Changes in Temperature and Precipitation Extremes in the IPCC Ensemble of Global
Coupled Model Simulations
VIATCHESLAV V. KHARIN AND FRANCIS W. ZWIERS
Canadian Centre for Climate Modelling and Analysis, Environment Canada, Victoria, British Columbia, Canada
XUEBIN ZHANG
Climate Data and Analysis Section, Environment Canada, Toronto, Ontario, Canada
GABRIELE C. HEGERL
Nicholas School for the Environment and Earth Science, Duke University, Durham, North Carolina
(Manuscript received 15 August 2005, in final form 5 September 2006)
ABSTRACT
Temperature and precipitation extremes and their potential future changes are evaluated in an ensemble
of global coupled climate models participating in the Intergovernmental Panel on Climate Change (IPCC)
diagnostic exercise for the Fourth Assessment Report (AR4). Climate extremes are expressed in terms of
20-yr return values of annual extremes of near-surface temperature and 24-h precipitation amounts. The
simulated changes in extremes are documented for years 2046–65 and 2081–2100 relative to 1981–2000 in
experiments with the Special Report on Emissions Scenarios (SRES) B1, A1B, and A2 emission scenarios.
Overall, the climate models simulate present-day warm extremes reasonably well on the global scale, as
compared to estimates from reanalyses. The model discrepancies in simulating cold extremes are generally
larger than those for warm extremes, especially in sea ice–covered areas. Simulated present-day precipita-
tion extremes are plausible in the extratropics, but uncertainties in extreme precipitation in the Tropics are
very large, both in the models and the available observationally based datasets.
Changes in warm extremes generally follow changes in the mean summertime temperature. Cold ex-
tremes warm faster than warm extremes by about 30%–40%, globally averaged. The excessive warming of
cold extremes is generally confined to regions where snow and sea ice retreat with global warming. With the
exception of northern polar latitudes, relative changes in the intensity of precipitation extremes generally
exceed relative changes in annual mean precipitation, particularly in tropical and subtropical regions.
Consistent with the increased intensity of precipitation extremes, waiting times for late-twentieth-century
extreme precipitation events are reduced almost everywhere, with the exception of a few subtropical
regions. The multimodel multiscenario consensus on the projected change in the globally averaged 20-yr
return values of annual extremes of 24-h precipitation amounts is that there will be an increase of about 6%
with each kelvin of global warming, with the bulk of models simulating values in the range of 4%–10% K
⫺1
.
The very large intermodel disagreements in the Tropics suggest that some physical processes associated with
extreme precipitation are not well represented in models. This reduces confidence in the projected changes
in extreme precipitation.
1. Introduction
Human activities and the environment are greatly
affected by climate and weather extremes. A growing
interest in extreme climate events is motivated by the
vulnerability of our society to the impacts of such
events. There is growing evidence suggesting that the
anthropogenic forcing is affecting the present climate
(International Ad Hoc Detection and Attribution
Group 2005) and will continue to do so in the future
(Cubasch et al. 2001). The impacts of the changing cli-
mate will likely be felt most strongly through changes in
intensity and frequency of climate extremes. It is there-
fore important to document future changes that might
be caused by anthropogenic activities.
Corresponding author address: Viatcheslav Kharin, Canadian
Centre for Climate Modelling and Analysis, University of Victo-
ria, P.O. Box 1700, STN CSC, Victoria, BC, Canada.
E-mail: slava.kharin@ec.gc.ca
V
OLUME 20 JOURNAL OF CLIMATE 15 APRIL 2007
DOI: 10.1175/JCLI4066.1
© 2007 American Meteorological Society
1419
JCLI4066
Simulations with global coupled ocean–atmosphere
general circulation models (CGCMs) forced with pro-
jected greenhouse gas and aerosol emissions are the
primary tools for studying possible future changes in
climate mean, variability, and extremes. Changes in
rainfall distributions have attracted much attention be-
cause of the particular vulnerability of human activities
to hydrological extreme events such as flood-producing
rains and droughts. The intensity of extreme precipita-
tion is projected to increase under global warming in
many parts of the world, even in the regions where
mean precipitation decreases (e.g., Kharin and Zwiers
2000, 2005; Semenov and Bengtsson 2002; Voss et al.
2002; Wilby and Wigley 2002; Wehner 2004). Future
increases in heavy precipitation are accompanied by
reduction in the probability of wet days, implying a
more extreme future climate with higher probabilities
of droughts and heavy precipitation events.
Changes in temperature extremes tend to follow
mean temperature changes in many parts of the world.
However, Kharin and Zwiers (2000, 2005) reported that
cold temperature extremes warm faster than warm ex-
tremes in mid- and high latitudes, mainly as a result of
snow and sea ice melting in winter under global warm-
ing. Increased temperature variability has been re-
ported in some studies over land in summer (Gregory
and Mitchell 1995; Kharin and Zwiers 2005), implying
potentially larger relative increases in warm extremes
than in mean summertime temperature.
The ability of the recent generation of atmospheric
general circulation models to simulate temperature and
precipitation extremes was recently documented by
Kharin et al. (2005) for models participating in the sec-
ond phase of the Atmospheric Model Intercomparison
Project (AMIP2). The purpose of the present study is
to document the performance of the current generation
of CGCMs in simulating present-day extremes of tem-
perature and precipitation and their potential changes
under different projections for the evolution of the an-
thropogenic forcing, using model output submitted to
the Program for Climate Model Diagnosis and Inter-
comparison (PCMDI; http://www-pcmdi.llnl.gov) in
support of the Intergovernmental Panel on Climate
Change (IPCC) Fourth Assessment Report (AR4).
The paper is organized as follows. Datasets are de-
scribed in the next section. Extreme value methodology
is summarized in section 3. The ability of the models to
simulate present-day precipitation and temperature ex-
tremes is documented in section 4. Their changes under
several emission scenarios are examined and discussed
in section 5. The paper is concluded by a summary in
section 6.
2. Datasets
The Working Group on Coupled Modeling (WGCM)
of the World Climate Research Program (WCRP) re-
quested that modeling groups submit daily model out-
put for a number of 20-yr time periods to PCMDI in
support of the IPCC AR4. In the present study we
analyze annual extremes of daily maximum and mini-
mum surface air temperature and of 24-h precipitation
amounts for the time period 1981–2000 from simula-
tions of the twentieth-century climate (20C3M), and for
two 20-yr time periods 2046–65 and 2081–2100 from the
Special Report on Emissions Scenarios (SRES) B1,
A1B, and A2 experiments. Figure 1 illustrates the time
evolution of carbon dioxide concentrations and sulfate
aerosol loadings in these three emission scenarios. The
gray shaded areas indicate the 20-yr time periods for
which daily temperature and precipitation output was
available for most of the models.
The B1 emission scenario, also known as the 550-
ppm stabilization experiment, envisions the slowest
growth of anthropogenic greenhouse gas concentra-
tions, followed by the A1B experiment, or the 720-ppm
stabilization experiment, with somewhat more rapid
forcing. Many groups continued these simulations up to
year 2300 with the concentrations held at the year-2100
level, but these stabilizations phases are not considered
in the present study. The fastest growing greenhouse
gas concentrations are specified in the A2 experiment
with roughly 1% per year of CO
2
increase in the second
half of the twenty-first century. The CO
2
concentra-
tions are similar in the A1B and A2 emission scenarios
up to the middle of the twenty-first century, but the A2
scenario also specifies somewhat greater sulfate aerosol
concentrations, which are thought to have a cooling
FIG. 1. The time evolution of the CO
2
concentrations (solid
lines, y axis on the left-hand side) and globally averaged sulfate
aerosol loadings scaled to year 2000 (dashed lines, y axis on the
right-hand side) as prescribed in the IPCC SRES B1, A1B, and
A2 experiments. The gray shaded areas indicate the time periods
analyzed in the present study.
1420 JOURNAL OF CLIMATE VOLUME 20
effect on surface temperature (e.g., Ramanathan et al.
2001).
The CGCMs that we analyzed are listed in Table 1
together with their horizontal grid resolutions and the
number of vertical levels in the corresponding atmo-
spheric components. Spectral atmospheric models are
also characterized by the spectral type and truncation.
Model output was available on a variety of grids with
resolution ranging from 72 ⫻ 45 to 320 ⫻ 160, with
the median resolution being 128 ⫻ 64. The vertical
resolution varies from 12 levels to 56 levels with the
median of 26 levels. Table 1 also lists estimates of equi-
librium climate sensitivities compiled from the PCMDI
IPCC model documentation Web site (http://www-
pcmdi.llnl.gov/ipcc/model_documentation/ipcc_model_
documentation.php and references therein). Equilib-
rium climate sensitivity is defined as the global surface
air temperature change under CO
2
doubling in slab
ocean experiments and ranges from 2.1 K in the Insti-
tute of Numerical Mathematics Coupled Model version
3.0 (INM-CM3.0) and National Center for Atmo-
spheric Research (NCAR) Parallel Climate Model
(PCM) to 4.3 K and larger in the L’Institut Pierre-
Simon Laplace Coupled Model version 4 (IPSL-CM4)
and Model for Interdisciplinary Research on Climate
3.2, high-resolution version [MIROC3.2(hires)]. A
number of modeling groups submitted daily output
from several ensemble members per scenario. These
models will be identified when the results of the ex-
treme value analysis are presented in the next sections.
Daily precipitation and daily temperature output for
all three scenarios was not available for all models
listed in Table 1. Daily model output from the A2 ex-
periment was not available for two models: the God-
dard Institute for Space Studies (GISS) Atmosphere–
Ocean Model (AOM) and the MIROC3.2(hires). Daily
temperature extremes were not available for the
NCAR Community Climate System Model version 3
(CCSM3). Daily temperature output from the NCAR-
PCM model was excluded from the analysis because
daily maximum and minimum temperature extremes
appear to be (erroneously) identical in 1981–2000. In
total, daily model output for years 1981–2000 was avail-
able from 14 models for temperature and from 16 mod-
els for precipitation.
To ensure consistency of the results for all three sce-
TABLE 1. List of IPCC global coupled climate models analyzed in the present study and their horizontal and vertical resolutions.
Model resolution is characterized by the size of a horizontal grid on which model output was available, and by the number of vertical
levels. Spectral models are also characterized by their spectral truncations. Equilibrium climate sensitivity is provided where available.
Model label and
climate sensitivity Resolution Institution and reference
CGCM3.1(T47) 3.6 K 96 ⫻ 48 L32 T47 Canadian Centre for Climate Modelling and Analysis
(http://www.cccma.ec.gc.ca/models/cgcm3.shtml)
CGCM3.1(T63) 3.4 K 128 ⫻ 64 L32 T63 Canadian Centre for Climate Modelling and Analysis
(http://www.cccma.ec.gc.ca/models/cgcm3.shtml)
CNRM-CM3 n/a 128 ⫻ 64 L45 T63 Centre National de Recherche Météorologique, France (Salas-Mélia et al. 2006,
manuscript submitted to Climate Dyn.)
ECHAM5/MPI-OM 3.4 K 192 ⫻ 96 L31 T63 Max-Planck-Institut für Meteorologie, Germany (Jungclaus et al. 2006)
ECHO-G 3.2 K 96 ⫻ 48 L19 T30 Meteorological Institute of the University of Bonn, Germany, Meteorological
Research Institute, South Korea (Min et al. 2005)
GFDL-CM2.0 2.9 K 144 ⫻ 90 L24 Geophysical Fluid Dynamics Laboratory (Delworth et al. 2006; Gnanadesikan
et al. 2006)
GFDL-CM2.1 3.4 K 144 ⫻ 90 L24 Geophysical Fluid Dynamics Laboratory (Delworth et al. 2006; Gnanadesikan
et al. 2006)
GISS-AOM n/a 90 ⫻ 60 L12 Goddard Institute for Space Studies Laboratory (Russell et al. 1995;
http://aom.giss.nasa.gov)
GISS-ER 2.7 K 72 ⫻ 46 L20 Goddard Institute for Space Studies Laboratory (Schmidt et al. 2006;
Russell et al. 2000)
INM-CM3.0 2.1 K 72 ⫻ 45 L21 Institute of Numerical Mathematics, Russia (Diansky and Volodin 2002)
IPSL-CM4.0 4.4 K 96 ⫻ 72 L19 Institut Pierre-Simon Laplace, France
(http://dods.ipsl.jussieu.fr/omamce/IPSLCM4/DocIPSLCM4)
MIROC3.2(hires) 4.3 K 320 ⫻ l60 L56 T106 Center for Climate System Research, Japan (Hasumi and Emori 2004)
MIROC3.2(medres) 4.0 K 128 ⫻ 64 L20 T42 Center for Climate System Research, Japan (Hasumi and Emori 2004)
MRI-CGCM2.3.2 3.2 K 128 ⫻ 64 L30 T42 Meteorological Research Institute, Japan (Yukimoto et al. 2001, 2006)
NCAR-CCSM3 2.7 K 256 ⫻ l28 L26 T85 National Center for Atmospheric Research (Collins et al. 2006)
NCAR-PCM 2.1 K 128 ⫻ 64 L26 T42 National Center for Atmospheric Research (Washington et al. 2000; Meehl
et al. 2006)
15 A
PRIL 2007 KHARIN ET AL. 1421
narios and to minimize possible effects of different mul-
timodel ensembles on the multimodel mean response,
the analysis of changes in climate extremes is per-
formed only for models for which daily model output
was available for all three emission scenarios. As a re-
sult, analysis of changes in precipitation extremes was
performed for 14 models [all models in Table 1 except
for GISS AOM and MIROC3.2(hires)]. Changes in
temperature extremes are analyzed for 12 models (ex-
cluding also NCAR-CCSM3 and NCAR-PCM). For
completeness, the analysis was repeated for all avail-
able models, but the conclusions of the study remained
essentially unaffected.
Several diagnostics describing simulated 1981–2000
climate extremes are compared to those derived from
four reanalyses. The two older reanalyses are the Na-
tional Centers for Environmental Prediction (NCEP)–
NCAR reanalysis (Kalnay et al. 1996) denoted hereaf-
ter as NCEP1, and the 15-yr European Centre for Me-
dium-Range Weather Forecasts (ECMWF) Re-
Analysis (ERA-15: Gibson et al. 1997). The two more
recent ones are the NCEP–Department of Energy
(DOE) AMIP-II reanalysis (Kanamitsu et al. 2002), de-
noted as NCEP2, and 40-yr ECMWF Re-Analysis
(ERA-40; Simmons and Gibson 2000). We also per-
formed an analysis of annual extremes of nonoverlap-
ping 5-day mean precipitation rates (pentads), and used
for verification the Climate Prediction Center (CPC)
Merged Analysis of Precipitation (CMAP) pentad
dataset that is a blend of gauge observations, satellite
observations, and precipitation fields from the NCEP–
NCAR reanalysis (Xie et al. 2003). These are essen-
tially the same validation sources that are used in the
recent atmospheric model intercomparison study by
Kharin et al. (2005) but updated for the 1981–2000 pe-
riod whenever possible.
3. Methodology
Climate extremes are multifaceted meteorological
phenomena and can be characterized in terms of inten-
sity, frequency, or duration of one or more climatologi-
cal parameters. To address the multitude of possible
extreme value statistics, the WCRP/WGCM also re-
quested that modeling groups submit a number of ex-
tremes indices, as described in Frich et al. (2002). These
indices are not analyzed here but are the subject of
several other diagnostic subprojects (http://www-
pcmdi.llnl.gov/ipcc/diagnostic_subprojects.php; e.g.,
Tebaldi et al. 2006).
Here we follow the approach of Zwiers and Kharin
(1998), Kharin and Zwiers (2000), and Kharin et al.
(2005) and analyze extremes of surface air temperature
and precipitation in terms of return values, or return
levels, of their annual extremes. Note that there seems
to be no universally agreed definition of return values.
A conventional but somewhat loose definition of a T-
year return level as the level that is exceeded on aver-
age every T years is problematic in a nonstationary en-
vironment. We more precisely define a T-year return
value as the threshold that is exceeded by an annual
extreme in any given year with the probability p ⫽ 1/T,
where T is expressed in years. In particular, a 20-yr
return value is the level that an annual extreme exceeds
with probability p ⫽ 5%. The quantity T ⫽ 1/p indicates
the “rarity” of an extreme event and is usually referred
to as the return period, or the waiting time for an ex-
treme event.
Return values defined as above are essentially the
quantiles of a distribution of annual extremes and are
estimated from a generalized extreme value (GEV) dis-
tribution fitted at every grid point to samples of annual
temperature and precipitation extremes. The “three
type” GEV distribution comprises the three classical
asymptotic extreme value models, Gumbel, Frèchet,
and Weibull (Jenkinson 1955). Its three parameters, lo-
cation, scale, and shape, are estimated by the robust
method of “L-moments” (Hosking 1990, 1992), also
known as the method of probability-weighted mo-
ments, with the minor modification of Dupuis and Tsao
(1998) to ensure the feasibility of the parameter esti-
mates (i.e., to ensure that all observed or simulated
annual extremes are in fact permitted by the estimated
GEV distribution). This method of return value esti-
mation is well documented in the aforementioned stud-
ies and is therefore not presented here.
We note that the GEV distribution theory is valid
only asymptotically, that is, when extremes are drawn
from increasingly larger samples. In the present study,
annual extremes are drawn from samples of size 365 (or
366 for leap years). However, serial correlation and the
presence of an annual cycle may substantially reduce
the effective sample size. Therefore, it is imperative to
evaluate whether the asymptotic GEV distribution pro-
vides a reasonable description of the behavior of a
sample of observed annual extremes by performing
goodness-of-fit tests. We routinely conduct standard
Kolmogorov–Smirnov goodness-of-fit tests (Stephens
1970) that measure the overall difference between the
empirical and fitted cumulative distributions for all
available samples. These tests indicate that a GEV dis-
tribution is generally a reasonable approximation for a
distribution of annual extremes of the considered vari-
ables in most models. The goodness-of-fit is diminished
for annual precipitation extremes in extremely dry re-
gions in some models, most notably in IPSL-CM4.0.
The GEV fit is also somewhat problematic for annual
1422 JOURNAL OF CLIMATE VOLUME 20
precipitation extremes in the Tropics in both GFDL
models. Tropical annual precipitation extremes in these
two models exhibit a somewhat intermittent behavior
when more moderate annual extremes in some years
are alternated with very large values in other years. As
an additional check, we routinely estimate empirical
quantiles of annual extremes for moderate return peri-
ods and compare them to the corresponding L-moment
return value estimates. In most cases regionally aver-
aged empirical and parametric return value estimates
compare reasonably well and are not overly too differ-
ent even in situations where a GEV fit appears to be
problematic.
The choice of the L-moment method over the fre-
quently used method of maximum likelihood for esti-
mating the parameters of a GEV distribution is primar-
ily dictated by relatively short 20-yr samples as are
available for analysis. The standard maximum likeli-
hood estimator is less efficient than the L-moment es-
timator in short samples for typical values of the shape
parameter (Hosking et al. 1985). Coles and Dixon
(1999) argue that this is mainly due to unreliable esti-
mates of the shape parameter that translates to poor
performance for return values. There have been efforts
to improve the efficiency of the maximum likelihood
estimator. For example, Martins and Stedinger (2000)
propose a generalized maximum likelihood analysis by
specifying a geophysical prior distribution to restrict the
shape parameter to a physically plausible interval
within a Bayesian framework. Coles and Dixon (1999)
modify the likelihood function by introducing a penalty
term to restrict the shape parameter values to the range
for which the GEV distribution has finite mean. Both
approaches require user decisions about the specifica-
tion of the prior distribution or the weight and form of
the penalty term. The benefits of these, more general
and potentially more powerful but also somewhat more
complex techniques, do not override, in our opinion,
the simplicity of the L-moment method in the present
setting.
A potential drawback of the L-moment method in a
transient climate change setting is that it assumes the
stationarity of annual extremes. Kharin and Zwiers
(2005) demonstrated that the violation of this assump-
tion may introduce bias in return value estimates that is
comparable to sampling variance. Their finding was
based on three-member ensemble simulations with a
single CGCM, but its significance is diminished for the
present multimodel study. First, as will be demon-
strated further on, sampling errors in local return value
estimates for moderate return periods are generally
smaller than discrepancies between individual models
and therefore do not represent the main source of un-
certainty. Second, the bias is minor as compared to
sampling variance when return values are estimated
from short 20-yr samples from a single realization that
are available for the majority of models in the present
study. Third, the short sample size prohibits the use of
more complex statistical models with time-varying
GEV distribution parameters, as was done by Kharin
and Zwiers (2005). Such models can be fitted with the
maximum likelihood method but are less competitive
than models with constant parameters in short samples.
Any benefits that might be gained in reducing the bias
by employing a more complex statistical model are
likely to be offset by increased sampling variance.
Overall, the L-moment method appears to be an ap-
propriate and viable technique for the task in the
present setting.
Alternatives to the annual extremes approach in-
clude peak-over-threshold techniques based on a gen-
eralized Pareto distribution, and r-largest extremes
analysis with a GEV distribution (e.g., Palutikof et al.
1999; Zhang et al. 2004). Successful implementation of
these methods generally requires more decisions from
the user (e.g., declustering of extremes, specification of
a sufficiently large threshold, dealing with the annual
cycle, etc.). Thus applying these techniques in an auto-
mated manner in a multimodel ensemble setting across
a variety of very different climatological zones is a
rather difficult task. The main argument for using one
of these alternative techniques is that they may use the
available information more efficiently, which could po-
tentially result in more accurate return value estimates.
However, as will be demonstrated further on, sampling
errors are not the main source of uncertainty in the
multimodel/multiscenario setting. We therefore do not
consider the use of other methods in the present study.
Most of the analysis that follows is performed for the
return period of 20 yr, or equivalently, for the exceed-
ance probability by annual extremes of 5%. Longer
return periods, such as 50 yr (exceedance probability of
2%), or even 100 yr (exceedance probability of 1%),
are less advisable given the relatively short 20-yr
samples and considering the fact that only one climate
simulation was available for each emission scenario for
most models. Estimating return levels for very long re-
turn periods is prone to larger sampling errors and po-
tentially larger biases due to inexact knowledge of the
shape of the tails of a distribution of annual extremes.
The GEV distribution methodology also allows us to
examine changes in the exceedance probability of
events of a certain size. In particular, we examine pro-
jected changes in the exceedance probability p of late-
twentieth-century 20-yr return levels and express these
changes in terms of changes in waiting times T ⫽ 1/p.
15 APRIL 2007 KHARIN ET AL. 1423
