Automated Mapping of Post-Storm Roof Damage Using Deep Learning and Aerial Imagery: A Case Study in the Caribbean

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. It is unclear whether the Mask2Former mentioned in the abstract is proposed by the authors or refers to prior work. Since the manuscript does not provide a description of this model, I recommend that the authors clarify this point and give a clear explanation.
2. The description of the dataset is currently too fragmented. In Section 2, nearly every subsection includes elements of data introduction and processing, which disrupts the overall structure of the manuscript. It is recommended that the authors consolidate the dataset description and processing into a dedicated subsection to improve clarity and organization.
3. OA and IoU are commonly used metrics in segmentation tasks. It is unclear why these metrics were not employed in the current work. These metrics can provide useful information on overall classification accuracy and the degree of overlap between predicted and ground truth regions, offering a more comprehensive assessment of segmentation performance.
4. Regarding the visualization of true positives, false positives, and false negatives, the current presentation appears to only display the predicted bounding boxes. A more informative approach would be to visualize both the ground truth and predicted results using masks, with different colors representing the three cases. This would provide a clearer and more intuitive understanding of the model’s performance.
5. At the beginning of Section 1.2, the sentence ‘while the other 42 performed localization of roof damage’ is presented without an accompanying reference. It is recommended that the authors provide the relevant citation to support this statement.
6. I have some concerns regarding the novelty of this manuscript, despite the research direction being highly practical and valuable. The paper primarily focuses on how multiple datasets are integrated, but it does not provide any analysis or improvement of algorithms, nor does it discuss potential limitations of existing methods. Addressing these aspects would strengthen the contribution of the work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Lines 34-35: Too many keywords, some of which are redundant or not highly relevant to the paper.

 

Line 89: The expression "10 RGB test images" is incorrect. Similar expressions appear throughout the paper and should be revised accordingly.

 

Lines 99-100: The highlights of this paper are unclear. The current description fails to convey the innovation and contributions of the work.

 

Figure 1: The test images are evaluated for accuracy by labeling or generating polygons. The flowchart includes both images and polygons, making it unclear how accuracy assessment is performed against predicted polygons.

 

Table 1: What is the unit for "Quantity"? Several tables in the paper include quantities, and the inconsistency between quantity, GSD, and area makes it impossible to calculate the covered area using quantity and GSD alone.

 

Lines 243-249: Polygon boundaries are resolution-independent, and their positions are unrelated to resolution. Why is it necessary to resample the images to 5 cm?

 

Lines 280-288: Internal inconsistencies in roof imagery, such as chimneys, may affect polygon recognition, training, and accuracy evaluation.

 

Reference 117 is an unpublished work by the author. The effectiveness and reliability of the method should be briefly introduced in the paper, as it is a critical support for the results.

 

Line 665: The paper analyzes the impact of positive sample diversity on results. Would the diversity of negative samples and the positive-negative sample ratio also affect the results?

 

Section 3.3: The analysis lacks a clear conclusion. The paper should explicitly state the findings of this section.

 

The resampling method uses cubic convolution interpolation, which blurs edges. Could this affect polygon extraction and accuracy evaluation?

 

Table 6: The accuracy of the method and the research results appear inconsistent with the data in Table 5。

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript submitted for evaluation addresses an issue of undoubted practical interest, namely the detection of damage to buildings (roofs) as a result of storms and hurricanes, based on the application of aerial image processing techniques (very high-resolution orthoimages generated from drones). In this regard, the topic is considered to be in line with the journal's themes and the current general interest of the scientific community in this subject. The fundamental contribution of the work lies not so much in an excessively novel methodological proposal as in its comprehensive application to real case studies. In any case, this application is considered to be a novel approach of general interest.

The manuscript is considered to be well structured and of appropriate content and length, perfectly suited to the needs of the work being evaluated. Furthermore, the quality of the figures and documents is also considered to be very good.

The presentation of the methodology applied (both in terms of methods—basically Mask2Former—and the input information used) is very interesting and sets out a framework that is considered to be fully reproducible. Similarly, the application is considered to be correct (and scientifically and technically rigorous) and the results and conclusions consistent with the work carried out. This fact, derived from the practical application to a real case study, significantly marks the interest of Remote Sensing readers.

All formal aspects are considered correct for publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf