Presentation on theme: "Heiko Paeth, Institute of Geography, University of Würzburg,"— Presentation transcript:
1Heiko Paeth, Institute of Geography, University of Würzburg, Regional Dynamical Downscaling of Mediterranean Climate – Climate Change PerspectivesMedCLIVAR Workshop 2007, La Londe les MauresIntroductionDynamical downscalingExtreme value statisticsSimulated extreme eventsSimulated changesPostprocessing of model dataConclusions
3How can we infer future changes in the frequency and intensity I. IntroductionHow can we infer future changesin the frequency and intensityof extreme events?dynamical aspect (climate modelling)statististical aspect (assessment of uncertainty)
4II. Dynamical downscaling low latitudes are dominated by convective rain eventsthe spatial heterogeneity of individual rain events is highregional rainfall estimates are subject to large sampling errors
5II. Dynamical downscaling station data global model regional model statist. interpol.day-to-day variability annual precipitationstation data are often too sparse to represent regional rainfallglobal models are too coarse-grid for regional detailsstatistically interpolated data sets fail in mountainous areasdynamic nonlinear regional models account for the effect of orography
6II. Dynamical downscaling 3 x CO2the rainfall trends predicted by the global model are barely relevant to political plannings and measuresthe rainfall trends predicted by the regional model are much more detailed and of higher amplitudemore detailed fingerprint or spatial noise added value ???
7II. Dynamical downscaling Temperaturedifferencesbetweenensemblemembersat certaintime scalesmeasureof internalvariabilitydifferentinitialconditions(stochastic)statisticalcomparisonPrecipitationvarianceof theensemblemeanmeasureof externalvariabilityconsideration of various ensemble members enables the statistical quantification of the human impact on climate in the climate model
9II. Dynamical downscaling The main features of Mediterranean climate are well reproduced by REMO.
10III. Extreme value statistics fclimate parameterThe processes, which cause climate extremes, are not necessarily the same as for weak climate variations.Hence, they usually do not obey a normally distributed random process.
11III. Extreme value statistics The Generalized Pareto Distribution (GPD) is a useful statistical distribution, since it is a parent distribution for other extreme value distributions (Gumbel, Exponential, Pareto).The quantile function x(F) is given by:= location parameter (expectation)= scale parameter (dispersion)= shape parameter (skewness)The parameters of the GPD can be estimated by the method of L-moments.Estimation of T-year return values (RVs):cumulative GPDsT=5aq99%RV43mmdispersion parameter: threshold quantile
12III. Extreme value statistics uncertainty of the RV estimate is inferred from bootstrap sampling:from fitted GPD b random samples of size N generatedfrom random samples b indi- vidual RVs estimatedthese b RVs are normal distri-buted such that STD is a mea-sure of the standard error of the RV estimatesignal-to-noise ratio is given by MEAN/STD over b RVs1cGPDNrandomnumbersnew samples of size NmmfSTD90% conf. interv.RVchange in RV is significant at the 1% level, if 90% confidence inter-vals of two PDFs of RVs over b bootstrap samples do not overlap:fpresent-dayclimateforcedclimateRV
13III. Extreme value statistics 100-year RV in mmThe 100-year RV estimate ranges between 200 mm and 800 mm, depending on the random sample.
14III. Extreme value statistics single estimate / simulationone predicted valuewithout uncertainty range:pretended precisionRV2000205099%Monte Carlo approachprobabilistic forecastwith mean and uncertaintyrange:more objective basis fordecision makers90%s+=84%RVx=50%securitycostss-=16%10%200020501%probabilistic forecast of future rainfall changes provides a reasonable scientific basis for political plannings and measures
15IV. Simulated extreme events 1-year return values of heavy daily rainfallThe occurrence of extreme rain events is a function of the land-sea contrast, orography, geographical latitude and seasonal cycle.
16IV. Simulated extreme events 1-year return values of high daily temperatureThe occurrence of high temperature is also a function of the land-sea contrast, orography, geographical latitude and seasonal cycle.
17IV. Simulated extreme events S/N ratio for1-year RVsof heavydaily rainfallThe estimate of extrem values is more robust in regions and seasons with large-scale rather than convective precipitation.The choice of long return times in the pre-sence of short time series is unappropriate.
18V. Simulated changes extremes (1y-RV) seasonal means PRECIPITATION 2025 minuspresent-dayextremes (1y-RV)α = 5%seasonal means
19V. Simulated changes extremes (1y-RV) seasonal means TEMPERATURE 2025 minuspresent-dayextremes (1y-RV)α = 5%seasonal means
20VI. Postprocessing of model data assessed variabilitydiscontinuitydaily precipitationThe assessment of changes in weather extremes is very sensitive to inhomogeneities in observational data.No problem with model data.
21VI. Postprocessing of model data differentinitialconditions(stochastic)radiationbudget andenergyfluxesatmosphericand oceaniccirculationinstabilityandconvectioncloudmicro-physicsprecipitationerrornonlinear error growthtimeprecipitation is the end product of a complex causal chaineach step imposes addititional uncertainty, particularly if it is based on a physical parameterization in the model
22VI. Postprocessing of model data observed stationtime series(local information)REMO grid box(50km x 50km)climate models:area-meanprecipitationobservations:localstation datacomparison ?model datastation data
23VI. Postprocessing of model data Weather Generatorsimulatedgrid-boxprecipitation(dynamical part)local topography(physical part)random distributionin space(stochastical part)virtual station rainfall(result)
24VI. Postprocessing of model data original REMO rainfallREMO rainfall:- wrong seasonal cycle- underestimated extremes- hardly any dry spellsWeather Generator:- statistical distributionas observed- individual events not inphase with observationsrainfall from weather generatorstation time seriesmodel datastation datamodel data postprocessed
25VII. ConclusionsRegional climate models are required in order to account for the spatial heterogeneity of Mediterranean climate.The estimate of extreme values and their changes requires appropriate statistical distributions and a probabilistic approach.When estimating EVs from short time series, it is necessary to restrict the analysis to short return periods.The occurrence of climate extremes is a function of land-sea contrast, orography, geographical latitude and seasonal cycle.REMO projects no coherent changes in heavy rainfall whereas warm temperature extremes clearly tend to increase.Systematic model deficiencies and the grid-box problem can be overcome by use of a weather generator.The model results now need to be corroborated by available homogeneized long-term observational time series.