Projected changes in Rhine River flood seasonality under global warming

warming Erwin Rottler1, Axel Bronstert1, Gerd Bürger1, and Oldrich Rakovec2,3 1Institute of Environmental Science and Geography, University of Potsdam, Karl-Liebknecht-Straße 24–25, 14476 Potsdam, Germany 2UFZ-Helmholtz Centre for Environmental Research, Permoserstraße 15, 04318 Leipzig, Germany 3Faculty of Environmental Sciences, Czech University of Life Sciences Prague, Kamýcká 129, Praha – Suchdol, 165 00, Czech Republic Correspondence: Erwin Rottler (rottler@uni-potsdam.de)

Idealized seasonal distribution of extreme discharge present future Figure 1. Idealised seasonal distribution of nival and pluvial flood frequencies and potential overlap due to climate change.

Data and Methods
The mesoscale hydrologic model (mHM) v.5.10 (Samaniego et al., 2010;Kumar et al., 2013;Samaniego et al., 2018a) is used to detect and assess projected changes in Rhine River floods under future climate conditions ( Fig. 2 and 3). mHM is a spatially distributed hydrologic model based on grid cells. Key feature of mHM is the Multiscale Parameter Regionalization (MPR) technique, which allows to account for subgrid variability (Samaniego et al., 2010(Samaniego et al., , 2017. During MPR, high resolution 5 physiographic land surface descriptors are translated into model parameters. A detailed description of the two phases of MPR, i.e., regionalization and upscaling, is given in Samaniego et al. (2010). In the framework of this study, the high resolution physiographical datasets describing the main features of the terrain, e.g., digital elevation model, aspect, slope, soil texture, geological formation type, land cover and leave area index (LAI), are in 500 m resolution. More information on underlying data sources is presented in Rakovec et al. (2016). 10 Meteorological forcing data of the model consists of daily average, maximum and minimum temperature and precipitation.
mHM forced with E-OBS meteorological data is calibrated against observed streamflow at the three gauges Lobith, Basel and Cochem during 1951-1975 using the Dynamically Dimensioned Search algorithm (DDS; Tolson and Shoemaker, 2007) and the Nash-Sutcliffe efficiency (NSE; Nash and Sutcliffe, 1970). In the framework of this multi-basin calibration, we attain one parameter set, which we apply to the entire basin. In order to evaluate the model performance in all important sub-regions 25 the flow direction in the lower resolution (routing resolution) is equal to the flow direction in the underlying high-resolution grid cell with the highest flow accumulation (Samaniego et al., 2010). The stream celerity is determined as a function of terrain slope (Thober et al., 2019).
All dominant hydrological processes are modelled at 5 km spatial resolution. We estimate reference crop evapotranspiration following the Hagreaves-Samani equation, an empirical approach using minimum climatological data (Hargreaves and Samani, 5 1985;Samani, 2000). The empirical coefficient of the equation is determined during calibration. The usage of this simple approach enables a consistent set-up across historical and future model space. The actual evapotranspiration is estimated based on the fraction of roots in the soil horizons and a stress factor for reducing potential values calculated based on the actual soil moisture. The stress factor is determined using the Feddes equation (Feddes et al., 1976). If the soil moisture is below the permanent wilting point, evapotranspiration is reduced to zero. In case the soil moisture is above field capacity, the 10 evapotranspiration equals the fraction of roots. If the soil moisture is in between the permanent wilting point and field capacity, evapotranspiration is reduced by the fraction of roots times the stress factor. Our model set-up distinguishes six soil layers up to a total depth of 2 m. Organic matter is possible until 0.3 m. In total, more than 2000 soil types with different clay content, sand content and bulk density are defined. Land surface with impervious cover are treated as free-water surfaces and actual evapotranspiration is estimated with an additional evaporation coefficient. More details of the soil parameterization in mHM 15 can be found in Livneh et al. (2015).
The canopy interception is modelled with a maximum interception approach. The maximum interception capacity is estimated based on the given LAI values. Water can leave the interception storage as throughfall, which is estimated as a function of the current and maximum canopy water content and the incoming precipitation. Evaporation from the canopy storage depends on the current and maximum canopy water content and the potential values of evapotranspiration. We simulated snow 20 using an empirical degree-day approach, whereas degree-day-factors differ depending on the dominant land use class. In order to account for snowmelt following the energy input from liquid rainfall, degree-day factors are increased depending on the amount of liquid precipitation. Degree-day factors only can increase to a certain threshold value. Surface runoff from impervious areas is calculated based on a linear reservoir exceedance approach. Interflow from the unsaturated zone is determined using a nonlinear reservoir with saturation excess. Groundwater is assumed as a linear reservoir.

5
The changes in mHM-based flood seasonality are further differentiated and scrutinised for three different warming levels:

10
In order to assess the changes in flood characteristics, we determine the timing and magnitude of annual and monthly maxima of streamflow, precipitation (total and liquid), snowmelt and actual evapotranspiration for the hydrological year starting on the 1st of October (Tab. 2). In case of precipitation, we investigate maxima of 5-day sums (P max5 ). Previous investigations indicate that precipitation accumulating a couple of days before the event is most relevant for flooding (Froidevaux et al., 2015). For snowmelt and evapotranspiration, we extend this time window to 14 days and assess the magnitude and timing of 14-day 15 sums (S max14 and ET max14 ). We assume that in order to have substantial impact on streamflow, meteorological conditions favouring snowmelt or evapotranspiration need to prevail longer than only a few days. According to our experience, a 14-day window width provides a good estimate to assess potential impacts on streamflow.
In the framework of the analysis, we focus on the three gauges: Basel, Cochem and Cologne (Fig. 3). Selected gauges and sub-basins enable a detailed insight into changes in pluvial and nival processes and changes in the main channel of the 20 Rhine River. Gauge Basel is located at the transition from High to Upper Rhine. The basin upstream gauge Basel encompasses large areas of high alpine character. Snowmelt during spring and early summer is an important runoff/flood-generating process (Wetter et al., 2011;Stahl et al., 2016). Runoff at gauge Cochem (Moselle River) is characterised by a pluvial flow regime with high runoff during winter and low runoff during summer (Fig. 4). Flooding typically occurs in winter and early spring due to large-scale advective precipitation (Pfister et al., 2004;Bronstert et al., 2007). The gauge Cologne is located in the Lower

25
Rhine region after the confluences of the main tributaries Moselle, Neckar and Main (Fig. 2).
In the case of annual maxima, we display the timing and magnitude as boxplots and histograms. The length of the boxplot whiskers is 1.5 times the interquartile range (IQR). However, if no data point exceeds this distance, the whiskers only reach until the minimum/maximum value. The notches extent to +/ − 1.58 · IQR √ n with n being the length of the data vector (McGill et al., 1978; R Core Team, 2019). The notches roughly represent 95% confidence intervals for the difference in two medians. For 30 visualisation purposes, we do not display whiskers and outliers of boxplots displaying monthly maxima values. Histograms always depict the probability density and have a total area of one. In order to estimate the importance of snowmelt with regard to runoff peaks, we calculate the ratio between snowmelt the preceding 14 days and snowmelt the preceding 14 day plus precipitation the preceding 5 days (melt fraction). We also determine the average annual cycle of this ratio. In addition  . Pardé-coefficients (ratio of average montly discharge and the mean annual discharge) (Pardé, 1933;Spreafico and Weingartner, 2005) for gauges Cochem, Basel and Cologne calculated based on measured discharge from the time frame 1917 to 2016.
to the average annual cycles of the melt fraction, we calculate the average elevation of the snowmelt and the fraction solid precipitation compared to the total precipitation.

Results
The magnitudes of annual streamflow maxima at gauge Basel increase with rising temperatures (Fig. 5 a). However, this increase is not linear with the magnitude of the warming. The most prominent increase shows up between the historic time frame and the 1.5 • C warming level. Among the different warming levels we distinguish marginal differences. In general, annual runoff maxima are recorded throughout the year (Fig. 5 b). In the historical period, runoff peaks cluster during spring 5 and early summer (snowmelt season). In a warming climate, this cluster is more and more dispersed and annual maxima are increasingly recorded during winter, in particular for the 3 • C warming level. At gauge Cochem, no clear signals of change are detected, neither for the magnitudes nor the timing of annual streamflow maxima (Fig. 5 b and e). At gauge Cologne, streamflow maxima tend to be stronger at the selected warming levels compared to the historic time frame. Again, differences among warming levels are only marginal.

10
For both gauges Basel and Cochem, the estimated contribution of snowmelt to annual streamflow maxima strongly decreases with rising temperatures (Fig. 6 a and b). At gauge Cochem, the number of streamflow maxima having an estimated runoff contribution of snowmelt of more the 20% is reduced by 45% between the historic time frame and the 3 • C warming level.
Magnitudes of S max14 diminish (Fig. 6 c and d). The median of S max14 for gauge Basel is around 40 mm in the historic time frame. At a 3 • C warming, it is almost halved. In the Rhine Basin upstream gauge Basel, S max14 do not only get weaker, they 15 also tend to be recorded earlier in the hydrological year (Fig. 6 e). In both sub-basin, liquid and solid P max5 increase with rising temperatures (Fig. 6 g, h, i, and j). At gauge Basel (Cochem), the median of liquid P max5 increases from 25.7 mm (17.4 mm) in the historic time frame to 31.3 mm (19.8 mm) at a 3 • C rise in temperature. Also magnitudes of ET max14 increase with rising temperatures (Fig. 6 k and l). At a 3 • C warming, the median of ET max14 magnitudes increases by 10% (7%) for the sub-basin upstream gauge Basel (Cochem) compared to the historic simulations. 20 Decreases in solid precipitation are most prominent in winter (Fig. 7 a and b). Our results indicate that at a 3 • C warming, on average, the fraction of solid precipitation will be reduced to less than 40% in the sub-basin upstream gauge Basel in winter.
The estimated fraction of snowmelt contributing to streamflow strongly decreases in the Moselle catchment during the cold   season (Fig. 7 d). At gauge Basel, strongest decreases in the melt fractions show up end of spring and in summer (Fig. 7 c).
The average melt elevation is moving upward the elevation range throughout the year (Fig. 7 e).
At gauge Basel, monthly streamflow maxima generally increase during winter and decrease in late summer (Fig. 8 a).
Streamflow maxima in May and June seem to increase in magnitude at the more moderate warming levels (up to a warming of 2 • C) and decrease as warming progresses. A similar pattern of initial increases in monthly maxima and a subsequent 5 stabilisation or even a decrease at higher warming levels shows up in December and January at gauge Cochem (Fig. 8 b) and in all winter months at gauge Cologne (Fig. 8 c). In general, patterns of change in monthly streamflow maxima at gauge Cologne seem to reflect an overlap of features visible at gauges Basel and Cochem.
Act. ET Magnitudes of snowmelt peaks remain fairly stable for gauge Basel during winter (Fig. 9 a). Strong decreases in S max14 show up in spring and are most pronounced from May to July. In the Moselle catchment upstream gauge Cochem, S max14 strongly decrease throughout the cold season (Fig. 9 b). P max5 tend to increase in intensity throughout the year (Fig. 9 c, d, e and f). In the Moselle catchment, no big differences between changes in liquid and total P max5 is detected. In the Rhine Basin upstream gauge Basel, rising temperatures seem to evoke changes from solid to liquid precipitation, which enhances the overall 5 increase in rainfall intensity, particularly in the cold season ( Fig. 9 c and  the Rhine Basin during winter ( Fig. 9 g and h). We detect highest values of ET max14 reaching up to 50 mm in the sub-basin upstream gauge Cochem during summer. Values of ET max14 increase with rising temperatures.

Discussions
Rising temperatures diminish seasonal snow covers (see also Bavay et al., 2009;Rousselot et al., 2012;Schmucki et al., 2015;Beniston et al., 2018). As a result, the importance of snowmelt as a flood-generating process decreases (Fig. 6 a, b, c and d). In 5 the Rhine Basin upstream gauge Basel, S max14 decrease for all months of spring and summer (Fig. 8 a). At no point in time during the snowmelt season, a warming climate results in an increase in risk of snowmelt-driven flooding. Our results indicate that the temporal shift forward of the annual snowmelt maxima (Fig. 6 e) is not due to an increase in snowmelt magnitudes earlier in the year. It rather seems that events early in the snowmelt season, even if weakened by rising temperatures, more often are the strongest of the year already, as snow packs are increasingly depleted within the course of the snowmelt season. 10 We can not confirm the hypothesis that an earlier snowmelt due to rising temperatures shifts the risk of snowmelt-driven flooding forward in time. Despite the temporal shift forward of annual snowmelt maxima, flood hazard seems to decrease, as the temporal shift concurs with a strong decrease in snowmelt magnitudes (Fig. 6 c). Our findings go along with results from Musselman et al. (2017), who suggest that a "shallower snowpack melts earlier, and at lower rates, than deeper, later-lying snow-cover". However, the disappearance of snow packs and glaciers is likely to favour low-flow conditions along the Rhine River (Junghans et al., 2011;Stahl et al., 2016). Another factor having the potential to initiate or reinforce low-flow situation are increasing values of evapotranspiration, particularly during summer ( Fig. 9 g and h).
Our results indicate that at in the sub-basin upstream gauge Basel during winter, the lack of snowmelt from lower elevations, at least partly, is compensated by snowmelt from areas located at higher elevations ( Fig. 7 e and Fig. 9 a). This compensation effect seems to be increasingly insufficient as the snowmelt season progresses and the snowline moves upward. We suggest 5 that in winter, the almost unchanged potential of snowmelt-induced runoff encounters more intense rainfall events ( Fig. 9 c), in turn, resulting in a strong increase in streamflow maxima (Fig. 8 a). Our results confirm previous studies suggesting that rising temperatures lead to more intense precipitation events (e.g., Lehmann et al., 2015;Alfieri et al., 2015;King and Karoly, 2017;Bürger et al., 2019;Rottler et al., 2020) and a shift from solid to liquid rainfall (e.g., Allamano et al., 2009;Addor et al., 2014;Davenport et al., 2020). In catchments having mixed hydrological regimes with rainfall and snowmelt, rising temperatures 10 seem to lead to a shift from snowmelt to rainfall as most important flood generating process (Vormoor et al., 2015(Vormoor et al., , 2016. Reconstructing the largest floods in the High Rhine since 1268, Wetter et al. (2011) indicate that about half of all large events occurred during summer due heavy precipitation combined with high baseflow from snow-and ice-melt. Our results indicate that with rising temperatures, most flood events will occur in winter. In March and April, the increase in rainfall intensity in the Rhine Basin upstream gauge Basel compares to increases in winter, the magnitudes of streamflow maxima, however, hardly change (Fig. 8 a). We suggest that the increasing potential of rainfall-induced flooding is counterbalanced by decreasing snowmelt (Fig. 9 a and c). Furthermore, our results hint at a transient increase in flood magnitudes during May and June (Fig. 8 a). It seems that during those two months, snowmelt is still strong enough to support an increase in discharge peaks due to more intense rainfall at moderate warming levels (1.5 • C 5 13 https://doi.org/10.5194/hess-2020-605 Preprint. Discussion started: 23 November 2020 c Author(s) 2020. CC BY 4.0 License. and 2.0 • C). With further rising temperatures, however, the magnitudes of streamflow maxima reduce along with declining snowmelt (Fig. 8 a).
For gauge Cochem and the associated sub-basin of the Moselle River, we detect a similar interaction between snowmelt and rainfall: an increasing flood potential due to more intense rainfall encounters declining snow packs. Again, decreases in snowmelt magnitudes seem to counterbalance increased precipitation intensity resulting in comparatively small and tran-

15
In Cologne, which is located at the main stream after the confluence of all major tributaries, signals emerging from the different sub-basin superimpose. Accordingly, we detect increases in runoff peaks during winter (Fig. 8 c). Detected increases seem to level off as temperature continue to rise beyond the 2 • C warming level. We do not find indications supporting the hypothesis describing the creation of a new flood type in the Rhine River Basin due to a transient merging of nival and pluvial flood types. 20

Conclusions
We investigate changes in flood seasonality in the Rhine River Basin under 1.5, 2.0 and 3.0 • C warming using the spatially distributed hydrologic model mHM. In order to improve our understanding of changes in rainfall-and snowmelt-driven runoff, we carried out a detailed inspection of the Rhine River Basin upstream gauge Basel and the Moselle River Basin upstream gauge Cochem. We detect significant changes in both rainfall-and snowmelt-driven runoff peaks. Rising temperatures deplete 25 seasonal snowpacks. As a consequence, the importance of snowmelt as flood-generating process diminishes. At no time during the year, a warming climate results in an increase in the risk of snowmelt-driven flooding. Furthermore, solid precipitation (snowfall) strongly decreases during winter. The shift from solid to liquid precipitation further enhances the overall increase in rainfall intensity.
Our results indicate, that in order to understand changes in annual and monthly streamflow maxima, the examination of