Small-scale floods are a consequence of high precipitation rates in small
areas that can occur along frontal activity and convective storms. This
situation is expected to become more severe due to a warming climate, when
single precipitation events resulting from deep convection become more
intense (super Clausius–Clapeyron effect). Regional climate model (RCM)
evaluations and inter-comparisons have shown that there is evidence that an
increase in RCM resolution and, in particular, at the convection-permitting
scale will lead to a better representation of the spatial and temporal
characteristics of heavy precipitation at small and medium scales. In this
paper, the benefits of grid size reduction and bias correction in climate
models are evaluated in their ability to properly represent flood generation
in small- and medium-sized catchments. The climate models are sequentially
coupled with a distributed hydrological model. The study area is the Eastern
Alps, where small-scale storms often occur along with heterogeneous rainfall
distributions leading to a very local flash flood generation. The work is
carried out in a small multi-model framework using two different RCMs (CCLM
and WRF) in different grid sizes. Bias correction is performed by the use of
the novel scaled distribution mapping (SDM), which is similar to the usual quantile
mapping (QM) method. The results show that, in the investigated RCM ensemble, no
clear added value of the usage of convection-permitting RCMs for the purpose
of flood modelling can be found. This is based on the fact that flood events
are the consequence of an interplay between the total precipitation amount
per event and the temporal distribution of rainfall intensities on a
sub-daily scale. The RCM ensemble is lacking in one and/or the other. In
the small catchment (
Floods in small- and medium-sized catchments are often triggered by atmospheric processes on small scales, i.e. small-scale frontal systems (Schemm et al., 2016) and convective storms. In the Austrian Alpine area, these types of small-scale storms cause millions of Euros in damage every year. This situation is expected to become more severe as a result of a warming climate and the Clausius–Clapeyron relationship. Single precipitation events are expected to become more intense (e.g. Allen and Ingram, 2002; Trenberth et al., 2003; Allan and Soden, 2008; Gobiet et al., 2014), and recent investigations have shown increases in deep convective precipitation can exceed the Clausius–Clapeyron relationship (known as the super Clausius–Clapeyron scaling effect, e.g. Lenderink and Van Meijgaard, 2009; Berg et al., 2013; Wang et al., 2017; Lenderink et al., 2017).
Regional climate models (RCMs) are valuable tools for studying climate
change effects on water resources. They are employed to generate climate
simulations at scales below a 50 km horizontal resolution, like in the
EU-FP7 project ENSEMBLES (Hewitt and Griggs, 2004) or the North American
Regional Climate Change Assessment Program (NARCCAP) (Mearns et al., 2009).
RCMs operating with 0.11
With improvements in numerical weather prediction (NWP) and computing technology, RCM grid spacing can now be further reduced to allow convection-permitting climate simulations (CPCSs). CPCSs benefit from two major advantages with respect to precipitation extremes: (1) deep moist convection, which is the most important process in the majority of extreme precipitation events, is physically resolved by the RCM; and (2) the representation of orography and surface fields is improved. Multiple studies have already demonstrated the added value of convection-permitting models (CPMs, Prein et al., 2015) in capturing extreme precipitation (e.g. Chan et al., 2013, 2014, 2016; Meredith et al., 2015; Zittis et al., 2017) and their frequency of occurrence (Ban et al., 2014; Knist et al., 2018). However, there are only a few future projections that use CPCSs, like Prein et al. (2017), Ban et al. (2015), Kendon et al. (2014), and Knist et al. (2018). Although processes are better represented in CPCSs, local biases are not necessarily being reduced. Their bandwidths are large and (spatial and temporal) correlation coefficients are poor when they are compared to highly resolved observation data (e.g. Prein et al., 2013; Ban et al., 2014; Knist et al., 2018). In particular, Ban et al. (2014) and Knist et al. (2018) found that their models (CCLM and WRF) increasingly overestimate extreme events in mountainous regions. This makes bias-correction techniques indispensable, even if deep convection becomes resolved by RCMs. Also, additional computational costs for the convection-permitting simulations are high, which can limit their application, in particular for climate change studies in decision making (e.g. uncertainty assessment by ensemble simulations).
Hinging on the scale of the driving data, climate change impact studies have often focussed on water balance in relatively large catchments (e.g. Fowler et al., 2007). Regarding floods, numerous studies were performed and pointed out the high uncertainties in the GCM–RCM–hydrological-model chain (e.g. Hennegriff et al., 2006; Dankers et al., 2007; Hanel and Buishand, 2010). Maraun et al. (2010) provided a comprehensive review on the requirements of hydrological models and their fulfilment via RCMs. They define the requirements in a correct representation of (1) intensities, (2) temporal variability, (3) spatial variability and (4) consistency between different local-scale variables. Köplin et al. (2014) used future climate change scenarios from the ENSEMBLES project to analyse the seasonality and magnitude of floods in Switzerland. They found that the simulated change in flood seasonality is a function of the change in flow regime type. Magnitudes of both mean annual floods and maximum floods (in a 22-year period) are expected to increase in the future because of changes in flood-generating processes and scaled extreme precipitation. Using the new EURO-CORDEX models Alfieri et al. (2015) assessed projected changes in flood hazard in Europe based on the RCP8.5 scenario and the hydrological LISFLOOD model. Their results indicate that the change in frequency of discharge extremes is likely to have a larger impact on the overall flood hazard as compared to the change in their magnitude. On average, in Europe, flood peaks with return periods above 100 years are projected to double in frequency within 3 decades. In an effort to sequentially couple convection-permitting RCMs with a hydrological model, first attempts have been made. For example, Kay et al. (2015) use results of a 1.5 km RCM nested in a 12 km RCM driven by European reanalysis boundary conditions to drive a gridded hydrological model. However, they found that the 1.5 km RCM generally performs worse than the 12 km RCM for simulating river flows in 32 example catchments.
In this study, two regional climate models (CCLM and WRF) with different grid
spacing (
Study area and station distribution
The study area is located in south-eastern Austria, at the border of the Eastern Alps (Fig. 1). Meteorological data of all available stations in the region were acquired from the Hydrographic Service of the provincial government of Styria and the Austrian Central Institute for Meteorology and Geodynamics (ZAMG). Figure 1 shows the distribution of the stations during the period 2000 to 2009, which corresponds to the calibration period of the hydrological model. Data coverage has improved through the years by installing new stations. Historically in Austria, the network of stations with daily data (ombrometer) is much denser than the network of stations with high temporal recording (e.g. every 15 min or hourly). In the bottom right plot the development of the station availability in southern Styria is shown. At the beginning of 2000 the number of stations with high temporal resolution significantly increased, whereas the number of stations with daily data was high since the beginning of the study period in 1989.
Interpolated fields of precipitation and air temperature are generated on an hourly basis. Stations with daily data are incorporated into the interpolation procedure to benefit from the dense network as follows (Reszler et al., 2006): first, daily data are interpolated on the model grid (1 km). Then, hourly data are interpolated on the same grid and the daily sum of the cells is calculated and scaled to the daily grid. Spatial distribution of daily precipitation is expected to be accurate even in the years before 2000, which is important for an accurate representation of the general water balance. However, due to the high spatial variability of precipitation in the region, hourly fields before 2000 contain more uncertainty. In contrast, uncertainty in interpolated hourly air temperature is generally much lower. The data were interpolated by a regression with station altitude and an interpolation of the residuals on the 1 km working grid. As an interpolation method for both variables, the inverse squared distance method was used. The interpolated fields for model calibration serve also as a reference data set for the RCM evaluation.
Stream gauges used for evaluation (Fig. 3).
Run-off data for a high number of stream gauges are available at an hourly
time step. These gauges are all used for model calibration (black triangles
in Fig. 1, data provided by the Hydrographic Service of Styria).
Representative gauges were selected in this study (labelled triangles with
corresponding catchment boundaries in Fig. 3, Table 1) in order to cover a
wide range of catchment sizes (75 to 1100 km
RCM domains. ERA-Interim is dynamically downscaled with CCLM and WRF
from its initial resolution of
In the eastern part (so-called Grabenland creeks) only one gauge (Fluttendorf – Fl) is selected because of the relatively homogeneous climate, geology and soils. The run-off record extends over the whole simulation period and data are assumed very reliable according to the data provider (Hydrographic Service). Geology and soils mainly consist of tertiary material. Influence of continental climate is increasing towards the east with values of annual precipitation in the order of annual evapotranspiration: mean annual precipitation (MAP) is 700 mm in the east, whereas in the western part MAP ranges from 1100 mm at the foot to 1500 at the high altitudes.
Spatial model structure (sub-catchments, nodes, routing reaches), available gauges for calibration and catchments of stream gauges for evaluation highlighted (nested catchments are shaded). Evaluation gauges see Table 1.
The RCMs we employ are the non-hydrostatic Consortium for Small-scale Modeling (COSMO) model in CLimate Mode (COSMO-CLM or CCLM) (Böhm et al., 2006; Rockel et al., 2008) version 4.8 clm 17 and the Advanced Research version of the Weather Research and Forecasting Model (WRF/ARW) (Skamarock et al., 2007) version 3.3.1. Both models are driven by the reanalysis data set ERA-Interim (Dee et al., 2011) and cover the period 1989 to 2010. The models' innermost domain, the Greater Alpine Region (GAR) with 3 km grid spacing, is reached via intermediate pan-European domains (without nudging) with 12.5 km grid spacing for CCLM and 50 and 12.5 km grid spacing for WRF. By doing so, we mimic a typical set-up as it is used in regional climate modelling applications and we do not run the risk of underestimating internal variability in our investigations. The simulations of the pan-European domains have contributed to the EURO-CORDEX initiative and have been evaluated in several studies, e.g. Katragkou et al. (2015), Kotlarski et al. (2014) and Prein et al. (2016). The model configurations for the convection-permitting (3 km grid spacing) simulations in the GAR are based on experiences from previous sensitivity experiments (Suklitsch et al., 2011; Awan et al., 2011; Prein et al., 2013, 2015). Our RCMs differ from their coarser resolved counterparts (EURO-CORDEX) insofar that the parameterisation for deep convection, the Tiedtke scheme (Tiedtke, 1989) in CCLM and the Kain-Fritsch scheme (Kain, 2004) in WRF has been turned off in the GAR. Overview of the model domains and simulations used are given in Fig. 2 and Table 2, respectively.
RCMs and their settings.
The novel method scaled distribution mapping (SDM) is used to bias correct the model precipitation and temperature data time series (Switanek et al., 2017). SDM is a parametric method, but it is nearly identical to that of quantile mapping (QM) when correcting the historical period. However, for a future period (or any period outside of calibration), the method scales the observed distribution by the relative (for precipitation) or absolute (for temperature) distances between the future and historical modelled cumulative distribution functions (CDFs). The commonly used bias-correction method of QM (Wood et al., 2004; Piani et al., 2010; Themeßl et al., 2011; Teutschbein and Seibert, 2013) assumes that error-correction functions can be treated as stationary from one time period to another. This assumption is responsible for altering the projected climate change signal. For example, a projected mean increase in precipitation of 20 % can be inflated to be 30 %, while extremes can be altered even more dramatically. However, Maraun (2012), Teutschbein and Seibert (2013), Maurer and Pierce (2014) and Switanek et al. (2017) showed this assumption of a stationary error-correction function to be invalid, and as a result, the altering of the raw-model-projected changes to precipitation and temperature was found to be unjustified. In addition, quantile mapping was found to overestimate values of low precipitation and underestimate high precipitation (Maraun, 2013). SDM, in contrast, does not rely on a stationary error-correction function, but rather attempts to best preserve the raw-model-projected changes across the entire distribution. However, the overestimation (underestimation) of low (high) precipitation intensities remains. Bias correction was performed on RCM precipitation and temperature data independently for each grid cell and calendar month. It was implemented on a 3-hourly window to more accurately capture the observed diurnal cycle.
The spatially distributed model KAMPUS (Blöschl et al., 2008) is used,
which is in operational use for flood forecasting in Austria. It contains
conceptual models for snowmelt, soil moisture accounting and flow routing.
The snow model is based on the degree-day approach,
which calculates snowmelt depending on the air temperature. For snow accumulation precipitation
is split into snow and rainfall by a lower and an upper threshold
temperature with a linear transition. Depending on the actual soil moisture,
rainfall and snowmelt are non-linearly partitioned into a component that
increases soil moisture and a component that contributes to run-off,
Total run-off on a grid cell is calculated as the sum of the outflows from all zones. It is then aggregated to sub-catchments and convoluted by a linear storage cascade which represents run-off routing in the stream network within each of the sub-catchments. Routing in the river reaches which connect model nodes is formulated by a cascade of linear reservoirs (Reszler et al., 2008b). Using a stepwise linear formulation, this model allows for incorporating non-linear effects in flood rooting, such as flood wave acceleration at high water levels and flood retention at flood plains. For the latter, discharge thresholds for flooding the banks and levees and existing 2-D hydrodynamic studies have been provided by the Hydrographic Service of Styria for calibrating the corresponding parameters. This is particularly important for a plausible representation of flood peak attenuation in very large floods. Since the hydrological model is also driven by simulated, often biased, precipitation input, flood peaks which exceed observations may be simulated.
The model domain extends over all of southern Styria (grey shaded window in Fig. 1). The western part has previously been calibrated (Ruch et al., 2012), as it is implemented for operational flood warning by the provincial government of Styria. The eastern part was extended in the current study. The model has a sub-catchment structure with 96 catchments and 152 internal model nodes (Fig. 3). The model is driven by precipitation and air temperature with an hourly temporal resolution and a 1 km gridded spatial resolution. No further climate variables are required; the potential evaporation is represented by the modified Blaney–Criddle method (Schrödter, 1985), which only requires air temperature as input.
The method of extending the model to the eastern domain followed the strategy outlined by Reszler et al. (2006, 2008a). This approach contains several steps for parameter identification based on the dominant processes concept (e.g. Grayson and Blöschl, 2000) and proposes the usage of auxiliary information and data (e.g. field surveys, snow depths, hydrogeological data) and the stratification into different event types (convective, advective and snowmelt events). Spatially distributed information is incorporated in a GIS framework, but the resulting hydrotope (i.e. areas with similar hydrological behaviour, also called hydrological response units) structure is manually fine tuned. The following hydrotope types were chosen (compare to Reszler et al., 2006): urban areas, low-density urban areas, steep slopes open, steep slopes forest, flat agricultural areas with porous aquifer, saturation areas and karstic areas. Hydrotope structure and parameter values are chosen in consistency with the existing model in western Styria, where in some catchments (e.g. at the foothills of the Koralpe massif) the physiographic situation is similar.
In order to combine quantitative and qualitative evaluation of the different
model simulations, the following measures are chosen:
catchment size as an indicator for general attenuation effects; frequency of floods, i.e. maximum annual floods (MAF); seasonality of floods; other variables, such as soil moisture (simulated by the hydrological model); event-based analyses (performance at particular events, event/weather types).
Catchment size is implicitly incorporated by the selection of the gauges
with a wide range of catchment areas from small to medium scale (75 to
200 km
Frequency of floods are analysed by typical statistics of maximum annual
flood peaks using the following plotting position (Weibull)
The seasonality of floods gives first insights into the main hydrological
drivers for flood occurrence (Parajka et al., 2010). It is the result of the
relative influences of soil moisture, evaporation and snow processes and
varies considerably in space. In their event-type analyses, Merz and
Blöschl (2003) used the seasonality of maximum annual flood peaks
as an indicator describing the timing of floods. Here, seasonality is first
analysed simply by counting MAF peaks in the four seasons: December, January
and February (DJF); March, April and May (MAM); June, July and August (JJA); and
September, October and November (SON). Second, in order to illustrate
seasonality for different simulation runs in the small multi-model
framework, circular statistics are performed. For each event the date of
occurrence of the MAF is transposed to an angle by
Using a hydrological model for an evaluation of climate model results also enables the incorporation of other hydrological quantities, which give indications about the performance of the climate model regarding the hydrological conditions. Soil moisture is an important variable to be analysed in terms of non-linearity and threshold processes in flood generation (e.g. Penna et al., 2011). It is continuously calculated by the hydrological model and, hence, can be used as a comparison between the different simulation runs.
Added value of using higher model resolution. The colour bar corresponds
to the correlation coefficients between the observed and the modelled spatial
fields of averaged precipitation. The
At last, mainly using the 3 km convection-permitting RCM results, run-off simulations at characteristic events are checked for their realistic event evolution and the plausibility of the corresponding atmospheric and hydrological conditions.
Model efficiency at the selected gauges in the calibration and historical (validation) period.
Catchment-averaged heavy (
In order to demonstrate added value due to a reduction in the model grid spacing, we derived averaged precipitation fields of the models and the observational data and calculated the spatial correlation coefficient between them. Figure 4 illustrates the resultant correlation coefficients for all models, months and hours of the day. Higher correlations for both models, illustrated by the warmer colours, are more clearly observed towards the left side of Fig. 4, the side where highest model resolutions are depicted. This shows that the RCMs improve, on average, in their ability to simulate precipitation fields across space as the resolution of the model increases.
Added value is also seen on catchment-averaged quantities. Generally, the
convection-permitting models increase precipitation intensities from heavy
(
The hydrological model was calibrated, for each sub-catchment, against run-off data of all available stream gauges in the period 2000–2009 (Fig. 1). Calibration results in western Styria are available, and the found parameters in catchments with similar soil and geological properties serve as a priori values for the catchments in the extended part. The historical data in the current study (1989–1999) are used for model validation. This allows also a validation of the existing model; these data were provided for the current study and had not been used for model calibration. Quantitative metrics such as the commonly used Nash–Sutcliffe efficiency (NSE, Nash and Sutcliffe, 1970), the bias based on mean run-off values and the root mean square error (RMSE) are used to measure model calibration. In Table 3 the results for the selected gauges are listed. As it is often the case, NSE is lowest in the smaller catchments, e.g. Schwanberg and Fluttendorf with 0.77 and 0.78 in the calibration period, respectively. In the validation period the NSE falls below 0.7 in these two catchments. The historical period also includes phases with poor data availability (see Fig. 1), which is also the reason for the drop in the NSE value in the validation period at the Gündorf gauge.
Examples of hydrographs in the calibration and validation period are
attached in the Supplement (Figs. S1 and S2). In addition to
flood peaks, run-off generation and rainfall response is represented very
well. Differences in the shape of hydrographs are also accurately simulated.
For example, the Schwanberg gauge shows short peaks due to short
concentration times in the small catchment but at the same time high
baseflow. The latter indicates a high fraction of slowly draining flow
components (groundwater) from long-term storage. On the other hand, in the
medium-sized Gündorf catchment, short peaks also indicate short response
times, but baseflow is significantly lower. This difference can be
attributed to the different geologic conditions in the area. In the
Schwanberg catchment the significant subsurface storage can be attributed to
a deep weathering zone overlaying schists and gneiss, and geology in the
Gündorf catchment consists mainly of tertiary material (silt, loam) with
very low storage capacities. In the larger catchments flood peaks are smooth,
lasting over several hours, which shows the attenuation effects (Tillmitsch,
Leibnitz). The resulting model parameter values representing timescales of
run-off response show the flashy character in the catchments: The time
constant of the fast flow component, i.e. surface run-off in open steep
slopes (
Simulated and observed maximum annual flood peaks vs. empirical return periods (Eq. 1, flood frequency plots) of the selected gauges in the period 1989–2010. The peaks in the validation period are marked with red colour.
For this study, representation of flood frequency is important.
Model-simulated maximum annual floods for the entire study period
(calibration and validation period combined, 1989–2010, 22 years) are
compared with observed flood peaks in Fig. 6 (flood frequency plot).
Although the MAF distribution was not explicitly subject to calibration, and
the data availability was relatively poor in the period 1990-2000, the model
accurately simulates observed flood statistics at the selected gauges. The
largest flood is simulated well at all gauges, while simulation results at
the smaller events are reasonable. Both in the calibration and validation
period, deviations in significant events are analysed in terms of probable
errors in input (precipitation), model structure, model parameters and/or
run-off data. At the exceptional events threshold processes are operative,
which are accurately simulated. For example, the largest flood at the
Schwanberg gauge in August 2005 (Fig. 6, plot above left, extrapolated RP
would be more than 100 years) was a very local event (see Fig. 9
lowest panel) and the interpolated rainfall is assumed to be relatively
uncertain. In order to simulate the observed flood peak, parameters which are not plausible and decrease model performance at other large
events would be needed. Also inundation occurred during the event in the Schwanberg town,
which likely led to uncertainties in the observed peak run-off data. At the
Fluttendorf/Gnasbach gauge, flooding occurred at the two events in 2009 (see
Fig. 5), and retention by inundation in flood plains was calibrated
successfully in the flood routing model. At the Voitsberg gauge the two
largest floods were slightly underestimated. The largest flood occurred in
October 1993, within the validation period, which was underestimated in
the simulation. The largest flood in the simulation is the 2009 flood, which
was represented very well (see Supplement). Data quality used to
be poor in 1993 in the high-altitude catchment in the north-western part.
Station density in this part today is still lower than for example in the
south-western part (see Fig. 1). The same situation can be stated for the
Tillmitsch gauge. At this gauge, the four medium event peaks (from 92 to 117 m
With the calibrated hydrological model, simulations are performed using the results of the RCMs as input. In Sects. 6.1 and 6.2, evaluation results are discussed in detail for the run-off simulations using the CCLM results. The same procedure has been applied using the WRF results (provided in Supplement). In a synthesis step (Sect. 6.3), all the results of the small multi-model framework are summarised and compared for formulating final conclusions.
Flood frequency plots for the selected gauges using uncorrected CCLM data
(and the ERA-Interim data) as input compared to the observations are shown
in Fig. 7. The figure illustrates the improvement of the results using the
3 km CCLM data, particularly for the smallest catchment, Schwanberg/S. Sulm,
with a size of approximately 75 km
Simulated maximum annual flood peaks using raw CCLM data as input and observed maximum annual flood peaks vs. empirical return periods (Eq. 1, flood frequency plots) of the selected gauges in the period 1989–2010.
Number of maximum annual floods in the four seasons (seasonality) from the simulation using raw CCLM data as input compared to the observation at the selected gauges in the period 1989–2010.
Example of two events (August 1996, above, and August 2005, below, in each case plotted with catchment precipitation above the run-off) simulated with raw CCLM 3 km data compared to the simulation with observed input and observed run-off data for the Schwanberg gauge.
Most of the RCM settings show negative biases regarding MAF peaks; however,
some are significantly positively biased, e.g. Voitsberg/Kainach in the
north-western Alpine part. At the Fluttendorf gauge (upper-right subplot)
the 0.44
The seasonal occurrence (winter: DJF, spring: MAM, summer: JJA and autumn:
SON) of the simulated MAFs is analysed in Fig. 8. The improvements are
evident when reducing grid size; the simulation with the uncorrected 3 km
CCLM data represents the observed seasonality very well. The figure further
shows that both the CCLM with 0.44
Simulated soil moisture on a monthly basis (attached in Fig. S5 above)
shows annual dynamics that are similar to the
seasonality of MAF. Also, the improvements using the 3 km CCLM
(0.03
Comparison between modelled (CCLM 3 km), bias-corrected and observed
precipitation characteristics tied to the 22 MAF events in the
Schwanberg catchment.
As the first results show, the CCLM 3 km setting yields a clear benefit
regarding magnitude and frequency of large floods particularly in small
catchments. As stated above, the floods of the simulations are not
necessarily aligned in time with observations. Figure 9 shows two simulation
periods for the Schwanberg/S. Sulm gauge (75 km
In the same way as for the raw RCM data, the hydrological model is driven using bias-corrected data. After bias correction, results of flood statistics using CCLM (Fig. 13) are improved, except for the smallest catchment, Schwanberg. Here in particular, the results deteriorate compared to the run using the uncorrected data (Fig. 6).
This can be explained by an interference of the temporal distribution of
precipitation intensities during the flood-generating rainfall events and
the bias correction that simply ignores such temporal relationships.
Figure 10 shows the precipitation intensities that
contribute to the maximum annual flood events in Schwanberg simulated by
CCLM 3 km, before and after bias correction. Each event is limited to a
duration of 2 days before the maximum peak flow is reached.
Figure 10a demonstrates the work of the bias
correction that removes severe underestimation (overestimation) of low (high)
intensities in the CCLM 3 km data, but leaves the total amount of
precipitation of these events largely unaffected so that a median underestimation
of
Same as Fig. 10, but for WRF 3 km.
Same as Fig. 10, but for ERA-Interim in the Gündorf catchment.
In contrast, the positioning of WRF 3 km peak flows in Schwanberg lies above the observations and the bias correction leads to a deterioration (Fig. S3). In this case, WRF 3 km overestimates precipitation intensities across the flood events and the bias correction changes this (due to the aggregation of single grid cells to catchments) into an underestimation (Fig. 11a). This leads to an overestimation (underestimation) of event-related precipitation amounts (Fig. 11b) for the uncorrected (corrected) data. In WRF 3 km the temporal distribution of the intensities is in much better agreement with the observations than in CCLM 3 km (compare Fig. 11c and Fig. 10c). However, since the total amount is overestimated, the peak flows are higher. The bias correction further deteriorates the temporal distribution of the intensities that lie closer to the flood event and together with the underestimation of the total amount this gives a rapid drop in the positioning of the peak flows (Fig. S3). Note that this good representation of the temporal distribution in WRF 3 km is a catchment-specific feature.
Also, in the small catchments, the aggregation to 3 h sums has an influence on the performance. We tested it by using the 3 h sums of the CCLM 3 km and comparing to the 1 h results (not shown). There is a decrease in flood peaks, but the main decrease in performance in the small Schwanberg catchment is due to the error correction explained above.
In some cases bias correction leads to overcompensating for the flood
peaks, particularly in the case of the ERA-Interim data. For instance in
Gündorf, flood-event-related precipitation intensities and amounts are
largely underestimated in ERA-Interim by more than
From a return period of 6–10 years the flood simulations are very sensitive to overestimations (e.g. Voitsberg and Gündorf gauges in Fig. 13) and underestimations (see Fig. 7) of the simulated rainfall, which is due to the non-linearity in the rainfall-run-off process (e.g. Komma et al., 2007; Rogger et al., 2012). This threshold is consistent with usual concepts in hydrology, such as the concept of the GRADEX method (e.g. Merz et al., 1999). At this size of floods the soils have been saturated by a high amount of precipitation and 100 % of the subsequent rainfall comes to run-off. This is vital to take into account when it comes to correcting high rainfall intensities within the bias-correction procedure.
Seasonal occurrence is improved for all CCLM settings after bias correction
(Fig. 14). In particular, the shift from summer to spring using the raw
0.11 and 0.44
Simulated maximum annual flood peaks using bias-corrected CCLM data as input and observed maximum annual flood peaks vs. empirical return periods (Eq. 1, flood frequency plots) of the selected gauges in the period 1989–2010.
Number of maximum annual floods in the four seasons (seasonality) from the simulation using bias-corrected CCLM data as input compared to the observation at the selected gauges in the period 1989–2010.
Run-off simulated with the uncorrected (1 h rainfall sums) and bias-corrected (3 h rainfall sums) 3 km CCLM data for the period with the largest floods at the Fluttendorf/Gnasbach gauge. Above: catchment precipitation.
As for the seasonality, the seasonal shift in the simulated soil moisture is
removed after bias correction, but the underestimation in summer and autumn
cannot be entirely compensated for (see Fig. S5 below).
This can be attributed to the fact that the modelled events are different in
size, shape and overall structure to those of observations. The SDM
methodology is performed independently for each grid cell and as a result
is not imposing the structure of typical broad-scale observed weather
events. Therefore, even though the distributions of bias-corrected
precipitation align to observations at individual grid cells, the average
precipitation amounts across multiple grid cells can differ from
observations. ERA-Interim results now lie exactly on the observation.
However, for the MAF performance using ERA-Interim data is not sufficient
(compare Fig. 13). This shows that using observed atmospheric conditions
with large grid size (
For an event-based illustration of the effect of bias correction, two events in 2009 at the Fluttendorf/Gnasbach gauge were chosen using the 3 km CCLM data as input (Fig. 15). The first event in June is the largest in the series and the second event in August is the second largest in the series. Synoptic forcing is different between the two events: the first event is controlled by a persistent upper-air cut-off low that is located over the Balkan region and brings warm and moist air towards the eastern Alpine region from the east (Godina and Müller, 2009). This led to floods in the whole southern Styria region, whereas the second flood is mainly driven by convective processes and concentrated on the eastern part. For the first event, the model with the uncorrected 3 km CCLM data simulates an event with the same order of magnitude, but slightly different timing, as the observation. After the bias correction, flood peak is decreased due to a general reduction of precipitation in the bias correction in this period. A reduction of rainfall in this period results from the bias correction as a consequence of the overestimation of the MAFs by raw CCLM data (compare Fig. 7, upper-right subplot). However, after bias correction, this is still the largest flood peak in the series (see Fig. 13, upper-right subplot). The second event is completely missed by the simulation run with the raw climate model data. No significant rainfall is simulated in the RCM and, hence, bias correction is totally ineffective. It is clear that at such missed events there is no possibility to correct raw RCM data using any statistical bias-correction method. Bias correction is not able to compensate for general uncertainties in representing convective situations. Note that bias-corrected intensities in the upper panel are aggregated 3 h sums.
Statistical measures of maximum annual flood peak distribution evolving
the different model runs.
The statistical measures of mean, standard deviation and skewness for the
22-year sample of maximum annual floods resulting from the 14 different
variants are illustrated in Fig. 16. The mean (left plot column) and the
standard deviation (middle plot column) are related to the catchment area in
order to compare these measures between the gauges. Results using the
ERA-Interim data are plotted in the centre, and the results using the
different RCM settings with decreasing grid size are plotted towards the
left (CCLM) and the right (WRF). The values with raw RCM data as input are
plotted as black points; the values with bias-corrected RCM data as an input
are plotted as red points. The observed measures are indicated with a thin
horizontal line for each gauge. The figure first clearly shows the decrease
in mean-specific run-off peaks and – in connection to this – the specific
standard deviation with the catchment sizes (S to L from above) for all
variants. This is mainly the consequence of a decrease in mean areal
precipitation for large rainfall intensities and short durations (e.g.
Hershfield, 1961; Lorenz and Skoda, 2000) but also of attenuation effects
through flood routing. As discussed in the previous section, in most of the
CCLM data-driven simulations the statistical properties are improved
reducing the grid size (black points) and further improved after bias
correction (red points). For the larger catchments, Tillmitsch and Leibnitz,
the differences between the model variants are small, which, again,
indicates the good performance of the coarser RCMs regarding general flood
statistics (particularly CCLM). This improvement is not always the case for
the WRF-driven runs. Particularly large biases from the uncorrected run are
either not compensated for (e.g. WRF 0.44
Results of seasonality (circular statistics of the maximum annual
floods) evolving the different model runs.
In order to summarise the performance of the small multi-model framework regarding seasonality, Fig. 17 shows the results applying Eq. (2) and Eqs. (3)–(6) on the simulated MAFs using the different RCM data, raw (above) and after bias correction (below). The observation is plotted with a green filled square. As discussed in Sect. 6.1, the results illustrate again the improvement of the seasonality using the 3 km CCLM data (full red squares) compared to the simulations with the coarser CCLM data for all gauges. For example, the highest concentration of timing, i.e. length of vector, of floods in a year in Voitsberg is represented well by the raw 3 km CCLM (upper middle subplot). However, this outstanding result of CCLM 3 km is the result of compensating for errors: the complex interplay between single precipitation intensities and their temporal distribution during flood-generating rainfall events is not correctly represented (Sect. 6.2). Either the total precipitation amount is properly captured but the temporal distribution is failed or vice versa. This also holds for the other RCM simulations, including WRF 3 km, and ERA-Interim. This figure summarizes again that the bias-correction method is not able to correct displacements in this complex interplay.
Using the coarser RCMs, both the timing and strength of seasonality of MAFs deviate significantly from the observations in all catchments. Moreover, the scatter between the different settings is large. However, to some extent all CCLM settings represent the weak seasonality in the eastern part (Fluttendorf catchment, upper-right subplot). The convection-permitting WRF 3 km setting does not provide any improvements compared to the coarser resolutions. Timing of MAFs tends to be concentrated in May/June for all catchments, whereas flood events occur mainly from July to September. This indicates that more or less all WRF settings fail in representing the general mechanisms for flood generation in this area and at this scale. Mostly, discrepancies can be compensated for by the bias correction in the CCLM case, but not for the WRF case. In some catchments using the WRF 3 km settings, the results are worse after bias correction. For example, at the Fluttendorf gauge (upper-right subplot in Fig. 17 below) the concentration of timing shifts from the beginning of May (with a low strength) to February (with a relatively high strength), a month when flood generation is also influenced by snowmelt processes.
This study implemented regional climate models sequentially coupled with a
spatially distributed hydrological model to be used for enhanced flood
modelling on small and medium spatial scales (up to approximately 1000 km
Evaluations using observed data in a historical period (1989–2010) showed
that, in the investigated RCM ensemble, no clear added value of the usage of
convection-permitting RCMs for the purpose of flood modelling can be found,
although CCLM 3 km outperforms in most flood statistics. This is based on
the fact that flood events are the consequence of an interplay between the
total precipitation amount per event and the temporal distribution of
rainfall intensities on a sub-daily scale. The investigated RCM ensemble is
lacking in one and/or the other. The seemingly good CCLM 3 km results
in the small catchment lie on an overestimation of the intensities and
underestimation of the total rainfall amount. This superposition is not
systematic across the catchments. From a statistical perspective, all RCMs
with all resolutions are able to produce precipitation rates that may cause
floods in the study area. In catchments with an area less than 100 km
The bias-correction-method scaled distribution mapping is able to
systematically reduce biases on a seasonal basis. SDM improves results in
magnitude and seasonality of maximum annual floods in all settings except
for the small catchment (
With respect to climate change applications of convection-permitting simulations for flood representation we can conclude that, despite the seemingly good results in the CCLM 3 km setting, attention has to be paid and the testing of the results against historical data is of utmost importance. On the other hand, deep convection parameterisations in coarser resolved standard RCMs have been shown to be a source of deep uncertainty. For instance, Kendon et al. (2014) found significant increases in summertime precipitation in convection-permitting climate simulations in the UK while the coarser resolved counterpart does not show any significant change. Ban et al. (2015) and Berthou et al. (2018) found similar results for short-term extreme precipitation events in the Alpine region and in the Mediterranean. In order to circumvent possibly misguided but far reaching climate change adaptation strategies, either convection-permitting RCMs or proper statistical convection emulators (that are currently discussed in the climate modelling communities) should be used. Coarser models could still be used in larger catchments for rough estimations, but they should not be taken for granted regarding local and/or regional flood change. Also, there is a trade-off in the additional costs of a 3 km simulation and the postulated (small scale) process description as long as the physical representation of such small-scale processes can be substituted by statistical ones. Regarding bias correction, the temporal dynamics of the rainfall have to be analysed; an application of a current error-correction method can be recommended only if RCM errors are found to be systematic.
The used hydrometeorological data can be obtained from the
Hydrographic Service of the province of Styria (for free) and from
the Central Institute for Meteorology and Geodynamics (ZAMG) (for a charge). They were not stored
in a separate study-related platform. CCLM simulations with 0.11
The supplement related to this article is available online at:
CR performed the hydrological model simulations, MBS implemented and applied the scaled distribution mapping method, and HT conducted the convection-permitting WRF simulations.
The authors declare that they have no conflict of interest.
The study was funded by the Austrian Climate Research Programme (ACRP 6th call, project number KR13AC6K11102 – CHC-FloodS) by the Austrian Climate and Energy Fund (KLIEN). Regional climate model output on the EURO-CORDEX domain was provided by Klaus Keuler (BTU Cottbus, Germany) and Klaus Görgen (Institute of Bio- and Geosciences, Agrosphere, Research Centre Jülich). The WRF simulations from Klaus Goergen used in this study were conducted at the Centre de Recherche Public – Gabriel Lippmann (now the Luxembourg Institute of Science and Technology) under grant FNR C09/SR/16 (CLIMPACT) from the Luxembourg National Research Fund. The convection-permitting simulations were conducted in the course of the project NHCM-2 (nhcm-2.uni-graz.at), funded by the Austrian Science Fund (FWF, project no. P24758-N29). We also thank Andreas F. Prein (National Center for Atmospheric Research, USA). He conducted the convection-permitting CCLM simulation and helped with fruitful comments. Computational resources were provided by the Jülich Supercomputing Centre (JSC) and by the Vienna Scientific Cluster (VSC). Additionally, we gratefully acknowledge the European Centre for Medium-Range Weather Forecasts (ECMWF) for providing ERA-Interim data. Hydro-meteorological data were provided by the Hydrographic Service of the province of Styria and the Central Institute for Meteorology and Geodynamics (ZAMG). Edited by: Joaquim G. Pinto Reviewed by: three anonymous referees