Regional Dynamical Downscaling of Mediterranean Climate Climate Change Perspectives

1
Regional Dynamical Downscaling ofMediterranean
Climate Climate Change Perspectives
Heiko Paeth, Institute of Geography, University
of Würzburg,
MedCLIVAR Workshop 2007, La Londe les Maures

• Introduction
• Dynamical downscaling
• Extreme value statistics
• Simulated extreme events
• Simulated changes
• Postprocessing of model data
• Conclusions

2
I. Introduction
industrial emissions
heat stress
traffic emissions
flood
biomass burning
drought
over- grazing
wind extremes
3
I. Introduction
How can we infer future changes in the frequency
and intensity of extreme events?

• dynamical aspect (climate modelling)
• statististical aspect (assessment of uncertainty)

4
II. Dynamical downscaling

• low latitudes are dominated by convective rain
  events
• the spatial heterogeneity of individual rain
  events is high
• regional rainfall estimates are subject to large
  sampling errors

5
II. Dynamical downscaling
station data global model
regional model statist. interpol.
day-to-day variability annual
precipitation

• station data are often too sparse to represent
  regional rainfall
• global models are too coarse-grid for regional
  details
• statistically interpolated data sets fail in
  mountainous areas
• dynamic nonlinear regional models account for
  the effect of orography

6
II. Dynamical downscaling
3 x CO2

• the rainfall trends predicted by the global
  model are barely relevant to political plannings
  and measures
• the rainfall trends predicted by the regional
  model are much more detailed and of higher
  amplitude
• more detailed fingerprint or spatial noise ?
  added value ???

7
II. Dynamical downscaling
Temperature
differences between ensemble members at
certain time scales
measure of internal variability
different initial conditions (stochastic)
statistical comparison
Precipitation
variance of the ensemble mean
measure of external variability

• consideration of various ensemble members
  enables the statistical quantification of the
  human impact on climate in the climate model

8
II. Dynamical downscaling
ECHAM5/MPI-OM 2001-2050 A1B (GHG) constant LC
Land degradation 2001-2050 FAO original
REMO 2001-2050 A1B (GHGLC)
REMO 1960-2000 observed GHG constant LC
ECHAM5/MPI-OM 1960-2000 observed GHG constant LC
REMO 2001-2050 B1 (GHGLC)
ECHAM5/MPI-OM 2001-2050 B1 (GHG) constant LC
Land degradation 2001-2050 FAO reduced

• dynamics hydrostatic
• physics ECHAM4
• sector 30W-60E 15S-45N
• resolution 0,5 20 hybrid levels
• validation good results

9
II. Dynamical downscaling

• The main features of Mediterranean climate are
  well reproduced by REMO.

10
III. Extreme value statistics
f
climate parameter

• The 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.

11
III. 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 GPDs
dispersion parameter threshold quantile
12
III. Extreme value statistics

• uncertainty of the RV estimate is inferred from
  bootstrap sampling

• from fitted GPD b random samples of size N
  generated
• from random samples b indi- vidual RVs estimated
• these b RVs are normal distri-buted such that STD
  is a mea-sure of the standard error of the RV
  estimate
• signal-to-noise ratio is given by MEAN/STD over b
  RVs

1
cGPD
N random numbers
0
mm
new samples of size N
f
STD
90 conf. interv.
RV

• change 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

f
present-day climate
forced climate
RV
13
III. Extreme value statistics
100-year RV in mm

• The 100-year RV estimate ranges between 200 mm
  and 800 mm, depending on the random sample.

14
III. Extreme value statistics
single estimate / simulation
one predicted value without uncertainty
range pretended precision
?
RV
2000
2050
99
probabilistic forecast with mean and
uncertainty range more objective basis
for decision makers
Monte Carlo approach
90
s84
security
costs
x50
RV
s-16
10
2000
2050
1

• probabilistic forecast of future rainfall
  changes provides a reasonable scientific basis
  for political plannings and measures

15
IV. Simulated extreme events
1-year return values of heavy daily rainfall

• The occurrence of extreme rain events is a
  function of the land-sea contrast, orography,
  geographical latitude and seasonal cycle.

16
IV. Simulated extreme events
1-year return values of high daily temperature

• The occurrence of high temperature is also a
  function of the land-sea contrast, orography,
  geographical latitude and seasonal cycle.

17
IV. Simulated extreme events
S/N ratio for 1-year RVs of heavy daily
rainfall

• The 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.

18
V. Simulated changes
PRECIPITATION 2025 minus present-day
extremes (1y-RV)
a 5
seasonal means
19
V. Simulated changes
TEMPERATURE 2025 minus present-day
extremes (1y-RV)
a 5
seasonal means
20
VI. Postprocessing of model data
assessed variability
discontinuity
daily precipitation
1840 1860 1880 1900
1920 1940 1960
1980 2000

• The assessment of changes in weather extremes is
  very sensitive to inhomogeneities in
  observational data.
• No problem with model data.

21
VI. Postprocessing of model data
different initial conditions (stochastic)
radiation budget and energy fluxes
atmospheric and oceanic circulation
instability and convection
cloud micro- physics
precipitation
error
nonlinear error growth
time

• precipitation is the end product of a complex
  causal chain
• each step imposes addititional uncertainty,
  particularly if it is based on a physical
  parameterization in the model

22
VI. Postprocessing of model data
observed station time series (local information)
REMO grid box (50km x 50km)
climate models area-mean precipitation
observations local station data
comparison ?
model data
station data
23
VI. Postprocessing of model data
Weather Generator
simulated grid-box precipitation (dynamical part)
local topography (physical part)
random distribution in space (stochastical part)
virtual station rainfall (result)
24
VI. Postprocessing of model data

• REMO rainfall
• - wrong seasonal cycle
• - underestimated extremes
• - hardly any dry spells
• Weather Generator
• - statistical distribution
• as observed
• - individual events not in
• phase with observations

original REMO rainfall
rainfall from weather generator
station time series
model data station data model data postprocessed
25
VII. Conclusions

• Regional 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.