AlleyFloodNet: A Ground-Level Image Dataset for Rapid Flood Detection in Economically and Flood-Vulnerable Areas

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduces AlleyFloodNet, a ground-level image dataset designed to support rapid and accurate flood classification in high-risk urban environments. It is interesting and significant, but there are some questions should be answered or improved:

1. a description of the dataset should be added in the abstract;

2. there are too few quantitative indicators in the abstract, it is suggested to add some quantitative descriptions;

3. the introduction is less logical, and it is suggested to reorganise it; this part should include the brief description about the current state of research on the dataset and deep learning monitoring methods, and finally summarise the research content and main contributions of this paper;

4. the Related works section lacks literature analysis and summary, and it is suggested to organise it according to the two parts of datasets and methods;

5. "3.2 Models" section is too detailed in its description of the models, and it is recommended to simplify it;

6. lines 330 and 332, there are spelling mistakes in ‘astly’ and ‘n this study’;

7. I think there are serious problems with the organisation of the experiments in section 3.6. The fact that both models were tested with AlleyFloodNet does not prove anything. Because the two datasets are inherently different, both models trained separately using the two datasets would be more likely to accurately predict images that are more similar to the original data. In other words, if both models were evaluated with the FloodNet test data, the results would certainly differ from the existing results.

Author Response

Comments 1: a description of the dataset should be added in the abstract
Response 1: We agree with this comment. Therefore, We have clearly added a detailed description of the AlleyFloodNet dataset in the abstract. Specifically, We clarified that AlleyFloodNet consists of ground-level images collected from diverse global regions, capturing economically and flood-vulnerable urban areas under realistic conditions such as varying water levels, colors, and lighting conditions.

[Updated text in the manuscript: Abstract, Page 1, Lines 10–15]
'To address this, we constructed AlleyFloodNet, a dataset designed for rapid flood detection in flood-vulnerable urban areas, with ground-level images collected from diverse regions worldwide. In particular, this dataset includes data from flood-vulnerable urban areas under diverse realistic conditions, such as varying water levels, colors, and lighting.'

Comment 2: There are too few quantitative indicators in the abstract; it is suggested to add some quantitative descriptions
Response2: We agree with this comment. Therefore, We have provided specific quantitative indicators in the abstract. In particular, We included the performance metrics of the ConvNeXt-Large model fine-tuned on AlleyFloodNet, clearly specifying accuracy, precision, recall, and F1-score.

[Updated text in the manuscript: Abstract, Page 1, Lines 15–17]
'By fine-tuning several deep learning models on AlleyFloodNet, the ConvNeXt-Large model achieved excellent performance, with an accuracy of 96.56%, precision of 95.45%, recall of 97.67%, and an F1-score of 96.55%.'

Comments 3: the introduction is less logical, and it is suggested to reorganise it; this part should include the brief description about the current state of research on the dataset and deep learning monitoring methods, and finally summarise the research content and main contributions of this paper
Response 3: Thank you very much for this insightful suggestion. We agree and have thoroughly reorganized the introduction to enhance its logical flow. Specifically, We first discussed the current state of research, highlighting limitations in existing datasets and deep learning-based methods. We then clarified the need for specialized ground-level image datasets focusing on economically and flood-vulnerable urban areas. Finally, the introduction now clearly summarizes the main research objectives and contributions of the study, highlighting the uniqueness and practicality of AlleyFloodNet.
[Please note that the changes related to this comment can be found throughout the highlighted parts of the Introduction section, excluding the last paragraph.]

Comments 4: the Related works section lacks literature analysis and summary, and it is suggested to organise it according to the two parts of datasets and methods
Response 4: We fully agree with your comment. Accordingly, We have reorganized the "Related Works" section into two clear subsections:
1. Dataset-based Research:  This subsection systematically analyzes and summarizes existing studies focused on flood detection datasets, clearly presenting their characteristics and scope.
2. Deep Learning-based Detection Techniques and Methodologies Research: This subsection comprehensively reviews existing studies employing various deep learning algorithms and methodologies for flood detection.

Comments 5: "3.2 Models" section is too detailed in its description of the models, and it is recommended to simplify it
Response 5: We agree and have simplified the descriptions of each deep learning model in section 3.2 by focusing only on the essential structural characteristics and rationale for using these models, reducing unnecessary details.

Comments 6: lines 330 and 332, there are spelling mistakes in ‘astly’ and ‘n this study’
Response6: Thank you very much for pointing out these errors. I have carefully corrected these spelling mistakes in the manuscript.

Comments 7: I think there are serious problems with the organisation of the experiments in section 3.6. The fact that both models were tested with AlleyFloodNet does not prove anything. Because the two datasets are inherently different, both models trained separately using the two datasets would be more likely to accurately predict images that are more similar to the original data. In other words, if both models were evaluated with the FloodNet test data, the results would certainly differ from the existing results.
Response 7: We fully agree with your concern. Initially, the experimental design indeed involved comparing AlleyFloodNet with the UAV-based FloodNet dataset, resulting in inherent validity issues due to the fundamental differences between these datasets.

To resolve this issue, the experimental dataset was changed from the UAV-based FloodNet to the Kaggle Ground-level Flood Classification Dataset, which consists of over 10,000 ground-level images, including many captured in general urban contexts. Using the Kaggle dataset allowed for a more valid and relevant comparison, as it contains images with viewpoints similar to AlleyFloodNet but without specialized focus on economically and flood-vulnerable urban environments. The updated experiments clearly demonstrate the effectiveness and necessity of specialized datasets like AlleyFloodNet for accurately detecting floods in economically vulnerable urban areas.

The original experimental contents previously in section 3.6 have now been reorganized and moved to sections 4.4 (Experimental Setup) and 5.3 (Results).
Please refer to sections 4.4 and 5.3 in the revised manuscript for further details. The limitations of the comparative approach are clearly stated in the Discussion section, and future research involving reciprocal cross-dataset evaluations is also proposed to strengthen the robustness and generalizability of the results.

Once again, thank you very much for your constructive comments and detailed suggestions, which have significantly improved the manuscript. I deeply appreciate your time and effort in reviewing my work.

Reviewer 2 Report

Comments and Suggestions for Authors

1. Even with the strong contribution of the dataset, there is no introduction of novel models, algorithmic advancements, or innovative technical methodologies in the manuscript. The manuscript is just relying on established architectures as well as well-known training procedures. With no methodological contribution, the study doesn't have enough originality to be published in Electronics.
2. The dataset contains less than 1,000 images, and this is insufficient to train and test deep models. The dataset mostly captures South Korea and lacks other environments (such as other floods, different weather, and nighttime scenes). This limits its applicability as a benchmark.
3. The article states that it has SOTA performance due to existing models such as ConvNeXt and ViT. The comparison to FloodNet is minimal and not well presented. No comparison is done with different datasets and no study is presented to establish that AlleyFloodNet has superior performance.

Author Response

Comments 1: Even with the strong contribution of the dataset, there is no introduction of novel models, algorithmic advancements, or innovative technical methodologies in the manuscript. The manuscript is just relying on established architectures as well as well-known training procedures. With no methodological contribution, the study doesn't have enough originality to be published in Electronics.
Response 1:  We fully understand your concern regarding methodological originality. In response, we have explicitly emphasized the originality and contribution of this research by adding specific new content to the Discussion and Conclusions sections of the manuscript. The revised manuscript clearly highlights our unique application of established deep learning architectures through successful fine-tuning specifically tailored to the newly developed AlleyFloodNet dataset. Moreover, we explicitly articulated the necessity and practical implications of our research for developing real-time flood detection systems to protect lives and property in economically and environmentally vulnerable urban areas.

In particular, the following points have been newly emphasized in the revised manuscript:
1. Highlighting Unique Contributions: We explicitly demonstrated how established deep learning architectures were effectively adapted and successfully fine-tuned to the specialized AlleyFloodNet dataset, directly addressing the important but previously underexplored area of flood detection in economically and environmentally vulnerable urban areas.
2. Necessity and Practical Implications: We clearly described why the AlleyFloodNet dataset and the successfully fine-tuned deep learning models are critical foundational resources for practical, real-time flood detection and alert systems. Furthermore, we emphasized how our research substantially contributes to realistic flood prevention and effective disaster response in vulnerable urban regions.

We fully understand and appreciate your concerns regarding methodological originality. As you pointed out, our study relies on established deep learning architectures and known training methods; therefore, we recognize that you may perceive a lack of completely novel algorithms or methodological innovations. However, we would like to gently mention that in the research areas of flood classification and segmentation, many studies propose novel models primarily by fine-tuning collected datasets with existing deep learning architectures, effectively demonstrating their practical applicability rather than developing entirely new algorithms from scratch. In addition, from a methodological perspective, our approach demonstrates significant originality in achieving high classification performance even with relatively low-resolution images (224×224 pixels). This allows substantial savings in computational resources, making the approach highly practical and applicable for real-time flood detection systems.

In line with this research practice, our study presents a newly developed dataset, AlleyFloodNet, specifically targeting flood detection in economically and environmentally vulnerable urban areas. We have successfully fine-tuned well-established models to this specialized dataset, demonstrating its significant potential for practical real-time flood detection applications. Considering these aspects, we sincerely hope you might positively acknowledge the practical originality and application-focused contributions of our research.

Comments 2: The dataset contains less than 1,000 images, and this is insufficient to train and test deep models. The dataset mostly captures South Korea and lacks other environments (such as other floods, different weather, and nighttime scenes). This limits its applicability as a benchmark.
Response 2: Thank you for highlighting this important issue. In response to your comment, we expanded the AlleyFloodNet dataset from fewer than 1,000 images to a total of 1,110 images. We have clearly indicated that AlleyFloodNet includes images collected from various countries worldwide in both the Abstract and Section 3.1 (Data Description). Moreover, we explicitly detailed the additional images we collected in response to your comment in the last paragraph of the Data Description section.

Comments 3: The article states that it has SOTA performance due to existing models such as ConvNeXt and ViT. The comparison to FloodNet is minimal and not well presented. No comparison is done with different datasets and no study is presented to establish that AlleyFloodNet has superior performance.

Response 3: We fully agree with your concern. Initially, as you correctly pointed out, the experimental setup involved evaluating models trained on FloodNet, which primarily consists of UAV-based aerial imagery. This indeed raised validity issues due to inherent differences between the datasets.

Therefore, following your suggestion, We revised the experimental setup by replacing FloodNet with a publicly available general ground-level Flood Classification Dataset from Kaggle, containing over 10,000 ground-level images. The Kaggle dataset shares similar ground-level viewpoints but generally covers broader urban scenarios without specific considerations for economically vulnerable contexts. Using this dataset, We fine-tuned deep learning models and evaluated their performance on the AlleyFloodNet dataset. The revised experiments quantitatively demonstrated a significant performance drop when the Kaggle-trained model was evaluated on AlleyFloodNet, clearly emphasizing the necessity and specialized advantage of the AlleyFloodNet dataset.

Detailed quantitative results and further discussions of these comparative analyses are explicitly included in sections 4.4 (Experimental Setup), 5.3 (Results), and the Discussion section of the revised manuscript. Please refer to these sections for further details.

Thank you again for your thoughtful comments and valuable suggestions, which have significantly improved our manuscript. We greatly appreciate your time and effort in reviewing our work.

Reviewer 3 Report

Comments and Suggestions for Authors

Did the AlleyFloodNet dataset introduced by the authors create it by the authors? What specific practical flood detection methods can be developed using accurate disaster image detection and classification technology? In addition, what is the motivation for fine-tuning the deep learning model? Are the Google and YouTube icons in Figure 1 copyrighted and can be used in the paper? The authors should list the contributions and the preview of the remaining chapters after line 71. In the related work section, the author's presentation of each literature in chapters is illogical. The authors should set up several different aspects for classification and centralized citation, and list tables and specific parameters to compare the differences between existing studies and this study. In Chapter 3, the authors should first state the purpose and motivation of the study. In Figures 2 and 3, the spliced ​​images should not be shown, but the meaning of each image should be specifically labeled. Furthermore, Chapter 3 is clearly a methodology chapter, but the authors introduce the experimental settings and methods in 3.3 and after, which is very inappropriate. The authors should move them all to Chapter 4. The title of Chapter 4 should be "Experimental Evaluation". In addition, the discussion chapter must be added and clearly explain the findings, advantages and limitations of the results. The annotations for Figures 4, 5, and 6 should be below the figures (or above the tables). What does Figure 6 want to convey? The authors’ roles and data, conflicts of interest, etc. must also be clearly stated at the end of the paper. Thanks. 

Author Response

Comments 1: Did the AlleyFloodNet dataset introduced by the authors create it by the authors?
Response 1: We collected flood and non-flood images from Google and YouTube, and then created the AlleyFloodNet dataset ourselves. To clarify this point, we have added the following sentence to Section 3.1 (Data Description):

"In this study, we developed an original dataset named AlleyFloodNet, composed of flood and non-flood images, to identify the occurrence of flooding using a binary classification approach."

Comments 2: What specific practical flood detection methods can be developed using accurate disaster image detection and classification technology?
Response 2: Using the AlleyFloodNet dataset, practical and effective flood detection and alert systems tailored to economically and environmentally vulnerable urban areas can be developed. Specifically, as detailed in the 6. Discussion section, deep learning models fine-tuned on AlleyFloodNet can be integrated into automated monitoring systems based on real-time ground-level imagery (e.g., CCTV or smartphone footage). These systems can rapidly classify images to detect localized flooding events and immediately issue location-specific alerts to emergency services and residents. This enables timely evacuations and efficient disaster response, thereby significantly enhancing public safety and community resilience in vulnerable urban areas. The introduction and the newly added 6. Discussion section have been reviewed, and the relevant content has been added as requested.

Comments 3:  what is the motivation for fine-tuning the deep learning model?
Response 3: Thank you for pointing this out. As per your request, we have clearly described our motivation for fine-tuning deep learning models in Section 3.2 ("Deep Learning Models"). Specifically, we adopted a fine-tuning approach using pre-trained weights from models trained on large-scale datasets such as ImageNet. This method was chosen to improve the generalization capability of the models and to enable the AlleyFloodNet dataset to effectively capture specific flooding patterns occurring in economically vulnerable urban areas. As per your request, we have clearly described the motivation for fine-tuning the deep learning models in section 3.2 ("Deep Learning Models") of our manuscript.

Comments 4: Are the Google and YouTube icons in Figure 1 copyrighted and can be used in the paper? 
Response 4: Thank you for your valuable comment. In response to your concern, we have replaced the original logos with copyright-free icons instead of the official Google and YouTube logos. Please refer to the revised figure at the end of the Introduction section.

Comments 5: The authors should list the contributions and the preview of the remaining chapters after line 71.
Response 5: Thank you very much for your valuable suggestion. Following your recommendation, we have significantly enhanced the logical structure of the Introduction by adding detailed descriptions of our research contributions. Additionally, we have included a comprehensive preview of the remaining chapters at the end of the Introduction section, as you specifically requested.

Comments 6: In the related work section, the author's presentation of each literature in chapters is illogical. The authors should set up several different aspects for classification and centralized citation, and list tables and specific parameters to compare the differences between existing studies and this study.
Response 6: We fully agree with your comment. Accordingly, We have reorganized the "Related Works" section into two clear subsections:
1. Dataset-based Research:  This subsection systematically analyzes and summarizes existing studies focused on flood detection datasets, clearly presenting their characteristics and scope.
2. Deep Learning-based Detection Techniques and Methodologies Research: This subsection comprehensively reviews existing studies employing various deep learning algorithms and methodologies for flood detection.

Comments 7: In Chapter 3, authors should clearly state the purpose and motivation of the study first.
Respnse 7: Thank you very much for your valuable suggestion. To clearly emphasize the purpose and motivation of our study, we have explicitly added a description in Section 3.1 (Data Description), stating the importance of training deep learning models on our newly developed dataset to overcome the limitations of existing datasets. In addition, similar content has been added to the Introduction to improve the overall logical coherence and completeness of the manuscript.

Comment 8:  In Figures 2 and 3, the spliced images should not be shown, but the meaning of each image should be specifically labeled.
Response 8: Thank you very much for your valuable comment. Figure 2 presents example images from the flood class, and Figure 3 shows example images from the non-flood class in the AlleyFloodNet dataset. Specifically, we manually examined images retrieved from Google and YouTube by searching for the keyword 'flood', classifying images as "flood" when structures or people were significantly submerged, and as "non-flood" when the ground was merely wet or there was heavy rainfall without noticeable water accumulation. This additional clarification has been incorporated into the final part of the Data Description section.

Comments 9: Chapter 3 is clearly a methodology chapter, but the authors introduce the experimental settings and methods in section 3.3 and beyond, which is very inappropriate. The authors should move all these contents to Chapter 4. The title of Chapter 4 should be "Experimental Evaluation".
Response 9: Thank you for your valuable suggestion. As you pointed out, Chapter 3 should exclusively cover the methodology. Therefore, we have respectfully relocated all experimental settings and procedures described from section 3.3 onward to Chapter 4, and clearly retitled Chapter 4 as "Experimental Evaluation."

Comments 10: The discussion chapter must be added and clearly explain the findings, advantages, and limitations of the results.
Response 10: Following your suggestion, we have newly added a Discussion section as Chapter 6, explicitly describing and discussing the main findings, advantages, and limitations of our experimental results.

Comments 11: The annotations for Figures 4, 5, and 6 should be positioned clearly below the figures (or above the tables).
Response 11: Thank you for pointing out this mistake. As indicated, we have corrected the placement of annotations for Figures 4, 5, and 6, positioning them clearly below each figure.

Comments 12: What does Figure 6 intend to convey? The authors should clearly describe its intended meaning.
Figure 6
Response 12: Thank you for highlighting this issue. Upon reconsideration, we have decided to remove Figure 6 and instead clearly explain its intended meaning within the text to enhance clarity.

Comments 13: The authors’ roles, sources of data, conflicts of interest, etc. must be clearly stated at the end of the paper.
Response 13: As requested, we have explicitly added a section at the end of our manuscript clearly stating the authors’ roles, sources of data, and conflicts of interest.


thank you very much for your thorough review and thoughtful suggestions, which have greatly improved the clarity and quality of our manuscript. We sincerely appreciate your time and effort dedicated to reviewing our work.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

1. ‘2. Related Works’ should not be a list of literature. Please summarise and describe the process of development;
2. Suggest adding some qualitative evaluation results, such as prediction results;
3. ‘6. Discussion’ suggests adding some graphs of results.

Author Response

Comments1: ‘2. Related Works’ should not be a list of literature. Please summarise and describe the process of development;
Response1: As suggested, we have revised Section 2 (Related Works) to summarise and clearly describe the development process of existing flood detection datasets and techniques, rather than simply listing relevant literature.

Comments2: Suggest adding some qualitative evaluation results, such as prediction results;
Response2: We have addressed this by including qualitative analyses of prediction results through specific examples of misclassified images in Section 5.3 (Results Comparative Analysis of Flood Datasets and Misclassification Patterns). This qualitative evaluation allows readers to better understand the model’s prediction characteristics and challenging scenarios.

Comments3: ‘6. Discussion’ suggests adding some graphs of results.
Response3: In response to your suggestion, we have added comparative graphs of experimental results clearly visualizing the differences in performance between AlleyFloodNet and the similar ground-level Kaggle Flood Classification Dataset. These graphs effectively demonstrate the superiority of AlleyFloodNet.

We sincerely appreciate all the reviewers for their valuable time and constructive comments. Their insightful feedback has significantly contributed to improving our manuscript. Thank you once again for your thoughtful consideration.

Reviewer 2 Report

Comments and Suggestions for Authors

No further comments need

Author Response

Thank you very much for your valuable comments and suggestions. We have carefully incorporated all of your feedback, which has significantly enhanced the quality of our manuscript. We sincerely appreciate your constructive guidance.

Reviewer 3 Report

Comments and Suggestions for Authors

In Section 3.1, the details of "binary classification method" should be explained. In addition, how to solve the problem that the ground photos collected by the author are not representative enough for narrow alleys? In addition, the reviewer cannot understand that the author has been introducing a lot of knowledge summaries of existing technologies in the methodology section of Chapter 3, such as deep learning models, AlexNet, VGG-19, etc. These contents as prerequisite knowledge should be introduced in a new section named "Preliminaries" before the current Chapter 3. The current Chapter 3 is the fourth chapter to truly introduce the specific sections or steps of the author's original method (Figure 1 can be moved here). The author must highlight the originality of the proposed method and the specific implementation steps and details (including the required hardware computing resources, detailed code of the software programming language, etc.). This allows readers to replicate and expand the author's contribution, which is a valuable reference for publication. In the experimental section, the purpose and motivation of the experiment should be clearly described first. How to perform the experiment and which steps in the proposed method are specifically referenced should also be clarified. The weakest part is Chapter 3, which makes it impossible for readers to find the original part of the author. Thanks. 

Author Response

Comments1: In Section 3.1, the details of the 'binary classification method' should be explained.
Response1: Thank you for pointing out the necessity of clarifying the binary classification method in Section 3.1. To address this, we have added explicit details at the end of Section 3.1 (Data Development and Preprocessing)

comments2: t is necessary to explain how the authors addressed the issue of collected ground-level photos not being sufficiently representative of narrow alleyways.
Response2: Thank you for highlighting the importance of adequately representing narrow alleyways in the collected ground-level images. As clearly stated in Section 3.1 (Data Development and Preprocessing) of the manuscript, we carefully addressed the representativeness by designing and constructing the AlleyFloodNet dataset with particular emphasis on narrow alleyways, semi-basement residences, and lowland areas. Specifically, images were intentionally collected from various economically and environmentally vulnerable urban regions worldwide—including South Korea, Japan, China, Vietnam, the United States, Spain, Italy, and India—and captured at low camera angles and close distances to realistically depict localized flooding scenarios. Furthermore, to enhance the representativeness and robustness of the dataset, we increased the total number of collected images, thus effectively overcoming the limitations mentioned in your comment.

Comment3:
Respse3: 

Thank you for your valuable suggestion regarding placing descriptions of existing deep learning models into a separate "Preliminaries" section. We sincerely appreciate and have carefully considered your comment. While creating a separate preliminary section is indeed feasible, many similar studies in the fields of deep learning and computer vision typically describe the details of model structures and characteristics directly within the "Materials and Methods" section. This approach is frequently adopted because the architectural properties of each model are closely tied to experimental design and data processing methods, enabling readers to more clearly understand the rationale behind selecting specific models in the context of the study.

Considering that we have already substantially revised our manuscript based on your constructive feedback, we kindly ask for your understanding regarding this matter, as we opted to retain this commonly used structure in line with related research in the field.

We agree on the importance of clearly emphasizing the originality of the proposed method and providing detailed implementation steps, including hardware resources, software programming languages, and specific code. The originality of our proposed method has been extensively discussed in the Introduction, Section 3.1 (Data Description), and the Discussion section. Furthermore, in the Experimental Evaluation (Section 4), we explicitly emphasized the originality by conducting comparative experiments demonstrating the superior performance of our dataset compared to existing datasets. Additionally, detailed implementation steps and specifics regarding hardware resources, software programming languages, and relevant code have now been further elaborated and explicitly included in the revised manuscript.

Comments 4: In the experimental section, the purpose and motivation of the experiments should be clearly described, and it should be explicitly clarified how the experiments were conducted and specifically which procedures of the proposed method were applied.
Response4: 

Thank you for your valuable suggestion regarding the clarity of the experimental purpose, motivation, and procedures applied. As respectfully pointed out, we believe that the purpose and motivation of the experiments have already been sufficiently described in the Introduction(Section1), Related Works (Section 2), and Materials and Methods (Section 3) of our manuscript. Specifically, our study explicitly stated that the primary goal of the experiments was to evaluate and validate the accuracy of deep learning models using AlleyFloodNet for detecting floods in localized urban environments, utilizing comprehensive performance metrics such as accuracy, precision, recall, and F1-score.

In addition, through comparative experiments with the Kaggle Flood Classification dataset, we distinctly emphasized the unique effectiveness and necessity of AlleyFloodNet. The detailed experimental procedures—such as dataset splitting, model fine-tuning, and performance evaluation—have been clearly outlined in the manuscript. Therefore, we respectfully believe that the experimental section, in conjunction with the detailed explanations provided in earlier sections, already sufficiently addresses the points raised.

Comments5: The weakest part is Chapter 3, which makes it difficult for readers to find the originality of the author.
Response5: Thank you for highlighting this important point. We respectfully clarify that the originality of our study has already been sufficiently emphasized throughout various sections of the manuscript, including the Introduction (Section 1), Related Works (Section 2), Data Description (Section 4), as well as the Discussion and Conclusion sections. Specifically, the originality of our research lies in constructing AlleyFloodNet, a specialized dataset explicitly tailored for localized and economically vulnerable urban flooding scenarios, distinct from conventional urban flood image datasets. Furthermore, we experimentally validated the effectiveness of AlleyFloodNet by fine-tuning and evaluating various state-of-the-art deep learning models, such as ConvNeXt-Large. Our experimental results clearly demonstrated the superior performance of models trained on AlleyFloodNet compared to those trained on general flood datasets, thereby visually and quantitatively reinforcing the unique contribution of our research. In addition, we have further clarified and reinforced this originality through comparative experiments with other existing datasets.

 

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

No other comments.

Author Response

Thank you very much for your helpful comments and suggestions on our research.

Reviewer 3 Report

Comments and Suggestions for Authors

In Chapter 2, the authors should list tables and set corresponding parameters to specifically compare the differences between existing studies and this study. In addition, the reviewers appreciate the authors' adjustment of the content structure of Chapter 3. As shown in Figure 1, how did the authors fine-tune the 6 deep models? Are these models executed simultaneously? What is the arrangement and distribution between them? The current Chapter 3 should comprehensively and systematically introduce the methods including image classification, comparative data, and multiple sets of classification parameters, rather than just introducing each deep model separately. In other words, the reviewers have been emphasizing that the author's original process and methods must be fully reflected. Also, why is there no introduction to ConvNeXt-Large in the methods section? Furthermore, in the results section, in addition to comparing the performance between different models, what specific new findings did the authors make about the specific data itself? Thanks. 

Author Response

Comments1: he authors should list tables and set corresponding parameters to specifically compare the differences between existing studies and this study.
Response1: Since this study was conducted using distinctly different data from previous research, I believe it is unnecessary to compare evaluation metrics directly. Additionally, comparisons with prior studies have been clearly described in a way that is understandable without explicitly using tables. Instead, there is a comparison section with other datasets.

Comments2: As shown in Figure 1, how did the authors fine-tune the 6 deep models?
Respnse2: I explicitly specified it in the introduction by revising it to "each model."

Comments3:  The current Chapter 3 should comprehensively and systematically introduce the methods including image classification, comparative data, and multiple sets of classification parameters, rather than just introducing each deep model separately.
Response3: Thank you for your valuable comment. I have  added the relevant content at the end of Section 3.2 ("Deep learning models").

Comments4: The authors need to clearly explain why the ConvNeXt-Large model was not introduced in the methods section (Chapter 3).
Response4: While making revisions based on another reviewer's suggestions, a typographical error occurred. I have now corrected it. Thank you for pointing it out.

Comments5: In the results section, besides simply comparing the performance between different models, authors should also present specific and new findings derived from the specific data itself.
Comments5: As you previously suggested, I have moved that part to the Discussion section.