EGU General Assembly 2020
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Surrogate climate change projections for the Lake Victoria region with a convection-permitting model.

Jonas Van de Walle1, Oscar Brousse1, Roman Brogli2, Matthias Demuzere3, Wim Thiery4, and Nicole P.M. van Lipzig1
Jonas Van de Walle et al.
  • 1KU Leuven, Heverlee, Belgium (
  • 2ETH, Zurich, Switzerland
  • 3RUB, Bochum, Germany
  • 4VUB, Brussels, Belgium

Extreme weather is posing constant threat to more than 30 million people living near Lake Victoria or depending on its resources. Thousands of fishermen die every year by severe thunderstorms and associated water currents, while hazardous over-land thunderstorms largely affect people living inland, continuously facing flood risks. These risks call for better understanding of such climate extremes over the region. Climate models are a useful tool to gain insight in the complex behaviour of thunderstorms, especially when simulated at convection-permitting resolution. Such simulations, explicitly resolving deep convection at fine resolutions, have been shown to improve the representation of extreme events in many parts of the world, also in equatorial East-Africa (Finney et al., 2019; Kendon et al., 2019; Van de Walle et al., 2019). As a response, the CORDEX-Flagship Pilot Study “climate extremes in the Lake Victoria basin” (ELVIC) initiative is currently setting up an ensemble of convection-permitting simulations over the region.

At this stage, future climate projections are needed to assess the impact of anthropogenic climate change on extreme weather the region. Therefore, a surrogate global warming approach following Schär et al. (1996), Kröner et al. (2016), Liu et al. (2016) and Rasmussen et al. (2017) has been applied to a convection-permitting COSMO-CLM simulation. In this approach, the lateral boundary conditions from the ERA5 (~31 km resolution) reanalysis are perturbed in accordance with the recent CMIP6 ensemble-mean end-of-century SSP5 8.5 climate change scenario. This approach confers three major advantages over the more conventional methods. First, by perturbing with the ensemble-mean, it excludes uncertainties of GCMs without the need for a time and computational intensive high resolution ensemble approach. Second, it avoids including present-day circulation biases. Third, no intermediate nesting steps are necessary, as the perturbed ERA5 allows a direct downscaling to the convection-permitting climate projection.

Besides the methodology, results for the Lake Victoria basin will be presented. Although the occurrence of extreme over-lake precipitation in the present-day climate is mostly controlled by large- and mesoscale atmospheric dynamics (Van de Walle et al., 2019), its future intensification is mainly attributed to increased humidity (Thiery et al., 2016). Furthermore, the effect of changed large-scale dynamics is assessed, as not only temperature and humidity, but also wind forcing is modified.

How to cite: Van de Walle, J., Brousse, O., Brogli, R., Demuzere, M., Thiery, W., and P.M. van Lipzig, N.: Surrogate climate change projections for the Lake Victoria region with a convection-permitting model., EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-13559,, 2020.


Display file

Comments on the display

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

displays version 1 – uploaded on 30 Apr 2020
  • CC1: Comment on EGU2020-13559, Carlo Cafaro, 08 May 2020

    Dear Jonas,

    thanks for your nice work and presentation. I am writing here because maybe it's easier than the chat.

    I am working for the African SWIFT project and one of areas of our interest is Lake Victoria. I am evaluating CP ensembles at short-range (hours to days) in comparison with lower resolution ensembles.

    I have some questions for you.

    A general comment:
    1) If I am correct you have not used all the ensembles, but just the ensemble mean. So you have assessed the value of CP forecasts, not of the ensembles. 
    Have you looked/will you look at the whole distribution as well ? Are you planning to perform any probabilistic forecast verification ?


    2) Fig:1. It seems that the CP is better in predict the location of rainfall. Have you looked at that ?

    3) Fig. 3: Nice graph, especially the second column.

    4) Fig. 5: I guess the difference is obs-models. Does this depend on the fact that CP overestimates rainfall?

    Carlo Cafaro

    • AC1: Reply to CC1, Jonas Van de Walle, 08 May 2020

      Dear Carlo,


      Thanks for your interest!

      Just to be clear, we are doing climate simulations, no weather forecasts.

      Hope the following answers your questions:

      1) We did look at each member individually as well, but showed only ensemble mean spatial plots. You can see the contributions of different members in the seasonal / diurnal cycles.

      2) As the title suggests, I concluded no added value of convection-permitting simulation over the parametrised one. The lower panels of Fig. 1 are biases against the observational mean, so there is no improvement.

      3) thanks :)

      4) No, all differences are model ensemble mean - observational mean. It’s clear from the figure that the number of rainy events is highly overestimated by the parametrised simulations (consistent for all the members by the way). There is a big improvement of the high resolution simulations.


      Let me know if something is unclear!



      • CC2: Reply to AC1, Carlo Cafaro, 08 May 2020

        Dear Jonas, 

        thanks for your answer. Yes, I realized they were climate projections.

        1) Sorry, I thought that you had an ensemble system for each of the model listed in the table.
        2) Yes, I can see that the CP model is biased as well. My question was about whether you have investigated also the location of the max rainfall, it seems that CP gives some value in predicting it.
        4) Thanks for clarifying the difference.
        Maybe I can rephrase the question: so you defined a rainy event with a fixed threshold. How sensitive are the results to this threshold, given that the model intensity is biased as well ?

        Hope it's clearer.