RACDNet: Resolution- and Alignment-Aware Change Detection Network for Optical Remote Sensing Imagery

Round 1

Reviewer 1 Report

In this paper, a resolution- and alignment-aware change detection network (RACDNet) is proposed for multi-resolution optical remote-sensing imagery change detection. Experimental results show the good performance of the proposed method. However, some issues should be addressed. 

Major issues: 

1) In the third part, the description of the proposed method is too simple. For example, the authors said the super-resolution network is improved, but the interpretability of this improvement is not clear. Why do the authors do this? And I think the super-resolution model proposed by the authors is not very novel. This structure has been proposed for a long time. I suggest the authors further explain the reason for this design including super-resolution part and change detection part. 

2) The introduction is too simple. This paper mainly discusses the change detection problem in optical image. In order to more comprehensively describe the related works, some other types of modes have also been proposed, such as spectral image and lidar image. It is suggested that the authors add the description of related work, such as 

[1] Super-Resolution Mapping Based on Spatial-Spectral Correlation for Spectral Imagery [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(3): 2256-2268. 

[2] Adaptive Target Profile Acquiring Method for Photon Counting 3-D Imaging Lidar[J]. IEEE Photonics Journal, 2016, 8(6): 6805510. 

[3] Improving Pixel-Based Change Detection Accuracy Using an Object-Based Approach in Multitemporal SAR Flood Images [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2015, 9(7): 3486 - 3496 

[4] A Binning Approach to Quickest Change Detection With Unknown Post-Change Distribution[J]. IEEE Transactions on Signal Processing, 2019, 67(3): 609 - 621 

In addition, the paper structure should also be given in detail in the introduction. 

3) The experimental part needs to be improved. In order to further prove the universality of the proposed method, can the authors add the updated super-resolution methods and change detection methods? The current comparison method is not very novel. It is also suggested that the authors add the evaluation index of operation time. 

Minor issues: 

1) The variables of some formulas are not explained. For example, in formula (3). 

2) The format of the paper also needs further improvement. For example, the sentence position after Figure 7 and Figure 6 is inappropriate. In addition, there are some grammatical errors in the article, which need further careful proofreading. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors of this manuscript propose a resolution-aware change detection and alignment network (RACDNet) for multi-resolution optical remote sensing images. In the first step, a lightweight super-resolution network is proposed to create high-quality bitemporal images, fully taking into account the complexity of the different regions, which facilitates detailed information recovery. Adversive losses and perceptual losses are used to improve visual quality. In the second step, deformable convolution blocks are embedded in the new Siamese-UNet architecture to align bitemporal deep features, which allows us to generate robust difference features to extract change information. The authors further use an atrose convolution module to extend the receptive field and an attention module to bridge the semantic gap between encoder and decoder. To test the effectiveness of our RACDNet, a new multi-resolution change detection dataset (MRCDD) was created using Google Earth. Quantitative and qualitative experimental results show that the author's RACDNet can significantly improve the details of the reconstructed images, and the CD performance outperforms other current methods by a wide margin. 

My comments are as follows

1) In formula (14), by which set is the summation performed?

2) In section 4.3 Training Details, why was this particular toolkit chosen? Were there any alternatives?

3) The authors analyze only the quantitative advantages in comparison with other methods. Were there any qualitative differences and advantages?

Overall, the manuscript is of pretty good quality. And a very detailed comparison with other methods is given. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thank the authors for detailed reply. I have no other questions.