A New Change Detection Method for Heterogeneous Remote Sensing Images Via an Automatic Differentiable Adversarial Search
, Jie Chen 1, Hongcheng Zeng 3, Wei Yang 3, Jie Chen 3, Zhixiang Huang 1 and Long Sun 4Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a heterogeneous remote sensing image change detection method based on the differentiable adversarial search (DFCS) strategy. By designing an adaptive discriminator and a Gabor and local normalized cross-correlation (G-LNCC) feature fusion module, and combining a geometry structure-based collaborative supervision (GSCS) loss function, it achieves efficient detection of changes in heterogeneous images. Although the manuscript demonstrates the significant advantages of this method, there are still some areas that need improvement, especially in terms of the method's innovativeness and the rationality of the experimental setup.
1. The authors only mention the application of NAS in other fields. So, what are its unique advantages in the Hete-CD field or the differences compared to existing methods (such as fixed-architecture methods based on GANs)?
2. The "temperature annealing mechanism" mentioned in Equation (3) is suggested to be further elaborated in the text regarding its function and specific implementation.
3. The introduction mentions that existing methods have problems of "blurred change boundaries" and "structural distortions" in heterogeneous image change detection, but these specific manifestations in existing methods are not clearly demonstrated in the experimental section (Section 4).
4. Dynamically combining filters may lead to instability in the training process. Since each convolutional kernel is composed of a linear combination of multiple basic filters, this approach seems to introduce excessive parameter redundancy.
5. The process of dynamically combining filters may make it difficult for the network to converge during training because this combination increases the complexity of the network. A theoretically rigorous proof is suggested to be provided.
6. Gabor filters are mainly used for texture analysis, while LNCC is used for measuring local structural similarity. How can it be ensured that these two types of features maintain consistent scale and orientation when fused?
7. Pay attention to formatting issues. The "where" following a formula should be indented.
8. It is suggested to increase the proportion of literature from the past three years.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsFor detailed reviews, please refer to the attachment.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsPlease see the attached file.
Comments for author File: Comments.pdf
The English language is fine.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe authors propose a differentiable neural architecture search (DFCS) method for heterogeneous remote-sensing change detection. By dynamically constructing the discriminator, front-end fusion of Gabor and LNCC, and a multi-level structural supervision loss (GSCS), the approach achieves state-of-the-art accuracy on five public datasets. Overall, the paper is well written and innovative. Before it can be considered for acceptance, however, the following issues must be addressed:
1. What is the role of the Gabor filter bank and what are its advantages for extracting features from heterogeneous images?
2. There should be no space before "where" in any equation.
3. Several figures have low resolution; vector graphics are recommended.
4. How is λ_struct in Equation (7) set?
5. More recent unsupervised heterogeneous change-detection baselines (e.g., RIEM, SDIR) should be included for comparison.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe author has carefully revised the existing issues.
Comments on the Quality of English LanguageThe English expression could be further strengthened.
Author Response
We sincerely appreciate your constructive comments and your explicit recognition of our previous revision efforts. In strict accordance with your suggestion, we have carefully addressed the remaining concern regarding language quality, dedicating this round of revision to further strengthening the English expression throughout the manuscript.
Response to Comments on the Quality of English Language
Point : The English expression could be further strengthened.
Response : We sincerely appreciate your constructive feedback on the English expression. In response, we have conducted a comprehensive line-by-line review of the entire text to further enhance linguistic clarity. We have meticulously corrected grammatical errors, optimized sentence structures for improved readability, and refined vocabulary to ensure a more rigorous academic tone. Additionally, we have verified consistency in terminology and formatting throughout the text. We hope these comprehensive revisions will significantly enhance the manuscript's readability and quality, bringing it up to publication standards.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have revised the manuscript according to my comments. But there are several problems should be solved:
1.Although the paper mentions the use of five-fold cross-validation, it does not elaborate on the specific randomness control employed in data partitioning nor clarify whether the same partitioning was applied across all comparison methods. This may compromise the comparability and reproducibility of the results.
2.Although the “Conclusions” section mentions slightly lower accuracy on the Gloucester I dataset, it does not further analyze whether this relates to the quality of the data itself, annotation consistency, or scene complexity. Nor does it discuss the method's applicability under extreme modal differences, such as optical-infrared.
3.This is one of the few authors I've encountered who doesn't prepare responses according to the official template.
I recomended accept after minor revision.
Author Response
Thank you very much for your positive recommendation and for giving us the opportunity to further refine our manuscript. We sincerely appreciate your constructive comments regarding the experimental details and the discussion of limitations. We would like to express our sincere apologies for not using the official response template in the previous submission. We have now strictly followed the journal's requirements and provided a point-by-point response below. Additionally, we have clarified the randomness control in our cross-validation and expanded the discussion on the method's performance on the Gloucester I dataset and its applicability to extreme modal differences. All changes have been marked in the re-submitted files.
Comments 1: Although the paper mentions the use of five-fold cross-validation, it does not elaborate on the specific randomness control employed in data partitioning nor clarify whether the same partitioning was applied across all comparison methods. This may compromise the comparability and reproducibility of the results.
Response 1: Thank you for pointing out this crucial detail regarding the experimental setup. We fully agree that ensuring reproducibility and fairness is essential for validating our method. In Section 4.1.2 (L648-L654) of the revised manuscript, we have explicitly clarified the data partitioning process to address your concern. Specifically, we utilized a fixed random seed to generate the data split indices, and importantly, we applied these identical indices across all comparison methods. This rigorous control guarantees that all models were trained and evaluated on the exact same subsets for every fold, ensuring strictly fair comparisons and the reproducibility of our results. We have added these specific descriptions in the experimental setup section of the revised manuscript.
Comments 2: Although the “Conclusions” section mentions slightly lower accuracy on the Gloucester I dataset, it does not further analyze whether this relates to the quality of the data itself, annotation consistency, or scene complexity. Nor does it discuss the method's applicability under extreme modal differences, such as optical-infrared.
Response 2: We greatly appreciate your pointing out the lack of further discussion on the Gloucester I dataset in the “Conclusions” section and have expanded the Section 5 (L1033-L1049) to provide a deeper analysis of these issues. regarding the performance on the Gloucester I dataset, we have clarified that the slightly lower accuracy is primarily attributed to the inherent annotation inconsistencies and severe geometric distortions present in the historical map data, rather than scene complexity alone. These factors introduce label noise that challenges fine-grained boundary detection. Furthermore, regarding the applicability to extreme modal differences such as optical-infrared scenarios, we have added a discussion acknowledging that while our G-LNCC module effectively handles structural disparities, the fundamental physical differences between thermal footprints and visual textures pose unique challenges. We have noted that while our method serves as a robust baseline, incorporating modality-specific thermal feature extraction could further enhance performance in such extreme cases.
Comments 3: This is one of the few authors I've encountered who doesn't prepare responses according to the official template.
Response 3: We offer our sincerest apologies for this oversight in our previous submission. We deeply respect the rigorous standards of the journal and the reviewers' time. For this revision, we have strictly adhered to the official template to prepare this response, ensuring that the layout and organization meet the required standards to facilitate your review. We appreciate your patience and for pointing this out. We assure you that we will pay strict attention to such requirements in future submissions and guarantee that this will not happen again.
