Reply to Anonymous Reviewer #1

We thank the reviewer for the thorough and helpful review of our manuscript and the generally positive feedback. Here below we explain how the comments of the reviewer will be addressed. Comments are quoted in italic.

1) Maybe the following lines (i.e., lines 278-280) could be somewhat extended for including as much information as possible from the analysis outputs: “In general, few models only predict the worst meteorological drought statistics during 2015-2100 with respect to 1850-2014 observations, and MME does not resolve the problem as it delivers a less conservative prediction with respect to past occurrences of multiyear droughts”. For instance, the “few models” and the “worst meteorological drought statistics” could be reported. Discussing the related analysis outputs in greater detail would be beneficial, to my view, as they consist one of the most interesting parts of the paper.

We will follow the suggestion of the reviewer thus explaining with more details the results from our analysis. Specifically, the text at lines 278-280 of the original manuscript will be revised as follows:

“In conclusion, the MME consistently predict less concerning meteorological drought statistics during 2015-2100 with respect to 1850-2014 observations, with the only exception of a slight increase of DF under SSP2.6. Furthermore, the worst drought predictions are provided by IPSL-CM6A-LR under SSP8.5, which however exhibits about 58% underestimation of DF.”

2) Additional recommendations for future research (aside from the general directions already provided in the paper) could be added (e.g., in the “Conclusions” section). These recommendations could include the application of the methodological framework of this work to other areas around the globe. Discussions on the minimum observed data availability requirements for such an application would be also beneficial, to my view.

We agree with the reviewer. Accordingly, the following sentence will be added to the conclusions of the paper:

Our results suggest that validation at local scale of GCM simulations is an essential step to inform downscaling procedures and correction techniques, to make sure that model predictions are consistent with the local features of climate. However, extreme events like multiyear droughts are unfrequent and therefore validating their predicted statistics is particularly challenging. Further research efforts are needed to condition climate models for drought simulation according to our up-to-date knowledge on the historical and recently observed features of drought events. For the actual generation of future drought scenarios and designing climate change adaptation actions, stochastic simulation aimed at replicating the statistics of historical time series, with appropriate updates to account for climate change, is still the most reliable option to make a synthesis of the available information.

3) In line 233, it is written that “the ensemble mean performs better than any individual model for all indices”. However, according to Figure 10, the ensemble mean performs better than the individual models for most, but not all, indices. Note that, for instance, it exhibits worse performance than GFDL-ESM4 in terms of the index referring to very heavy rainfall days (R20mm).

Thanks for the comment. To be more precise, we will change the last sentence into “The ensemble mean performs better than any individual model for most of the indices. In general, the ensemble mean better reproduces the extreme indices than individual model.”

4) Figure 9 could be extended for providing information about the seasons December-January-February (DJF) and June -July-August (JJA) as well.

We agree with the reviewer. We will extend the figures for seasons DJF and JJA and also add the corresponding description for the result in the main text.

5) Some amendments could be applied to Figures 4, 5 and 7 for improving their readability. Specifically, for all the sub-figures belonging to each of these figures, the axes limits could be set the same. Moreover, a note could be added to the caption of Figure 8 for explaining that the thick lines represent the observed time series and the ensemble mean or, even better, the legend could be amended for providing this information. Lastly, the text size in Figure 6 could be increased.

Thanks for the suggestion. For Figures 4, 5 and 7, we will adjust the axes for sub-figures to be more readable. We will modify the legend of Figure 8 as suggested and will increase the text size in Figure 6.

Reply to Anonymous Reviewer #2

We thank the reviewer for the thorough and helpful comments. They are very useful to improve the clarity of our manuscript. Here below we explain how the comments of the reviewer will be addressed. Comments are quoted in italic.

Page 1, line 16 and page 6, line 122: Check for the references so that they appear in chronological order. 

We will check the references and make the changes suggested by the reviewer.

Page 14, Figure 8: The lines representing both, the observed and ensemble data, should be thicker in the legend to really appear as they have been plotted in the figure. 

Thanks for your suggestion. We will modify the legend for observed and ensemble data according to the suggestion of the reviewer (see also our reply to comment #5 by reviewer #1).

Page 15, Figure 9: I suggest changing the symbol of the ensemble set (maybe a star or a square) to be easily identified. As it is now it is difficult to differentiate it from the ACCESS-CM2 set. 

We agree with the reviewer and therefore will change the symbol of the ensemble set accordingly.

It would be interesting to know R_LT values by maybe including them in Table 4. 

Thanks for your suggestion. R_LT values represent the long-term mean value of annual rainfall for each model and observed data. However, in Table 4 we want to show the statistics for multi year drought events rather than rainfall to be more focused on drought. Therefore, we prefer to remain the current format of Table 4.

Reply to Anonymous Reviewer #3:

We thank the reviewer for the thorough and helpful comments. They are very useful to improve the clarity of our manuscript. Here below we explain how the comments of the reviewer will be addressed. Comments are quoted in italic.

The manuscript is well written and the tables and figures are useful and adequately presented. My only comment refers to the fact that the use of historical time series to predict the future occurrence and characteristics of multiyear drought is not really included in the work, which should be more explicitly said when retrieving conclusions on the comparison of historical time series and future projections by GCMs.

The comment of the reviewer is appropriate. The address it, we will add the following text in the conclusions of the paper:

For the actual generation of future drought scenarios and designing climate change adaptation actions, stochastic simulation aimed at replicating the statistics of historical time series, with appropriate updates to account for climate change, is still the most reliable option to make a synthesis of the available information.

Lines 149-150. Please, include some comment on the choice of these threshold values. Are they scaled in Figure 3?

To address the comment of the reviewer the following sentence will be added at the end of line 150 of the original manuscript: “The threshold values have been identified with a trial and error procedure by verifying that relevant droughts observed in the past have been consistently identified.”

The threshold values will be scaled in Figure 3.

Lines 193-196. This paragraph starts by including all models in the same category, with poor capability to represent annual rainfall during the historical period, including their ensemble result, but the final sentence points out to the latter reproducing the mean of the observations. I suggest to redact this more clearly.

The text in lines 193-195 of the original manuscript will be corrected as follows:

All models and the ensemble show negative values of NSE which indicate an overall poor ability in replicating historical annual rainfall. ACCESS-CM2, MPI-ESM2-2-LR, and GFDL-ESM4 depict relatively poorer performances for both periods with the ACCESS-CM2 showing the less satisfactory ones. The multi-model ensemble and FGOALS-g3 exhibit a higher NSE with a value close to zero.

Figures 6 and 7. Please, add “of annual rainfall” in the captions.

Thanks for the suggestion, we fully agree with it and apologise for the lack of clarity. We will modify the captions accordingly.

Figure 8. Please, increase the width of the lines in the legend for observations and ensemble to facilitate their identification in the graph.

We agree with the reviewer. We have will modify the legend for observed and ensemble data accordingly.

Table 4. Are DD and DI mean values during the studied period? I suggest writing “Some statistics” in the caption, instead.

We will modify the caption of Table 4 as follows:

Table 4. Mean values over the considered period of drought frequency (DF), duration (DD), intensity (DI) and maximum deficit (MD) for multiyear meteorological droughts exhibited by observed data (1850-2014) and reproduced by models for the historical (1850-2014) and future (2015-2100) periods under the two considered scenarios.

Line 245. I would drop the use of “significantly” here, since no significance test is really done, even if the values show this apparent difference. This also holds in other places in the text (e.g. line 267).

We agree with the reviewer. Accordingly, we will drop the word “significantly” in line 195, 212, 245 and 291 and will replace “significant” with “evident” in line 257, 265 and 267. We will drop “significant” in line 227. We will also replace “significant” with “considerable” in line 221, “significantly” with “considerably” in line 258. We will finally replace “significant” with “much” in line 289.

Figure 11. The 30-yr moving average for the projections under the different future climate scenarios could also be added as in the historical observations.

We agree with the reviewer that Figure 11 is not easy to read. To avoid including additional lines that may distract the reader, we will remove the moving average line for historical data.

Line 285. I would write “ of SOME statistics”, not so general as it is in the text.

We agree with the reviewer and will amend the text accordingly.

Lines 290-294. Related to my previous comment on Table 4, these sentences would then refer to mean values and, thus, these comments should clarify that less critical mean behaviour are produced by models, although extremes are not assessed. This might also affect the run theory application if alternating extremes take place, resulting in less drought events being identified in the future projections.

We agree with the reviewer and therefore will change the text as follows:

For multiyear meteorological droughts, that is our main focus, we pointed out above that the multi-model ensemble can satisfactorily simulate their mean frequency while significantly underestimating mean duration, intensity and maximum deficit. Additionally, the mean duration, intensity and maximum deficit predicted by models in the future are generally less critical than what was observed in the past. Some models may individually and occasionally predict more hazardous values for some drought statistics under climate change. However, it is not possible to consistently identify a more precautionary model and therefore no indication can be derived for making predictions when no observed data are available for model validation.

Reply to Anonymous Reviewer #4:

We thank the reviewer for the thorough and helpful comments. They are very useful to improve the clarity of our manuscript. Here below we explain how the comments of the reviewer will be addressed. Comments are quoted in italic.

Section 3.1 Authors describe the cumulative annual rainfall comparison. I would avoid to mention “hyetograph” since, for a moment, I was disoriented thinking that the single hyetograph within the years were included in the analysis. Maybe simply “time series” would be appropriate.

We agree with the reviewer and will then change “hyetograph” with “time series”.

Section 3.2. While in Section 3.1 detailed information and equations are provided (Eq. 1-6), in this Section details are not provided as well. I would give more details on the Taylor diagram, on how to read it and explaining the three considered parameters.

We will add few more details to interpret the Taylor diagram and will invite the reader to refer to Taylor (2001) for more information.

Section 3.3. Authors compare eight extreme rainfall indexes using RMSE, however they limit the evaluation in comparing each model respect to the others not offering the single performance. I would suggest to add a plot to the Figure 10 that allows the reader to figure out the single model attitude to correctly simulate extreme values. A Relative Absolute error could be appropriate.

We understand the concern of the reviewer but we would prefer not to include an additional plot. Therefore, we propose to indicate in the text the median root mean square error (RMSE), which will then allow to infer the RMSE of each model basing on the results presented in Figure 10.

Section 5, 6 , and 7. I am impressed by the results. Above all by Figure 4, 5, 6, and 7. Table 3 as well. Authors well commented the results, however in the conclusion they could be clearer. Indeed the sentences present in lines 284, 286, and 302 seem in contradiction to the obtained results. From Figure 4 and 5, ensemble plots can not deserved any reliability, from Figure 6, the minimum MARE is more than 20%, finally looking Figure 7 and Table 3, moments and autocorrelation are always misunderstood. It is clear to me that, for Bologna time series, the GCM are not capable to reproduce rainfall and this confirmed in the drought analyses and related conclusions.

We agree with the reviewer and therefore changed the text in the conclusions as follows:

The present study refers to the region of Bologna, where the availability of a 209-year-long daily rainfall series allows us to make a unique assessment of GCM reliability and their predicted changes in rainfall. GCMs provide a satisfactory simulation of rainfall seasonality but other statistics are not consistently reproduced. Statistics are not well reproduced by the mean ensemble simulation and most individual models, which predict very different conditions and therefore the identification of the future climatic forcing remains a challenge. With respect to droughts, GCMs’ predictions for the future generally deliver a worse picture with respect to present day simulations, but our results suggest carefully considering the impact of uncertainty when designing climate change adaptation policies.

For some situations, classical engineering methods for critical event estimation under the assumption of stationarity may turn out to be more precautionary. Therefore, rigorous use and comprehensive interpretation of the available information are needed to avoid mismanagement, by also taking into account that the impact of multiyear meteorological droughts is likely to be exacerbated by further pressure on water resources due to increasing population and water demand.

Our results suggest that validation at local scale of GCM simulations is an essential step to inform downscaling procedures and correction techniques, to make sure that model predictions are consistent with the local features of climate. However, extreme events like multiyear droughts are unfrequent and therefore validating their predicted statistics is particularly challenging. Further research efforts are needed to condition climate models for drought simulation according to our up-to-date knowledge on the historical and recently observed features of drought events. For the actual generation of future drought scenarios and designing climate change adaptation actions, stochastic simulation aimed at replicating the statistics of historical time series, with appropriate updates to account for climate change, is still the most reliable option to make a synthesis of the available information.

We thank very much Dr. Yves Tramblay for providing a community comment. The opportunity of getting a feedback from peers other than the reviewers is a distinguished feature of HESS. Therefore we are extremely grateful to colleagues that take this opportunity. The critical part of the comment reads as follows:

However, there is a problem with one aspect of the methodology:  GCM simulations driven by radiative forcings are not intended to reproduce the chronology of internal variability. As a consequence, it is irrelevant compute a correlation coefficient, or Kling-Gupta efficiency (KGE) and Nash-Sutcliffe Efficiency (NSE) criterions, between time series of observed precipitation and precipitation simulated by individual GCM runs (Jiang et al., 2013). Then, Figures 4 and 6 and lines 177-203 are not appropriate. The conclusion that “historical sequence of annual rainfall is not satisfactorily reproduced” is know even before reading the paper, since climate models are not deterministic models intended to reproduce a chronology.

We agree with Dr. Tramblay that previous contributions highlighted that it is not fair to compare GCM simulations of past climate with observations, as GCMs are not meant to reproduce the observed climate. However, we think that a comparison between the chronological orders of simulation and data is still a useful information, as we believe that it is not yet clear to the community what performances can be expected from GCMs for specific statistics like the chronological sequence of annual rainfall. Therefore, we believe that such a comparison is an added value to our paper and thus would prefer to keep it.

Reply to the Editor’s comment

We would like to thank the Editor for the careful assessment of our contribution. It is constructive and helpful to improve our presentation. The editor raised two major concerns that are summarised by the following quotes from the Editor’s comment:

-    “I suggest the authors to re-consider the sections of their work that focus on comparing different statistics between the historical observations and GCM simulations, rather than comparing concurrent time series.”

-    “They should also carefully consider the potential biases introduced by the area to point estimation issue.”

Regarding the first issue, we fully agree with the Editor that GCMs should be evaluated by assessing their capability of reproducing the statistics of observed data, including the progress of statistics in time. Particularly in our case, we believe it is important to assess whether the evolution along time of rainfall statistics in Bologna is well reproduced by GCMs. In fact, the only comparison of the probability distribution of annual data over the full observation period does not provide enough information on the capability of models to predict how climate will change in the future. For instance, the probability distribution would not change if the sequence of annual rainfall is shuffled therefore eliminating change and persistence. To make a comprehensive assessment of the capability of models to reproduce change, it is also necessary to present a comparison of statistics for common subperiods.

However, we recognize that the annual subperiod may be too short for a meaningful assessment of statistics (note: annual rainfall is a statistic computed on the observed and simulated daily observations), and therefore we recognize the potential weakness of our approach in this respect.

We believe that such weakness can be resolved by substituting in the revised manuscript the comparison of annual rainfall with the comparison of cumulative rainfall over longer subperiods of 10 years (we are willing to also make the analysis for the 20-year window if the Editor believes that this would be a useful addition). The resulting plots of 10-year concurrent rainfall would be affected by less variability with respect the plots of concurrent annual rainfall and would therefore be easier to read. For the quantitative assessment of model performance to predict change, we would prefer to maintain the use of NSE, KGE and MARE to evaluate the goodness of the fit. These indexes can be fruitfully used for comparing series of concurrent statistics. However, we are willing to use other indexes if the Editor has a different suggestion.

We also recognize that the wording through the paper needs to be revised to better emphasise that we are comparing statistics for common subperiods  and not observations. For instance the concluding sentence of the abstract (quoted by the Editor in their comment) which reads:

“The results prove that rainfall statistics, including seasonal patterns, are fairly well simulated by models, while the historical sequence of annual rainfall is not satisfactorily reproduced” 

will be changed to 

“The results prove that rainfall statistics for the full observation period, including seasonal patterns, are fairly well simulated by models, while the progress of 10-years rainfall along time is…..” (to be completed).

Regarding the second issue, we decided to use bi-linear interpolation of rainfall from the 4 grid points around the location of Bologna to estimate point rainfall. We selected bi-linear interpolation after trying different spatial interpolation methods such as weighted inverse distance and nearest-neighbour interpolation and checking that the results did not change much. Bilinear interpolation is also common used to alleviate the scale problem of a mismatch between the coarse grid and station point. (Bracegirdle and Marshall, 2012; Zhang et al., 2022). We are willing to incorporate these details in the revised version of the paper.

We agree with the Editor that subgrid spatial variability may be underestimated by interpolating grid rainfall, as convective rainfall may be not well reproduced at the grid scale. However, we analyse annual rainfall for detecting drought frequency, whose variability in space for the Bologna region and the considered grid size can be assumed to be negligible. Support to the above assumption is provided by the annual climatic reports by the Regional Agency of Environmental Protection, which are presented at https://www.arpae.it/it/temi-ambientali/meteo/report-meteo/rapporti-annuali for the past 5 years. Each of these reports presents maps of the spatial distribution of each year’s cumulative rainfall over the region. Such maps show that the variability is essentially governed by ground elevation, which is similar for the considered grid points in the bi-linear interpolation. Therefore we believe that our procedure does not introduce a systematic bias. We also would like to refer to the maps presented by Antolini et al. (2016) which confirm the low variability in space for spatial and long-term seasonal rainfall as well.

We are willing to include in the revised version of the paper a discussion on the potential impact of subgrid processes (note that convective processes do not contribute significantly to annual rainfall), a discussion of our assumption above and its justification for the purpose of our analysis.

We agree with the Editor that neglecting subgrid variability may be not justified for high rainfall events, which are originated by convective processes. Therefore, in the revised version of the paper we propose to remove the analysis of the extremes (sections 3.3 and 5.1.3), which we also recognize is not very relevant for drought risk assessment.

To summarise, we propose to make the following changes to the revised manuscript to resolve the concerns of the Editor:

1)    Substitute the comparison of annual rainfall statistics with 10-years rainfall statistics;

2)    Provide a discussion of subgrid variability and an expanded description and discussion of our assumption in this respect and the interpolation method that has been used;

3)    Remove the analysis of rainfall extremes;

4)    Revise the wording in the paper according to the above changes, by in particularly emphasizing the value and interest of comparing rainfall statistics computed over the whole observation period and common subperiods.

Of course, we will also incorporate the necessary changes to address the concerns presented in the reviewers’  and community comments we received. We hope that the present reply better clarifies the content and purpose of our analysis. Once again, we are grateful to the Editor for the useful review.

References

Antolini, G., Auteri, L., Pavan, V., Tomei, F., Tomozeiu, R., & Marletto, V. (2016). A daily high‐resolution gridded climatic data set for Emilia‐Romagna, Italy, during 1961–2010. International Journal of Climatology, 36(4), 1970-1986.

Bracegirdle, T. J. and Marshall, G. J.: The Reliability of Antarctic Tropospheric Pressure and Temperature in the Latest Global Reanalyses, Journal of Climate, 25, 7138–7146, https://doi.org/10.1175/JCLI-D-11-00685.1, 2012.

Zhang, H., Zhang, F., Zhang, G., and Yan, W.: Why Do CMIP6 Models Fail to Simulate Snow Depth in Terms of Temporal Change and High Mountain Snow of China Skillfully?, Geophysical Research Letters, 49, e2022GL098888, https://doi.org/10.1029/2022GL098888, 2022.