research proposes new siamese network for the change detection problem as a binary semantic segmentation task and learns to extract features from the image pairs directly. Furthermore, authors collected 862 labeled image pairs for change detection. The research is well structured, but some parts should be explained/corrected.
Section 2 - Related works - I think that other siamese neural networks from the previous research should be mentioned in Related works and discussed in Discussion section. Such as:
Chen, H., Wu, C., Du, B., Zhang, L., & Wang, L. (2019). Change detection in multisource VHR images via deep siamese convolutional multiple-layers recurrent neural network. IEEE Transactions on Geoscience and Remote Sensing, 58(4), 2848-2864. - which you already mentioned before and
Zhan, Y., Fu, K., Yan, M., Sun, X., Wang, H., & Qiu, X. (2017). Change detection based on deep siamese convolutional network for optical aerial images. IEEE Geoscience and Remote Sensing Letters, 14(10), 1845-1849.
Figure 2 - The part of image is missing.
Lines 276-277 - Why didn't you design your network for 480x480 size with adding additional layer.
The 480x480 input image is then unnecessary and overkill for network - you could do your research with 240x240.
And the final result will always be better with lower resolution output than input - because you down-sample original image.
Section 4.7.4 (Comparison with other segmentation models) - This segment should be strengthen. "Why" is the most important question which should be answered.
Section 5 - The conclusion should be strengthen with more concise conclusion. The conclusion written in this way is too general.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
The authors propose a Change Detection dataset including changes in urban construction-related areas. The authors also propose a new model based on a Siamese Neural Network in the feautre extractor subsystem.
The new model is also compared to other existing models on the presented datasets by using different methods. An ablation study is also performed to investigate the need for the different model sublocks.
Methdos, results and discussions are well detailed. The quality of presentation is good (except for Figure 2 that does not fit well in the page).
However, since the model presented is compared only on the if dataset, this work deserves for a publication only if the authors plans to release the dataset (and code, maybe) enabling future research on the topic.
Otherwise, it would affect the reproducibility of the work.
In this case, please add a link to the paper where the dataset can be found.
Author Response
We would like to express our gratitude to you for your clear remarks and valuable suggestions.
Point 1: Since the model presented is compared only on the if dataset, this work deserves for a publication only if the authors plans to release the dataset (and code, maybe) enabling future research on the topic. Otherwise, it would affect the reproducibility of the work. In this case, please add a link to the paper where the dataset can be found.
Response 1: Thanks for your comment. We are preparing to make our code and dataset open at https://github.com/yangle15/Deep-Siamese-Networks-based-Change-Detection.
Reviewer 4 Report
--Authors present a semantic segmentation-based method for change detection using the siamese network concept. An extensive ablation study in terms of various factors like network architecture, backbone structure with the comparison with various other possible variations, the proposed method is shown to be logical. The results presentation and conclusion are presented clearly and correspond to the original hypotheses.
--Comments:
-- In introduction, what does line 54-55 ( Although...) mean?
--Collecting large amount of paired images is good, if it is done. However, author say previous methods use to rely on smaller no. of paired images. In deep learning designing the network N/Ws requiring the lower no. of images to train is taken as the contribution. So this concept should be considered in the paper while bragging about the dataset, in relation to the proposed network's capability.
--In line 45-46, the it is noted that the existing paired optical images are dependent largely on spectral information influenced by weather, sunlight etc., and is posed as a limitation of existing dataset. What measures author use to eradicate this limitation?
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The article is much better
Reviewer 3 Report
The authors added a link to the dataset to the manuscript as suggested. The repository is not accessible for me. However, this might be due to the fact the authors have still some work to do to release the dataset, and it is still private. Given that and the others improvements, I propose the acceptance of the work in this form.
