EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Can the latest generation of regional climate models reproduce European snow conditions and how do biases translate into uncertainties of snow cover projections?

Katharina Bülow1, Sven Kotlarski2, Christian Steger3, and Claas Teichmann1
Katharina Bülow et al.
  • 1Climate Service Center Germany (GERICS), Helmholz Zentrum Geesthacht, Hamburg, Germany (
  • 2Federal Office of Meteorology and Climatology MeteoSwiss, Zurich, Switzerland
  • 3Institute for Atmospheric and Climate Science, ETH Zurich, Zurich, Switzerland

Snow cover is a crucial part of the climate system due to its distinctive alteration of surface reflectance (snow-albedo-feedback) and its influence on further physical surface properties (e.g. heat conduction and water storage). These effects are particularly relevant in alpine areas and high latitude regions, where snow coverage prevails for a significant part of the season. In addition, various human activities rely on snow cover duration and/or snow amounts, such as winter tourism, agriculture and hydropower production.

The EURO-CORDEX project provides an RCM ensemble with a horizontal resolution of ~50 and ~12 km for both present-day and future climates assuming different emission scenarios. These simulations present a potentially valuable information source for the future snow cover evolution. Prerequisite, however, is the ability of RCMs to reproduce historical snow cover conditions. These issues are addressed in the present work on a European scale. A horizontal resolution of ~12 km allows for an improved representation of topography and is thus particularly interesting for snow cover studies, as snow in alpine regions strongly correlates with elevation. We therefore only consider the high-resolution EURO-CORDEX RCMs and, for the climate projection part, simulations for RCP2.6, RCP4.5 and RCP8.5.

To assess the RCMs’ ability of reproducing current snow cover conditions in Europe, we evaluate simulated snow water equivalent and snow cover duration/extent by comparison against different reanalysis data (e.g. ERA5, UERRA MESCAN-SURFEX) and snow products derived from remote sensing. Regarding the spatial domain, we consider entire Europe with a focus on four mountainous regions (Alps, Norway, Pyrenees and Carpathians). The evaluation reveals that, on an European scale, mean yearly snow cover duration is well captured by the ensemble mean of the models. However, the majority of the RCMs underestimates snow cover extent throughout the season. This bias is more pronounced in the reanalysis (ERA-Interim) driven set of simulations than in the GCM-driven runs. In regions with complex topography, winter snow water equivalent is distinctively overestimated in some simulations - whereas certain grid cells reveal glaciation (i.e. year-round snow coverage). A comparison with E-OBS data indicates that biases in snow cover duration and amount are, besides arising from inaccurate snow schemes, linked to mismatches in simulated air temperature and precipitation patterns. Scenarios for the 21st century show a distinctive reduction in snow cover duration for low-elevation regions, whereas the magnitude of this decrease depends, amongst other factors, on the climate scenario. Projected decreases in the snow cover are less pronounced for medium to high-elevation regions.

How to cite: Bülow, K., Kotlarski, S., Steger, C., and Teichmann, C.: Can the latest generation of regional climate models reproduce European snow conditions and how do biases translate into uncertainties of snow cover projections?, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-9964,, 2020

Comments on the presentation

AC: Author Comment | CC: Community Comment | Report abuse

Presentation version 2 – uploaded on 01 May 2020
Change of the corresponding email
  • CC1: Comment on EGU2020-9964, Michael Matiu, 04 May 2020

    Dear Katharina,
    nice display, and impressive body of work! Looking forward to seeing the final paper!

    I have a few questions:
    1. Do you have an idea about the accuracy of the SWE in the reanalysis & remote sensing products you use?
    2. In your methods you use some constants to convert SWE to SND, and then SND to SNC. I imagine you do this to remove the effects of the model parametrization (and to increase your RCM sample size). However, this has a strong effect on your results. Have you also tried other thresholds e.g. for snow days?
    3. Related to 2.: It's because we did some similar (yet less extensive) validation of the RCMs for the Alps, but used directly the snow cover as parametrized in the models. And the specific RCM results are somewhat different (e.g. SMHI had low bias, but has very high in yours - slide 9).


  • CC2: Comment on EGU2020-9964, Samuel Somot, 04 May 2020

    details: when you mention CNRM on the curves, is it CNRM-ALADIN53 or CNRM-ALADIN63 ? evaluation of the new ALADIN63 is (at least for us as model develpers) more relevant as we can still improve it. The ALADIN53 version being dead. Thanks

    • AC2: Reply to CC2, Katharina Bülow, 05 May 2020

      We used ALADIN63 forced with era-interim for the Evaluation and also ALADIN63 for the scenario runs.

  • CC3: Comment on EGU2020-9964, Samuel Somot, 04 May 2020

    Question: Between the evaluation step and the projection step, did you decide to exclude some models because too bad ? I'm expecting that not all RCMs are good in representing snow variables. If yes, did you measure the impact of the exclusions ?

    • AC1: Reply to CC3, Katharina Bülow, 05 May 2020

      Bonjour Samuel, Indeed the evaluation indicates several poor-performing models. However, as long as their output does not appear to unphysical we leave these models in the ensemble for the scenario analysis. However, several models that show unrealistic snow accumulation issues need to be removed. We have experimented with just excluding accumulating grid cells, not entire simulations, but this does not seem to be an option and it seems we need to exclude these models entirely (see the lines marked with an asterisk on slide 13, for instance). Good idea to check the impact of the exclusion on the final signals. We did not do so for the time being (at least not explicitly).


Presentation version 1 – uploaded on 01 May 2020 , no comments