the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Jul 2026
Rapid flood damage estimation tool for urban pluvial floods with scarce data
Abstract. Reliable and rapid flood damage estimation is crucial for both disaster risk reduction and crisis management. Yet, existing models primarily focus on riverine floods, neglecting urban pluvial floods – a substantial gap, as heavy rainfall can lead to flooding virtually anywhere. Here, we present FlooDEsT, a new machine learning (ML)-based tool to rapidly estimate building-level damage from urban pluvial flooding with four key improvements, compared to traditional models. First, the model was trained specifically on damage data from urban pluvial flood events. Second, the tool utilises XGBoost, a ML technique capable of capturing complex non-linear data relationships. Third, the tool efficiently utilises geographical information only as necessary, reducing pre-processing time. Fourth, to address the common challenge of missing data, the tool uses smart random sampling techniques to impute building-level features that are representative to buildings affected by this flood pathway, reducing exposure bias. The tool’s computational performance was evaluated in two German case studies, involving about 2300 and 440,000 buildings. The tool provided damage estimates in respectively 2.1 to 5.6 seconds per thousand buildings, representing a 2.7- to 6.6-fold improvement in speed over a baseline approach. FlooDEsT tackles critical gaps in damage modelling, offering valuable support for disaster preparedness and response.
- Preprint
(1691 KB) - Metadata XML
-
Supplement
(189 KB) - BibTeX
- EndNote
Status: open (until 24 Aug 2026)
- RC1: 'Comment on egusphere-2026-3462', Yamile Villafani, 21 Jul 2026 reply
-
RC2: 'Comment on egusphere-2026-3462', Anonymous Referee #2, 28 Jul 2026
reply
The manuscript presents FlooDEsT, a new model for estimating damage caused by pluvial floods. After reviewing the state of the art, the authors present the main components of the model and describe its training, testing, and validation using real-world data. The computational performance of the model is also evaluated, highlighting its potential to support real-time applications. The manuscript is well aligned with the scope of the journal and contributes to addressing an important research gap in flood damage modelling, thereby supporting more effective flood risk management. The paper is well written, the methodology is clearly described, and the results are critically discussed. I do, however, have a few suggestions for improving the robustness and applicability of the model, which should be addressed before the manuscript is suitable for publication
Major and minor comments
- Lines 40–52 are somewhat repetitive. Moreover, it is not entirely clear how the models based on insurance claim counts operate. Since this is not a standard approach in flood damage modelling, I recommend providing a more detailed description of these models and explaining how they differ from more conventional damage estimation approaches.
- Training & validation. I would expect the Amsterdam dataset, which comes from a different country with distinct architectural characteristics, to be less comparable to the other datasets. I therefore wonder why the authors decided to retain this dataset and, in particular, why they used it for the validation phase rather than for training. Did the authors investigate whether the model's performance changes significantly when different combinations of training datasets are used? Such an analysis would help assess the robustness and transferability of the proposed approach.
- Lines 155-156 “To overcome this challenge, we impute the missing data based on the distribution of each 155 variable observed in the survey data (Figure 1), replicating the patterns observed in SWF events” → While this assumption is reasonable for hazard variables, it may introduce substantial biases for exposure and vulnerability variables, as the buildings included in the datasets may not be representative of the entire building stock potentially affected (i.e., all buildings in Germany). I therefore suggest replacing the exposure and vulnerability distributions derived from the datasets with distributions obtained from national databases, which would likely provide a more representative characterization of the exposed building stock.
- Lines 172-176 “To overcome this, considering that the application (e.g. a forecast) will always have a pre-defined domain, and that the building data basis does not need frequent updates, the geographical information may be removed; thus, the following step will only process a raw array data (a simple track max 175 values per index, hereafter named “raster screening”; Figure 3)” → the simplification introduced is not clear; please add more details.
- Lines 199-200 “Based on observed loss ratios across our dataset, we defined the damage classes as: Low: 0 < rloss ≤ 3%; Medium: 3% < rloss ≤ 9%; High: 9% < rloss ≤ 16%; Very high: 16% < rloss” → Could the authors explain in greater detail how the classes were defined? For example, were they based on quartiles or another classification criterion? In addition, it would be useful to discuss how the choice of class boundaries may influence the model's performance and whether the results are sensitive to different classification schemes.
- Section 3.2. I would expect the estimated effect of each variable to depend also on its variability within the dataset (i.e., on the distribution from which the samples are drawn). Could the authors comment on this aspect? This issue becomes particularly relevant if the currently adopted exposure and vulnerability distributions are replaced with those derived from national datasets, as suggested above, since different underlying distributions may affect both the estimated variable importance and the overall model performance.
- Figure 8. The figure legend is not immediately clear. In particular, it is not obvious that the gray bars represent the native approach, whereas the orange bars correspond to the distribution-based sampling approach. I suggest revising the legend to make this distinction more explicit.
- Lines 267-268 “This might be due to a surprising moderate water depth (up to 80 cm), which might have caused the model to predict rather lower damages” → Do the authors mean that the observed water depths are lower than would be expected given the reported damage? If so, this raises an additional issue that does not appear to be addressed in the manuscript: how was the reliability of the survey data assessed? Since the model relies on observed damage data, it would be important to describe any quality control or validation procedures applied to the survey information.
- Lines 284-285 “The raster screening (without counting the pre-processing of the basis data, which is fixed for all runs afterwards) performed 6.1 times faster than the baseline function” →Which is the baseline function? The zonal statistics?
- Conclusions: “This demonstrates the tool’s practical applicability beyond the training data regions, offering confidence in its performance for operational use” & “The current dataset also reflects a limited geographic scope, primarily focusing on German urban contexts, which may affect its transferability to regions with different architectural characteristics” → Transferability across different geographical contexts is a critical aspect of flood damage modelling. I therefore suggest briefly describing the urban and architectural settings in which FlooDEsT can currently be expected to provide robust predictions. This would help readers better understand the intended domain of applicability of the model and its potential limitations when transferred to different regions or building stocks.
Citation: https://doi.org/10.5194/egusphere-2026-3462-RC2 -
RC3: 'Comment on egusphere-2026-3462', Anonymous Referee #3, 31 Jul 2026
reply
I appreciate the opportunity to review this manuscript. The manuscript presents FlooDEsT, a machine-learning-based tool for rapid building-level flood damage estimation in urban pluvial flood events. The proposed methodology combines survey-derived damage data with an XGBoost model to estimate relative building losses while addressing common operational challenges, particularly incomplete exposure information through a probabilistic imputation framework. The model is calibrated using survey data from several pluvial flood events in Germany and evaluated using independent datasets from Germany and the Netherlands. Overall, I found the manuscript well organised and relevant to the scope of the journal. The topic is timely, as flood damage models specifically developed for urban pluvial flooding remain relatively scarce despite the increasing frequency and impact of these events. I also appreciate that the authors have focused not only on predictive performance but also on developing a methodology that is computationally efficient and suitable for operational applications.
In my opinion, the manuscript is suitable for publication after addressing a few comments that mainly concern the clarification of some methodological aspects and the interpretation of several results.
Overall comments
The manuscript presents a clear workflow, and the validation using independent datasets provides confidence in the proposed methodology. Most of my comments are intended to improve the transparency of the methodological choices and to further discuss the applicability, interpretation and limitations of the proposed model. I do not believe that additional analyses are necessary to support the conclusions; rather, I think that a few aspects would benefit from further clarification and discussion.
Specific comments
Page 1, Lines 16–18. The concluding sentence of the Abstract states that FlooDEsT "tackles critical gaps in damage modelling." While the Abstract introduces several contributions, including the development of a model specifically for pluvial flooding, the treatment of missing exposure data, and improved computational efficiency, it is not entirely clear which of these are referred to by the expression "critical gaps." I would appreciate a slightly more explicit statement of the main scientific and operational gaps addressed by the proposed methodology so that the principal contribution of the manuscript is immediately clear to the reader.
Page 8, Lines 155–162. I found the probabilistic imputation framework to be one of the most interesting aspects of the proposed methodology. Nevertheless, I would appreciate some additional justification regarding a few implementation choices. In particular, it would be useful to explain the rationale behind selecting Beta, Gamma and Multinomial distributions to represent the different predictor variables, the motivation for performing exactly 100 imputations, and the criterion used to determine whether the errors introduced by the imputation procedure were considered "acceptable". Although these choices may have been empirically justified during model development, a brief discussion would improve the transparency of the methodology and facilitate its application by other researchers.
Page 11, Lines 209–212. One result that particularly caught my attention is that floor space emerges as the most influential predictor, even above water depth. Since the model estimates relative building losses, I found this result somewhat difficult to interpret from a physical perspective. It would therefore be helpful if the authors could further discuss why floor space plays such an important role in the model. For example, does floor space itself influence the relative loss, or is it acting as a surrogate for other characteristics that are not explicitly represented in the dataset, such as building value, occupancy, construction type or the value of the exposed contents? Intuitively, larger buildings may experience greater absolute losses, but it is less evident why floor space alone would strongly influence the relative loss. Clarifying this point would help readers better understand the physical interpretation of the feature importance analysis.
Page 14, Lines 255–259. The analysis shows that, despite the variability introduced by the probabilistic imputation, 80% of the estimates remain within the same damage class. I found this result encouraging. However, I wondered whether these stable classifications are evenly distributed across the different damage classes or whether they are mainly associated with particular categories. For example, are the most consistent predictions concentrated in low-damage cases, or do they also include the highest damage classes? If this information is readily available, a brief discussion of this point would help readers better understand whether the reported stability is consistent across the full damage spectrum or mainly reflects the predominance of certain classes.
Page 15, Lines 265–269. The manuscript states that the seven very high damage cases were not correctly predicted, possibly because they were associated with relatively moderate water depths (up to 80 cm). As a reader, I would appreciate a slightly more detailed discussion of these cases. If relatively moderate water depths alone do not fully explain the observed damage, what building characteristics or vulnerability factors do the authors believe contributed to such high relative losses? For example, were these buildings characterised by particular vulnerability or exposure conditions that are not fully represented by the current predictor set, or do these cases simply reflect rare situations within the available dataset? A brief discussion of this point would help readers better understand the limitations of the proposed methodology when predicting extreme losses.
Page 15, Lines 275–276. The discussion on model transferability could be clarified further. In Section 3.3, the manuscript states that the validation dataset is limited to western European building stocks with similar building standards and that broader transferability remains to be tested. However, the Conclusions (Page 17, Lines 315–322) also state that the validation demonstrates the tool's practical applicability beyond the training data regions, before acknowledging that the current dataset may limit its transferability to regions with different architectural characteristics. I would therefore appreciate a brief clarification of the intended domain of applicability of FlooDEsT. In particular, under which conditions do the authors consider that the model can currently be transferred with confidence, and when would recalibration or additional validation be advisable? A short discussion reconciling these statements would help readers better understand the current capabilities and limitations of the proposed methodology.
Citation: https://doi.org/10.5194/egusphere-2026-3462-RC3
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 66 | 32 | 12 | 110 | 16 | 14 | 9 |
- HTML: 66
- PDF: 32
- XML: 12
- Total: 110
- Supplement: 16
- BibTeX: 14
- EndNote: 9
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This work aiming presents FlooDEsT, an XGBoost-based tool for rapid damage estimation from pluvial floods. The training data is based on household survey data from Germany and validated on independent datasets from Amsterdam and Karlsruhe. It addresses an important innovation for under-modelled hazard pathway (pluvial floods). Yet, major revision and clarification is needed. Please see my comments and suggestions below.
Introduction
1. The motivation that pluvial floods are under-represented in damage modelling is clearly stated. However, why impacts are not systematically documented (line 21) required additional information that further supports this as the gap.
2. Consider reviewing the damage models and present them in a more structured manner. For example, distinguishing them between type of model, and summarising key limitations and constraints before proposing multivariable models. Additionally, it is not clear how multivariable modelling improve damage estimations. In Line 50 you introduce this as a solution to the deficiencies in other models. Provide some references to this.
3. While the four improvements are clear, consider clarifying how FlooDEsT differs from earlier work by Bronstert et al., Lindenlaub et al., and Dobkowitz et al. beyond the extended dataset and improved computational performance. Specify the novelty of the work. Consider explicitly stating this in the text as well, as it would help to articulate the main scientific contribution.
Data and methods
4. Have you considered using other tree-based methods, such as random forest or lightGBM? How is XGBoost particularly suitable for your method?
5. People tend to forget and/or underestimate the level of impacts for diverse reasons and adding a temporal aspect could increase this effect. You have validation points from 10 years after the first points, and 14 years after the training set. Are you accounting for this or handling this in any way? If so, how? What happens when households do not repair/replace in that span of time?
6. The training and test sets split is described, but details on the splitting procedure are missing. Please specify whether the split was random, stratified, or followed another method.
7. In your methods you are working with data exclusively from Germany, and have external data points from Amsterdam for the validation. This could result in more localised/regional differences inherent to the affected households after pluvial flooding events. Have you considered this as a limitation for the work? This already led to removing entries (see specific comments below).
8. The predictor set points need further clarification:
9. The filtering of Amsterdam data by removing entries with water entry via windows, roofs or downpipes should be described more explicitly by clarifying how many points were removed. Is this introducing any selection bias?
10. The estimation of building value with linear mixed-effect models is a critical step to obtain absolute losses. Its performance is not reported. Please provide performance metrics for the building value model. Discuss how uncertainty propagates to relative loss and damage class assignment.
11. In Line 139 you mention a case was dismissed because the building value was "much lower that reported losses". Could you explicitly provide information as to which case that was? Please specify the used criterion. Are there other points close to the threshold? Was this done before or after fitting the model? Additional explanation is needed for the Amsterdam case. Building values there would be based and modelled on the German insurance market standards.
12. To improve clarity and robustness of this section, some gaps need to be filled.
13. In your sensitivity analysis, have you considered testing scenarios with multiple missing variables simultaneously? This could provide a more realistic setting in operationalising realistic data gaps. Report not only RMSE but changes in damage class under these scenarios.
Results and Discussion
14. The independent validation datasets are valuable, but with a small total sample size. The manuscript acknowledges this limitation, but the implications for the robustness are not fully explored. Discuss how sensitive RMSE and the damage classes are to individual points, given the small sample size. Confidence intervals or other uncertainty estimates could be a valuable addition.
15. The model fails to predict "very high" damage cases in the validation set. This is a critical limitation for a tool intended for flood damage estimation, where extreme losses are key.
16. You provide references mostly for the results in the calibration performance. Please provide additional references in the following subsections of your results. These provide a more in-depth understanding and discussion of your approach, strengths and weaknesses.
17. Are there other ways to optimise the procedure? How can reproducibility be ensured? Please consider including a short paragraph on the main text.
18. Figure 10 is a very useful way of showing visually the estimated damage classes. Consider adding a brief interpretation of patterns if known, to connect the damage outputs to the drivers. It could be connected to what point 15 mentions.
Conclusions
19. Please consider adding recommendations for practitioners, including testing under different infrastructure, etc.
General remarks:
Consider reviewing the language and other minor details. Examples include: