The authors present a process-based framework that integrates temperature-conditioned rainfall projection, stochastic storm transposition, and high-resolution hydrodynamic flood modelling to assess future pluvial flood hazards. The topic is relevant to the cope of HESS, and the manuscript is well-written, concise, and straightforward to understand.  The work is methodologically innovative and supported by an impressive modelling effort. However, the following points need to be addressed before publication. My suggestion is for a revision of minor to moderate extent. 

1.    The framework explicitly assumes that changes in short-duration extreme rainfall are primarily thermodynamic, with limited influence from dynamical changes (Section 2). While this assumption is discussed later, it underpins the entire DSQM–SST-AUTOSHED cascade.  Can the authors more rigorously justify this assumption for the Beijing region, particularly given evidence of circulation-driven changes in East Asian summer rainfall? A sensitivity or discussion contrasting thermodynamic vs. dynamic contributions would strengthen the credibility of the projections.

2.    In this study, 1-km CMORPH-based rainfall is employed to drive street-scale pluvial flood simulations. With emerging radar datasets at 500 m or 100 m that more finely capture convective structures, please clarify how the proposed framework would be extended to incorporate these higher-resolution inputs for future flood projection.

3.    In the absence of a detailed sewer network, a uniform drainage rate is subtracted over road surfaces. How realistic is this assumption under extreme rainfall events at the 100-year return level? The authors should clarify whether drainage saturation or failure could alter the relative changes in inundation depth and duration under warming.

4.    The results indicate that storms become more intense but spatially smaller, yet flood depth and velocity still increase. It deserves deeper discussion why reduced storm areas result in more severe pluvial flood hazards.

5.    The framework is positioned as broadly applicable, yet it is tailored to convective storms and pluvial flooding. Please more clearly delimit the domain of applicability. In particular, under what climatic and surface conditions would the current DSQM–SST–AUTOSHED cascade not be expected to perform well, and how might this influence use by practitioners?

6.    The terms ‘regional warming levels’ and ‘warming scenarios’ are used interchangeably. Please choose one primary term, such as ‘regional warming levels’, and define it clearly once (relative to 1998-2019 baseline). Use it consistently in figures, captions, and text. 

The paper provides an interesting case study of the impacts of climate change on flooding behaviour in a large urban area (300km², Beijing), where the key highlight of the paper is the explicit consideration of the changes in spatial characteristics of the rainfall fields under climate change. This is an important topic: the incorporation of morphed rainfall fields derived from SST rainfall fields is a clever approach to a thorny problem, and the case study nicely demonstrates how such information can be used to assess the impacts of climate change on floods over a large urban area. The paper is also well written.

Overall, I think the paper makes a good contribution and would be of interest to readers of this journal. However there are three key aspects that should be given additional consideration by the authors: 1) there is a need to further qualify the scope and utility of the proposed “framework” (ie in places there is a tendency for over-statement), 2)  the methodological limitations for estimating floods of a given return period needs qualification and further discussion, and 3) the impracticality of the computational burden needs to be more thoughtfully discussed. I provide some justification for these points below (and given I am citing some of my own co-authored papers in support of these statements it is appropriate that I am identified).

With regard to my point 1), the title of the paper and high-level messaging presents the work as a “computationally efficient process-informed framework”, where the key results are presented as the impacts of climate change on floods of a specific return period. I don’t agree that the approach can be referred to as being computationally efficient, at least not for practical purposes, and that references to the approach as being “a proposed framework” is an overstatement, at least as regards the estimation of floods of a given return period. The assessment of the impacts of climate change are restricted to assessing changes in average areal rainfall intensity and the spatial extent of storms, and does not include the impacts of temporal patterns or antecedent soil moisture. Stating these limitations is not intended to be criticisms of the paper – it is entirely appropriate that the authors set the bounds of their research as suits their purposes – but rather these aspects point to the need for the key contributions of the paper to be more clearly stated and qualified.

In my view the strength of the paper is as a case study which demonstrates the impacts of climate change on large floods in a large urban area, where specific attention is given to characterising the impact of higher temperatures on the average areal intensity and spatial extent of rainfall events. The stochastic characterisation of the rainfall fields under climate change (steps I and II in Figure 1) provide a valuable framework for examining the impacts of climate change on the intensities and extents of storms rainfalls of a given return period (as already published in an earlier Water Resour Res paper by Zou et al, 2025), but I don’t think the deterministic transformation of these rainfalls into floods using a rain-on-grid hydrodynamic (RoGH) model in the manner demonstrated (Step III in Figure 1) is sufficient for this contribution to be regarded as a “computationally efficient process-informed framework” for assessing flood impacts. I provide reasons for this below.

With regard to my point 2), there are two key limitations to the adopted approach which need to be addressed in some fashion, at least in the discussion: a) the weakness of the validation of the RoGH modelling component, and b) the restrictive assumptions used to estimate floods of a specified return period.

In terms of scientific defensibility of the results, the weakest aspect of the paper – whether regarded as a case study or a methodological framework - is the lack of appropriate validation of the RoGH model. The evidence for validation of the SST rainfall model is provided in Table S3 in supplementary material (this gives good confidence in the SST methodology), but no equivalent standard of evidence is provided for the RoGH model. There are two elements to this, namely parameterisation of the routing scheme, and the manner in which flood return periods are estimated. With respect to parameterisation of the routing scheme, the paper compares simulated inundation levels against maximum levels recorded at six underpasses for a single historic event. While these comparisons against maximum levels for a single historical event are very good, particularly for what is assumed to be the largest areas (points 5 & 6), the slight problem here is that the routing parameters were verified using rainfall with a spatio-temporal resolution of 500m and 5min, which is substantially finer that what is used for the SST simulations. Given that model results are dependent on the resolution of both the rainfall inputs and the routing scheme, it would have been better to demonstrate model performance using historical information aggregated to match what is used to derive the key results of interest. Thus, could the authors provide comment on, or evidence for, the dependence of these calibration results on the finer resolution adopted.

Of more importance is that no attempt has been made to demonstrate that the adopted modelling scheme is able to reproduce flood frequency quantiles derived from gauged streamflows (ie that the scheme is suited to the estimation of floods with known return periods). We know from Table S3 that the rainfall inputs are suitable for this, but it is a big leap of faith to assume that this guarantees that the adopted scheme can reproduce flood quantiles. The key problematic – and implicit – assumption being made here is that a Y-yr rainfall input to the model will yield a Y-yr flood peak. The limitations of this assumption have been recognised for some decades, and for a recent detailed discussion of this in the context of event-based modelling under climate change see Nathan (2025). The rigorous solution to this problem is to ensure that the SST inputs sample an appropriately wide range of rainfall return periods and to take explicit account of the joint probabilities involved, where the most appropriate validation test is to compare the derived flood quantiles with those based on observed flood data (which is the same comparison as undertaken in Table S3, but for flood maxima). To this end, is there any gauged flood information available in the study area that could be used to compare flood quantiles? (Such data could also be used to assess the simulation of additional historic events). It would be surprising if not, but this is a point that needs to be discussed. If no information is available to do this, then perhaps the easiest solution within the scope of this paper is to make it clear that the flood results are results are based on an ensemble of 100-yr rainfall intensities, they do not necessarily represent 100-yr floods. That is, it should be discussed that the adopted modelling approach is not well suited to estimating floods of a given return period, but rather the results are assumed to be indicative of the expected changes in floods of a notional return period. In terms of assessing defensibility, the nature and availability of gauged streamflow information across the study area should thus be discussed to provide context for the apparent paucity of relevant information. Also, while I appreciate that the code for AUTOSHED is publicly accessible, there does not appear to be any prior published paper or source that full describes the model or evaluates its application to a real world event – the provenance of this model is cited as Li et al (2023b) but as far as I can tell this is merely a conference abstract?

With regard to point 3), my reasons for stating that the scheme should not be considered as being computationally efficient is that it is stated that the model takes 1 hour to simulate a single 24-hour storm. Thus, with the adopted approach it thus presumably took 50 hours of computational time to derive results for an ensemble of 50 sets of rainfall realizations to derive an approximate estimate of a flood of a given return period (line 268 – I think that’s right?). However, to correctly estimate the return period of a 100-yr flood using this scheme it would be necessary to use all 500 rainfall SST realisations which would take 500 hours (~21 days). These additional simulations are required as it is expected that rainfalls with average areal intensities ranging from around 10-yr to 500-yr will contribute to a 100-yr flood due to the influence of different spatio-temporal patterns (the authors could easily check this statement by inspecting a simple scatter plot of average rainfall intensity against flood peak for a given location – Fig 6a in the manuscript shows that rainfalls of different spatial attributes with average intensities varying between 30-yr and 200-yr contribute to a the expected probability of a 50-yr rainfall quantile (200mm), and this range will increase markedly when propagated through to flood quantiles). Again, I stress that I am not being critical of what the authors have done – I appreciate the effort and quality of this work – I am making these points so that the authors don’t undermine the value of their contribution by making claims that extend beyond the boundaries of their research question. From my perspective the authors have demonstrated just how computationally impractical it is to rely solely on the use of RoG hydrodynamic models – the run times are just too great to solve these problems in a practical fashion with scientific defensibility – and the only useful solutions that I am aware of is to use these models in combination with some form of surrogate model to derive a result (there has been a sharp increase in the number of paper on this topic in recent years, see for example the examples and literature cited in Fraehr et al, 2024), or use low fidelity hydrodynamic models. It would thus be useful if the authors could comment further on the implications and limitations of the computational burden of their RoGH scheme for use in the SST stochastic framework, particularly if used to estimate floods with a given return period.

Following on from this point, one aspect of the paper that I think warrants more nuanced discussion is the exploration of the influence of spatial resolution presented in Section 4.1. This kind of investigation is a valuable addition to the paper as it provides the means to assess the trade-off between computational burden and model accuracy. As present the authors conclude that it is necessary to retain the original high-resolution results. This may be the case, but it could be argued that using a coarser model resolution yields a more accurate estimate of the 100-yr flood because the lower computational overhead makes it is feasible to properly resolve the joint probabilities involved in the making statistical inferences using the ensemble of spatio-temporal patterns over the required range of average rainfall intensities. That is, when a coarse model is deployed in a more rigorous stochastic joint-probability framework the resulting estimate of the 100-yr flood based on the comprehensive analysis of a wide range of rainfall quantiles may be more accurate than the one obtained using a fine resolution model based on only 50 simulations forced by a single rainfall quantile. Thus, it would be helpful if the authors could offer any insights about the relative magnitude of variability represented by the ensemble of spatio-temporal SST rainfall fields compared to the additional epistemic uncertainty associated with the use of a coarser hydrodynamic model. Some indication of this could be obtained (without undertaking additional simulations) by comparing the range of inundation levels associated with the ensemble of SST rainfall fields with the difference in levels associated with different model resolutions.

Lastly, I note that the morphing of rainfall fields used in SST considers the influence of thermodynamic components on areally averaged rainfall intensities and spatial extents. This approach has a lot of merit and is a valuable contribution to the problem, though as far as I can tell the approach, while providing useful evidence for changes in spatial extents of storms (a topic which has received little attention in the past), it doesn’t consider changes in the temporal distribution of rainfalls. I can see how this aspect could be incorporated into the morphing scheme, but this aspect does not yet appear to have been explicitly characterised. Temporal distributions have a major influence on flood response (eg see Hettiarachchi et al, 2018) and there is evidence (eg Fadhel et al 2018; Wasko and Sharma, 2015; Visser et al, 2023) that temporal patterns are becoming more front-loaded with higher temperatures. The authors could perhaps comment on this additional limitation in Section 4.3.

Fadhel, S., Rico-Ramirez, M., Han, D. (2018): Sensitivity of peak flow to the change of rainfall temporal pattern due to warmer climate, J Hydrol Volume 560, 2018, Pages 546-559

Fraehr, N., Wang, Q. J., Wu, W., & Nathan, R. (2024). Assessment of surrogate models for flood inundation: The physics-guided LSG model vs. state-of-the-art machine learning models. Water Research, 252, 121202.

Hettiarachchi, S., Wasko, C., and Sharma, A.: Increase in flood risk resulting from climate change in a developed urban watershed – the role of storm temporal patterns, HESS, 22, 2041–2056.

Nathan, R. (2025). Improving event-based methods for modelling flood risk in a variable and non-stationary climate. Phil Trans Royal Society A, 383: 20240292.

Visser, J. B., Wasko, C., Sharma, A., & Nathan, R. (2023). Changing Storm Temporal Patterns with Increasing Temperatures across Australia. J Clim  36(18), 6247–6259.

Wasko, C., & Sharma, A. (2015). Steeper temporal distribution of rain intensity at higher temperatures within Australian storms. Nature Geo, June, 8–11.