Robust ensemble selection by multivariate evaluation of extreme precipitation and temperature characteristics

Abstract

Extreme hydrometeorological events often cause severe socioeconomic damage. For water resources assessments and policy recommendations, future extreme hydrometeorological events must be correctly estimated. For this purpose, projections from Regional Climate Models (RCMs) are increasingly used to provide estimates of meteorological variables such as temperature and precipitation. The main objective of this study is to investigate whether a full ensemble or a subset of RCMs reproduces the spatiotemporal variability of observed extremes better than single models. The implications for policy recommendations and impact assessments are then discussed. In particular, the key conditions under which a subset of RCMs could be used for impact assessments are examined. Temperature and precipitation fields of 13 ENSEMBLES RCMs are compared against observations from Germany between 1961 and 2000. Eleven indices characterizing extreme meteorological events were selected for this comparison. The ability of the individual RCMs is estimated based on an overall score and a rejection rate. The former quantifies the biases of these indices. The latter estimates the mean statistical significance quantified by the Wilcoxon rank-sum test. The performance of all possible combinations of RCMs is investigated. Computationally feasible algorithms for finding the best-performing subensemble are also presented and evaluated. One of the proposed algorithms is able to find subensembles with the lowest rejection rate, which are useful for either policy recommendations or impact assessments. These subsets of RCMs showed smaller and less significant bias than single RCMs or the full ensemble over several regions.

Stochastic temporal disaggregation of monthly precipitation for regional gridded data sets

“… quantifies the upper tail of the precipitation cdf whereas CDT indicates the duration of extreme dry spells. Both of these indices have been frequently used for the evaluation of modeled precipitation [ Sillmann and Roeckner , ; Herrera et al ., ; Thober and Samaniego , ].…”

“…Coarse global climate model variables can be dynamically downscaled using regional climate models [ Maraun et al ., ]. This approach considers the physics of atmospheric processes but often fails to reproduce extreme precipitation amounts that are important for hydrologic modeling [ Thober and Samaniego , ]. Additionally, dynamical downscaling is computationally very expensive, and the obtained spatial resolution is still too coarse for impact modeling [ D'Onofrio et al ., ].…”

“…This study is aiming at the temporal disaggregation of these variables, which is equally important for two reasons. First, temporal disaggregation of monthly CM outputs to daily values offers the possibility to obtain realistic extremes, which is often not achieved by daily CM outputs [ Thober and Samaniego , ], and simultaneously preserves the low‐resolution characteristics given by CMs. Second, multimodel ensemble (MME) approaches have been promoted in the past years to gain a better understanding of the uncertainty associated with modeling climate change [ Taylor et al ., ].…”

Ensemble‐based Reconstructed Forcing (ERF) for regional climate modeling: Attaining the performance at a fraction of cost

“…For example, to complete a 2‐D ensemble matrix of 10 RCMs running with IBCs from 12 GCMs (as in the European Coordinated Regional Climate Downscaling Experiment (EURO‐CORDEX) project), 120 individual simulations must be conducted [ Giorgi and Gutowski , ]. To lower the cost, subsets of MME simulations are often selected through processes that require detection and down weighting (or eliminating) of poor‐performing models [ Giorgi and Mearns , ; Tebaldi and Knutti , ; Pierce et al , ; Weigel et al , ; Thober and Samaniego , ; Xie et al , ]. Properly evaluating the models and designing optimal ensembles pose yet another challenge [ Gleckler et al , ; Knutti et al , ; Schaller et al , ; Pennell and Reichler , ], as a model may perform well for one variable but not another.…”

“…For any given emission scenario, model structure uncertainties dominate, resulting in substantial discrepancies among GCM projections [Cook, 2008;Räisänen, 2007;Schaller et al, 2011;Shepherd, 2014]. RCMs are often more desirable for future climate projections due to their generally higher spatial resolution (needed for vulnerability assessment and adaptation studies) [Schaller et al, 2011;Thober and Samaniego, 2014;Giorgi and Gutowski, 2015;Xie et al, 2015] but are subject to additional uncertainties. As limited area models, RCMs rely on GCM output to set initial and boundary conditions (IBCs) [Giorgi and Gutowski, 2015;Xie et al, 2015].…”