Dear Salvatore,

We would like to sincerely thank you for your thorough and constructive evaluation of our manuscript. We appreciate your positive feedback on our work and your valuable suggestions for improving its clarity.

Following your general suggestion, we will add an Appendix to the revised manuscript defining all acronyms, indices, units, and technical terms to enhance readability. We will also address all of your specific comments, and the corresponding changes will be incorporated into the new version of the manuscript.

While the revised manuscript will incorporate all of these changes, we would like to take this opportunity to provide extended responses to the three specific points you raised for further discussion below.

1. Regarding the advantage of using GCPs instead of directly reading water levels from measuring tapes (Page 9, Line 246).

We appreciate this insightful question. From the outset, our project’s aim was to develop a truly non-invasive, contactless method for river monitoring. We believe that our GCP-based approach offers several key advantages over the use of measuring tapes, which will be described more explicitly and concisely in the revised manuscript, following these main points:

• Installation flexibility: Measuring tapes require installation in specific locations that are compatible with both low and high water levels, and must be in direct contact with the water. This significantly limits the possible deployment sites. In contrast, GCPs are placed outside the water and allow for greater flexibility, as georeferencing can be achieved from almost any stable feature within the camera’s field of view.
• Long-term reliability: Field equipment such as measuring tapes or even GCPs can degrade over time due to environmental factors like vegetation growth, freeze up, debris, sedimentation, or vandalism. In our multi-year study, some GCPs became partially covered or deteriorated, but the system continued to function optimally as long as a sufficient number of GCPs remained visible. In contrast, if a measuring tape becomes obscured or lost, water level readings are no longer possible.
• Methodological scalability: While measuring tapes only support water level estimation, GCP-stabilized imagery enables advanced hydrological measurements beyond water levels (e.g., surface flow velocity quantification)
• Towards a non-invasive development: Importantly, our methodology is not inherently dependent on the use of physical GCPs. In future implementations, we envision leveraging virtual GCPs (vGCPs)—that is, using fixed features already present in the scene with known 3D coordinates. This would eliminate the need for any physical installation (neither GCPs nor measuring tapes), further reducing the environmental impact and making the approach even more non-invasive.

1. Concerning the generalization of our model to sites with more diverse riverine environments and whether full retraining would be needed (Page 15, Line 391).

Thank you for raising this point. We realize this aspect may not have been fully explained in the original manuscript, so we will make it clearer in the revised version. For training, we used the RIWA dataset, which contains a wide range of images with water in many different contexts—not just rivers (1163 images). This means that the model was already exposed to a lot of variety during training. In fact, only a small part of the training images came from the rivers analysed in our study (145 images, covering 4 study sites), and even with just a few of these, the model was able to segment river water correctly.

Of course, we can't guarantee exactly how the model will perform in every possible new river environment, but based on our results, we are confident that it generalizes well and can be adapted to new river sites only adding few well labelled images into the dataset. Although we never claim full transferability to new study areas, we appreciate the reviewer’s remark and will revisit the use of this term in the revised version.

1. On the potential complexity of our approach for water level estimation and the possible additional uses of our image-based 3D model in hydrological monitoring.

Thank you for this observation. Although this aspect is not addressed in detail in the current article, it is true that image-based systems for 3D river monitoring have a broader scope within hydrological monitoring. The images and video sequences captured by the cameras are also used to estimate surface flow velocity and, subsequently, river discharge, allowing for hydrological monitoring without any direct contact with the river and by only one continuously operating sensor system.

However, these advanced measurements rely on having a robust photogrammetric setup and accurate water level estimation. For this reason, the focus of this article is specifically on these two foundational aspects. Since this point has also been raised by other reviewers, we will include a paragraph in the Introduction to clarify how our approach fits into the broader objectives of the KIWA project (Grundmann et al. 2024) and its potential for future applications in hydrological monitoring.

Once again, I would like to personally thank you for your thoughtful comments, which will undoubtedly help make the revised manuscript clearer and more accessible to a broader audience. I hope that our detailed responses have addressed your concerns and serve to better demonstrate the scope and significance of the work we have carried out.

Dear Reviewer #2,



We would like to sincerely thank Reviewer #2 for the time and effort dedicated to reviewing our manuscript. We regret that you consider the article unsuitable for publication in HESS. However, we are confident that our work is innovative and represents a significant advancement in automatic water level monitoring. As we will detail further, this work is a foundational step towards comprehensive, non-intrusive, image-based river monitoring (including water level, flow, and discharge), and is therefore worthy of publication in HESS. In the revised manuscript, we will make an effort to highlight this innovation more explicitly. To our knowledge, existing literature on optical water level monitoring typically relies on hybrid approaches that combine image data with additional instrumentation such as LiDAR or radar, or use direct visual readings of measuring tapes captured in the images. Moreover, most of these studies are constrained to short time periods and highly controlled environments, where channel geometry and external conditions are well known in advance.

While we respect your critical perspective, we believe some of your requests—such as a "comprehensive state-of-the-art review of non image-based methods" or a "detailed cost comparison"—fall outside the scope of a research article focused on a specific methodological contribution. While these are absolutely valuable topics, they are better suited for specific review paper. Additionally, we do not share the view that our manuscript resembles a technical report. Although it necessarily contains highly technical sections due to the novel technology involved (AI and photogrammetry), we maintain that the paper demonstrates substantial scientific analysis and depth.

The vast majority of your minor comments will also be addressed in the revised manuscript. For specific requests, such as listing the 32 CNN architectures, we believe that the source is clearly cited, and reproducing material from another published work would not be appropriate.  Below, we will provide detailed responses to your major comments, hoping our responses and revisions will clarify the value of our contribution

Comment #1

The core innovation of our work is the development and long-term testing of a fully non-invasive (contactless) optical monitoring system over more than two years in real river environments. This represents an important first step towards further advancements that can be achieved with image-based methods in hydrological monitoring. Unlike most previous studies—which often rely on controlled settings or require equipment placed directly in the water—our approach achieves high accuracy without any in-stream devices and has been validated across several real-world sites, not just in laboratory conditions or controlled water bodies. Honestly, authors consider that this is enough relevant for HESS publication.

We deliberately benchmarked our system against co-located, in-situ physical sensors, as this provides the most direct and meaningful validation of real-world accuracy. Comparing our results to other optical systems from the literature can be misleading, as performance is highly dependent on site-specific factors (e.g., camera resolution, lens type, distance). Our continuous, multi-year comparison against a physical reference offers a far more robust characterization of reliability. The water levels at the reference gauges are recorded using a float-type recorder and a pressure level sensor. It is a redundant monitoring system and is operated, maintained and quality controlled by the national monitoring agency. In our opinion it is the best reference you can get and it is independent from our work. We will add this information in the manuscript

Finally, our work focuses on integrating established deep learning models into a robust hydrological pipeline. We used validated models for specific sub-tasks: water segmentation (Wagner et al., 2023) and GCP detection (Blanch et al., 2025). These cited papers already provide detailed discussions of alternative architectures (including models like SAM) and the specific AI procedures used. Therefore, we do not believe it is necessary to reproduce that already-published content in this manuscript.

Comment #2

We thank the reviewer for this constructive suggestion. We agree that analyzing performance across different scenarios is essential for demonstrating robustness. While the reviewer suggests accomplishing this by creating explicit data subsets (breaking the long time series analysis), we argue that our study achieves this through a more holistic and, we believe, more realistic approach: analyzing the model's performance over a continuous, two-and-a-half-year time series.

This long-term evaluation inherently captures the very scenarios of concern—from seasonal vegetation growth to snow cover and low-flow conditions—and the manuscript already discusses their impact on accuracy. We maintain that this continuous analysis provides a more stringent and honest test of real-world operational reliability than evaluating isolated, pre-selected subsets. The seasonality of these challenges (e.g., vegetation in summer) acts as a natural data partitioning, and by publishing the daily data, we allow for a transparent assessment of how precision varies under these naturally occurring conditions. And we decided to specifically plot this variation of the absolute errors on Figure 6.

By reporting the mean absolute errors over the entire period (both for the first year and for the full study duration), we are providing the actual performance that would be expected in a real-world application. Artificially partitioning the dataset could, in optimal conditions, have led to reporting lower absolute errors than those published in the paper, but authors honestly believe that this would have created unrealistic expectations and not reflected the true overall performance achievable by the method. And we wouldn’t be able to detect changes of accuracies over time.

Comment #3

The authors would like to clarify that, in our view, “robustness” refers to the reliability and consistency of the method once it is installed at a given site, rather than to its direct transferability from one location to another. While transferability is certainly an interesting and relevant topic—one which we mention and briefly discuss in line 390 of the manuscript—we do not claim that the method can be deployed at new sites without any retraining.

What we do highlight is that the model achieved strong performance in each study area using only a small proportion of locally annotated images (about 11% of the total dataset, distributed across the four sites). Given the wide range of possible water bodies, installation types, environments, camera perspectives, and focal lengths, we consider it scientifically limited to present a transferability analysis based on just a few locations—a similar point is also addressed in our response to Reviewer 1, comment 2. In our view, it is much more transparent and scientifically honest to clearly state how many site-specific images were actually needed to achieve a high-performing AI model, rather than implying universal transferability without evidence.

Authors agree that it would be very interesting to analyze the transferability of this part of the process in detail, but such a specific analysis would require testing the method across 10?, 20?, or even more different types of rivers and camera combinations, which is clearly—once again—beyond the scope of this article.

Comment #4

We appreciate this comment and fully acknowledge the need to improve the motivation and context in the introduction of the manuscript. We also thank the reviewer for recognizing the level of expert knowledge required for this approach. As already discussed in our responses to Reviewer #1 (comments 1 and 3), the use of non-invasive, contactless optical methods for water level estimation is part of a broader research project (KIWA project, cited in the paper (Grundmann et al. 2024)). Once a robust and precise photogrammetric setup is established—as detailed in this article—the same camera system can also be used to record video clips, which enables the estimation of river flow velocity and, ultimately, discharge. These developments will be addressed in future publications, completing a comprehensive set of non-invasive, image-based tools for river monitoring.

The benefits and rationale for using contactless methods, as opposed to in-stream equipment such as scale bars, have already been detailed in our previous responses (reviewer #1). In summary, our approach is designed not only for water level estimation but also as a foundation for more advanced hydrological monitoring, all while minimizing environmental impact and the need for physical intervention in the river. Finally, while we recognize that a certain level of technical development is required, but in scientific research this should never be considered a limitation if it leads to benefits or significant advances in monitoring capabilities -without in any way diminishing the value of other methods.

Dear Riccardo,

We sincerely thank you for your thorough and detailed evaluation of our manuscript. We appreciate the constructive feedback and we will address all minor comments in the revised version, including improvements to figures, terminology, and manuscript structure.

However, we would like to clarify that we are not comfortable with the implication of potential “commercial efforts” regarding code availability. The code is publicly accessible upon request and is already being used in various projects and by other groups. The author team will further discuss whether to make the full codebase openly available.

Given the overlap between some of your comments and those raised by Reviewer #2, we kindly suggest referring to our responses to Reviewer #2 for additional context and clarification regarding several key aspects of the manuscript. Additionally, below we provide specific responses to the main issues you have raised.

Scientific Article vs. Scientific Report

We appreciate the reviewer’s recognition that the technical work is solid. As we have mentioned in responses to previous reviewers, we will work in future versions of the manuscript to ensure it does not read as a project report. However, we do not fully agree that this characterization is fair. We believe the methodology is clearly explained, the results are tailored to the requirements of the scientific discipline (going beyond what would be expected in a project report), and the discussion is appropriate. The descriptive tone may reflect a more personal narrative of the work carried out, and we will review the manuscript to adjust verb tenses and improve scientific framing.

As we have also stated in response to other reviewers, it is not the objective of this publication to discuss the generalization of the AI components. These aspects are thoroughly detailed and discussed in the respective articles dedicated to these developments, which are properly cited in the text (Wagner et al., 2023; Blanch et al., 2025).

Transferability is Overstated

The authors are concerned about this impression, which we believe does not accurately reflect the content of the manuscript and is not entirely fair. At no point do we state that the objective of our development is the transferability of the system from one study area to another. In fact, the term "transferability" appears only four times in the manuscript:

• Line 202: To indicate that the lack of IR datasets means our dataset is NOT transferable (AI segmentation).
• Line 225: To state that the use of data augmentation facilitates the transferability of results (AI segmentation).
• Line 258: To indicate that the GCP detection method shows good transferability (as discussed in Blanch et al., 2025a).
• Line 390: "suggesting the model's high transferability," referring—as we have explained in previous responses—to the use of only 11% local images.

Transferability is not mentioned in the title, abstract, or conclusions. Suggesting possible high transferability of the AI segmentation model does not, in any way, directly or indirectly imply that the entire system is transferable, since, as the reviewer correctly points out, site-specific elements are required for each study area. Nowhere in the manuscript do we claim high system-level transferability, and in future versions we will review the content to avoid any such impression even more.

Ambiguity of Goal

We appreciate this comment and excuse for misleading you. We will rephrase this sentence in an updated version of the manuscript to be scientifically correct. The original objective of the project supporting this publication (KIWA: Artificial Intelligence for Flood Warning) was to explore the potential of AI tools for flood warning. In this regard we recognized quite early that we can achieve measurements with an accuracy in the centimetre range with the optical monitoring system, which allows for using the optical measurements for state updating of forecast models. In fact, because the original goal was on floods, we could afford not to focus on achieving high precision for the lowest water levels—these are not critical for high flood water levels. This is the reasoning behind our statement in line 455 that low-flow conditions do not represent a significant limitation for the system as a whole. Nevertheless, we will remove this specific reference to flood warning in order to avoid any confusion and better distinguish the scope of this publication from that of the broader project. As mentioned in our responses to previous reviewers, we will ensure that the motivation and scope of this work are better introduced and contextualized in a future revision.

Comparisons

We respectfully disagree with the suggestion that our comparisons are unfair. Based on the reviewer’s comment, it might seem that our manuscript criticizes or diminishes other methods, which was never our intention. We will review the manuscript to ensure this is not the case and will clarify in the introduction that our aim is to present a new method that addresses some of the limitations of previous approaches—particularly by offering a fully contactless solution.

We do not believe it is appropriate to compare our method directly with approaches focused solely on flood detection, as our development clearly extends beyond that specific application. As noted in our response to Reviewer #2, a comprehensive comparison with all non-image-based methods would be more suitable for a review article than for a research paper focused on a specific methodological contribution.

We would see the need for exhaustive comparisons if our work were limited to laboratory development under controlled conditions. However, we believe that benchmarking our system against official gauge stations, operating in the field for over two years, offers the most meaningful and rigorous evaluation.

Evaluation Metrics and Goals

We thank the reviewer for this technically focused comment. In the revised version of the manuscript, we will consider including an additional relative error metric and a discussion of operational thresholds, especially for low-flow conditions. However, we do not share the view that centimetre-level optimization is unnecessary for flood detection. We appreciate the recognition that the method could be used for flood detection without requiring the high level of precision we achieve. Nevertheless, our scientific approach is to optimize the photogrammetric process as much as possible, always aiming for the highest quality and accuracy. We believe that this pursuit of precision does not detract from the final objective. Furthermore, if the cameras are to be installed at sites with very shallow, wide river cross-sections centimetre accuracy would be necessary. Another scenario are cameras installed at larger distances; also, then highest precision would be needed to get sufficient water level values. Since we recognize that explicitly framing “flood detection” as the primary goal may be confusing, and that focusing on “water level monitoring” would provide a clearer context, we will reconsider how this aspect is presented in the revised manuscript.