Spatial extremes in the hydro- and atmosphere: understanding and modelling



Hydro-meteorological extremes such as floods, droughts, storms, or heatwaves often affect large regions therefore causing large damages and costs. Hazard and risk assessments, aiming at reducing the negative consequences of such extreme events, are often performed with a focus on one location despite the spatial nature of extreme events. While spatial extremes receive a lot of attention by the media, little is known about their driving factors and it remains challenging to assess their risk by modelling approaches. Key challenges in advancing our understanding of spatial extremes and in developing new modeling approaches include the definition of multivariate events, the quantification of spatial dependence, the dealing with large dimensions, the introduction of flexible dependence structures, the estimation of their probability of occurrence, the identification of potential drivers for spatial dependence, and linking different spatial scales.

This session invites contributions which help to better understand processes governing spatial extremes and/or propose new ways of describing and modeling spatial extremes at different spatial scales.

Target audience: hydrologists, climatologists, statisticians, machine learners, and researchers interested in spatial risk assessments.

Co-organized by AS4/NH1
Convener: Manuela Irene BrunnerECSECS | Co-conveners: András Bárdossy, Philippe Naveau, Simon Michael PapalexiouECSECS, Elena Volpi
vPICO presentations
| Thu, 29 Apr, 14:15–15:00 (CEST)

vPICO presentations: Thu, 29 Apr

Chairpersons: Manuela Irene Brunner, Elena Volpi, András Bárdossy
Deeksha Rastogi, Danielle Touma, Katherine Evans, and Moetasim Ashfaq

An intensification of hydroclimate extremes in response to increase in radiative forcing has the potential to cause severe and widespread socioeconomic damages. Therefore, a comprehensive evaluation of projected changes in the characteristics of these extremes in a warming climate is necessary for emergency preparedness and planning. While the intensity and frequency of these extremes have been thoroughly investigated, the efforts on understanding their spatial characteristics are still limited. To this end, we use an ensemble of high-resolution regional climate simulations to investigate the spatial characteristics of daily-scale precipitation events across the United States, in addition to other features. The simulations cover 1966–2005 in the historical period and 2011–2050 in the future period under Representative Concentration Pathway 8.5 (RCP 8.5) scenario. The simulated ensemble compares well with observations in the historical period, and project further intensification of widespread extremes in the near future. Further, our results demonstrate that the projected changes in the characteristics of precipitation events are associated with more frequent occurrences of extreme years where contributions from intense and widespread events to the annual precipitation is unprecedently high. These findings highlight the need for more rigorous investigations of changes in the spatial characteristics of extremes to prepare for potential future changes and associated risks.

How to cite: Rastogi, D., Touma, D., Evans, K., and Ashfaq, M.: Investigating Future Changes in the Spatial Characteristics of Precipitation Extremes over the United States, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-6096, https://doi.org/10.5194/egusphere-egu21-6096, 2021.

Jordan Richards, Jonathan Tawn, and Simon Brown

Fluvial flooding is not caused by high intensity rainfall at a single location, rather it is caused by the extremes of precipitation events aggregated over spatial catchment areas. Accurate modelling of the tail behaviour of such events can help to mitigate the financial aspects associated with floods, especially if river defences are built within specification to withstand an n-year event of this kind. Within an extreme value analysis framework, univariate methods for estimating the size of these n-year events are well studied and cemented in asymptotic theory.

 To complement these techniques, we develop a high-resolution spatial model for extreme precipitation by providing a fully spatial extension of the conditional approach for modelling multivariate extremes. We simulate realistic precipitation fields from this model and use univariate techniques to make inference about the extremal behaviour of aggregates over specified spatial domains. The challenge of zero precipitation data is overcome and further applications of the model are discussed. The model is fit to data from a convection permitting forecast model within the 2018 UK Climate Projections (UKCP18).

How to cite: Richards, J., Tawn, J., and Brown, S.: Modelling the tail behaviour of precipitation aggregates using conditional spatial extremes., EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-472, https://doi.org/10.5194/egusphere-egu21-472, 2021.

Chandra Rupa Rajulapati, Simon Michael Papalexiou, Martyn P Clark, Saman Razavi, Guoqiang Tang, and John Pomeroy

Assessing extreme precipitation events is of high importance to hydrological risk assessment, decision making, and adaptation strategies. Global gridded precipitation products, constructed by combining various data sources such as precipitation gauge observations, atmospheric reanalyses and satellite estimates, can be used to estimate extreme precipitation events. Although these global precipitation products are widely used, there has been limited work to examine how well these products represent the magnitude and frequency of extreme precipitation. In this work, the five most widely used global precipitation datasets (MSWEP, CFSR, CPC, PERSIANN-CDR and WFDEI) are compared to each other and to GHCN-daily surface observations. The spatial variability of extreme precipitation events and the discrepancy amongst datasets in predicting precipitation return levels (such as 100- and 1000-year) were evaluated for the time period 1979-2017.  The behaviour of extremes, that is the frequency and magnitude of extreme precipitation, was quantified using indices of the heaviness of the upper tail of the probability distribution. Two parameterizations of the upper tail, the power and stretched-exponential, were used to reveal the probabilistic behaviour of extremes. The analysis shows strong spatial variability in the frequency and magnitude of precipitation extremes as estimated from the upper tails of the probability distributions. This spatial variability is similar to the Köppen-Geiger climate classification. The predicted 100- and 1000-year return levels differ substantially amongst the gridded products, and the level of discrepancy varies regionally, with large differences in Africa and South America and small differences in North America and Europe. The results from this work reveal the shortcomings of global precipitation products in representing extremes. The work shows that there is no single global product that performs best for all regions and climates.

How to cite: Rajulapati, C. R., Papalexiou, S. M., Clark, M. P., Razavi, S., Tang, G., and Pomeroy, J.: Reliability of global gridded precipitation products in assessing extremes, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-3246, https://doi.org/10.5194/egusphere-egu21-3246, 2021.

Paola Mazzoglio, Ilaria Butera, and Pierluigi Claps

The intensity and the spatial distribution of precipitation depths are known to be highly dependent on relief and geomorphological parameters. Complex environments like mountainous regions are prone to intense and frequent precipitation events, especially if located near the coastline. Although the link between the mean annual rainfall and geomorphological parameters has received substantial attention, few literature studies investigate the relationship between the sub-daily maximum annual rainfall depth and geographical or morphological landscape features.
In this study, the mean of the rainfall extremes in Italy, recently revised in the so-called I2-RED dataset, are investigated in their spatial variability in comparison with some landscape and also some broad climatic characteristics. The database includes all sub-daily rainfall extremes recorded in Italy from 1916 until 2019 and this analysis considers their mean values (from 1 to 24 hours) in stations with at least 10 years of records, involving more than 3700 stations.
The geo-morpho-climatic factors considered range from latitude, longitude and minimum distance from the coastline on the geographic side, to elevation, slope, openness and obstruction morphological indices, and also include an often-neglected robust climatological information, as the local mean annual rainfall.
Obtained results highlight that the relationship between the annual maximum rainfall depths and the hydro-geomorphological parameters is not univocal over the entire Italian territory and over different time intervals. Considering the whole of Italy, the highest correlation is reached between the mean values of the 24-hours records and the mean annual precipitation (correlation coefficient greater than 0.75). This predominance remains also in sub-areas of the Italian territory (i.e., the Alpine region, the Apennines or the coastal areas) but correlation decreases as the time interval decreases, except for the Alpine region (0.73 for the 1-hour maximum). The other geomorphological parameters seem to act in conjunction, making it difficult to evaluate, with a simple linear regression analysis, their impact. As an example, the absolute value of the correlation coefficient between the elevation and the 1-hour extremes is greater than 0.35 for the Italian and the Alpine regions, while for the 24-hours interval it is greater than 0.35 over the coastal areas.
To further investigate the spatial variability of the relationship between rainfall and elevation, a spatial linear regression analysis has been undertaken. Local linear relationships have been fitted in circles centered on any of the 0.5-km size pixels in Italy, with 1 to 30 km radius and at least 5 stations included. Results indicate the need of more comprehensive terrain analysis to better understand the causes of local increasing or decreasing relations, poorly described in the available literature.

How to cite: Mazzoglio, P., Butera, I., and Claps, P.: How landscape and climate affect the spatial variability of the Italian rainfall extremes? Some initial clues based on I2-RED, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7159, https://doi.org/10.5194/egusphere-egu21-7159, 2021.

David Cross and John Paul Gosling

Assessment of both localised and widespread flooding is vital for flood insurance to ensure adequate financial protection for businesses and property owners alike. But modelling precipitation and catchment response on very large spatial scales remains a challenge because of the availability of data and the high dimensionality of the problem. Modelling flood risk for insurance requires spatially coherent estimation of extremes which go beyond the historical record. At the national and continental scale, it can be difficult to apply models which maintain both the dependence structure of the precipitation field and the marginal distributions which determine local impacts. Recent research into spatiotemporal random fields modelling is highly promising. Numerical weather prediction is also an attractive prospect because correlations are implicitly captured in physical processes, but the computational demand and the uncertainty of perturbed physics ensembles can limit its usefulness.  

We introduce a data driven approach for widescale flood risk assessment based on modelling extreme precipitation fields. Using gridded reanalysis precipitation data, we identify extreme precipitation events in space and time using a measure of correlation in the tails of the marginal distributions. The simulation of extreme precipitation follows two main processes. First, the timing and extent of events are modelled using a Poisson distribution for event triggers, and a spatial Poisson process perturbs event footprints for observed events in the neighbourhood of the trigger location. The second stage is to model the extreme precipitation field within the event footprint. A Copula process is used to estimate extreme precipitation quantiles for all simulation points within the event ensuring internal spatial coherence. Our method has the flexibility to model extreme precipitation with any underlying physical conditions using computationally efficient models which facilitate widescale risk assessment.

How to cite: Cross, D. and Gosling, J. P.: Modelling extreme precipitation fields for large scale flood insurance, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10095, https://doi.org/10.5194/egusphere-egu21-10095, 2021.

Bora Shehu, Winfried Willems, Luisa Thiele, Henrike Stockel, and Uwe Haberlandt

Rainfall intensity-duration-frequency (IDF) curves are required for the design of several water systems and protection works. These curves are typically generated from the station data by fitting a theoretical distribution either to the annual extremes (AMS) or partial extremes (PE) series. Nevertheless, two main problems arise: i) for generating intensity depth for high return periods, long time series are needed (more than 40 years). While this is the case mainly for daily recordings, for sub-hourly time series only few point measurements are available. ii) as the station data are only local measurements, there is a need for regionalization of the of IDF curves to ungauged locations. Thus, the aim of this study is to investigate the use of different data types and methods in generating reliable IDF curves for ungauged locations.

For this purpose, the available gauge data from the German Weather Service (DWD) in Germany are employed, which include: 5000 daily stations with more than 40 years available, 1100 sub-hourly (5min) recordings with observations period shorter than 20 years, and finally 89 sub-hourly (5min) recordings with 60-70 years of observations. Annual extremes are extracted for each location for different durations D=5, 10, 15, 30, 60, 120, 180, 240, 360, 720, 2880 minutes, and a Generalized Extreme Value (GEV) probability distribution is fitted to each duration level as well as across all duration levels by the methods of the L-moments and Maximum-Likelihood, in order to derive the intensity quantiles for the given return periods Ta=2, 10, 20 and 100 years. First, a disaggregation scheme to 5 min resolution is performed on the daily recordings in order to investigate if disaggregated daily data can be useful for the IDF estimation of sub-daily durations. Then, the rainfall extremes of short observations are corrected by a correlation-based augmentation method. Finally, as the extreme intensities and durations are co-dependent, a normalization of the AMS over all the durations is performed.

To evaluate the regionalization of the IDF curves to ungauged regions, three methods are investigated: i) flood index method ii) regionalization with normalization of extremes over the durations and ii) kriging interpolation (ordinary and external drift kriging) of local AMS quantiles or parameters of the fitted distribution. The performance of these regionalization techniques is then evaluated by cross-validation, where the local IDF from the long sub-hourly time series are considered the true reference. Based on the relative bias, rmse and correlation the best method is selected and used for the regionalization of the IDF curves in Germany. Different data products are fed in the regionalization methods to answer the following questions: are the disaggregated long time series useful in regionalizing sub-hourly IDF? Can space be traded for time (and vice versa) when regionalizing IDF? What is the best incorporation of different data sets for the regionalization of the IDF? Lastly, a bootstrap method is as well employed to account for the uncertainties in estimation intensity-duration extremes for the given return periods.

How to cite: Shehu, B., Willems, W., Thiele, L., Stockel, H., and Haberlandt, U.: Regionalization of Intensity-Duration-Frequency Curves for different data types in Germany, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7622, https://doi.org/10.5194/egusphere-egu21-7622, 2021.

Katharina Schroeer, Cornelia Schwierz, Simona Trefalt, Alessandro Hering, and Urs Germann

Hailstorms and associated hail stone sizes are a tricky atmospheric hazard to assess, because the processes leading to severe convective weather are complex and the spatiotemporal scales of the impacts are often small. The high natural variability of hail requires expensive high-resolution, area-covering measurements to establish robust statics. Weather radars help to achieve this, but despite growing data archives, records usually do not yet extend to climatological time scales (≥30y), and reference ground observations to calibrate hail algorithms are still fragmentary. Consequentially, there remain substantial uncertainties regarding the long-term hazard of hail. Nevertheless, stakeholders require estimates of return periods for preventive regulations or as input to downstream impact models, e.g., in the insurance and engineering sector.

In the project “Hail climatology Switzerland” MeteoSwiss partnered up with three federal offices, the insurance and engineering sectors to establish a common national reference of the occurrence of hail in Switzerland. The deliverables include developing return period maps of extreme hail events. However, the definition of such extremes varies across sectors. For example, stakeholders from damage prevention require impact probabilities of the largest hailstorm onto an average rooftop, whereas reinsurance stakeholders are interested in nation-wide worst-case events. Here we report on the approaches we took in deriving the frequencies of severe hail considering the different stakeholder demands and the challenges and uncertainties we thereby encountered.

Using newly reprocessed gridded radar hail data, we assess frequencies of observed hail occurrence in Switzerland over 19 years (2002-2020). We further developed a probabilistic hazard model using stochastic resampling of hailstorms, driven by large-scale environmental boundary conditions. In order to take a storm-object perspective on extremes, we isolate more than 40’000 individual hailstorm footprints. This allows us to consider local storm properties such as the distributions of hail stone sizes by storm area and duration. In addition, we identify region-dependent extreme storm properties, which is specifically relevant in the Alpine region, where high and complex topography creates sharp climatic gradients and results from other regions are often not easily transferable.

Results show that observed storm tracks vary strongly between years, and hail footprints vary substantially by storm type. Comparing our results obtained from the longest radar-based hail record so far, we find that the spatial patterns of hail agree well with existing hazard maps derived, i.a., from damage claims. However, we also find that frequencies of local extreme hail stone sizes may have been underestimated in the past. This is further corroborated by a regionally aggregated comparative analysis of the radar record to historical records of very large hail in Switzerland over the past century.

How to cite: Schroeer, K., Schwierz, C., Trefalt, S., Hering, A., and Germann, U.: Estimating the hazard from extreme hail events across regional to local scales, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-16123, https://doi.org/10.5194/egusphere-egu21-16123, 2021.

Ngoc Tran, Johannes Buck, and Claudia Kluppelberg

Causal inference for extreme aims to discover cause and effect relation between large observed values of random variables. This is a fundamental problem to in many applications, from finance, engineering risks, nutrition to hydrology, to name a few. Unique challenges to
extreme values are lack of data and lack of model smoothness due to the max operator. Existing methods in extreme value statistics for dimensions d ≥ 3 are limited due to an intractable likelihood, while techniques for learning Bayesian networks require a large amount of data to fit nonlinear models. This talk showcases the max-linear model and new algorithms for fitting them. Our method performs well on real data, recovering a directed graph for both the Danube and the Lower Colorado with high accuracy purely through extreme measurements. We also compare our method to state-of-the-art algorithms for causal inference for nonlinear models, and outline open problems in hydrology, extreme value statistics and machine learning.

How to cite: Tran, N., Buck, J., and Kluppelberg, C.: Causal inference for extremes on river networks, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-341, https://doi.org/10.5194/egusphere-egu21-341, 2021.

Peng Zhong, Raphael Huser, and Thomas Opitz

The modeling of spatio-temporal trends in temperature extremes can help better understand the structure and frequency of heatwaves in a changing climate, as well as their environmental, societal, economic and global health-related risks. Here, we study annual temperature maxima over Southern Europe using a century-spanning dataset observed at 44 monitoring stations. Extending the spectral representation of max-stable processes, our modeling framework relies on a novel construction of max-infinitely divisible processes, which include covariates to capture spatio-temporal non-stationarities. Our new model keeps a popular max-stable process on the boundary of the parameter space, while flexibly capturing weakening extremal dependence at increasing quantile levels and asymptotic independence. It clearly outperforms natural alternative models. Results show that the spatial extent of heatwaves is smaller for more severe events at higher altitudes and that recent heatwaves are moderately wider. Our probabilistic assessment of the 2019 annual maxima confirms the severity of the 2019 heatwaves both spatially and at individual sites, especially when compared to climatic conditions prevailing in 1950-1975. Our applied results may be exploited in practice to understand the spatio-temporal dynamics, severity, and frequency of extreme heatwaves, and design suitable regional mitigation measures.

How to cite: Zhong, P., Huser, R., and Opitz, T.: Statistical Modeling of Non-Stationary Heatwave Hazard, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-56, https://doi.org/10.5194/egusphere-egu21-56, 2021.

Andrea Böhnisch, Elizaveta Felsche, and Ralf Ludwig

Heat waves are among the most hazardous climate extremes in Europe, commonly affecting large regions for a considerable amount of time. Especially in the recent past, heat waves account for substantial economic, social and ecologic impacts and loss. Projections suggest that their number, duration and intensity increase under changing climate conditions, stressing the importance of quantifying their characteristics. Yet, apart from the analysis of single historical events, little research is dedicated to the general propagation of heat waves in space and time.  
Heat waves are rare in their occurrence and limited observational data provide little means for robust analyses and the understanding of dynamical spatio-temporal patterns. Therefore, we seek to increase the number of analyzable events by using a large climate model ensemble. The use of several model members of comparable climate statistics allows to robustly assessing various spatial patterns of heat waves as well as their typical temporal evolutions.  
Here, we explore a data-driven approach to infer cause-and-effect relationships from, in this case, regional climate model ensemble data in order to analyze the spatio-temporal propagation of spatially distributed phenomena. Our aim is to investigate specifically the transitions and inter-dependencies among heat waves in Europe. The approach includes the identification of most frequent heat wave patterns by clustering and the derivation of directed links between core regions of these heat wave classes using causal discovery in a data set of high spatial resolution. 
We present the setup of our framework, including clustering results of heat waves and first results of our analysis.

How to cite: Böhnisch, A., Felsche, E., and Ludwig, R.: Detecting the spatio-temporal propagation of heatwaves, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7391, https://doi.org/10.5194/egusphere-egu21-7391, 2021.

Vera Melinda Galfi and Valerio Lucarini

We analyse persistent temperature events, like heat waves or cold spells, by applying large deviation theory (LDT), and show that events with a long duration have also a substantial spatial extension. We point out that by using LDT one finds typical spatial patterns related to the persistent temperature extremes. Based on the output of a state-of-the-art climate model, we define the climatology of persistent heatwaves and cold spells in some key target regions of the planet by constructing empirically the corresponding rate functions for the surface temperature, and we assess the impact of increasing CO2 concentration on such persistent anomalies. In particular, we notice the increasing hazard associated to heatwaves in a warmer climate. We show that two 2010 high impact events - summer Russian heatwave and winter Dzud in Mongolia – are associated with extended atmospheric patterns that are exceptional compared to the typical ones, but typical compared to the climatology of extreme events.

How to cite: Galfi, V. M. and Lucarini, V.: On spatial patterns of heat waves and cold spells from a large deviations perspective, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15371, https://doi.org/10.5194/egusphere-egu21-15371, 2021.