Summary

The manuscript “The value of ultra-detailed survey data for an improved flood damage modelling with explicit input data uncertainty treatment: INSYDE 2.0” proposes a tailored flood impact modeling framework INSYDE to account for the lack of information/uncertainty with regards to the required micro-scale vulnerability and exposure characterstics of buildings. In this study, the framework and process of data-preparation are discussed alongside three test cases to show the benefits.

The authors have done a good job. The authors are addressing a very relevant need for better fit tools to support decision-making acknowledging uncertainty. The authors offer a comprehensive idea how to deal with limited knowledge and give insights into the sensitivities. The manuscript is generally well written and makes particular good use of Tables. The overall structure of the manuscript could be improved alongside revising some of the figures and reflecting on their use. Strengthening the discussion of the benefits and learnings for a decision-maker from using INSYIDE 2.0 compared to other models, could make this manuscript a very strong and relevant addition to the scientific community.

General comments

• The authors do a great job in justifying the two research questions to explore in this study. While this reviewer can clearly identify the evidence presented to answer the second research question. However, the first question on arguing the added value in terms of output quality and usefulness seems to be addressed marginally and should receive more attention in the results and discussion section. What can decision-makers specifically learn from INSYDE 2.0? In that context, this reviewer observed that the Authors seem not to discuss any limitations of this approach.
• Line 95: This reviewer thinks that the description of the methodological approach could benefit from improved visualization and explanation as well as restructuring the sections. Figure 1 seems complete but complex. The reviewer cannot recognize the elements mentioned in the figure in the accompanying text (or the subsequent subsection titles). This makes it very difficult to follow. Using more descriptive subsection titles aligned with the steps in an (updated?) Figure 1 could avoid this challenge. If others were to use INSYDE 2.0, would they use this methodological approach as well? If so, making a clear distinction between describing INSYDE 2.0,  Preparing data for INSYDE 2.0 and Applying INSIYDE 2.0 could be beneficial in the method section.  
•  This reviewer is wondering whether the title (particularly: 'The value of ultra-detailed survey data')  of this manuscript captures the main purpose of this study. Firstly, because the data-sources discussed in this paper are not limited to survey data. Secondly, the paper mostly focuses on exploring the effects of uncertainty.

Specific comments

1. Line 23: these references seem outdated to confirm the Author’s claim regarding development and remaining limitations of flood risk modelling over the past decade.
2. Line 103 – Line 106: What benefits did the Authors see in using data from a specific region to explore the sensitivities of the impacts/feature importance? Would the findings from a sensitivity analysis not offer similar insights, perhaps even more generalizable?
3. Line 108 – Line 119: Did the authors generate the EDF’s? How many data points were available for fitting the distributions mentioned in Tab.1 and Tab.2 ? Since this study is addressing effects of uncertainty, it would be interesting to know on what basis these distributions are developed . This reviewer thinks it could be a good idea to add these EDF’s + fits in the Supplemental Material.
4. Line 120: How were different return periods included in the EDF’s?
5. Meanwhile the current text in section 2.3 is interesting and probably relevant for the functioning of the model, this reviewer does not see the direct link between the purpose of this study (include uncertainty into INSYDE) with this fix addressing the scalability problem. Since this bias seems to be mentioned in 2.1 (line 83), this reviewer would suggest to mention the change of the model in 2.1 and/or put the detailed elaboration in the Supplementary Material. Given that the change already had been applied when using INSYDE in Belgium, this seems not innovative or relevant to report here. Instead, this reviewer was expecting an elaboration of the statement in line 100 elaborating on the process of translating mixed source data into distributions or required adjustments to the (existing) probabilistic framework.
6. Line 148 – line 150: Why did the authors choose to build one dataset containing two different flood types (riverine and flash floods)? Would it not be more accurate to have two separate ones, one per flood type?
7. Line 152: These distributions are generated based on the Po river modelling exercise only? Or also for the two historic case studies separately?
8. Line 155 – Line 163: The study would benefit from additional elaboration how these correlations are built. This reviewer can understand why authors built a synthetic correlation between inundation depth and duration (does it work differently for riverine floods and flash floods?), but much less with regards to the flow speed and flood duration. Can the Authors elaborate on these decisions? How accurate was the fitting of the correlation function based on the sample values? In case of the Po river modelling, flow depth and flow velocity were modelled and thus directly correlated already? What does d / max(d) stand for? Why did the Authors not use Copulas to account for joint probabilities?
9. Line 163 – Line 165: This is not clear. It seems that he and v are the leading parameters and d, s, q are depending on these parameters. Why do we lose information? And what can this reviewer picture under “[…] the values of d*, v* and s* were then replaced with the correspondent percentiles from the datasets of d, v and s”? What is the effect of this?
10. Line 174: How did the authors end up with the number of 5000 hypothetical buildings? Did the authors explore the convergence behavior? When looking into uncertainty and Monte Carlo sampling justification of such choices should be provided.
11. Line 179: In the results, readers are presented with the damage difference as the metric to explore the feature importance. Information on how this metric is calculated (e.g. aggregated vs averaged over the 5000 buildings) would clarify how to read the results.
12. Line 195: per house 1000 replicates were generated to account for uncertain combination of ~ 20 parameters. Did the authors explore the convergence behavior of the results to confirm that this choice is reliable (see comment regarding line 174)?
13. Line 215 – Line 220: Here the Authors mention INSYDE and its benefits. Showcasing the utility of INSYDE 2.0 would not only be towards a decision-maker but also towards the previous version INSYDE. It would be interesting to see the results of the old INSYDE alongside INSYDE 2.0 .
14. Line 220: Table 3 is very helpful. This reviewer would suggest to add the details regarding the synthetic case study in that table as well for overview purposes.
15. Line 224: The benefit of section 3.1 is not clear to this reviewer. It seems to focus on the pairwise occurrence of parameters. It is unclear how it offers evidence to answer the initially proposed research questions. This reviewer thus suggests to either incorporate it in the methodology section or place it in the Supplemental Material. For this reviewer, pairwise occurrence is just one of the different elements in the data sampling process. For example, the pairwise occurrence in the collected information used for the sampling might be of additional interest (to gain insight into uncertainty progression). At the same time, Figures 2 and 3 seem to have an incorrect design: the subplots on the diagonal seem to be histograms, but the y-axis labels are not correct.
16. Line 253: Here, an extended dataset is mentioned the first time (Elaboration in the Method section needed!). So authors are using 4 case studies? In general, what is the justification for the three different cases? What added benefit do the authors see by adding a second stylized case here?
17. Figure 4: How is the damage difference calculated? The medians are barely visible. In general this figure is very colorful, while some other elements are not visible (whiskers, box whiskers for LM to YY in upper plot). The bars seem to go beyond the chosen y-axis limits on some occasions , e.g for BE (upper plot).
18. Line 260: how do the correlations between he, d, v, and s play into these results?
19. Line 261: 10,000 EUR per house or averaged across the entire set of buildings? Are these higher damages or lower damages? A permutation of only 10,000 EUR for each house would lead to a difference of up to 5 million EUR, which is significant again? What are the uncertainty bounds for this difference?
20. Figure 5: The left plot is very useful and clear to see. Why did the Authors choose to combine the scatter plots for the other BT’s?
21. Figure 7: While Table 4 is very helpful and supports the reasoning in the text, this reviewer has doubts regarding Figure 7. First of all, the use of log-log scales makes it very difficult to interpret the results since distance is not constant at different positions in the figure. As such, diverging from the diagonal has much more severe implications towards increasing Observed Damages. Secondly, this reviewer is wondering whether making use of the damage difference as in previous figures might be more informative and supportive regarding the research questions. What is a decision-maker learning from this visualization about uncertainty in modelling?
22. Table 4: What is the computed expected damage? Other point: As a decision-maker, a conclusion I would draw form this Table is that Lodi and Caldogno are both cities that are on average much more vulnerable (because of building properties) than other cities in Northern Italy. Or that INSYDE underestimates damages. Can a decision-maker get any insights into how much of the model gap with reality can be attributed to the uncertain input data vs. other sources of uncertainty/error (e.g. the hazard-damage relations of the model). What lessons can a decision-maker learn from this? Elaborating a bit more on this could benefit the significance of this work.
23. Line 350-352: This reviewer has not seen any evidence that confirms this claim. The advantage can only be compared to the original INSYIDE set up, or the learnings of analysing the uncertainty bands.

The paper introduces INSYDE 2.0, a flood damage modeling tool that integrates ultra-detailed survey and desk-based data to enhance the reliability and informativeness of flood damage predictions. By incorporating a probabilistic module, it addresses missing input data and provides transparent information about uncertainties arising from limited knowledge of damage explicative variables. This integration significantly improves the reliability and robustness of flood damage assessments. While the extensive use of ultra-detailed data contributes to the model's reliability and the importance of accurate damage estimation, the logical coherence and readability of the manuscript still require significant improvement. Therefore, a major revision is necessary to strengthen the manuscript for potential publication.

Major comments:

1. One of my main comments is regarding the dataset used in this study. The manuscript introduces various datasets, such as synthetic datasets, observed datasets, survey datasets, auxiliary datasets, specific datasets, and empirical datasets. However, it is unclear what each dataset represents and how they are related. Are these datasets intended to establish the functional relationships between building parameters, serve as reference data, or facilitate model validation? The lack of clarity regarding the nature and relationships of these datasets, as well as their correspondence with the terms used in Figure 1, significantly hampers readability. I think it’s good to provide a bit more information on the construction of all these datasets. Particularly since this manuscript focuses on uncertainty analysis.
2. The description of building damage in the paper is not very intuitive, and the location distribution information of buildings is hardly mentioned, so it is difficult for readers to apply INSYDE 2.0 to their examples. In my opinion this question - how do we sample complex and spatially distributed variables in a meaningful way? - is one of the key research questions that the uncertainty and sensitive analysis community will need to work on if we want to move on to the next stage of applying this type of technique to complex models. I am not suggesting the authors should solve this issue, but I think they should point out this is a very important and critical step and a big area for future research and development.
3. In Section 2.2, the meaning of "specific catchments of the district representing the distinctive inundation types occurring in rural or urban areas as well as in flat or steeper regions within the Po River district" is unclear. It would be beneficial to include a spatial distribution map or a dem illustrating the Po River district to provide a better understanding of the study area.
4. While it is described that nearly 5000 hypothetically flooded buildings are analyzed, the manuscript does little to explain or describe the building characteristics involved. Numerous contributing factors affect economic loss. Given the primary objective of this manuscript is to quantify loss, it would be necessary for the authors to describe the 5000 hypothetically flooded buildings in more detail.
5. In lines 298-301, I guess the authors mean that high variability in certain parameters leads to significant deviations in loss estimation. If so, this issue represents a substantial portion of the cases and should be addressed. The authors may consider discussing potential measures to reduce such errors.

Minor comments:

1. Language and Sentence Structure: Please carefully review the manuscript for sentence clarity and readability. Some long sentences, such as lines 301-303, are difficult to follow. Consider breaking them into two sentences or rewriting them for better readability.
2. Line 154: Typo - "250.000" should be corrected.
3. Line 156-158: Please describe the meaning of values "c₁-c₅."
4. Line 163-164: Clarify the meaning of "to obtain the correct distributions without losing information on the interdependence among the variables, the values of d*, v*, and s* were then replaced with the correspondent percentiles from the datasets of d, v, and s."
5. Table 2: Define "ECDF" in the table or provide the abbreviation at the end of the table for better understanding.
6. Line 174 and 177: Clarify the difference between hypothetically flooded buildings and the generated synthetic dataset. Consider listing the involved datasets in a table or figure and explaining their significance.
7. Figure 3: Provide further explanation for the labels in the figure.
8. Line 235: Explain the diagonal axis of Figure 2, including the value of the y-axis.
9. Line 255: In Figure 4, clarify the meaning of the bar colors and the red dots. Additionally, explain the color progression from dark blue to light blue and indicate if it represents a rank sequence. Complete all figure labels or provide the necessary information in the figure caption.
10. Figure 7: Increase the resolution of the figure.