Simulating river discharges variations and flood events from large-scale atmospheric information with statistical and dynamical downscaling models: Example of the Upper Rhône River

Floods are highly destructive natural hazards causing widespread impacts on socio-ecosystems. This hazard could be further amplified with the ongoing climate change, which will likely alter magnitude and frequency of floods. Estimating how flood-rich periods could change in the future is however challenging. The classical approach is to estimate future changes in floods from hydrological simulations forced by time series scenarios of weather variables for different future climate scenarios. The development of relevant weather scenarios for this is often critical. To be adapted to the critical space and time scales of the considered basins, weather scenarios are thus typically produced from climate models with downscaling models, either dynamical or statistical.

In this study, we assessed the ability of two typical simulations chains to reproduce over the last century (1902-2009) and from large-scale atmospheric information only observed temporal variations of river discharges and flood events of the Upper Rhône River (10,900 km²). The modeling chains are made up of (i) the atmospheric reanalysis ERA-20C, (ii) either the statistical downscaling model SCAMP (Raynaud et al., 2020) or the dynamical downscaling model MAR (Gallée and Schayes, 1994), and (iii) the glacio-hydrological model GSM-SOCONT (Schaefli et al., 2005).

The daily Mean Areal Temperature (MAT) and Precipitation (MAP) time series were compared to the observed ones over the period 1961-2009. The meteorological results highlight the need for a bias-correction for both downscaling models. To avoid irrelevant simulations of the snowpack dynamics, especially for high elevations, the bias-correction was needed not only for the precipitation and temperature scenarios but also for the lapse scenarios of the dynamical downscaling chain. Simulated discharges are globally in very good agreement with the reference ones in the bias-corrected simulations. Whatever the river basin considered, the multi-scale observed variations of discharges are well reproduced (daily, seasonal and interannual). The reconstruction power of the chains is lower for low frequency hydrological situations, namely low flow sequences and annual discharge maxima. Flood events tend to be underestimated by each simulation chain.

Flood activity was also estimated from the discharge time series using the Peak Over Threshold (POT) method. The results over the last century are very promising, and encourage us to continue towards simulations over the last millennium. Outputs from the PMIP4 experiments (CESM1 Last Millennium Ensemble) will be statistically downscaled with the SCAMP model (for reasons of computation costs) and used as forcings in the GSM-SOCONT model.

References: 
- Raynaud et al. (2020) HESS doi.org/10.5194/hess-24-4339-2020 
- Gallée and Schayes, 1994 MWR doi:10.1175/1520-0493(1994)122<0671:DOATDM>2.0.CO;2
- Schaefli et al. (2005) HESS doi.org/10.5194/hess-9-95-2005