Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article is interesting and holds significant potential.

 

The main subject area - pavement issues detection via UAV usage is worth attending and needs deeper investigation. Developing half-automatic methods via external programs and scripts is worth pursuing to save time compared to current manual detection techniques. However, using half-automatic methods is not entirely accurate compared to traditional detection, but it is still worth developing.

 

The authors' determination to consider many variables is commendable. This study's relevance to a crucial area is evident, and its findings could profoundly impact it, inspiring further research and development in the field.

 

Your use of up-to-date references. It not only adds credibility to your work but also shows your respect for the current state of research in the field. The article's simple, direct construction further enhances its readability, making it a strong point of this paper. Chronological review is worth appreciating.

 

Your article's thoughtful use of images and charts enhances readability and aids comprehension of complex concepts, making it a valuable contribution to the paper. Your efforts in this regard are appreciated and recognized.

 

The English language is acceptable. Easy to read and understand.

 

The article could be published in its existing form, but minor revisions are strongly expected.

 

 

Remarks/questions:

 

1. lines 19-25 - what i u-net and resnet34? You place a shortcut in the abstract. Readers will not understand you. Please try to fix it, and present this info more generally because it is not the time to place such info in the abstract... If you make it more general, it could provide that someone will understand what you mean and go deeper into your article. You can optionally add the sentence “These frameworks are named....”.

 

2. lines 22-26 + 27-34 - this part of information should be briefly explained in the methods section header.... placing info about frameworks threshold in abstract, creating mess what article is about, what was did , what fundamental results have you gathersd, and how can this info be usefull,

Please reconstruct the abstract to emphasize quickly what did you do, what the main results are, where can it be usefull/or what you will do next

 

3. line 84, which scenarios - give an example ;)

I am curious about two things. - Please answer me.

First - how your detection framework would be accessible for institutions - where to download it, use it, and combine with some aerial imagery gathered as an ortophotomap from software like Pix4D. Second: What is a limitation of your frameworks? What about rutting issues and other?

 

4. Figure 103 - good overview, tell me why you proposed the projection to WGS84? Will Your framework work at other projections -local one like EPSG 2180 for Poland conditions? Aerial mapping is typically done in WGS 84, but finally product is further applied in a local projection.

 

5. Lines 220-221 please explain more widely in the text why did use sigmoid function to decoder activation

 

6. lines 646-662 conclusion part - its too general. In this header, you are expected to discuss your results. Please extend this paragraph (You can use a list to simplify your thoughts). Write what the limitations of your frameworks, what your results show - how accurate it is, present some mathematical values which would create a precise remark/conclusion about your study like “our framework was better than coming by 30% so it can be used here and there (practical application)”

 

Generally - nice try and magnificent work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper presents a practical pipeline that turns crack segmentation masks into vector polygons and exports them to common GIS formats. The topic is relevant for inspection workflows, and the paper is easy to follow in parts. However, the current work does not provide evidence strong enough for publication, mainly because the geospatial part is not validated in a way that matches the paper’s own positioning, and because the end to end demonstration is built on evaluation steps that depend on ground truth in a way that cannot be reproduced at deployment time.

The technical problem is the geospatial mapping setup. The paper states that the original images do not include navigation metadata, so the image locations are emulated using synthetic waypoints from a road model, and the ground sampling distance is assumed rather than measured. The paper also notes that the resulting coordinates and GIS measurements are illustrative. In this situation, it is not possible to demonstrate accurate geospatial localization, because there is no real sensor based reference for position, scale, altitude, camera calibration, or pose error. The work therefore demonstrates a coordinate conversion procedure, not localization accuracy in a real UAV inspection setting.

A second problem is that the GIS output is produced only after filtering predictions using a ground truth dependent instance validation step. Only predicted instances matched to ground truth with IoU at least 0.50 are retained, and the paper reconstructs masks that contain only those validated pixels before extracting contours. This makes the final GIS layer look clean, but it is not a deployable pipeline because ground truth is not available at inference time. It also hides false positives, which are often the most important failure mode for an operational GIS layer.

There are also concerns about evaluation rigor and internal consistency. The reported instance level statistics are computed on the validation set, and it is not clear that final metrics are reported on the held out test set. The dataset is small at 400 images with a test split of 20 images, which makes strong conclusions fragile. In addition, some reported counts appear inconsistent with the stated split sizes, for example the number of images with at least one true positive is larger than the validation set size, which reduces confidence in the reported results.

The instance level evaluation design is also not well aligned with crack geometry. Matching is based on axis aligned bounding box IoU with one to one Hungarian assignment. For thin, elongated, and fragmented cracks, bounding box overlap can be misleading and can penalize reasonable predictions when fragmentation differs between annotation and prediction. The paper also applies morphological operations before component extraction, but key parameters and ablations are limited, so it is hard to separate what comes from the network versus heuristic post processing. Finally, the paper uses the term lightweight for the vector outputs, but it does not quantify size, vertex counts, or simplification trade offs, so that aspect remains unsupported.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The article is interesting and relevant from a practical point of view. It considers a system for visual detection of road cracks and geospatial mapping of road surface defects based on UAVs. However, I have a number of comments and responses to which will significantly improve the perception of the results obtained.

1.Justify why the ResNet-34 architecture was used and not the more modern ResNet-152 with convolutional layers with different resolutions.
2.In addition to the accuracy of crack detection, it is necessary to specify the time costs for making a decision.
 3.In your work, you perform image segmentation in a wide range of thresholds. In view of this, in the Introduction and References you need to mention Features extraction from multi-spectral remote sensing images based on multi-threshold binarization.Such multi-threshold binarization allows to significantly reduce the training time while ensuring high accuracy.

4.It is necessary to supplement how artifacts (affine transformations, changes in lighting and contrast, weather conditions) will affect the accuracy of crack detection.
5.Is ±15° sufficient?
6.Justify why the Dice metric was used instead of the F1 score.
7.Supplement References with articles that will enhance the relevance of the problem.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The revision moves in a better direction by limiting the scope and clarifying that the geospatial mapping is intended only as an illustrative component. However, the manuscript still contains internal inconsistencies that make the scope and the deployable pipeline unclear and therefore difficult to trust.

Figure 1 still presents an instance filtering step based on IoU with a threshold of 0.50 as part of the pipeline. This contradicts the textual explanation stating that the GIS output is generated from the complete set of predictions without any ground-truth-dependent selection.

Related to this issue, the use of wording such as _true positive mask_ in the illustrative pipeline section is problematic. Even if the intention is merely to show an example that happens to correspond to a correct detection, this phrasing implies that true-positive selection is an actual processing step within the pipeline. This part should be renamed and rewritten to make it explicit that the example is chosen purely for illustration and that the pipeline output itself does not depend on ground truth.

The manuscript also requires stronger consistency regarding post-processing. The paper states that the final configuration relies on raw thresholding without post-processing, yet the contour extraction section still applies a minimal area threshold to remove small connected components. Either the pipeline is genuinely raw, or it includes area filtering. If an area threshold is retained for geometric validity, this should be stated explicitly, aligned with the ablation terminology, and reflected consistently in the description of the final selected variant.

The added stability analysis is useful, but the conclusions drawn from it are currently overstated. The selected operating threshold varies substantially across different test set sizes, so the text should not suggest that the threshold itself is stable. Moreover, because all splits are generated using a fixed random seed, the analysis does not capture variability due to different data splits or training randomness. If the term _robust_ is to be used, the analysis should include multiple random seeds or repeated splits; otherwise, the conclusions should be softened to match the scope of what is actually evaluated.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

I am satisfied with the responses to my concerns. The changes and additions made have significantly improved the perception of the results obtained.

Author Response

We thank the reviewer for their positive feedback and for confirming that the revisions satisfactorily address their concerns. We are pleased that the changes and additions have improved the clarity and perception of the results.