Dear Referee,

        We sincerely apologize for not responding to your message at the earliest opportunity and offer our heartfelt apologies!

        We are truly grateful for your meticulous review and valuable suggestions on our manuscript. Based on your comments, we have revised the manuscript point-by-point and provided detailed explanations for the relevant issues. We have strived to enhance the rigor and clarity of the paper through these improvements and sincerely hope to gain your approval. The specific responses are as follows:

• Regarding the validation set:

        Thank you for your valuable comment. We would like to further clarify that the 10% validation set mentioned in our study can indeed be considered equivalent to a "test set" as you interpreted. Throughout the entire training process, the validation set was strictly separated from the training set, and the model never accessed any original or augmented images from the validation set.

        During data processing, data augmentation techniques were applied exclusively to the training set, while the validation set remained as untouched original images. The validation set was used solely for performance evaluation after each training epoch, primarily to monitor the training process, prevent overfitting, and ensure a fair assessment of the model’s generalization ability.

• Regarding hyperparameter tuning:

        Thank you for raising this point. In this study, we employed the official default hyperparameter settings of YOLOv8 without any additional tuning. Parameters such as learning rate, batch size, and confidence threshold were maintained at their default values to ensure the consistency and reproducibility of the model results.

• Regarding the evaluation of complex scenes:

        Thank you for your insightful question regarding the evaluation of complex scenes. Currently, there is no publicly available large-scale dataset of bus flood inundation images, and the images retrieved from social media predominantly depict regular scenes rather than extreme complex scenarios. Due to the scarcity and dispersion of such complex scene images, we were unable to construct an independent large-scale evaluation dataset.

        The two examples were selected based on their representative difficulty and relevance to the target application. The selection criteria included: (1) ensuring diversity by covering common interference factors such as low-light nighttime conditions, object occlusion, and multiple object overlaps, thereby avoiding bias caused by a single disturbance type; (2) prioritizing scenes presenting compound challenges that were not sufficiently covered during training, to better assess the model’s adaptability to unseen complex environments.

        During the training process, we employed multi-level and diverse data augmentation strategies to significantly increase the complexity and diversity of the training data, encouraging the model to learn more robust feature representations. Experimental results demonstrate that the augmented model achieved notable improvements in complex scene detection. Based on the current results, we believe that the model has good generalization potential and can effectively handle unseen complex scenes. In future work, we plan to further validate the model performance on a larger-scale dataset.

• Response to detailed revision suggestions:

        We have addressed all suggested textual revisions item by item.

        Once again, we sincerely thank you for your careful review and valuable feedback on our work! Should there be any points we have not addressed adequately, we are willing to provide further clarifications.

Dear Referee,

        We sincerely apologize for the delay in responding to your comments and offer our heartfelt apologies for the oversight. We greatly appreciate your detailed review and valuable suggestions. Based on your feedback, we have revised the manuscript accordingly and provided detailed explanations for the relevant issues. Our aim is to enhance the rigor and clarity of the paper, and we sincerely hope that the revisions will meet your expectations. Below are our specific responses:

• On the Practical Significance of Water Depth Estimation:

        We are grateful for your constructive suggestion. The practical significance of water depth estimation in emergency response certainly deserves further clarification. In the original manuscript, our primary focus was on estimating flood water depth distributions from social media images to help identify potential flood risk areas in urban settings.

        In practical applications, even a difference of just a few centimeters in water depth can have varying impacts. For example, 20–30 cm of flooding may cause vehicle stalls or obstruct pedestrian movement, while water levels approaching or exceeding 40–50 cm are more likely to enter residential or commercial areas, posing a significant threat to personal safety and property. By referencing objects like buses and bus door steps in the images, we attempt to provide a visual risk indicator. We hope that the predicted depth ranges will assist emergency response teams in assessing flood severity, identifying high-risk areas, and planning priority responses. In the revised manuscript, we will carefully consider your suggestions and provide further clarification on this point. Thank you again for your valuable feedback.

• On Using Buses as Reference Objects:

        We appreciate your attention to this issue. During the model development process, we paid particular attention to the reliability of the reference objects. All images used for training and testing were manually screened to ensure that the buses in the images were either stationary or moving normally in the floodwaters, and that there was no buoyancy or floating involved, ensuring that buses remained stable reference points for accurate measurements.

        We standardized the height of the buses in our model to approximately 3 meters, referencing typical urban bus models seen in current social media images. Based on a systematic evaluation of the images, we classified flood water depth into several levels, with the majority of images showing water levels between 0–50 cm. Only in a few rare cases did the water level exceed 100 cm. Given the rarity of high-water scenes (i.e., >100 cm) in urban environments, we chose to focus on the more common water levels for classification, without setting higher water levels, as we believe this is more aligned with realistic application needs and image distribution characteristics.

• On the Description of Data Augmentation:

        Thank you for your comment regarding data augmentation. We plan to modify and simplify this section in the revised manuscript.

• On Providing a Brief Introduction to the YOLO Model:

        We appreciate your suggestion. We understand that some readers of NHESS may not be very familiar with machine learning and object detection frameworks. Therefore, in the revised manuscript, we will add a brief introduction to the YOLO model in the introduction section, as you suggested.

• On Established Methods for Estimating Flood Water Depth:

        We appreciate your suggestion for expanding the background of the study. In fact, we briefly reviewed the main methods for urban flood water depth estimation in the original manuscript, including water level gauges, remote sensing technologies, and hydrodynamic models.

        We pointed out that while water level gauges provide accurate point measurements, their high deployment and maintenance costs limit their widespread use. Remote sensing methods are mainly used to identify flood inundation areas, but they do not yet have the capability to directly estimate water depth. While hydrodynamic models can estimate water depth, they require high-quality input data, are computationally complex, and have slower response times, making them less suitable for real-time emergency response. We believe that the background provided supports the validity of our method using social media images for flood water depth estimation and shows that it can serve as an effective complement to traditional methods. Due to space limitations, the discussion of these methods in the original manuscript was relatively brief, but it covers the core points.

• On the Discussion Section:

        Thank you for your valuable suggestion regarding the discussion section. We fully agree with your point that we need to more comprehensively assess the limitations of our method and compare it with existing research, particularly studies using other reference objects.

        In the revised manuscript, we will modify the discussion section in accordance with your suggestion, providing a more detailed analysis of the method's limitations. We will also add comparisons with other studies to help readers better understand the relative strengths and weaknesses of our method.

Once again, thank you for your valuable suggestions. We are confident that these revisions will further improve the quality of the manuscript.