Review of Ruff and Pfahl (2022): What distinguishes 100-year precipitation extremes over Central European river catchments from moderate extreme events.

In this paper, the authors study very extreme precipitation events over Central Europe river catchments, and the associated atmospheric dynamical conditions, in a robust way, thanks to the use of a very large ensemble of operational weather predictions. This clever and original approach allows them to study a much larger sample of extreme events than it would be possible with standard approaches based on observations or reanalyses.

The paper is generally well written, and discusses interesting results. But there are also some potential methodological issues, the methodological choices are not always well justified, and the implications of these choices are not always discussed.  Therefore, I think that major revisions are needed.

General comments

The authors could have done the same analyses with climate models outputs: large Single Model Initial-condition Large Ensembles (SMILEs) exist and could be used to obtain large sample sizes to do robust analyses of extreme precipitation events. The implicit hypothesis of the authors is, I think, that a weather prediction system provides more accurate representation of extreme precipitation events than climate models. I think this implicit hypothesis should be stated, and discussed, using references.

Precipitation observations are not assimilated in weather prediction systems, I think (is it the case here? It should be discussed), and, in the end precipitation is strongly the result of the atmospheric model, especially at day 10. So, what is really the advantage of a weather predictions system compared to a climate model? Only resolution?

As discussed below, there are also a few important drawbacks to the approach they follow compared to using SMILEs, and therefore I think it is important to discuss these points.

The authors use a clever and original approach to obtain a very large sample of extreme precipitation events to study, thanks to the use of the results of a large ensemble of weather predictions.  But there are some unacknowledged limitations. Even if the ensemble of weather predictions is large, it spans a very short time period for climatological studies (2008-2019) and therefore samples a small sample of sea surface temperature (SST) conditions, for example. It means that it does not sample correctly interannual and low-frequency climate variability e.g. ENSO variations, decadal variations in NAO, AMV etc. These modes of variability or others may impact precipitation extremes. For example, what if the link between atmospheric circulation and extreme precipitation events is different during el Nino and la Nina events? Also, the 2008-2019 period is strongly impacted by anthropogenic forcings. This issue, and how it may impact the results of the study, should be discussed.

More generally (as noted in the specific comments), the authors should sometimes better describe their analyses, better explain why they do them, how they reach their conclusions based on these analyses, and discuss their limitations. A part in the conclusion section should be dedicated to the discussion of the limitations of the analyses.

The physical analysis of extreme precipitation events could have been more developed. For example, the cyclone tracking algorithm is only based on SLP, which might not allow to capture the potentially complex atmospheric circulations associated with extreme precipitation. Also, the authors do not look at atmospheric stability, convective precursors etc., which are likely to play a very important role regarding these events.

Other comments

Section 2.1

Very little is said on the weather prediction system. There are almost no references on the model, assimilation system, on the skill of the prediction system etc. Please add some information and references.

I suppose that the model has evolved during the period studied by the authors, with also changes in the assimilation system and observation networks etc. Am I right?  We really need to have information on the evolutions of the system during the period studied by the authors. 

Given these evolutions, there could be potential issues with the temporal homogeneity in the dataset studied by the authors and the analysis in Fig S1 is far from sufficient to show that it is not the case. At the very least, a discussion is needed on this point.

L116. The authors cite Breivik (2013) to support the hypothesis that precipitation on day 10 of forecasts is independent. But this was with a previous version of the weather prediction system, I suppose. The skill of weather prediction systems increases with time and maybe it is not the case anymore?

Also, what the authors say, i.e. that precipitation from the different members on day 10 is independent implies that there is no predictability of precipitation at 10 days. Is it true? This should be discussed. Is it consistent with what we know about the skill of the weather prediction system? And even if it is true for precipitation, I’m quite sure it is not true for atmospheric circulation and that there is skill at day 10 for SLP, geopotential etc. So maybe precipitation itself from the different members is “independent “at day 10, but it is not the case for the associated atmospheric circulation, which is studied by the authors. What are the implications? I think it may be problematic, for example, to assess the equivalent sample size for the composites of large-scale circulation leading to extreme precipitation events. How is it done? In any cases, this general issue should be discussed.

Figure 2a and b, and near line 205.

It is not totally clear to me how exactly the correlations and auto-correlations are computed. Are they computed at each point and then averaged on the catchment, or is precipitation spatially averaged before computing the correlations and auto-correlations? Is the annual cycle removed before computing the correlations and auto-correlations?

Also, are the correlations and auto-correlations calculated on the complete precipitation series or on the series with only the 10^th forecast day? Based on section 2.1, I assume that it is the second possibility, but it is not so clear in section 3.2. Also, for each day there are two values, corresponding to two initializations, right? How is taken into account in the calculation of daily correlations and auto-correlations? 

By the way, are the auto-correlations significant?

Around L223. The reasoning behind the analyses in Figure 2a and 2b is not clear, and how exactly these analyses are linked to the hypothesis that precipitation from different members on the 10^th forecast day is independent is not very clear. There are no real conclusions regarding this hypothesis.

For example, the authors write “to put the correlation coefficients between the times series into context and also to evaluate the auto-correlation of precipitation time series obtained from one ensemble member” as only justification to the analyses in Figure 2b, with no explicit connection with the previous hypothesis. And in the end, they don’t conclude on the implications of the results for the hypothesis they want to prove.

L252.  “can be considered independent”

What are exactly the criteria to consider them as independent? Could you describe the exact reasoning?

L253. “the data is thus suitable for systematic analysis of very extreme, 100-year precipitation events”. Even if we consider precipitation on the 10^th forecast day from the different members as independent, all data still only come from a 12-year period, and only sample 12 years of SST variability. As said in general comments, this is really problematic.

Note also that the period studied is strongly impacted by climate changes.

These points should absolutely be discussed, and the potential impacts on the conclusions of the paper clearly stated.

L259. What is the algorithm used to fit the parameters of the GEV distribution? Maximum likelihood?

L260. How do the authors deal with the non-stationarity due to climate change? Over the period of the interest climate trends are very likely to be strong, and therefore the simple GEV model used by the authors, which makes a stationarity assumption, is likely to be quite inaccurate. Some approaches to take into account climate trends with GEV statistical models exist. Why didn’t the authors use such an approach? The limitations of their method and its implications should at least should be discussed.

L270-271.

Are all the daily precipitation events greater than the 100-year return level really used for the composite analysis (as it could be understood from the text) or only the events corresponding to block-maxima are used? I.e. if two consecutive days, or days in the same week (or month or semester) are above the 100-year return level, are they all used in the composite analyses? It does not really impact the composites, but it impacts their statistical significance, as it impacts the effective sample size.

L285. The cyclone tracking algorithm is based only on SLP, which is quite basic regarding this kind of algorithm. Is it not problematic to track potentially complex situations leading to extreme precipitation events with such an algorithm? Is SLP not too “smooth” to capture correctly the complex dynamics associated with extreme precipitation events? Could the authors discuss the limitations of such approach or cite studies that show that it is OK to use such tracking algorithm for this kind of events?

Figure 5. It is necessary to add statistical significance in the figure with composites, to demonstrate that the days with extreme precipitation events are really different from the other days. Without significance testing, we don’t really know whether the authors discuss real signals or just statistical noise.

L392-393. How do the authors know that “most of the individual events develop in a similar way as shown in these composites”? Is this based on a sort of test or analysis (e.g. clustering?) or just by looking at all events? If the second option is right, is it really sufficient?  

L 392-407.

It is somewhat strange to spend a long paragraph describing these specific events without showing them. They could be shown in SI.

Fig 461. “in an area that is favourable for cyclogenesis”. Why? The authors could cite some papers.

L514. The authors discuss “Rossby wave breaking” at several places in the paper, even in the conclusion, but show no analysis of Rossby wave breaking.

L545. How do the authors explain the differences with Pfahl and Wernli (2012)?

For five Central-European catchments, authors study the characteristics of extreme precipitation events, mainly from the viewpoint of their causal atmospheric conditions. The study is focused on extreme events with the return period of areal daily precipitation total over 100 years, which the authors compare with a set of events of lesser intensity. For this, they use an innovative approach based on the processing of outputs from the ECMWF ensemble system.

In general, I find the study very interesting, innovative and well-written. Thus, I recommend it for publication in WCD after revisions with respect to the following comments.

Since operational outputs from the ECMWF model are used, I am not sure of the homogeneity of the input data (i.e. whether there was no major change in the model settings during the considered decade). This question needs to be discussed in more detail than only presenting time series of extreme percentile values of daily precipitation totals (S1). It would be necessary to verify whether the extreme events analyzed in the study were randomly distributed within the entire set of model outputs.

On the scale of large Central-European catchments, extreme precipitation events producing large floods regularly last more than one day. Authors mention this fact in conclusions but it is not enough in my opinion. Because accumulation of precipitation is a crucial factor of river floods, it should be mentioned already in the introduction as well as in the discussion at least. It could also happen that two days of extreme precipitation follow one after another in the dataset (as it was e.g. in July 1997) – in such a case, atmospheric conditions are certainly very similar on both days and it could influence the results. 

Moreover, extreme events usually affect more than one of the studied catchments (as e.g. in August 2002) but there is no mention of it in the article. It could be actually interesting and useful to have a 5x5 table giving the frequency of such events, with additional comments of possible cases when even more than two catchments were hit.

Finally, authors distinguish between MEPEs and LEPEs to present what makes the events really extreme; however, there are also events with return periods between 50 and 100 which are worth noting in my opinion. Do they exhibit any "transitional patterns" between MEPEs and LEPEs?

Detailed comments:

l. 17-20: The March/April 2006 flood was not due to an extreme precipitation event; it was a typical snow-melt flood, with the melting process accelerated by the rain-on-snow process.

l. 21: The sentence seems to mean that each of the mentioned flood events had such a large impact. Is it truth? I guess that less people died at least in March/April 2006 because snow-melt floods use to be well predicted in general.

l. 28: In my opinion, the situation described is for Donau in Passau, not Elbe in Dresden – please, check Merz et al. (2014) again.

l. 42: The surface cyclogenesis is due to the upper-level circulation, not vice versa. Thus, I suggest to change the beginning of the sentence as follows: “With such an upper-tropospheric configuration, … is often associated.”

l. 184: It should be mentioned that only a rather small part of the whole Danube catchment is considered.

l. 315: It is not clear whether the presented values are maximum values reached during the maximum event in each catchment or mean values calculated from the MEPEs. I fully agree that the highest values in the Weser/Ems region are due to the rather small area. To compare the catchments, it would be fine to present also maximum pixel values – I guess that the ranking of the catchments would be very different in this case.

Fig. 2a: I guess that the non-zero correlations between the entire daily time series of ensemble members could be due to the seasonal distribution of daily precipitation totals. Thus, the correlation analysis for two main seasons (DJF, JJA) would probably prove the independence of data better than the presented one.

Fig. 2c: I do not fully understand why where is only one simulated total over 40 mm in comparison with REGEN and E-OBS but at least four of them in case of HYRAS. I guess that the simulated precipitation is the same for all datasets, so there should always be three marks at the same height in the graphs here and in S2.

Fig. 4: I find these results really interesting because extreme floods at Rhine, Weser and Ems usually appear in winter. Thus, it would be interesting to compare the monthly distribution of simulated extreme events with the monthly distribution of historical events, if there are data for this at least for some basins (it would be particularly interesting for the Rhine basin, where less extreme events extend into the cold part of the year).

Fig. 8: I agree that the differences between MEPEs and LEPEs in 500 hPa geopotential height and SLP are less significant for Elbe, but what about the horizontal gradients of these variables? This is in my opinion much more important factor influencing the extremeness of subsequent precipitation.

Technical comments:

Terms “database” and “dataset” should be written as single words in my opinion.

If more than one MEPE or LEPE is considered, I recommend to use the abbreviations in the form MEPEs or LEPEs, respectively.

l. 20: “when” should be written instead of “where”;

l. 285: “known” should be written instead of “know”;

l. 432: “catchment” should be written instead of “catchments”;

l. 479 and 481: “on the day” should be written instead of “at the day” in my opinion;

Caption of Fig. 3: I suggest to mention already at the beginning of the caption that it is for all MEPEs but only in Danube catchment.

Caption of Fig. 5: Because it is rather long, I would prefer if you start with a general sentence like “Composites of atmospheric conditions 12 hours after MEPEs started in the Danube catchment”, followed by detailed description of individual maps.

Fig. 6: The color scale should have the same intervals as in Fig. 5a.

Please, check titles of all referenced journals if they are correctly written (l. 656 – Monthly Weather Review; l. 631 – Quarterly Journal of the Royal Meteorological Society should be without the subtitle).

Final comment: because the supplement is online only, it could be longer – readers would certainly appreciate graphs and maps for all considered catchments.