the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 25 Aug 2025
Uncertainty sources in a large ensemble of hydrological projections: Regional Climate Models and Internal Variability matter
Abstract. Multi-scenario, multi-model ensembles of hydrological projections are widely used to describe possible futures of regional hydrology and inform adaptation strategies. The Explore2 dataset is such an ensemble of river flow projections in Metropolitan France. It provides future simulations for 1,735 catchments with modeling chains composed of different hydrological models forced by 36 regional climate projections based on bias-adjusted EUROCORDEX simulations. This study assesses the uncertainties of this ensemble with QUALYPSO, a method specifically designed to deal with incomplete ensembles and to disentangle and quantify all uncertainty sources, including that due to internal variability.
Focusing on results obtained at the end of the century, this study shows a strong agreement between modeling chains towards decreases in low flows in a large southern part of France for a high-emission scenario, and very uncertain changes for the annual mean and high flows. Emission scenario uncertainty is the dominant source of uncertainty for low flows over the whole of France, and for mean annual flows in southeastern France. The contribution of the global and regional climate models is important for mean and high flows, especially in rainfall-dominated areas. Regional climate models contribute considerable uncertainty to low flows, much more than global models. The contribution of hydrological model uncertainty is large for low flows, moderate for mean annual flows, and small for high flows. For all climate and hydrological indicators, internal variability is often large and cannot be overlooked. It is often of the same order and sometimes larger than the uncertainty on the climate change response.
- Preprint
(18033 KB) - Metadata XML
-
Supplement
(23090 KB) - BibTeX
- EndNote
Status: open (extended)
- CC1: 'Comment on egusphere-2025-2727', Rasmus Benestad, 02 Sep 2025 reply
-
RC1: 'Comment on egusphere-2025-2727', Anonymous Referee #1, 06 Oct 2025
reply
This study explored the sources of uncertainty in different components of the model chain and investigated their contributions to two climate indicators and three hydrological indicators. The variation in performance was evaluated based on different regional characteristics. I consider the motivation of this paper very good, especially in the context of using ensembles for climate projection. The structure of the paper is well organized, and the presentation is good as well. However, I have a few concerns about the calculation methods that need to be resolved before the paper can be accepted.
- What is the purpose of applying cubic splines to the projection and what are the effects on the trend analysis (Line 583)? What is the meaning of the smooth trend denoted as CRi(t)? Please elaborate on the calculation method. Additionally, is the smoothing suitable for precipitation and hydrological indicators (especially max1D)? The authors may need to showcase some results from this step.
- In the estimation of internal variability (Lines 593–600), why does the method first estimate Di(t) as the difference between the raw projection (Yi(t)) and CRi(t), rather than directly simulating variability from the raw projection over the target period? Does this step reduce or increase the internal variability? Based on the results, the internal variability is super large—could this be because the smoothing is not applicable?
- Please elaborate on the calculation of ESi(t), using one example (e.g., RCP, s). Why is a linear regression model applied, and how is it used (Line 608)?
- For Equation (A7), does this equation still work if incomplete or unbalanced ensembles are used? How are the effects of incomplete ensembles reflected in the results? Authors failed to explain this in detail since this is the second major question to be solved.
- What is the difference between IV, RV, and FV? Should they use the same definition but with different superscripts/subscripts?
- Does the selection of the time span length (i.e., 30 years in this study) affect the results, since a longer time span would likely lead to larger internal variability?
Citation: https://doi.org/10.5194/egusphere-2025-2727-RC1
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 2,120 | 63 | 13 | 2,196 | 40 | 36 | 33 |
- HTML: 2,120
- PDF: 63
- XML: 13
- Total: 2,196
- Supplement: 40
- BibTeX: 36
- EndNote: 33
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
I think this paper is very interesting and a welcome contribution. I also appreciate the opportunity to discuss some of of the points made herein.
One point raised is "Model uncertainty arises from model imperfections" which is important, but this paper neglects uncertainties connected with the downscaling approach because it fails to mention others then dynamical downscaling (aka regional climate models, abbreviated as 'RCMs'). There are also other ways of downscaling global climate models (GCMs) which are based on entirely different assumptions and come with different strengths and weaknesses. We expect them to produce similar results if thay all are skillful, independently of each other. Hence, if dynamical and empirical-statistical downscaling give similar outlooks, then the results can be considered as being more robust. Therefore, I recommend that the paper includes some discussion on empirical-statistical downscaling in order to get a more complete picture on uncertainties associated with modelling.
The need of bias-adjustment also introduces uncertainties. It's in a fashion similarto 'sweeping the problem under the carpet', but also it assumes that the present biases are similarto those in a changed climate.
There is at least one example of downscaling precipitation statsistics large multi-model CMIP ensembles that may be of relevance: https://doi.org/10.5194/hess-29-45-2025. However, this example focuses on downscaling daily precipitation statistics and may require an additional step using weather generators to produce time series needed as input for hydrological models. On the other hand, the downscaled precipitation statistics provides a rule-of-thum estimate for number of days per year with heavy rainfall. The method described in this paper will provide a basis for studying the connection between climate change and water-born diarrhoea outbreak in the EU-SPRING project (https://www.springsproject.eu/).
One motivation for downscaling statistical properties (e.g. parameters of statistical distributions) is that statistical properties often are easier to predict/quantify than individual outcomes.
In some cases, climate internal variability (IV) actually provides some useful information about inter-annual variability and the range of plausible outcomes. For example, downscaled results of large ensembles provide a band of plausible temperatures in https://doi.org/10.1073/pnas.2503806122 that can be compared with historical temperatures, and such an evaluation reveals whether the downscaled results match the observed inter-annual variability. The mean of the model spread can be be interpreted as the climate normal, whereas upper and lower limits represent hot and cold years. It is laso interesting to note that the ensemble spread in some cases is close to being normally distributed.
The statement "To our knowledge, the Explore2 MME is the largest ensemble of hydrological projections ever produced from regional climate experiments at the scale of a country" is probably true - see https://doi.org/10.5194/hess-29-45-2025 where MMEs were downscaled for SSP370, SSP126, SSP245, and SSP585, ech with ~30 ensemble members (there are also unpublished results (work in progress) with downscaling total annual precipitation of 200-300 ensembles for each SSP).
When it comes to evaluation, it is not clear if the results are evaluated involving the complete chain of models. I.e. is the downscaling combined with hydrological modelling of GCM historical runs able to reproduce observed trends and inter-annual variability? Also, are the RCMs able to repeoduce past variability and trends?