CLIFF: A Multi-Modal Remote Sensing Model for Geological Hazard Monitoring Based on Bitemporal UAV Images

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a geological hazard monitoring remote sensing model named CLIFF, which integrates the multi-modal large vision-language model CLIP, the change detection network Bitemporal Image Transformer and the classification network EfficientNet. Corresponding datasets and model fine-tuning strategies are also designed in this work.
Overall, the research task and proposed method are interesting. The detailed revision suggestions are listed as follows:
1. This method adopts large vision-language models for feature extraction. Please illustrate whether the model complexity increases significantly after introducing such modules.
2. The framework combines features derived from large vision-language models and learnable features. It is necessary to explain the rationality of this hybrid structure and verify that it achieves better performance than single feature utilization schemes.
3. This work applies pre-trained models to change detection and classification tasks. Numerous existing studies have adopted pre-trained models to extract features or conduct fine-tuning for various visual tasks and achieved promising performance. Relevant literatures should be introduced to clarify the differences and unique characteristics of the pre-trained model application pattern in this paper.
-Vision-language model for object detection and segmentation: A review and evaluation
-Leveraging vision-language models for improving domain generalization in image classification
-M2ist: Multi-modal interactive side-tuning for efficient referring expression comprehension
-Mapper: Multimodal prior-guided parameter efficient tuning for referring expression comprehension
-Taskclip: Extend large vision-language model for task oriented object detection
-VLDadaptor: Domain adaptive object detection with vision-language model distillation
-Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation

4. According to the framework shown in Figure 2, classification is implemented after change region detection. Please discuss the advantages of this classification strategy based on detected change areas compared with direct classification methods.
5. More experimental results are recommended to be supplemented, including the visualization outcomes of change detection and hazard classification.

Author Response

Comment 1: [This method adopts large vision-language models for feature extraction. Please illustrate whether the model complexity increases significantly after introducing such modules.]

Response 1: [We thank the reviewer for this valuable comment. We have added a discussion on model complexity in Section 7 (highlighted in blue). In brief, since the CLIP encoder is frozen during training and its features can be pre-computed and cached at inference, the additional overhead is limited and well justified by the accuracy improvement shown in Table 1.]

Comment 2: [The framework combines features derived from large vision-language models and learnable features. It is necessary to explain the rationality of this hybrid structure and verify that it achieves better performance than single feature utilization schemes.]

Response 2: [We thank the reviewer for this insightful comment. We have explained the rationality of the hybrid structure at the beginning of Section 5.3.1, and added an ablation study in Section 6.4.3 to verify its effectiveness. ]

Comment 3: [This work applies pre-trained models to change detection and classification tasks. Numerous existing studies have adopted pre-trained models to extract features or conduct fine-tuning for various visual tasks and achieved promising performance. Relevant literatures should be introduced to clarify the differences and unique characteristics of the pre-trained model application pattern in this paper.-Vision-language model for object detection and segmentation: A review and evaluation-Leveraging vision-language models for improving domain generalization in image classification-M2ist: Multi-modal interactive side-tuning for efficient referring expression comprehension-Mapper: Multimodal prior-guided parameter efficient tuning for referring expression comprehension-Taskclip: Extend large vision-language model for task oriented object detection-VLDadaptor: Domain adaptive object detection with vision-language model distillation-Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation]

Response 3: [We thank the reviewer for recommending these relevant works. We have added a new paragraph at the end of Section 3, which introduces all the recommended literatures, covering the review of VLMs for detection and segmentation, VLM-based domain generalization, parameter-efficient side-tuning, task-oriented detection, and VLM distillation. On this basis, we have clarified the differences of our application pattern: unlike these works targeting single-image detection, segmentation, or grounding tasks in natural scenes, this paper transfers a frozen CLIP encoder to bitemporal remote sensing change detection through a lightweight transfer fusion module, combined with a cascaded hazard-type classification stage under scarce hazard-labeled UAV data.]

Comment 4: [According to the framework shown in Figure 2, classification is implemented after change region detection. Please discuss the advantages of this classification strategy based on detected change areas compared with direct classification methods.]

Response 4: [We thank the reviewer for this question. We have added a discussion in Section 5.4. In brief, compared with direct whole-image classification, the detect-then-classify strategy decouples the two tasks under scarce hazard-labeled UAV data so that each stage can exploit complementary data sources, filters out the static background to reduce false alarms, effectively increases the resolution of small hazard regions by classifying the cropped detection boxes, and naturally provides the spatial location and extent of hazards, which direct classification cannot offer.]

Comment 5: [More experimental results are recommended to be supplemented, including the visualization outcomes of change detection and hazard classification.]

Response 5: [We thank the reviewer for this suggestion. For change detection, the visualization results are provided in Fig. 10, which presents the bitemporal images, the change mask images, and the corresponding statistics such as the number of changed buildings and the change percentage. For hazard classification, we have supplemented a new visualization in Section 6.4.2 (Fig. 11}, highlighted in blue), which presents the original UAV images and the recognized hazard regions of representative debris flow cases.]

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

(1) The proposed CLIFF framework is essentially a cascaded combination of CLIP, BiT, and EfficientNet. However, the manuscript does not sufficiently elaborate on the theoretical innovation underlying the collaboration mechanism among these components. In particular, the authors should provide a more rigorous explanation of how the Transfer Fusion Module (Section 4.3.1) mitigates the feature distribution discrepancy between CLIP and BiT representations.

(2) There might be an inconsistency in terminology throughout the manuscript. The Abstract refers to the “Transfer Fusion Module,” whereas Section 4.3.1 uses the term “Migration Fusion Module.”

(3) The references cited in the Introduction part are relatively outdated. It is recommended that the authors incorporate the following recent studies to provide up-to-date context:

[1] X. Fu, X. Song, P. Pace, G. Aloi and G. Fortino, "Reinforcement Learning-Based Low-Delay Data Collection in UAV-Assisted IoT for Secondary Geological Hazard Monitoring," in IEEE Transactions on Vehicular Technology, vol. 74, no. 8, pp. 12935-12950, Aug. 2025, doi: 10.1109/TVT.2025.3554815.

[2] Li, H., Lin, B. & Zhu, H. UAV Photogrammetry with Low Image Overlap for Panoramic Imaging and River Morphology Monitoring. Journal of Geovisualization & Spatial Analysis, 2025,9(2):34. https://doi.org/10.1007/s41651-025-00235-2

(4) Section 4.4 states that “identifying instances of differing regions through connected component analysis,” however, the manuscript does not explain how minor variations are handled. The authors should clarify whether small connected regions below a predefined pixel threshold are removed and provide the rationale for the selected threshold. e.g., The authors should clarify whether small connected regions below a predefined pixel threshold are removed and provide the rationale for the selected threshold.

(5) In Section 5.3, the comparison model “CDNet” is cited as Reference [43]. However, the cited reference appears to correspond to a Content Delivery Network Emulator, which is unrelated to change detection. The authors should verify the citation and correct it if necessary.

(6) Since the proposed framework is intended for real-time UAV monitoring applications, the manuscript should report engineering performance metrics such as inference time per image pair, model parameter count, FLOPs, GPU memory consumption, and hardware configuration. Without these metrics, it is difficult to assess the practical dependability of the proposed system.

(7) Additional ablation studies are recommended to better validate the contribution of each component. For example, experiments such as “Without CLIP,” “Without Transfer Fusion Module,” and “Alternative Feature Extraction Networks” would provide a clearer understanding of the effectiveness of the proposed architecture.

(8) The manuscript does not specify whether all comparison methods were evaluated using the same image registration and alignment preprocessing procedures adopted by CLIFF. To ensure a fair comparison, the authors should clarify these things.

(9) Section 4.1 employs the ORB algorithm for keypoint matching and sets the FAST corner detection threshold to t = 40. However, no justification is provided for this parameter choice.

Author Response

Comments 1: [The proposed CLIFF framework is essentially a cascaded combination of CLIP, BiT, and EfficientNet. However, the manuscript does not sufficiently elaborate on the theoretical innovation underlying the collaboration mechanism among these components. In particular, the authors should provide a more rigorous explanation of how the Transfer Fusion Module (Section 4.3.1) mitigates the feature distribution discrepancy between CLIP and BiT representations.]

 

Response 1:[We thank the reviewer for this insightful comment. Regarding the collaboration mechanism, we have summarized the original designs beyond a simple integration of existing components in the Discussion (Section 7), and distinguished our pre-trained model application pattern from existing VLM transfer works in Section 3. Regarding the transfer fusion module, we have added an explanation in Section 5.3.1 on how it mitigates the feature distribution discrepancy between CLIP and BIT representations.]

 

Comments 2: [There might be an inconsistency in terminology throughout the manuscript. The Abstract refers to the “Transfer Fusion Module,” whereas Section 4.3.1 uses the term “Migration Fusion Module.”]

 

Response 2:[We thank the reviewer for pointing out this inconsistency. We have unified the terminology as "transfer fusion module" throughout the revised manuscript, including the Abstract, Section 5.3.1, all other related descriptions in the text, and the corresponding figures (Figs. 2 and 7), which have been redrawn accordingly.]

 

Comments 3: [The references cited in the Introduction part are relatively outdated. It is recommended that the authors incorporate the following recent studies to provide up-to-date context: [1] X. Fu, X. Song, P. Pace, G. Aloi and G. Fortino, "Reinforcement Learning-Based Low-Delay Data Collection in UAV-Assisted IoT for Secondary Geological Hazard Monitoring," in IEEE Transactions on Vehicular Technology, vol. 74, no. 8, pp. 12935-12950, Aug. 2025, doi: 10.1109/TVT.2025.3554815. [2] Li, H., Lin, B. & Zhu, H. UAV Photogrammetry with Low Image Overlap for Panoramic Imaging and River Morphology Monitoring. Journal of Geovisualization & Spatial Analysis, 2025,9(2):34. https://doi.org/10.1007/s41651-025-00235-2]

 

Response 3:[We thank the reviewer for recommending these recent studies. We have incorporated both references into the Introduction of the revised manuscript to provide up-to-date context on UAV-based geological hazard monitoring.]

 

Comment 4: [Section 4.4 states that “identifying instances of differing regions through connected component analysis,” however, the manuscript does not explain how minor variations are handled. The authors should clarify whether small connected regions below a predefined pixel threshold are removed and provide the rationale for the selected threshold. e.g., The authors should clarify whether small connected regions below a predefined pixel threshold are removed and provide the rationale for the selected threshold.]

 

Response 4:[We thank the reviewer for this careful observation. Yes, small connected regions below a predefined pixel threshold are removed, and we have clarified this along with the rationale in Section 5.4. Such small regions typically originate from residual registration errors, vegetation sway, and illumination differences between the bitemporal images, rather than from actual geological hazards, whereas the hazards of practical concern along pipelines occupy considerably larger ground areas. The threshold is therefore set according to the spatial resolution of the UAV images, so that noise-level variations are suppressed while hazard-scale regions are fully preserved.]

 

Comment 5: [In Section 5.3, the comparison model “CDNet” is cited as Reference [43]. However, the cited reference appears to correspond to a Content Delivery Network Emulator, which is unrelated to change detection. The authors should verify the citation and correct it if necessary.]

 

Response 5:[We thank the reviewer for catching this citation error. The reference has been corrected. We have also unified the model name as "LightCDNet" throughout the manuscript, including the Abstract, Sections 6.3 and 6.4.1, Table 1, and the Conclusion.]

 

Comment 6: [Since the proposed framework is intended for real-time UAV monitoring applications, the manuscript should report engineering performance metrics such as inference time per image pair, model parameter count, FLOPs, GPU memory consumption, and hardware configuration. Without these metrics, it is difficult to assess the practical dependability of the proposed system.]

 

Response 6: [We thank the reviewer for this suggestion. We have supplemented the model parameter count, FLOPs, and hardware configuration in the Experimental Setup (Section 6.3) and Table 3. Wall-clock inference time and memory consumption depend on the specific deployment hardware; given the near-real-time nature of UAV pipeline inspection, we will report deployment-specific measurements in our future work.]

 

Comments 7: [Additional ablation studies are recommended to better validate the contribution of each component. For example, experiments such as “Without CLIP,” “Without Transfer Fusion Module,” and “Alternative Feature Extraction Networks” would provide a clearer understanding of the effectiveness of the proposed architecture.]

 

Response 7: [We thank the reviewer for this insightful comment. We have explained the rationality of the hybrid structure at the beginning of Section 5.3.1, and added an ablation study in Section 6.4.3 to verify its effectiveness. ]

 

Comments 8: [The manuscript does not specify whether all comparison methods were evaluated using the same image registration and alignment preprocessing procedures adopted by CLIFF. To ensure a fair comparison, the authors should clarify these things.]

 

Response 8:[We thank the reviewer for raising this important point about fairness. Yes, all comparison methods were evaluated under exactly the same preprocessing: the key point-based registration and alignment are applied as a dataset-level preprocessing step, and the identical aligned dataset with the same data split is used for CLIFF and all baselines in Table 1. We have clarified this in Section 6.3.]

 

Comments 9: [Section 4.1 employs the ORB algorithm for keypoint matching and sets the FAST corner detection threshold to t = 40. However, no justification is provided for this parameter choice.]

 

Response 9:[We thank the reviewer for pointing this out. We have added the justification for this parameter choice in Section 5.1.2: t = 40 balances the quality and quantity of the detected corners, as a smaller threshold produces excessive spurious key points from noise and vegetation textures in UAV images, while a larger one retains too few corners in weakly textured regions and destabilizes the affine estimation. This value also lies within the commonly adopted range of the FAST detector in practice.]

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors should clearly mention their affiliation, with name of their department and lab as per journals standard guidelines.

In abstract, the authors should need to improve abstract by clearly highlighting shortcoming of single date deep learning methods and how the new proposed method CLIFF is significant in terms of other accuracy metrics.

In title or abstract, the authors should clarify why they propose CLIFF name for new methods. Is it acronym of some word. Although authors clarify it in introduction section.

In the last paragraph of introduction, the authors clearly mention the main problem and gap and what are the main objectives of current study.

I would suggest to change organization of paper:

The authors randomly placed figures without mentioning in text.

Lines 87-98 should be part of results section, or otherwise line 87-91 can be part of methodology.

Figure 1 should be part of study area and results section, it should not placed under introduction section.

Section 2 and 3 about literature review should be part of introduction.

Section 4 should be written as Materials and Methods, where authors clearly describe the datasets and CLIFF was formed by combining different methods.

Section 5 should be written as results and discussion and authors clearly write main findings of CLIFF and their significance by relating to previous methods in discussion.

The discussion section is weak and authors need to expand it by citing similar studies in similar context.

Author Response

Comment 1:[The authors should clearly mention their affiliation, with name of their department and lab as per journals standard guidelines.]

 

Response 1:[We thank the reviewer for this reminder. We have revised the author affiliations following the journal's standard format.]

 

Comment 2:[In abstract, the authors should need to improve abstract by clearly highlighting shortcoming of single date deep learning methods and how the new proposed method CLIFF is significant in terms of other accuracy metrics.]

 

Response 2:[We thank the reviewer for this suggestion. We have revised the Abstract accordingly. First, we have clearly highlighted the shortcoming of single-date deep learning methods: they can only analyze the static appearance of a scene, and thus can neither distinguish newly occurring hazards from pre-existing similar land covers nor capture the dynamic evolution of hazards, which motivates the bitemporal design of CLIFF. Second, we have enriched the reported results with additional accuracy metrics, including the F1 score of 0.8689 for change detection and the per-class accuracies of 87.32% and 86.17% for crude oil spills and landslides, respectively.]

 

Comment 3:[In title or abstract, the authors should clarify why they propose CLIFF name for new methods. Is it acronym of some word. Although authors clarify it in introduction section.]

 

Response 3: [We thank the reviewer for this suggestion. Yes, CLIFF is an acronym derived from the three core components of the proposed framework: CLIP, BIT, and EfficientNet. We have clarified this in the Abstract and at the first occurrence of CLIFF in the Introduction, where the acronym is spelled out as CLIFF (CLIP-BIT-EfficientNet) with the constituent letters underlined.]

 

Comment 4: [In the last paragraph of introduction, the authors clearly mention the main problem and gap and what are the main objectives of current study.]

 

Response 4: [We thank the reviewer for this suggestion. We have added a dedicated paragraph at the end of the Introduction, right before the contribution list, which explicitly states the main problem, the research gap, and the three main objectives of this study.]

 

Comment 5:[I would suggest to change organization of paper: The authors randomly placed figures without mentioning in text.

Lines 87-98 should be part of results section, or otherwise line 87-91 can be part of methodology.

Figure 1 should be part of study area and results section, it should not placed under introduction section.

Section 2 and 3 about literature review should be part of introduction.

Section 4 should be written as Materials and Methods, where authors clearly describe the datasets and CLIFF was formed by combining different methods.

Section 5 should be written as results and discussion and authors clearly write main findings of CLIFF and their significance by relating to previous methods in discussion.]

 

Response 5: [We thank the reviewer for these constructive suggestions. We have relocated all figures close to their first mention in the text, in particular, Fig. 1 has been moved to Section 4.1 where the study area is described, and Fig. 2 to Section 4.2 where the overall structure is introduced. We have also removed the use-case paragraph and the specific experimental numbers from the Introduction, and strengthened the Discussion with a new paragraph relating the main findings of CLIFF to previous methods. Regarding the section organization, since the Related Works section was substantially expanded during this revision at the reviewers' request, and the current structure follows a top-down presentation that is also common in recent articles of this journal, we respectfully prefer to retain the stand-alone Related Works and methodology sections. We hope the reviewer finds this arrangement acceptable.]

 

Comment 6:[The discussion section is weak and authors need to expand it by citing similar studies in similar context.]

 

Response 6:[We thank the reviewer for this suggestion. The Discussion Section 7 has been substantially expanded in this revision.]

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Dear authors,

I would like to share with you, some elements that I consider relevant about your manuscript, hoping to be helpful.

First, you indicated that there are so many external factors affecting the bitemporal UAV images, but I suppose you can trace a flight plan, 
and take aerial photos at the same place, heigth and attitude. This to avoid the perform geometric correction and image alignment. 

Second, you applied a sorting process to the all datasets (5) used, involving delete duplicate, incomplete, or poor-quality images. I would like to know how this affected each dataset, I mean in terms of % or sizes. Also, you over-sample the images in the dataset to balance their category distribution. Again, how this affected each dataset? %? sizes?

Third, you convert image data into a unified file format (such as JPEG, PNG), which one? what color space do you choose?. Please, be specific about these.

Fourth, you adjust all images to the same resolution, what final resolution was that?. Also, on interpolation methods used for scaling (bilinear interpolation or nearest neighbor interpolation), which one did you applied?

Fifth, you normalized the pixel values of the images, for accelerating model training and improving stability; which I think is plausible.

Sixth, you used data enhancement techniques, such as random rotation, flipping, cropping, and color jittering, how much? %? same for all datasets?

Seventh, why you did not compare your model's performance against the proposed models for other authors? ex. FC-Siam-diff-PA?. Also, you choose models such as STANet, which was proposed for predict transient variations and cumulative trends embedded in image sequences; TINYCD is a model smaller than the compared change detection models, and less computational complexity, no just performance.

Lastly, the core modules of the CLIFF model consist of the multi-modal LVLM, CLIP, the change detection network, BIT, and the classification network, EfficientNet. Specifically, CLIP is used to extract features, BIT is used to locate geological disasters, and EfficientNet is used to determine the type of geological hazards. So, your model add some existing methods for this particular type of change, so in particular, I would like to know what original part was developep by you, especifically.

Just a minor mistake in the references section, the 22 and 28 are the same.

That is all.

Best regards, Jaime

Author Response

Comment 1: [First, you indicated that there are so many external factors affecting the bitemporal UAV images, but I suppose you can trace a flight plan,

and take aerial photos at the same place, heigth and attitude. This to avoid the perform geometric correction and image alignment.]

 

Response 1: [We thank the reviewer for this practical suggestion, and we would like to clarify that pre-defined flight plans with fixed waypoints, heights, and attitudes are indeed adopted in our data collection, as the reviewer suggested. However, residual errors from the onboard positioning system, wind-induced attitude jitter at the moment of capture, and altitude drift across weather conditions inevitably translate into slight offsets, rotations, and scale differences between the bitemporal images, to which pixel-level change detection is sensitive. The lightweight key point-based alignment is therefore retained as a low-cost safeguard against these residual errors. We have clarified this in Section 5.1.]

 

Comment 2: [Second, you applied a sorting process to the all datasets (5) used, involving delete duplicate, incomplete, or poor-quality images. I would like to know how this affected each dataset, I mean in terms of % or sizes. Also, you over-sample the images in the dataset to balance their category distribution. Again, how this affected each dataset? %? sizes?]

 

Response 2: [We thank the reviewer for this clarification request. The five public change detection datasets are used with their official splits without modification and without oversampling. The sorting process applies to the self-built classification dataset: approximately 252,000 raw images were collected, and 209,154 (about 83%) were retained after cleaning. Fig. 9 shows the cleaned distribution before oversampling, which is applied only to the training split. These details have been added to Section 6.2.]

 

Comment 3: [Third, you convert image data into a unified file format (such as JPEG, PNG), which one? what color space do you choose?. Please, be specific about these.]

Response 3: [We thank the reviewer for pointing out this ambiguity, and we have specified it in Section 6.2. All optical images are converted to the PNG format (lossless compression) in the 8-bit three-channel RGB color space; grayscale images are replicated across channels, and alpha channels are discarded. The annotation masks are saved as single-channel binary PNG files, where 0 denotes background/unchanged and 255 denotes changed/hazard regions.]

 

Comment 4: [Fourth, you adjust all images to the same resolution, what final resolution was that?. Also, on interpolation methods used for scaling (bilinear interpolation or nearest neighbor interpolation), which one did you applied?]

 

Response 4: [We thank the reviewer for pointing out the ambiguity. We have specified this in Section 6.2: all images are converted to the PNG format (lossless compression), and the color space is unified to the 8-bit three-channel RGB space, where grayscale images are replicated across channels and alpha channels are discarded.]

 

Comment 5: [Fifth, you normalized the pixel values of the images, for accelerating model training and improving stability; which I think is plausible.]

 

Response 5: [ We thank the reviewer for this positive comment on the normalization step.]

 

Comment 6: [Sixth, you used data enhancement techniques, such as random rotation, flipping, cropping, and color jittering, how much? %? same for all datasets?]

 

Response 6: [We thank the reviewer for requesting the exact settings, which have been added to Section 6.2. Augmentation is applied to the training set only: random horizontal and vertical flipping (each with probability 0.5), random rotation by 0/90/180/270 degrees (probability 0.5), random cropping to the training patch size, and color jittering (probability 0.3, brightness/contrast/saturation within ±20%) for RGB images only. The same geometric transformations are applied synchronously to the two temporal images and the corresponding mask, with nearest-neighbor interpolation preserving the binary labels. The same strategy is used across all datasets.]

 

Comment 7: [Seventh, why you did not compare your model's performance against the proposed models for other authors? ex. FC-Siam-diff-PA?. Also, you choose models such as STANet, which was proposed for predict transient variations and cumulative trends embedded in image sequences; TINYCD is a model smaller than the compared change detection models, and less computational complexity, no just performance.]

 

Response 7: [We thank the reviewer for these careful observations. (1) Table 1 already compares CLIFF with nine methods from other authors; the entry previously labeled "FC-EF" corresponds to the pyramid-attention model of Li and Huo, to which FC-Siam-diff-PA belongs, and we have corrected the entry name accordingly. (2) We thank the reviewer for catching the STANet citation error, it has been corrected to the spatial-temporal attention network of Chen and Shi. (3) We agree on TinyCD's efficiency-oriented positioning and have added a clarification on the baseline coverage in Section 6.3.]

 

Comment 8: [Lastly, the core modules of the CLIFF model consist of the multi-modal LVLM, CLIP, the change detection network, BIT, and the classification network, EfficientNet. Specifically, CLIP is used to extract features, BIT is used to locate geological disasters, and EfficientNet is used to determine the type of geological hazards. So, your model add some existing methods for this particular type of change, so in particular, I would like to know what original part was developep by you, especifically.]

 

Response 8: [We thank the reviewer for this question, which is also addressed in the revised Section 7. The original parts developed by us are: (1) the transfer fusion module bridging the frozen CLIP and BIT feature spaces (Section 5.3.1; supported by the ablation in Section 6.4.3); (2) the fine-tuning strategy for scarce hazard-labeled data; (3) the multi-source bitemporal UAV and hazard classification datasets for pipeline monitoring in the Sichuan Basin; and (4) the end-to-end monitoring pipeline validated on a real-world use case. The pre-trained components are intentionally reused, whereas the way they are bridged, fine-tuned, and deployed constitutes our contribution.]

 

Comment 9: [Just a minor mistake in the references section, the 22 and 28 are the same.]

 

Response 9: [We thank the reviewer for catching this. The duplicated reference (the BIT paper appearing twice) has been merged into a single entry, and all citations now point to the same reference.]

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

My concerns have been addressed.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have significantly improved manuscript by incorporating my comments.

Reviewer 4 Report

Comments and Suggestions for Authors

Dear authors, 

I consider your manuscript's update took into account the main elements that I highlight during my revision, so I am glad you made clearer some assumptions for the reader.

I just want to congratulate you for your research, and recommend it for publish.

A minor thing, the fourth point was not answered really, I think you just copied the same answer from third point. I was meaning about pixel spatial resolution, not radiometric resolution, but I think you will take note on this.

Best regards,