2.1 Catchment: Gardon d'Anduze

The study is performed over a 543 km² Mediterranean catchment, the Gardon d'Anduze, located in the Cévennes region, on the southern slopes of the Massif Central mountain range. The Gardon d'Anduze has a complex topography ranging from 130 m to 1200 m a.s.l. Consequently, it is rather natural, forested and lightly urbanized. Indeed, most of its surface is covered by typical Mediterranean vegetation. The soils are relatively thin, from 10 cm at the top of the hillslopes to 100 cm close to the river bed (Vannier et al., 2014). It is considered a highly reactive Mediterranean catchment, known for experiencing some of the most destructive flash floods in France (Delrieu et al., 2005; Toukourou et al., 2011). The Gardon d'Anduze catchment has been extensively studied due to its location of particular interest in hydrology, especially for flash flood modeling and forecasting (Alfieri et al., 2011; Roux et al., 2011; Moussa, 2010). Figure 1 displays the location of Gardon d'Anduze in France (a) and orography as shown by an 80 m resolution DEM (b) and by the RCM and CPM (c and d).

Figure 1Location of the Gardon d'Anduze catchment (in red) over France (a) and orography of this catchment represented by an 80 m DEM (b), by the 12 km CNRM-ALADIN RCM (c) and by the 2.5 km CNRM-AROME CPM (d).

2.2 Climate models

In this study, we use climate simulations from a CPM and its forcing RCM, both developed and released at the Centre National de Recherches Météorologiques (CNRM), in Toulouse, France. Information about these two climate models is summarized in Table 1.

Table 1Information about the two climate models and their associated simulations used in this study.

Download Print Version | Download XLSX

CNRM-ALADIN is a 12 km grid cell RCM that has been run over the continental EUR-11 domain through the EURO-CORDEX initiative. Retrospective simulations are driven by the ERA-Interim global reanalysis dataset (Dee et al., 2011), while historical and future scenarios are forced by the CNRM-CM5 global model (Voldoire et al., 2013). This version is derived from the development of the NWP model ALADIN, thanks to the ACCORD research centers consortium. CNRM-ALADIN has been extensively used in the CORDEX framework over Europe, the Mediterranean, North America and Africa. For more details about the parameterization schemes and configurations of the last version of ALADIN, see Nabat et al. (2020) and Lucas-Picher et al. (2023).

CNRM-AROME is a CPM, which is adapted from cycle 41 of the NWP AROME formerly operating for numerical weather prediction. Simulations used in this study are produced, as part of the EUCP project, over the NW domain, covering France, the UK, the north of Spain and most of Germany with a 2.5 km mesh. Some papers have already evaluated this model, under a former release (Fumière et al., 2020), a different domain and forcing RCM version (Caillaud et al., 2021; Monteiro et al., 2022), or the same version and domain used in this study (Lucas-Picher et al., 2023). All these evaluations have established the added value of CNRM-AROME in the reproduction of extreme rainfall and HPEs compared to CNRM-ALADIN over the domains. To our knowledge, no published study has used CNRM-AROME projections for climate change assessment. In this paper, ALADIN and AROME refer to the version of CNRM-ALADIN and CNRM-AROME described above, respectively.

2.3 Observations

In this study, we compare the hourly simulated data with a high-resolution observed precipitation dataset called COMEPHORE. COMEPHORE is an hourly 1 km resolution gridded product covering Metropolitan France during the 1997–2019 period. COMEPHORE was built by merging weather station rainfall measurements and different radar sources (Laurantin et al., 2012; Caillaud et al., 2021). The dataset used approximately 3000 daily rain gauges and 1200 hourly rain gauges. The number of radars has more than doubled during the first decade of data. In 2019, COMEPHORE was built using data from 29 radars comprising the French radar network (ARAMIS), in addition to radars from the Swiss network and another one on Jersey Island. Even though the quality of the dataset is not temporally and spatially homogeneous and it has a tendency of underestimating some extreme values (Laurantin et al., 2012), COMEPHORE can be considered the best national precipitation product for studying hourly rainfall at high spatial and temporal resolutions. Furthermore, the Gardon d'Anduze river catchment is located in a region of high rain gauges and radar density, raising the confidence in this dataset for a benchmark in this study (Caillaud et al., 2021).

Temperature is the main variable to compute potential evapotranspiration (PET). As a reference, we extracted 3 h temperature measurements from six weather stations in the basin. Temperature was then linearly downscaled to the hourly time step and interpolated over the catchment using an inverse-distance-weighting method. The COMEPHORE rainfall dataset and PET derived from interpolated temperature are the best observation-based estimates in this area and are therefore considered the reference observational datasets in this paper.

Numerous methods exist to compute PET from climatic variables for hydrological modeling perspectives (Oudin et al., 2005). Here, we compute PET using the Hargreaves–Samani (HS) method (Hargreaves and Samani, 1982). We chose a reliable, widespread applied method that requires the minimum amount of climate variables at the daily time step. Daily PET is then disaggregated to the hourly time step using a standard distribution curve as done in Lobligeois (2014). We applied the same methodology for simulated temperature.

2.4 Hydrological models

Two hydrological models with different physical concepts were used in this framework. This choice was made to consider the potential uncertainty related to hydrological modeling that can affect hydrological projections (see, e.g., Lemaitre-Basset et al., 2021, which discusses this issue for the neighboring Hérault catchment).

The GR5H model is a lumped, conceptual rainfall–runoff model that runs hourly (Ficchì et al., 2019). Based on several conceptual reservoirs, such as a soil moisture accounting reservoir, and a unit hydrograph, this model uses catchment-aggregated hourly precipitation and potential evapotranspiration data to simulate hourly discharge. The GR5H model takes into account the interception of rainfall by vegetation, which was proven important for flood events (Ficchì et al., 2019). The GR5H model parameters were calibrated against observed discharge using the Nash and Sutcliffe efficiency (NSE) objective function. This model or close versions belonging to the so-called GR family of hydrological models have been used both for flood simulation (Ficchì et al., 2019; Astagneau et al., 2021) and climate change applications (Chauveau et al., 2013; Lemaitre-Basset et al., 2021). The GR5H model was run using the open-source airGR R package (Coron et al., 2017, 2020), which also provides the parameter calibration algorithm that was used.

The Ensemble Framework For Flash Flood Forecasting (EF5) is a software for distributed water balance modeling (Flamig et al., 2020), including different schemes for runoff production and routing. As part of EF5, we used the Coupled Routing and Excess STorage (CREST) model and the kinematic wave routing scheme, at the hourly time step. CREST is a fully distributed and process-based hydrological model. It can be defined as a hybrid between a conceptual and a physics-based model (Wang et al., 2011). All hydrological processes, such as runoff production, evapotranspiration and sub-grid cell routing are computed at each grid cell. CREST is composed of two excess storage reservoirs, one for interception by the vegetation canopy and one representing a layer of soil. For each cell, runoff and infiltration are separated using a variable infiltration curve. Therefore, this model takes into account the two main runoff production mechanisms: saturation excess and infiltration excess. The subsurface routing is done with a linear reservoir model. Total runoff, composed of surface and sub-surface runoff water, is then routed to the outlet following the orography provided by a digital elevation model (DEM) with the kinematic wave routing scheme. Actual evapotranspiration is determined in the model from PET input and the water content of the cell. The CREST model is composed of 13 parameters: 6 for runoff production and 7 for the routing. Here we will refer to this adaptation of CREST in EF5 simply as CREST for succinctness. The DiffeRential Evolution Adaptive Metropolis (DREAM) scheme is used for automatic calibration to estimate the best parameter set (Vrugt et al., 2009). A complete description of parameters is provided in Flamig et al. (2020). CREST is a model initially developed to respond to the need of forecasting floods at the global scale (Wang et al., 2011) but is perfectly suitable to simulate flash floods (Flamig et al., 2020) through EF5. Indeed, CREST has been used to study extreme hydrological events, for example, to reproduce floods from a major hurricane event (Li et al., 2021) or floods and flash floods in a warmer future (Li et al., 2022b, a). In this study, topographic data from version 1.1 of the HydroSHEDS database are used for CREST. Resolution of the DEM is 15 arcsec; hence, for this latitude it is around 300 m in longitude by 450 m in latitude. Drainage direction maps and flow accumulation maps are then produced using the QGIS software and packages.

The CREST and GR5H hydrological models have been calibrated using the hourly observed discharge from 2002–2018 and using the COMEPHORE hourly rainfall dataset and PET computed from observed air temperatures. For GR5H, all the model parameters have been calibrated. For CREST, most parameters have been fixed (Li et al., 2022a), and the calibration has been performed on a few sensitive parameters. The NSE criteria are used as an objective function for the calibration process. We initialized the models during a 1-year warmup period before the starting date. Both hydrological models have been evaluated using the following metrics (see Sect. 4.2) over the 2002–2018 period:

• Nash and Sutcliffe efficiency (Nash and Sutcliffe, 1970);

• KGE (Gupta et al., 2009);

• bias on mean flows;

• bias on quantile 99.9;

• bias on peak over threshold (POT) distributions – we compared the mean of simulated POT distribution to the mean of observed POT distribution.

The same period (2002–2018) is kept to run the hydrological models driven with retrospective climate simulations. For the scenario datasets, using a 1-year warmup period, we computed simulations over two epochs of 19 years for historical (1987–2005) and future (2081–2099) periods.

3.1 Bias correction

Climate simulations and notably precipitation often show significant biases that prevent their direct use in impact studies such as for hydrological impact modeling. Indeed, hydrological models are calibrated over climatic conditions that can differ strongly from the raw historical climate simulations. The use of bias correction methods on climate simulations is an open debate (Addor and Seibert, 2014; Huang et al., 2014; Maraun, 2016). However, correction of biased climate simulations is widely used for hydrological impact studies and future projections (Reszler et al., 2018; Giorgi, 2019; Lucas-Picher et al., 2021).

In order to correct simulated climate data biases, we used an univariate statistical bias correction method. CDF-t is a statistical bias correction method specifically developed for correcting biases of climate variables (Michelangeli et al., 2009). The basis of this method is an extension of a quantile mapping method that allows a change in distribution statistics for corrected variables. Consequently, CDF-t is particularly appropriate to correct future climate datasets in a non-stationary climate (Michelangeli et al., 2009; Vrac et al., 2012), and it is largely used to correct future climate projections for the sake of climate change impact studies. In this study, we applied the CDF-t correction on hourly precipitation and temperature variables. Due to the distinct seasonality of precipitation and a strong spatial variability of precipitation in this catchment, we corrected simulated datasets for each calendar month and over each grid cell separately. To take into account differences in the ratio of wet and dry hours, the Singularity Stochastic Removal (SSR) preprocessing method is applied for precipitation simulations (Vrac et al., 2016). Historical simulations are corrected against the corresponding observed period (2000–2018). With the common period of data between observations and historical AROME simulations being relatively short (8 years), we chose to correct observations and the historical simulation over two asynchronous periods of same length, respectively 2000–2019 and 1986–2005. This correction is then applied over the end-of-century RCP8.5 projections (2080–2099). All these operations are performed through the R package SBCK (Robin, 2022). Raw and bias-corrected hourly precipitation patterns have been compared for some extreme rainfall events, and spatial correlation is little affected by the bias correction (not shown). For the GR5H model, corrected climate outputs are then averaged over the Gardon d'Anduze catchment.

3.2 POT extraction

This study focuses on floods. To extract floods from discharge time series, two approaches are widely adopted: the annual maximum flood (AMF) and a peak over threshold (POT) extraction (Lang et al., 1999). The POT selection method consists of the extraction of all flood peaks exceeding a threshold. Contrary to the AMF method, the POT method preserves the hydrological information especially in Mediterranean catchments where several flash floods can occur every year. Hence, we selected this method for flood extraction in this paper. The POT method requires temporal independence between flood peaks. To ensure this assumption, a declustering algorithm has been developed and applied to evaluate temporal dependencies between POTs, adapted from Lang et al. (1999). The declustering applied herein considers that two floods are independent if there is a minimum duration of 96 h between the two events and if the discharge between different events must drop below 75 % of the minimum value of the flood peaks. Cunnane (1973) found that the POT method minimizes the sampling variance for a threshold producing on average 1.65 flood peaks per year, compared to the AMF. To get a sufficient sample of simulated POT (in particular for ALADIN), we fixed here the discharge threshold to get closer to 2 floods per year for the observed discharge (corresponding to a threshold equal to 265 m³ s⁻¹), after declustering.

3.3 Flood characteristics

For each event, several flood characteristics and their associated rainfall events are also analyzed. These metrics aim to understand future changes in flood mechanisms as projected by climate simulations. For instance, changes in flood peak baseflow can give information on change in antecedent soil moisture. The lag-time and flashiness indexes describe the intensity and speed of the rise of the river flow, a signature of potential catastrophic impacts (flash flood). At the same time, characteristics of the rainfall events (duration, intensity, maximum) in tandem with the flood characteristics help elucidate the driving processes and causes of future changes in the basin's hydrometeorology. Rainfall events associated with each POT (Sect. 3.2) are defined by a sequence of hourly rainfall prior to the flood peak (i) interrupted by no more than 2 dry hours and (ii) yielding at least 30 mm to trigger the hydrological event. These two conditions have been tested for different values with observed datasets and have shown very little sensitivity to defining the flood characteristics. The rainfall thresholds are related to our knowledge of the river basin dynamics and hydrological expertise. Table 2 summarizes metrics names, meaning and computation details.

Table 2Flood characteristics definition and details.

Download Print Version | Download XLSX

4.1 Added value of CPM at the catchment scale

We first analyze the climate datasets aggregated at the scale of the Gardon d'Anduze catchment. Figure 2 shows the annual cycle for daily minimum (T_min) and maximum (T_max) temperature and hourly precipitation before (raw) and after bias correction (BC-CDFT). AROME and ALADIN in retrospective mode (evaluation) show correct annual cycles despite an overestimation of summer temperature both for T_min and T_max. Results in this area are in line with the recent evaluation of the ALADIN and AROME climate models (Lucas-Picher et al., 2023). The ALADIN RCM is generally colder than AROME, which can be explained by an effect of resolution. A strong cold bias (3–5 °C) is visible over all seasons for both models in historical scenarios over the period 1986–2005. This cannot be explained by climate variability or by climate change signals between the decades 1986–2005 and 2000–2018. This is a known cold bias of climate models driven by the GCM CNRM-CM5 (Vautard et al., 2021). Annual cycles for T_min and T_max for future projections (RCP8.5) are similar to the evaluation period; i.e., climate change signals for temperature are almost of the same magnitude as the cold bias. As expected, this cold bias disappears after bias correction for both AROME and ALADIN historical scenarios. The bias-corrected temperatures under the RCP8.5 scenario strongly increase for all seasons. However, the strongest signal is seen for the summer months.

Figure 2Annual cycle of daily-aggregated minimum and maximum 2 m temperature (T_min and T_max) and hourly precipitation for raw (a, c, e) and bias-corrected (b, d, f) climate datasets. A 30 d rolling mean has been applied in order to smooth the high-frequency natural variability.

Download

Annual cycles of precipitation are well reproduced for ALADIN and AROME for the evaluation period. Both models are a bit too wet in spring and too dry in summer, but the wet season (October to December) on simulations is clearly distinguishable and identical to COMEPHORE dataset (observations). The wet bias in spring can deeply affect soil moisture state in the hydrological models and therefore probably lead to a change in the behavior of the first floods occurring during the autumn months. The two climate models are able to reproduce the precipitation seasonality. There is no evidence of added value from the CPM on the simulation of the annual cycle of precipitation on this catchment. Results are similar for AROME and ALADIN for historical projections: spring and summer seasons are too wet. This causes a weaker annual cycle of precipitation. The wet fall season, however, is correctly reproduced, but with too early a start and plateau. The CDF-t bias correction method is able to correct the annual cycle of precipitation. The dry season is therefore consistent with the reference. In terms of intensity, ALADIN corrected shows an increase in intensity of the wet season peak (>0.4 mm h⁻¹). We do not find this signal with AROME after correction. There is a dry signal for AROME corrected for the summer months. Results show no signal of wet season lengthening between the historical period and the future RCP8.5 scenario.

In terms of precipitation distribution, for the evaluation (retrospective) simulations, AROME performs better than ALADIN, especially for the precipitation extremes (Fig. 3, Table 3). The most extreme events (> quantile 99.9) remain underestimated for AROME. Fumière et al. (2020), Caillaud et al. (2021) and Lucas-Picher et al. (2023) have already shown the added value of the AROME CPM compared to its driving RCM, ALADIN, for modeling extreme hourly precipitation. This result confirms these past studies at the scale of a small catchment in the Cévennes region where AROME shows a clear added value (Caillaud et al., 2021).

Figure 3Distribution of wet hourly rainfall (>0.1 mm h⁻¹) for raw (a, c) and bias-corrected (b, d) precipitation datasets. Panels (a) and (b) show the retrospective simulations (2000–2018). Panels (c) and (d) depict the historical (1986–2005) and future (2080–2099) simulations under the RCP8.5 scenario. COMEPHORE observed hourly rainfall is appended to all panels (2000–2018). Probabilities under the abscissa axis are shown under a Gumbel transformation.

Download

Table 3Evaluation of raw and corrected simulated precipitation with respect to the COMEPHORE observed dataset.

Download Print Version | Download XLSX

The future ALADIN projected rainfall is lower than that of AROME, for the whole distribution. Extreme projected precipitation for ALADIN is lower than for the historical AROME data, partially due to the persistent bias. However, the projected hourly precipitation shows an increase in extreme hourly rainfall (> 95th percentile) under a warmer climate (RCP8.5 projection) for both models. This signal is stronger for the most extreme hourly rainfall (> 99.9th percentile) and for the AROME simulation. This local-scale result is in agreement with the broader multi-model ensemble study of Pichelli et al. (2021), which compared an ensemble of CPMs and their driving RCMs over 10-year periods (historical and future). They found a consistent signal of hourly rainfall extremes over southern France between CPMs and RCMs for the wet season (fall). For the dry season, a slight decrease in the 99.9th percentile of hourly precipitation of CPMs is shown, consistent with the annual cycle projection of Fig. 2. The large increase in the hourly extreme simulated precipitation is maintained after the bias correction. The signal of the projected precipitation, after bias correction, is largely positive for all probabilities exceeding 95th percentiles for both models. While ALADIN shows a stronger increase for the 95th and 99th percentiles than those of AROME, AROME produces the most positive trend for the most extreme hourly corrected rainfall (see Table 4). Indeed, the most extreme projected hourly rainfall is therefore expected to reach hourly rainfalls that have never been observed at the catchment scale (>40 mm h⁻¹).

Table 4Changes in extreme hourly corrected precipitation (%) under the RCP8.5 scenario relative to historical simulation for AROME and ALADIN.

Download Print Version | Download XLSX

4.2 Hydrological models evaluation

Table 5 presents a summary of the performance metrics of the hydrological models (GR5H and CREST) calculated over the evaluation period. GR5H reproduces well the observed discharge with KGE and NSE scores higher than 0.75. High flows and mean flows are correctly simulated even if they are slightly underestimated. The CREST performance is lower for all the metrics. It shows an acceptable global efficiency and an overestimation of both mean flows and high flows. Both hydrological models underestimate flood peaks.

Table 5Evaluation of hydrological model simulations against observed discharge (2002–2018). Results for general efficiency scores (NSE: Nash–Sutcliffe efficiency; KGE: Kling–Gupta efficiency) and relative biases on mean flows, high flows (quantile 99.9) and flood peaks (observed flood peak over threshold).

Download Print Version | Download XLSX

Figure 4 shows the annual cycle of discharge at the outlet of the Gardon d'Anduze catchment. Both models are able to reproduce the annual cycle, in particular the high-flow season caused by HPEs and flood events. As seen on Table 5, CREST tends to produce a more excessive discharge response to rainfall than GR5H. GR5H tends to overestimate low flows for summer months, which is, however, not the focus of the present work. One should keep in mind that the two hydrological models have been calibrated using a different strategy; all parameters of GR5H have been calibrated while most of CREST parameters used a priori estimates, since it is a common strategy for fully distributed physically based models (Clark et al., 2017).

Figure 4Annual cycle of discharge for observation, GR5H and CREST evaluation simulations over the 2002–2018 period. Multiyear regime of 8 d averaged flows at Anduze (Gardon d'Anduze catchment).

Download

These results must be moderated looking at Fig. 5, which shows the cumulative distribution of the observed discharge along with the GR5H and CREST simulated discharge for the evaluation period. The study indeed focuses on the most intense flood events in this catchment, i.e., flood peak above 265 m³ s⁻¹, leading to two floods per year on average. Biases in CREST simulations (Table 5, Fig. 4) are not necessarily reflected in the flood distribution since both models manage to reproduce the observed flood distribution on this small catchment. We can see a slight underestimation of the most intense flood by both hydrological models. The most severe flood corresponds to the major flood event of September 2002, one of the most damaging flash floods recorded in France (Delrieu et al., 2005; Vinet, 2008). Even if the return period of this observed flood exceeds 50 years, this outlier value of the distribution of observed POT has to be carefully interpreted. The peak discharge value of a flood of this magnitude is likely inaccurate because of large uncertainties related to the measurement of the water level, the extrapolation from the rating curve, and possible modifications of river bed topography and flow dynamics (Neppel et al., 2010). However, in terms of flood frequency, the number of POTs differs between the two models. While CREST simulates on average 3.1 floods per year, GR5H is closer to observation with on average 1.5 floods per year.

Figure 5Cumulative distribution functions (CDFs) of observed and simulated flood peaks over the evaluation period (2002–2018). Both models have been calibrated and run with observed precipitation (COMEPHORE) and PET derived from the observed temperature. The threshold is a fixed discharge value giving two floods per year from the observed discharge (265 m³ s⁻¹).

Download

4.3 Reproducing floods with the climate datasets

The flood distributions of the AROME- and ALADIN-driven simulations with GR5H and CREST are presented in Fig. 6. The green line on the Fig. 6 represents the hydrological simulations forced by the COMEPHORE observed dataset that has been used for the calibration and could be considered as the reference simulation. The AROME and ALADIN datasets are retrospective simulations (evaluation) with (CDF-t) and without bias correction. We used the same threshold as in Fig. 5, with the discharge value (265 m³ s⁻¹) leading to two observed floods per year after the declustering. Above all, only uncorrected retrospective climate simulations can be used in the scope of assessing the performance of climate models for reproducing floods. Firstly, Fig. 6 shows a large underestimation of POT distributions by CREST and GR5H. The hydrological models forced with the AROME CPM reproduce flood intensity better than when forced with ALADIN. This can be observed by the shape of the distributions, which is almost flat for ALADIN. Then, the ALADIN-driven flood distributions fail to reproduce the observed flood frequency with 0.4 and 0.9 floods per year for GR5H and CREST, respectively, compared to AROME, which simulates an acceptable number of floods (1.2 for GR5H and 2.3 for CREST). Consequently, AROME seems more reliable than ALADIN in simulating floods with both hydrological models. The results shown in Fig. 6 confirm CREST over-reactivity with a higher number of POTs simulated than for GR5H. This model behavior does not impact the flood intensity. Conversely, discharges in the upper half of the flood peak distribution (cumulative frequency > 0.5) are slightly lower for CREST than for GR5H.

Figure 6POT CDF for GR5H (a, b) and CREST (c, d) forced by the observed and retrospective climate simulations (2001–2018). POT from raw climate datasets is on panels (a) and (c), and bias-corrected climate datasets are on panels (b) and (d). The threshold is the same as in Fig. 5.

Download

After the CDF-t bias correction, all POT distributions are closer to the hydrological simulation forced by the observed data (green curve). All simulated datasets are compared here with the simulated discharge driven by the observed precipitation (green curve). The ALADIN-driven hydrological simulation shows a slightly better POT distribution than the AROME-driven one for both models, even if the most intense flood event (September 2002) is better simulated with AROME. An improvement in flood frequency modeling is noticed after bias correction for all simulations. However, the CREST simulations still have a positive bias for flood frequency, with higher values than the COMEPHORE-driven simulation (2.9 and 3.3 floods per year for AROME and ALADIN, respectively). To summarize, the added value found for the CPM compared to the RCM in extreme precipitation in previous sections seems to be transmitted for flash flood simulation. Bias correction reduces the difference between climate models, with no remnant bias on POT distribution. The choice of the hydrological model therefore does not seem to impact the results between the forcing climate datasets.

4.4 Climate change (flood intensity and frequency)

This section aims to provide an overview of the flood signal suggested by future climate projections and whether this projected evolution is impacted by the different hydrological models and climate models used. We now analyze the flood distributions from the hydrological models forced with the AROME and ALADIN climate simulations under the historical and the future RCP8.5 scenario (Fig. 7). In the first place, the two hydrological models show flood peak discharge weaker for ALADIN than for AROME. CREST simulates higher floods for historical and future projections than GR5H. The shape of the distribution of POT simulated with CREST is less flat than for GR5H distribution, reflecting a behavior similar to a Pareto distribution, hence a tail of the distribution associated with more extreme, i.e., rare events. The number of floods is higher for the CREST-driven hydrological simulations than GR5H ones. These simulations can reach a flood frequency exceeding four floods per year on historical AROME- and ALADIN-driven discharge simulations.

Figure 7POT distributions for GR5H (a, b) and CREST (c, d) forced by historical and future climate simulations. POTs from raw climate datasets are on panels (a) and (c), and those from bias-corrected climate datasets are on panels (b) and (d). The threshold is the same as for Fig. 5.

Download

Flood distributions from the bias-corrected historical and future projections show a good consistency between hydrological models, but higher differences among the climate simulations. Figure 7 highlights a major difference between the AROME CPM and its driving RCM, ALADIN. While the ALADIN-driven simulations indicate for both hydrological models a generalized increase in magnitude of floods, the AROME-driven simulations show a different signal depending on the probability of occurrence. Moderate floods are projected to be weaker in a warmer future when using AROME, but there is a positive increase for the most extreme floods. The threshold related to this change in signal seems to be located between 0.7 and 0.75 for both hydrological models; i.e., 25 % to 30 % of the most extreme floods are projected to be stronger in a future climate. Negative changes for CREST are less pronounced than for GR5H, but on the contrary, positive changes above this threshold are stronger, especially for the most extreme projected flood events.

The different changes in moderate and large floods could be explained by a decrease in future soil moisture, which can compensate for the increase in heavy precipitation for small to moderate floods. This result is in agreement with the studies of Brunner et al. (2021) and Wasko et al. (2023). This is consistent as well with Tramblay et al. (2023) and Bertola et al. (2021), who showed the importance of antecedent soil moisture in modulating changes in floods. This threshold behavior is not present for the ALADIN-driven flood simulations. Indeed, the signal for the ALADIN flood projections is positive for almost all the flood distribution, except for the weakest floods where no clear signal is projected. Both bias-corrected ALADIN and AROME flood projection distributions show a strong increasing signal for the distribution tail. ALADIN-driven simulated floods in a future climate reach higher peak values than AROME when simulated by the GR5H model. For CREST, the distribution tail of future floods for AROME and ALADIN are of the same order of magnitude.

The different behavior between the AROME- and ALADIN-driven flood simulations questions the reliability of low-resolution RCMs for hydrological impact studies related to extreme events. The poor representation of topography along with the parameterization of deep convection for RCMs leads to strong biases on HPE's intensity and temporal distribution. Indeed, extreme rainfall events for ALADIN are generally composed of long-lasting moderate hourly rainfall rather than a more realistic convective precipitation peak. The bias correction method works on each individual hourly time step, but it does not influence the temporal distribution of rainfall events (i.e., hyetograph shape), leading to an over-correction of the HPEs and probably to excessive hydrological reactions (not shown here). This strong bias correction probably prevents ALADIN from adequately simulating future changes on local processes that are highlighted by the AROME CPM. After bias correction, projected changes in flood frequency also depend on the forcing climate simulation. While AROME shows no change in flood frequency, ALADIN simulates a large increase in the number of floods in future climatic conditions. This trend remains consistent among the two hydrological models, with only little differences between CREST and GR5H simulations (Fig. 7).

4.5 Climate change impact on flood characteristics

We analyzed the flood and associated rainfall event characteristics (Fig. 8) simulated by historical and future bias-corrected climate simulations. Every box plot is made of all values of a specific metric relative to all extracted flood events or the associated rainfall events for one dataset. Firstly, most simulated datasets show a negative future signal for the baseflow component of future floods. Only the GR5H model driven by ALADIN shows a little increase in the median, but this signal is not clear because the distribution becomes wider for future baseflows. The same future positive signal can be found on the ratio metric: all datasets yield a negative trend, reflecting an increase in the runoff part of the streamflow during floods. There is no clear signal about the changes in the length of precipitation events (n_p) and the total precipitation (P_tot) events associated with floods. However, the maximum hourly precipitation of the largest rainfall event preceding the floods is projected to strongly increase. This result is coherent with Sect. 4.1 and confirms that hourly precipitation extremes can yield severe floods in this basin. The most intense increases seem to concern ALADIN-driven simulations with shifts in distribution exceeding 25 % (the median in historical simulations corresponds to the first quartile in the future). These results strengthen the confidence of the hypothesis of the decrease in soil moisture, leading to the threshold effect in AROME-forced simulations. In this study we do not explicitly simulate soil moisture, but the evolution of the baseflow can be considered as a proxy for the changes in soil moisture (Massari et al., 2023). The negative trend in the baseflow compensates for the increase in precipitation extremes until a threshold is reached, when the most intense hourly rainfall exceeds the infiltration capacity.

Figure 8Box plots of historical (light orange) and RCP8.5 future (dark orange) flood event characteristics (b) and related rainfall events (a) using the bias-corrected simulations. Each characteristic is detailed in Sect. 3.3.

Download

For the ALADIN-driven simulations, the constant increase in POT (Figs. 7 and 8) is a combination of two processes. Firstly, decreases in baseflow are less pronounced for ALADIN than for AROME, in particular for the GR5H hydrological model, reflecting a smaller infiltration capacity and the prevention of the damping role of soils in a warmer future. Secondly, as explained above, this trend is likely an artifact of the bias correction causing stronger HPEs in a warmer climate. Indeed, this can be highlighted in Fig. 8, where P_tot shows a little increase for the ALADIN-driven simulations and a slight decrease in the medians of the AROME rainfall events. Outliers of P_tot reach the highest values for the ALADIN-driven simulations. The length of future floods (n_Q) decreases for all simulations except for GR5H ALADIN, which shows a slight increase in the median and a lower spread for future scenarios. The lag time (LT) is projected to decrease for all simulations. This consistent signal of a shorter period between rainfall centroid and flood peak indicates a projected increase in the flood flashiness in this basin (FI). In more detail, the median of the flashiness index, representing reactivity and intensity of floods, is projected to increase for all simulations. The smallest increase in the flashiness index is shown for GR5H-AROME, while the other simulations show noteworthy positive trends and higher spreads. Some outliers reach extremely high values in the future projections, warning of potentially rare, but very extreme, flood events occurring in the future. This result is in agreement with the study of Li et al. (2022b), who found an increase in flash flood potential over the USA in the high-end emission scenario, notably in southern regions that have a wet convective season such as this Mediterranean catchment. However, the projected changes over this catchment should be interpreted with caution as they come from a single pair of RCM and CPM simulations. With these preliminary results, the different impacts of climate change on flood characteristics highlight the fine-scale benefits of CPM in simulating underlying hydrological processes. The robustness of the climate change impact on flash floods still needs to be confirmed with a comprehensive study that includes an uncertainty assessment.