HS7.7

The potential for changes in hydrometeorological extremes is routinely investigated by fitting change-permitting extreme value models to long-term observations, allowing one or more distribution parameters to change as a function of time or some physically-motivated covariate. In most practical extreme value analyses, the main quantity of interest though is the upper quantiles of the distribution, rather than the parameters' values. This study focuses on the changes in quantile estimates under different change-permitting models. First, metrics which measure the impact of changes in parameters on changes in quantiles are introduced. The mathematical structure of these change metrics is investigated for several models based on the Generalised Extreme Value (GEV) distribution. It is shown that for the most commonly used models, the predicted changes in the quantiles are a non-intuitive function of the distribution parameters, leading to results which are difficult to interpret. Next, it is posited that commonly used change-permitting GEV models do not preserve a constant coefficient of variation, a property that is typically assumed to hold and that is related to the scaling properties of extremes. To address these shortcomings a new (parsimonious) model is proposed: the model assumes a constant coefficient of variation, allowing the location and scale parameters to change simultaneously. The proposed model results in more interpretable changes in the quantile function. The consequences of the different modelling choices on quantile estimates are exemplified using a dataset of extreme peak river flow measurements.

How to cite: Prosdocimi, I. and Kjeldsen, T.: Understanding change in hydrometeorological extremes with statistical models - the importance of model parametrization, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1071, https://doi.org/10.5194/egusphere-egu21-1071, 2021.

Curves of rainfall intensity at various scales and for various return periods, else known as ombrian (or IDF) curves, are central design tools in hydrology and engineering. Construction of such curves often relies heavily on empirical or semi-empirical approaches, which hinder their applicability over large scales, and preclude simulation. Recent work by Koutsoyiannis (2020) has advanced these curves to theoretically-consistent stochastic models of rainfall intensity (ombrian models) extending their applicability to the full range of available scales, e.g. from minutes to decades. We present an open-source python toolbox implementing these advances in a straightforward and user-friendly manner and prove its applicability. The toolbox also employs advanced statistical fitting methods for extremes (K-moments), accounts for bias induced by temporal dependence, and allows optional blending of daily-scale data to reduce uncertainty of sub-daily records. The end result is the parameterization of the ombrian model and the graphical representation of rainfall intensity for any range of scales (supported by the data) and return periods.

Reference: Koutsoyiannis, D. 2020. ‘Rainfall extremes and Ombrian modelling’ in Stochastics of Hydroclimatic Extremes - A Cool Look at Risk (ed 0), National Technical University of Athens, Athens, pp 243-273, http://www.itia.ntua.gr/en/docinfo/2000/.

How to cite: Iliopoulou, T. and Koutsoyiannis, D.: PythOm: A python toolbox implementing recent advances in rainfall intensity (ombrian) curves, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-389, https://doi.org/10.5194/egusphere-egu21-389, 2021.

Intensity-Duration-Frequency (IDF) curves describe the main statistical characteristics of extreme precipitation events. Providing information on the exceedance probability or return period of certain precipitation intensities for a range of durations, IDF curves are an important tool for the design of hydrological structures.

Although the Generalized-Extreme-Value (GEV) distribution is an adequate model for annual precipitation maxima of a certain duration, the core problem of extreme value statistics remains: the limited data availability. Hence, it is reasonable to use a model that can describe all durations simultaneously. This reduces the total number of parameters and a more efficient usage of data is achieved. The idea of implementing a duration dependence directly into the parameters of the extreme value distribution and therefore obtaining a single distribution for a range of durations was proposed by Koutsoyiannis et al. (1998). However, while the use of the GEV is justified by a strong theoretical basis, only empirical models exist for the dependence of the parameters on duration.

In this study, we compare different models regarding the dependence of the GEV parameters on duration with the aim of finding a model for a wide duration range (1 min - 5 days). We use a combination of existing model features, especially curvature for small durations and multi-scaling for all durations, and extend them by a new feature that allows flattening of the IDF curves for long durations. Using the quantile score in a cross-validation setting, we provide detailed information on the duration and probability ranges for which specific features or a systematic combination of features lead to improved modeling skill.

Our results show that allowing curvature or multi-scaling improves the model only for very short or long durations, respectively, but leads to disadvantages in modeling the other duration ranges. In contrast, allowing flattening of the IDF curves leads to an improvement for medium durations between 1 hour and 1 day without affecting other duration regimes.

How to cite: Fauer, F., Ulrich, J., Jurado, O. E., Ulbrich, U., and Rust, H. W.: An Extended Model in Estimating Consistent Quantiles for Intensity-Duration-Frequency Curves, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7562, https://doi.org/10.5194/egusphere-egu21-7562, 2021.

Determination of peak flow or flow hydrograph in ungauged basins can be affected by considerable degree of uncertainty. Despite the considerable efforts to overcome this challenge, current methods provide design flood estimates that are still highly uncertain in ungauged catchments, even in the UK where the gauged network is relatively dense. A possible solution may be found in stochastic approaches and more specifically in the Derived Flood Frequency method, which gives the possibility to decompose runoff response effects dictated by the dominant hydrological processes for a catchment under study. Data scarcity can be then circumvented by application of UK-specific stochastic models, from which rainfall events and their relevant features are sampled. In this work, the latter rainfall model will be presented as a joint distribution function of spatial and temporal moments of catchment rainfall, along with their Intensity and Total Depth. The marginal distributions for each rainfall characteristic are studied through the L-moment method, which was previously developed for regional frequency analysis. The multivariate distribution of these rainfall characteristics will be described through the Vine Copula method, which can account for dependence very flexibly among several variables. Parameterisation procedures still require more development to allow application over ungauged case of studies.

How to cite: Quaglia, R., Woods, R., and Han, D.: Joint Distribution of Rainfall Characteristics: Intensity, Total Depth, Spatial and Temporal Moments, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8961, https://doi.org/10.5194/egusphere-egu21-8961, 2021.

To estimate return levels of wave heights (Hs) induced by tropical cyclones at the coast, a commonly-used approach is to (1) randomly generate a large number of synthetic cyclone events (typically >1,000); (2) numerically simulate the corresponding Hs over the whole domain of interest; (3) extract the Hs values at the desired location at the coast and (4) perform the local extreme value analysis (EVA) to derive the corresponding return level. Step 2 is however very constraining because it often involves a numerical hydrodynamic simulator that can be prohibitive to run: this might limit the number of results to perform the local EVA (typically to several hundreds). In this communication, we propose a spatial stochastic simulation procedure to increase the database size of numerical results with synthetic maps of Hs that are stochastically generated. To do so, we propose to rely on a data-driven dimensionality-reduction method, either unsupervised (Principal Component Analysis) or supervised (Partial Least Squares Regression), that is trained with a limited number of pre-existing numerically simulated Hs maps. The procedure is applied to the Guadeloupe island and results are compared to the commonly-used approach applied to a large database of Hs values computed for nearly 2,000 synthetic cyclones (representative of 3,200 years – Krien et al., NHESS, 2015). When using only a hundred of cyclones, we show that the estimates of the 100-year return levels can be achieved with a mean absolute percentage error (derived from a bootstrap-based procedure) ranging between 5 and 15% around the coasts while keeping the width of the 95% confidence interval of the same order of magnitude than the one using the full database. Without synthetic Hs maps augmentation, the error and confidence interval width are both increased by nearly 100%. A careful attention is paid to the tuning of the approach by testing the sensitivity to the spatial domain size, the information loss due to data compression, and the number of cyclones. This study has been carried within the Carib-Coast INTERREG project (https://www.interreg-caraibes.fr/carib-coast).

How to cite: Rohmer, J., Pedreros, R., and Krien, Y.: Spatial stochastic simulation to aid local extreme value analysis of cyclone-induced wave heights when numerical hydrodynamic simulations are scarce, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-2081, https://doi.org/10.5194/egusphere-egu21-2081, 2021.

We demonstrate that in data sparse environments, model parameter uncertainty is not the only cause of concern. To get a meaningful outcome, input data uncertainty has to be taken into account as well. The procedure involved calibration of a hydrological model using recent daily data rich time period along with validation. A historical flood was simulated (after warmup) for which the input data were relatively sparse in space, namely precipitation and temperature, using the calibrated model parameters. Precipitation was assumed to be the main driver of this event. Results showed that by only using interpolated precipitation (e.g. IDW or Kriging), the magnitude and timing of the peak were incorrect, even after using very many different parameter vectors that performed equally well for the recent times. Subsequently, the model was inverted for precipitation i.e. input fields that produced the correct timing, magnitude, dependence in space and distributions were searched for. This was done using a previously developed simulation algorithm. The new fields showed that the same hydrograph could have been produced by two main types of conditions, namely, early snow cover that melted and heavy rain. The plausibility of the simulated fields was also assessed by comparing their structure in space to events in recent times.

How to cite: Anwar, F., Bárdossy, A., and Seidel, J.: Hydrological modeling in data sparse environments, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-5040, https://doi.org/10.5194/egusphere-egu21-5040, 2021.

High-intensity and short-duration storms can generate pluvial floods in urban areas. Currently, 2D hydrodynamic models are recognised to be the best tool to simulate pluvial floods. The T-year synthetic design storm is usually assumed to generate the T-year pluvial flood. However, synthetic design storms cannot represent the variability in duration, precipitation and intensity temporal distribution of real storms that should be considered to account for their influence on water depths in pluvial floods. A more sound approach consists in estimating the T-year water depth in a given location from the frequency curve of water depths generated by a long series of possible rainfall events similar to the real storms.

However, 2D hydrodynamic models require high computation times that are not well suited with stochastic simulations. The Safer_RAIN tool is a rapid hydrostatic flood model based on a filling-and-spilling technique that has been developed within the SAFERPLACES project funded by the EIT Climate-KIC (Samela et al., 2020). Depressions and links between them are identified from a digital terrain model. The continuity equation is used to simulate how depressions are filled and spill to downstream depressions. Infiltration is simulated by using a distributed implementation of the Green and Ampt model that accounts for ponding time.

In this study, a stochastic methodology to delineate pluvial flood hazards is proposed in the Pamplona metropolitan area in Spain. First, the Safer_RAIN tool has been benchmarked by using spatially distributed high-resolution quantitative precipitation estimates (QPE) at time steps of 10 minutes for three real pluvial flood events. QPEs were obtained merging the data recorded at a set of automatic weather stations from the Spanish State Meteorological Agency (AEMET), the Regional Government of Navarre and crowdsourced networks, with continuous fields of radar observations. The Safer_RAIN tool has been benchmarked with the 2D hydrodynamic IBER model. In Barañáin, the results show a bias of -0.17–0.18 m and a RMSE of 0.22–0.49 m between water depths, as well as an accuracy correlation coefficient (ACC) of 0.87–0.99. In Zizur Mayor, the bias is -0.19–0.20 m, the RMSE is 0.29–0.55 m and the ACC is between 0.88 and 0.98.

Second, a long set of 10 000 synthetic storms has been generated by using a stochastic rainfall generator based on a bivariate copula approach fitted to data recorded at four rainfall-gauging stations located close to the case study. The 10 000 synthetic storms generated with a Gumbel copula fitted to the real rainfall events have been used as input data of the Safer_RAIN tool. Safer_RAIN preprocessing was done in 112 seconds and each simulation lasted around 45 seconds. A Generalized Pareto distribution function was fitted to the 10 000 water depth values in each cell of the grid. Pluvial flood hazard maps were obtained by estimating the T-year water depth in each cell of the grid.

Samela et al. (2020). Safer_RAIN: A DEM-Based Hierarchical Filling-&-Spilling Algorithm for Pluvial Flood Hazard Assessment and Mapping across Large Urban Areas, Water, 12, 1514.

How to cite: Mediero, L., Soriano, E., Oria, P., Bagli, S., Castellarin, A., Garrote, L., Mazzoli, P., Mysiak, J., Pasetti, S., Persiano, S., Santillán, D., and Schröter, K.: A stochastic methodology for pluvial flood mapping in urban areas with a fast-processing DEM-based flooding algorithm, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10187, https://doi.org/10.5194/egusphere-egu21-10187, 2021.