Feature-Differencing-Based Self-Supervised Pre-Training for Land-Use/Land-Cover Change Detection in High-Resolution Remote Sensing Images

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

See the attached.

Comments for author File: Comments.pdf

Author Response

See the attached.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript entitled "Feature-Differencing-Based Barlow Twins Self-supervised Pretraining for Land-Use/Land-Cover Change Detection in High-Resolution Remote Sensing Images" aimed to presents a new framework termed Barlow self-supervised pre-training and fine-tuning CD (BTCD), which is based on feature-differencing. The proposed method utilizes absolute feature differences to directly acquire distinct representations linked to regions that have changed from unlabelled bi-temporal remote sensing images in a self supervised manner. In addition, it incorporated invariant prediction loss and change consistency regularization loss to improve the alignment of images taken at different times, considering both the decision and feature space during the training of the network.

 

The manuscript is interesting, well written and organized and it is comprehensive. In my opinion it is suitable to publish in "Land".

 

I have few comments.

 

1) Most of the experiments were applied on small areas, it will be interesting to see how the developed model will handle larger areas (10 km x 10 km) using fine spatial resolution data (less than 1 m)

 

2) I propose to apply your novel algorithm to sentinel-2 satellite data in two different periods (2015 and 2023) in an entire city to see how this model deal with medium spatial resolution data.

 

3) Discuss the limitation of your model and draw future directions.   

Author Response

See the attached.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

 

This paper introduces a novel feature-differencing-based framework called Barlow Twins self-supervised pre-training and fine-tuning change detection network. This framework incorporates absolute feature differences and novel loss functions to enhance image alignment and mitigate the impact of environmental variations. However, there are areas for improvement and certain issues that need addressing.

1.   Abstract. The problem description is rather redundant, it is suggested to compress the statement. It would be beneficial to place greater emphasis on elucidating the framework's core concepts and novel contributions.

2.   Figure. The two proposed losses, invariant prediction (IP) loss and change consistency regularization (CCR) loss, are not mentioned in the descriptions in Figure1 and 2.1 Materials and Methods.

3.   Page10, The manuscript mentions the superior performance of using self-attention mechanisms compared to other CNN-based methods. However, it would be beneficial to explain why UNet++ is chosen and why it outperforms these methods.

4.   Font colors are not displayed in Table1.

5.   The results of the second line of (j) in Figure5 indicate that your method is still affected by noise or other factors, which is not as good as (d) in this respect. Please analyze the reasons.

6.   Many improper expressions, please pay attention to the logic of the article.

 

 

Comments on the Quality of English Language

1.   Abstract. L15-L19: The description of the difficulty in obtaining annotated data from lines is redundant. The repetition of "However" creates redundancy in the text.

2.   Introduction. There are several problems with logical expression in the article, such as Page2, L44-L4, L49-50 and L58-59.

Author Response

See the attached.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank the authors for incorporating my comments in the revised manuscript. I do not have any comments on the required changes. However, I found that the authors used a new data set, WHU-BCD, in their experiment. When I read this part and recalling the Abstract, I wounder if there is any other work with quantitative performance measures. Again, I am not asking for any change but just wonder (suggestion for making the conclusion more convincing).

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript has been improved.