Review of egusphere-2025-4502: "Climate and impact attribution of compound flooding induced by tropical cyclone Idai in Mozambique"

General Assessment

This manuscript addresses an important gap in attribution science by developing a storyline framework for tropical cyclone-induced compound flooding in data-sparse regions. The authors demonstrate technical sophistication in coupling multiple state-of-the-art models (SFINCS, wflow, D-Flow FM) to resolve all flood drivers dynamically. The application to TC Idai in Mozambique is particularly valuable given the underrepresentation of African cyclones in attribution literature. The work makes meaningful contributions to understanding how climate signals propagate from hazard to impact through nonlinear damage relationships.

However, the manuscript requires strengthening in several critical areas before publication. The counterfactual design lacks sufficient scientific justification for key parameter choices, the validation strategy needs refinement given data limitations, and the uncertainty quantification is inadequate for the compounding uncertainties inherent in this multi-model framework.

Major Comments

1. Counterfactual Design and Scientific Justification

The manuscript's attribution conclusions depend entirely on robust counterfactual scenarios, yet the justification for key parameters is insufficient:

•    Rainfall reduction (8%): While the Clausius-Clapeyron relationship is cited, the manuscript should explicitly demonstrate this calculation rather than only asserting it. Recent literature supports 8% reductions for ~1.1°C warming, but this warrants a dedicated methods subsection showing this applied to the case here and a discussion of uncertainties in this approach, particularly for tropical cyclones, where dynamic effects may deviate from thermodynamic expectations.

•    Wind speed reduction (10%): This is more problematic. Knutson et al. (2020) report median projections of 1-10% intensity increases for 2°C warming, suggesting 0.5-5% for current ~1.1°C warming. A 10% reduction appears to overestimate the counterfactual change, potentially inflating the attributed impact from wind-driven processes. The manuscript cites Mester et al. (2023), who used regional observed trends, but doesn't establish why this particular value is appropriate.

•    Sea level rise component: The methodology for SLR estimation needs clarification:

o    Authors use the Treu et al. (2024) dataset but this contains systematic biases. The manuscript must explicitly address whether such biases affect factual and counterfactual equally (canceling in differences) or differently (amplifying attribution error)

o    Figure S10 shows only 2015 data, yet the authors extrapolate to 2019, assuming linear trends.

o    Long-term tide gauge observations from the region should be incorporated to validate the 14 cm estimate

•    Bankfull discharge assumption: The 2-year return period assumption for bankfull discharge should scale between scenarios. If climate change increases discharge, the effective channel capacity may differ between factual and counterfactual. Holding this constant creates an inconsistent comparison -the counterfactual world would have had different equilibrium channel geometries. This deserves explicit discussion or sensitivity testing.

2. Methodological Concerns

Model simulation duration: How long were the simulations run? The manuscript doesn't specify the total simulation period or spin-up time. For compound flood modeling, the synchronization of multiple drivers is critical. Eilander et al. (2023) found surge peaked 3-5 days before discharge for TC Idai, producing limited compound interaction. Does this framework capture such timing effects? 

Boundary condition consistency: How are the multiple drivers synchronized temporally across model domains? The wflow warm-up (365 days) is mentioned, but what about SFINCS, D-Flow FM initialization? Are antecedent soil moisture conditions consistent between scenarios? How is infiltration handled in SFINCS relative to wflow?

Validation metrics: The validation should report separate performance metrics for: coastal zones (surge-dominated), fluvial zones (discharge-dominated)  and compound zones (driver interaction).  This would strengthen confidence that the model captures the different flooding mechanisms appropriately. 

Meteorological forcing quality: ERA5 underestimates TC intensity (appropriately addressed by Holland parametric winds), but the manuscript should clarify: Was the TC parametric wind model used throughout the entire modeling chain? If so, how do biases propagate? The Holland model performs well when fitted to observations or additional empirical relationships, but "out of the box" applications can have significant errors and miss asymmetric TC shapes. The 0.75 TC radius merging also needs explanation—how sensitive are results to this choice?

3. Wave Setup and Coastal Process Assumptions

The assumption that wave setup remains constant across counterfactual scenarios is scientifically invalid and potentially introduces substantial error since wave setup scales with wave breaking intensity. A 10% wind reduction would generate lower wave heights, producing correspondingly lower setup—potentially 15-20% reduction if setup scales as Hs².

Also, the SnapWave approach over transects is problematic since wave conditions aren't properly downscaled (ERA5 offshore waves are coarse), transects are poorly connected to actual water levels, and local nearshore processes (refraction, shoaling, breaking) may be inadequately represented in a transect approach

I strongly recommend: (1) fully integrate SnapWave 2D within SFINCS for dynamic wave-surge coupling in both scenarios, OR (2) remove the wave coupling component entirely, acknowledge this as a limitation, and note it may introduce ~10 to 20% uncertainty in coastal flood depths. The current approach undermines the compound flooding framework's credibility.

4. Flood Damage Modeling 

Depth-damage curve validation: The continental curves from Huizinga et al. (2017) assume European-style construction. Snel et al. (2019) showed Ethiopian traditional buildings experience 100% damage at 2m depth versus 5m for concrete structures. Mozambique's post-Idai assessments reported 111,163 completely destroyed houses, but no published studies validate these damage functions against actual losses. I recommend: 1) comparing aggregate model damage against reported sector-specific losses, 2) conducting sensitivity analyses with alternative damage curves for informal/traditional construction and 3) discussing this as a major uncertainty source.

5. Missing Elements: Uncertainty Quantification

This is the review's most critical concern. Every component carries substantial uncertainty:

-    Counterfactual parameter choices (rainfall: ±2-3%, wind: ±5%, SLR: ±5cm)

-    Meteorological forcing (ERA5 vs. parametric TC model inconsistencies)

-    Hydrological model calibration 

-    Missing river bathymetry (bankfull approximation)

-    Wave setup assumptions 

-    Exposure data completeness

-    Damage function transferability

These uncertainties compound multiplicatively, not additively. The total framework uncertainty is substantial.

The manuscript MUST include:

1.    Uncertainty quantification at each modeling step

2.    Formal uncertainty propagation through the attribution framework  

3.    Comprehensive sensitivity analysis on key assumptions

4.    Probabilistic framing of attribution statements with confidence intervals

Recent attribution papers explicitly report uncertainty ranges. Does climate change contribute 5-15% or 25-35% to damages? Without uncertainty bounds, readers cannot properly interpret the "31% damage attributable to climate change" conclusion.

Minor Comments

Exposure and Vulnerability Treatment (L285-290)

The manuscript states that exposure/vulnerability is held constant, but this critical assumption deserves more prominent discussion. The counterfactual answers: "What would 2019's exposed population experience under a pre-industrial climate?" not "What would the pre-industrial population experience?" This is correct for physical attribution, but should be explicitly stated in Section 2.3.2 rather than buried in the discussion. 

Supplementary Figure S2 (Return Period Analysis)

The disconnection around 1-2 years is striking and unexplained. How was the fit performed? Was this a standard GEV/Gumbel distribution? The discontinuity suggests potential issues with the extreme value analysis or the underlying wflow discharge distribution. Please add a methods subsection describing the EVA approach and discuss this feature.

Line 175 (Holland Model Implementation)

"linearly fading the data at 0.75 fraction of the TC radius" - Why 0.75? This appears arbitrary and could significantly affect results. Show sensitivity or cite precedent. Also, how is the TC eye resolved in terms of rainfall distribution? The Holland wind model has asymmetric components—were these included?

Recommendations

Despite the substantial revisions required, this manuscript represents important and novel work. The technical execution is sophisticated, the application to Mozambique addresses a critical gap, and the compound flooding attribution framework is genuinely innovative. With careful attention to the major comments—particularly uncertainty quantification, counterfactual justification, and wave setup treatment—this can become a strong contribution to NHESS and the broader attribution literature.

Priority actions:

1.    Add comprehensive uncertainty analysis (Monte Carlo or ensemble approaches)

2.    Revise or remove the wave setup coupling

3.    Strengthen counterfactual justifications with sensitivity analyses

4.    Improve damage model validation against reported losses

The open-source, globally-applicable framework you've developed has significant potential for advancing attribution science in vulnerable, data-poor regions. I look forward to seeing the revised manuscript.

This study proposes and applies a storyline attribution framework, combining a multi-model chain to conduct climate attribution and impact attribution analysis for the compound flooding event induced by Tropical Cyclone Idai (2019). The research design is relatively complete, the methodology is advanced, and the attempt in a data-scarce region demonstrates certain innovation. However, the manuscript exhibits the following issues:

Major Points：

1. Regarding the construction of the counterfactual scenarios, a uniform reduction of 8% in rainfall, 10% in wind speed, and the removal of 14 cm of sea-level rise were applied. While this method is operationally convenient and has some basis, it potentially overlooks the spatiotemporal heterogeneity of climate change. For tropical cyclones, climate change may affect not only their intensity but also their track, structure, translation speed, and spatial distribution of precipitation. The adjustment of only intensity parameters in this study may lead to an underestimation of the true impact of climate change.

1. The author employs GRDC discharge data from 1954–1984 for wflow model validation. Based on Figure S6, the simulated discharge performance is not satisfactory. And the simulated discharge exhibits a systematic overestimation, attributed by the authors to overestimation of ERA5 rainfall and the model being uncalibrated. This explanation is overly simplistic. Does the lack of model calibration imply substantial parameter uncertainty, thereby casting doubt on the results. Furthermore, the factual event occurred in 2019, the hydrological response of the catchment may have undergone significant changes due to factors such as climate change and land-use alteration. Consequently, the representativeness and reliability of using data from over three decades ago to validate simulations of the current extreme event are questionable.

1. Due to the lack of river bathymetry data, the authors adopted a simplified approach: directly subtracting the estimated bankfull discharge (approximated by the 2-year return period discharge) from the discharge boundary conditions

• However, has the applied semi-empirical relationship been validated locally in Mozambique? The relationship between discharge and return period can vary significantly across rivers under different climatic, geomorphological, and vegetation conditions.
• As shown in Table S1, the estimated bankfull discharge for Gauge 1 (Buzi River) is 3887 m³/s, but its 95% confidence interval spans [2935, 5140] m³/s, representing a relative uncertainty exceeding ±30%. Using a parameter with such high uncertainty as a decisive "subtractor" directly transfers and amplifies its error into the input boundaries of SFINCS.
• This method neglected the mechanisms of channel storage and flood attenuation. This stands in fundamental contradiction to the complex two-dimensional flood dynamics that SFINCS aims to simulate. Although the authors mention in the supplementary material that this method may lead to inaccuracies in input discharge, its scientific rationale and practical implications have not been sufficiently justified.

1. The simulated water levels are consistent with values from the literature but do not provide quantitative comparison metrics. This makes it difficult to objectively assess the simulation accuracy.

1. In the flood extent validation, although the Hit Rate is >76%, the False-Alarm Ratio is as high as 60% (based on the CEMS product), and the Critical Success Index is only 0.40, indicating limited overall predictive capability of the model. The significant discrepancy between the two satellite products (UNOSAT and CEMS) also highlights the lack of a reliable ground truth flood extent benchmark, further increasing the uncertainty in model evaluation.

1. Changes in rainfall lead to the largest increase in flood volume and extent (9% and 2%, respectively), while the impacts of sea-level rise (SLR) and wind speed changes on volume and extent are less than 1%. However, SLR and wind jointly contribute to a 27% increase in damages, whereas rainfall contributes only 4%. Why do SLR and wind, which have minimal impact on inundation extent, result in such a significant rise in damages? It is recommended to enhance the process-based explanation of the attribution conclusions, rather than merely presenting statistical results

Minor Points

1. The sentence on line 30, "Compound flooding from tropical cyclones (TCs) is one of the most damaging climate extreme events and is exacerbated by climate change (Frame et al., 2020; Smith and Katz, 2013)," is somewhat broad. To aid reader comprehension, especially for those not specialized in this specific sub-field, it would be beneficial to briefly define the core concept of "compound flooding" immediately following its first mention.

1. In lines 125, the description of removing permanent water bodies, based on the Global Surface Water dataset, is provided. However, the temporal relevance of this dataset during the flood event is not addressed. The static nature of the dataset may lead to the misclassification of dynamic water features, such as temporarily expanded rivers during the peak flood period, as permanent water bodies. Please provide more details.

1. The authors highlight the applicability of their framework to global TC events and data-scarce regions as a strength. However, they do not explicitly acknowledge that the global data products used in this study may inherently carry greater uncertainty in these very regions. Furthermore, the discussion insufficiently addresses the model uncertainties that are particularly pronounced in data-scarce settings. These omissions may weaken the perceived reliability of the conclusions when the framework is applied in such contexts.

1. (Lines 320–325) The discussion emphasizes that "focusing on a single flood driver may not give a good representation of the total impact of climate change." However, the results of this study itself demonstrate that for flood volume and extent, rainfall is the overwhelmingly dominant driver (9%), while the individual contributions of SLR and wind are negligible (<1%). Conversely, for damages, the combined effect of SLR and wind (27%) far exceeds that of rainfall (4%). Therefore, whether an impact is "underestimated" depends entirely on the metric of concern. The discussion should more dialectically acknowledge that in compound flooding events, different impact metrics may be governed by distinct key drivers. Consequently, attribution statements should explicitly specify the targeted metric (hazard metric vs. impact metric) to avoid potential misinterpretation.

Summary:

This article presents the results of an application of a global modeling framework that leverages multiple physics-based models to attribute increases in flood extent and damages from TC Idai to climate change (sea level rise, wind, rain). I particularly like that the authors show how climate change has a nonlinear impact on flood hazards and these nonlinearly impact damages. I commend the authors for all the modeling work they did and believe that it can be a valuable contribution after revisions.

Major Comments:

The goal of the paper is currently weakly stated (L75). The authors should rearticulate the gap or contribution of the paper to better frame the introduction and discussion. What is the new or useful information that the authors think this work contributes? I think the work has merit and has the potential to showcase how global models and frameworks can be leveraged in data scare regions. That said, there is almost no consideration for uncertainty, and it would be helpful if the authors could explain why they did not consider this in their analysis. The authors even suggest this in L336 where they state that “For areas with limited observational data, attribution statements should go hand in hand with a thorough uncertainty analysis.”

• What are the ranges of possible climate change factors for SLR, rain, wind? How would your results differ if you considered an ensemble of these? Is the modeling framework not suitable for conducting many simulations?
• How much uncertainty in the results comes from model errors? Is this due to the coarse model inputs or model structure? There are mentions of this throughout the paper, but the discussion would benefit from a structured layout that addresses how these would alter the findings from the study (increase/decrease/no change)? You could conduct sensitivity tests to support these statements or evidence from other articles that use the same models.

I like that this paper extends the analysis to exposure and shows how there is not a direct one-to-one relationship between changes in flood hazards and damages due to climate change. However, these exposure estimates are subject to the uncertainty in the flood hazards, and the authors have an opportunity to dive into the interplay more. Where are the buildings across the study area and how many are there? Is the fraction of the total buildings inundated in the coastal city higher or lower than the buildings impacted along the major rivers (you mentioned bias in L300)? Is fluvial flooding impacting 90% of the households in or near the floodplain even though the damages are small compared to the coastal area? You might consider breaking this down by political or watershed boundaries.

Minor:

Are the flood extent/volumes calculated at the grid or subgrid resolution?

In L125 you mention you remove permanent water bodies… does this include the river network or just the major coastal water bodies? Could this lead to an overestimation of the contribution of rain/river to the flood extents?

Are buildings usually elevated off the ground (L155)? Would it be worthwhile to account for this uncertainty by using a range of depth thresholds (say 0.05-1m) before estimating exposure/damages? 

L171 – what are the other global data sources? Maybe list in parenthesis with citations.

Line 181, you could strengthen this section by adding trends and citations for each flood driver mentioned – especially if they are trends for your study area/region.

Table 1 – Including information on the storm rain rates and wind speeds (like the mean, max, 90 percentile) would be helpful

L215 – The authors state that a lot of flooding is coming from the rivers which are being modeled using wflow which tends to perform poorly for extreme events (Fig S6). If I am understanding correctly, it seems that there are observational gauges near the SFINCS boundary. Why not use these as inflow to SFINCS to see how well the model performs compared to the satellite images and also to get an idea of the buildings/flood extent would be with a more precise discharge boundary condition? It might be a useful experiment especially given how impactful discharge (and the assumptions or error) are for this specific storm and study area.

L234 – any idea why the KGE is so low (0.28) for the Pungwe compared to the Buzi?

L236 – How much is ERA5 overestimating and how do you expect this would cascade to the runoff attribution? Is this mostly a problem for the discharge from wflow or is it also an issue for the rainfall directly on SFINCS?

Figure 4 – does this figure include the water bodies cells that are removed from the flood extent calculation? I would outline these so the reader knows what areas are excluded from the calculation.

Figure 5 – add the number of buildings exposed to this would be helpful

L310 – I would avoid saying “accurately reproduces” here unless you have defined what this is and have statistics as proof. Consider rephrasing.

If you wanted to save some space, you could reduce the text that mentions the vulnerability aspect of flood risk (i.e., L290 and the last paragraph in the discussion). This is important to consider and should be mentioned as something that does matter when conducting risk assessments. However, this paper primarily focuses on the hazard and exposure (buildings) and focusing on the findings and key takeaways that we can gleam for your results regarding these components would help streamline the paper.