Journal of Hydrology (2007) 347, 197– 210

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/jhydrol

Assessment of future change in intensity–duration–

frequency (IDF) curves for Southern Quebec using the
Canadian Regional Climate Model (CRCM)
a,*
Alain Mailhot , Sophie Duchesne a, Daniel Caya b, Guillaume Talbot a

a
INRS-Eau, Terre et Environnement, 490 de la Couronne, Québec (Québec), Canada G1K 9A9
b
Consortium Ouranos, 550 Sherbrooke Ouest, 19e étage, Tour Ouest, Montréal (Québec), Canada H3A 1B9

Received 16 March 2007; received in revised form 6 July 2007; accepted 10 September 2007

KEYWORDS Summary Intensity–duration–frequency curves are used extensively in engineering to

Extreme rainfall; assess the return periods of rainfall events. The estimation and use of IDF curves rely on
Climate change; the hypothesis of rainfall series stationarity, namely that intensities and frequencies of
Regional climate extreme hydrological events remain unchanged over time. It is however expected that glo-
model; bal warming will modify the occurrence of extreme rainfall events. In order to assess how
Spatial correlation; extreme rainfall events will be modified in a future climate, an analysis of the Canadian
Return period Regional Climate Model (CRCM) simulations under control (1961–1990) and future (2041–
2070) climates was performed. May to October annual maximum (MOAM) rainfall for 2-, 6-,
12- and 24-h durations were extracted and analyzed using regional frequency analysis for
grid boxes covering the Southern Quebec region. Comparison with available rainfall records
shows that CRCM estimates are consistent with those based on observed data considering the
different spatial scales related to observed data (meteorological stations) and to simulated
ones (grid boxes). Comparison of regional estimates in control and future climate at the grid
box scale reveals that return periods of 2- and 6-h events will approximately halve in future
climate while they will decrease by a third for 12- and 24-h events. Regional IDF curves at the
grid box and the station scales are proposed. The analysis of spatial correlation of simulated
MOAM series in control and future climates for the region under study suggests that, for a
given duration, spatial correlations will decrease in a future climate suggesting that annual
extreme rainfall events may result from more convective (and thus more localized) weather
systems. Multi-model ensemble systems (different GCMs with different RCMs) as well as
multi-member ensembles (investigation of possible sensitivity to initial conditions) are
needed to investigate the impact of model structures on future change in extreme rainfalls.
ª 2007 Elsevier B.V. All rights reserved.

* Corresponding author. Tel.: +1 418 654 3821; fax: +1 418 654 2600.
E-mail addresses: [email protected] (A. Mailhot), [email protected] (D. Caya).

0022-1694/$ - see front matter ª 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2007.09.019
198 A. Mailhot et al.

Introduction More specifically for Canada, the analysis of precipitation

indices (e.g. number of days with precipitation, maximum
It is expected that global warming will lead to an increase in total precipitation for a five-day interval) related to daily
intense rainfall events (intensity and/or frequency), since a rainfalls using historical records from all over this country,
warmer atmosphere can contain more humidity (and there- has revealed an increase in the annual total precipitation
fore more energy) and therefore produce a more active during the second half of the past century, mainly due to
hydrological cycle (Trenberth, 1999; Trenberth et al., an increase in the number of days with precipitation, while
2003). Allen and Ingram (2002) have suggested, based on a no consistent pattern is observed for extreme wet events
globally integrated view, that the extreme precipitation is (Vincent and Mekis, 2005). Stone et al. (2000) also reported
constrained by moisture availability and would therefore in- seasonally increasing trends in total precipitation during the
crease faster than the mean precipitation, constrained by 20th century for southern areas of Canada. These authors
the energy balance. Trend analysis suggests that it appears also mentioned that heavy and intermediate events are
to be the case in several regions (e.g. South Africa, Siberia, responsible for the increases observed during the past
eastern Mediterranean; see Groisman et al., 2005). Such 50 years.
modifications in rainfall events can result in significant im- The estimation of future modifications in precipitation
pacts over many sectors (e.g. urban drainage, dam design, indices due to variations in greenhouse gas concentrations
definition of flood plains, etc.; Kunkel et al., 1999). In urban must rely on climate models. In previous studies, global cli-
hydrology, for example, the design of drainage infrastruc- mate model (GCM) simulations have been used to assess
tures (whose life expectancy is comparable to the time changes in extreme rainfall under enhanced greenhouse
scale associated to the induced climate change (CC)) has conditions (e.g. Zwiers and Kharin, 1998). In more recent
been traditionally based on statistical analyses of historical studies, regional climate models (RCMs) have been used
records, assuming that the intensity and frequency of past (e.g. Ekström et al., 2005; Fowler et al., 2005; Jones and
events are statistically representative of what could happen Reid, 2001; Semmler and Jacob, 2004). RCMs present the
in the near future (Mailhot et al., 2006). However, in a con- advantage, over GCMs, that finer spatial information can
text of CC, this hypothesis must be revisited and the design be derived from these physically based models. The finer
criteria for drainage infrastructures should be revised to spatial resolution of RCMs is expected to be a more trust-
take into consideration the expected changes in the inten- worthy representation of processes involved during heavy
sity and frequency of heavy rainfall events (He et al., precipitations and therefore more adequate for water man-
2006; Grum et al., 2006; Papa et al., 2004). agement applications which involve smaller spatial scales
It is generally difficult to detect a clear trend in extreme (e.g. urban drainage).
rainfall intensity and/or frequency since, by definition, Recently, Frei et al. (2006) reported the results of an in-
these events rarely occur, that the inter-annual variability ter-comparison of daily (and multi-day) precipitation ex-
of heavy rainfalls is high and that records for rainfall series tremes as simulated by six different European RCMs.
are usually rather short. Despite these constraints, many Evaluation of the performance of these models to simulate
studies have reported an increased frequency of heavy pre- extreme precipitation in present climate was assessed
cipitation events in the USA (Karl et al., 1995; Karl and (European Alps data were used) and led these authors to
Knight, 1998), Australia (Plummer et al., 1999), Japan the conclusion that RCMs were capable of representing
(Iwashima and Yamamoto, 1993), Germany (Hundecha mesoscale spatial patterns in precipitation extremes that
and Bárdossy, 2005), Denmark (Arnbjerg-Nielsen, 2006), are not currently resolved by GCMs. Frei et al. (2006) also
Switzerland (Schmidli and Frei, 2005), China (Zhai et al., showed that a general tendency is observed between the
2005), southern and western African countries (New et al., 1961–1990 and the 2071–2100 periods towards an increase
2006), United Kingdom (Fowler and Kilsby, 2003) and India in the five-year return period value of one-day precipitation
(Goswani et al., 2006). intensity over northern and eastern Europe and a decrease
Alexander et al. (2006), while analyzing daily precipita- over southern Europe. Using the MM5 simulation model to
tion and temperature data from many sources over the generate daily precipitation at a 27-km horizontal resolu-
globe, have concluded that most precipitation indices show tion during the period of 1971–2100, Boo et al. (2006)
a tendency towards wetter conditions throughout the last showed that Korea will experience very important increases
century even though not all of them are statistically signif- in heavy precipitation events. The change in the pattern of
icant. These authors also observed that, when average con- heavy precipitation for Korea reported by Boo et al. (2006)
ditions over the globe are considered, the percentage is consistent with previous results observed for Japan and
contribution from the most extreme precipitation events China (Easterling et al., 2000).
to the annual precipitation total has been increasing. A sim- Grum et al. (2006) used simulation results from the re-
ilar analysis, realized by Groisman et al. (2005), showed a gional climate model HIRHAM to assess possible changes in
widespread increase in the frequency of very heavy precip- amplitude and frequency of maximum 1-h events (corre-
itations (99–99.7 percentiles) during the past 50–100 years sponding to the maximum rainfall depth to fall during one
(see also Tebaldi et al., 2006; Groisman et al., 1999 who clock-hour time; as far as we know, this is the only study
performed trend analyses of heavy rainfall using multi- considering sub-daily extreme events). An approach was
national data sets). Finally, the last report from the Working proposed to estimate the expected changes at point mea-
group 1 (WG1) of the International Panel on Climate Change surement according to the simulated changes in heavy pre-
(Summary for Policy Makers, SPM WG1-IPCC 2007) reported cipitation at the grid box size (25 km · 25 km) over
that heavy precipitations have increased on most of the Denmark. According to the results of Grum et al. (2006), re-
planetary land surface during the 20th century. turn period for a 1-h maximum intensity in present climate
Assessment of future change in IDF curves for Southern Quebec using the CRCM 199

(1979–1996) will approximately halve in future climate the computational rigid lid was located near 29 km. The
(2071–2100). Important research programs in the European CRCM time step at this 45 km resolution is 15 min. In their
Community have generated numerous simulations over Eur- analysis of the present climate as simulated by CRCM-
ope using many RCMs for both current and future climate. 3.7.1, Plummer et al. (2006) showed that the main
Unfortunately, such structured collaborations are still in characteristics of observed precipitation are reproduced
their infancy in North-America and the number of available by the model over most North America. This is true for both
investigations of change in extreme precipitations is very mean amounts and coefficient of variation. Plummer et al.
limited. (2006) also showed that the climate change signal in the
Even if the RCMs grid box size is much smaller than the precipitation is robust between versions of the model. This
GCMs’ (e.g. for the Canadian Regional Climate Model finding improves our confidence in the climate change signal
(CRCM), the grid box size is 45 km · 45 km while, typically, in the simulations used in the present investigation.
it is in the order of 300 km · 300 km for GCMs), the compar- Two simulations were considered in the present investi-
ison with rainfall data from stations is still challenging espe- gation. This pair of simulations makes a climate projection
cially when short duration and very localized intense events over the AMNO domain with the CRCM being driven by the
are considered (Osborn and Hulme, 1997; Osborn, 1997). Canadian global coupled climate model (CGCM2). The first
Therefore, in order to use RCMs output to compute statisti- simulation covers a period representative of the actual cli-
cal indices related to extreme rainfall events in a future cli- mate (GHG concentrations for the 1961–1990 period) and
mate, it is crucial to investigate how these estimated the second one the future climate (GHG concentrations
indices compare with those based on available rainfall data. for the 2041–2070 period).
This will be achieved in this paper using the areal reduction The CRCM simulations were driven from CGCM2 (Flato
factor (ARF). et al., 2000) data, a model that uses spectral dynamics
The objectives of this paper are twofold. First, statistical and was run, for the simulations presented here, at a reso-
characteristics of extreme rainfall events for the period lution of T32 with 10 vertical levels between the surface and
spanning 1961–1990, computed from a simulation of the 10 hPa. The ocean model is a three-dimensional grid-point
Canadian Regional Climate Model (CRCM), are compared model based on the GFDL MOM1.1 code (Pacanowski
to the statistical characteristics computed from rain-gauge et al., 1993) and uses a horizontal resolution of
observations. The region analyzed covers the southern part 1.875 · 1.875 and 29 vertical levels. Sea surface tempera-
of the Province of Quebec (Canada). Maximal rainfall depth tures and sea-ice cover, as simulated by the CGCM2, were
(period from May to October) for 2-, 6-, 12- and 24-h events interpolated in space and time to serve as time-dependent
are considered. Secondly, similar analyses for the future cli- lower boundary conditions for the CRCM over oceans. As
mate (2041–2070) are used to estimate the expected the North American Great Lakes are not resolved by the
changes in extreme rainfall events. CGCM2 at its resolution, lake-surface temperatures and
ice cover were calculated from the Goyette et al. (2000)
lake model. Geophysical fields over land points, such as li-
Climate models (CRCM_3.7.1 and CGCM2) and quid and frozen soil water content, snow amount and
experimental set-up ground temperature, were initialized with monthly mean
values from CGCM2 climatology. The CGCM2 ocean sea-sur-
Model data used in this investigation were taken from simu- face temperature and sea-ice distribution were used as sur-
lations generated by the Ouranos’ Climate Simulation Team face boundary conditions over the CRCM ocean grid cells.
as part of the Canadian Regional Climate Projections Pro- The CRCM simulations were driven by data from a CGCM2
gram. The regional model used for this study is the Canadian transient greenhouse gas (GHG) concentration simulation
Regional Climate Model. CRCM_3 utilizes most of the sub- covering the period from 1850 to 2100 (Flato and Boer,
grid-scale physical parameterization package of the sec- 2001). In both the CRCM and the CGCM2, the greenhouse
ond-generation Canadian Atmospheric General Circulation gas and aerosol concentrations follow the SRES-A2 scenario
Model (GCM2: McFarlane et al., 1992; Boer et al., 1992). (Houghton et al., 2001). According to historical levels of
The reader is referred to McFarlane et al. (1992) for an greenhouse gases, the effective concentration of CO2 has
extensive description of the physical parameterizations, to increased from 408 to 476 ppmv (parts per million by vol-
Caya and Laprise (1999) for a complete description of the ume) during the 1971–1990 period. The SRES-A2 scenario in-
model, to Laprise et al. (1998, 2003) for modifications creases the effective CO2 concentration from 714 to
appearing in CRCM_3 and to Plummer et al. (2006) for the 950 ppmv for the years 2041–2070.
additional modifications in CRCM_3.7.1. used for this study.
The CRCM simulations were performed at a horizontal reso-
lution of 45 km (true at 60N). The domain covers Canada, Available rainfall data sets
the continental United States and most of Mexico, with
201 · 193 grid points, and is referred to as the Amérique The area under study covers the southern part of the Prov-
du Nord, or simply the AMNO domain. The domain and the ince of Quebec (Canada), as shown in Fig. 1. This area was
topography, as resolved by the CRCM, is very large in the selected because it covers the three most important urban
context of regional climate simulations, which typically areas of the province (Montreal, Québec City and Sher-
use domains in the order of 100 by 100 grid points. In the brooke) and also because the region is relatively well cov-
vertical, 29 unequally spaced Gal-Chen scaled-height levels ered by a reasonably dense network of rain-gauge stations.
were used (Gal-Chen and Sommerville, 1975); the lowest Over Southern Quebec, 25–35% of the annual total pre-
thermodynamic level is about 25 m above the surface, and cipitation falls as snow, mainly from December to April.
200 A. Mailhot et al.

Figure 1 Area under study. Dashed lines represent the contour of the 61 CRCM grid boxes; the 48 circles are rain gauge stations
used for the analyses of rainfall on 2-, 6-, 12- and 24-h durations; the triangle is a station used for the analysis of rainfall on the 24-h
duration only; the three squares are rain gauge stations used for the analysis of rainfall on the 2-, 6- and 12-h durations.

Due to the cold weather conditions during winter, most of duration and in 1995 for the 1- to 12-h durations. This is
the tipping bucket rain-gauge stations are inoperative from due to the fact that the more recent data are not yet avail-
early November each year to late April of the following year; able at the time this paper is being written.
the exact dates vary for each year and for each station. In or-
der to achieve a uniform analysis, maximal annual values Statistical analysis of extreme rainfall
over moving windows of 2, 6 and 12 h were extracted from
the May to October period for each year and for each station. The statistical analysis of the MOAM series for a given dura-
Therefore, all statistical analyses carried on these duration tion (observed and simulated) allows the estimation of the
windows are defined as May to October Annual Maxima MOAM rainfall intensity of a given return period at a given
(MOAM). However, the 24-h duration MOAM values were ob- site (observed) or for a given grid box (simulated). Although
tained from a different database comprising only 24-h annual many distributions were considered in this study, two statis-
maxima (AM) values (i.e. no daily or monthly information are tical distributions were selected for use: the Generalized
available to get the exact timing of the maximum event). Extreme Value (GEV) and the Generalized Logistic (GLO) dis-
To be retained in the analysis, MOAM values had to be ta- tributions (see Hosking and Wallis, 1997 for a complete
ken from a May to October period having no more than description of these distributions).
34 days of missing data (out of a possible total of 184 days). The proposed statistical approach supposes that the data
For the 24-h AM values, however, this verification was not are independent, homogeneous and stationary (this last
possible since only annual values were available. From all property is important as it is the one that will presumably
Environment Canada rain gauge stations located in the study be affected by climate change). Tests were then performed
area, only the stations having precipitation records for at on simulated and observed MOAM series. Wald-Wolfowitz,
least 15 valid years were selected. This resulted in a total Wilcoxon (Mann-Whitney) and Mann-Kendall tests were sys-
of 51 stations for the analyses on 2- to 12-h durations, which tematically used to test respectively data independency,
have record lengths varying from 15 to 38 valid years (aver- homogeneity and stationarity. Acceptance level was fixed
age 24 years), and of 49 stations for the analyses on the 24-h at the 5% confidence level.
duration, records for these stations varying from 16 to 57 Distribution parameter sets have been first estimated for
years (average 23 years). Fig. 1 shows the location of the each duration at every grid point (simulated) and station
selected stations. Most records end in 1990 for the 24-h (observed) for comparison purposes. The L-moment method
Assessment of future change in IDF curves for Southern Quebec using the CRCM 201

(Hosking et al., 1985) was preferred to the maximum likeli- the GEV distribution was preferred and selected as the
hood estimation because it induces lower bias for small- regional distribution for observed data in Southern Quebec.
sample data sizes (Hosking et al., 1985). Regionalization The corresponding regional parameter values were esti-
of rainfall extremes was then realized following the meth- mated using the average l1 , tR, tR3 values (note that a region-
odology proposed by Hosking and Wallis (1997). Many ally averaged value of l1 was considered).
authors have previously reported works using the regional
analysis (see e.g. Alila, 1999; Alila, 2000; Onibon et al.,
2004). Regionalization procedures present numerous advan- CRCM simulated data
tages over a single-site frequency analysis since it is a con-
venient way of pooling statistics from different sample CRCM results for the grid boxes covering the region (see
data. More reliable estimations of the long return-period Fig. 1 for the location of the grid boxes) under study were
rainfall events, specifically when only single-site short re- analyzed. The MOAM series for each grid box for 2-, 6-,
cords are available, can be achieved using Regional Fre- 12- and 24-h precipitations were extracted from the simula-
quency Analysis (RFA). In the present study, the initial tion results for control and future climates using ‘‘sliding
region under consideration is Southern Quebec and simula- windows’’ of corresponding duration. It is important to note
tion results as well as station data were pooled to seek for that, since the simulation time step of the CRCM is 15 min,
a regional statistical description of rainfall extremes. The maximum values extracted from the simulation results may
regionalization technique supposes that frequency distribu- not exactly correspond to maximum values that would be
tions for n sites are identical apart from a site-specific scal- obtained if precipitation data were simulated using a time
ing factor which can be related to the average value or the step much smaller than the duration considered (this is
median value at that site. The approach used here encom- especially true for 2-h duration). No correction factor was
passes three steps: (1) the statistical testing of the regional applied to adjust this distortion, meaning that the simulated
homogeneity, (2) the selection of the regional distribution 2-h values for example potentially underestimate the
and, (3) the estimation of regional quantiles. A brief ‘‘true’’ values.
description of these steps is given in Appendix A. The reader A preliminary analysis, similar to the one achieved for
is referred to Hosking and Wallis (1997) for further details the observed data, was first realized to test data homogene-
regarding RFA. ity, independency and stationarity. These results globally
validate the hypothesis of independent and identically dis-
tributed (IDD) series (in general, no more than 4 grid boxes
Station records series do not satisfy these tests for each duration), the only
exception being 12-h MOAM series in control climate where
The data for 51 (or 49 for 24-h accumulations) meteorolog- 9 grid boxes series are diagnosed as non-homogeneous and 8
ical stations were compiled and analyzed (see Fig. 1). The as non-stationary. These series were considered in the fol-
methodology was applied to the MOAM for 2-h, 6-h, 12-h lowing analysis despite these litigious results. It is important
and 24-h accumulated precipitations. A preliminary analysis to note that, since it is expected that climate change will
of these data reveals that a only few MOAM series (between induce trends in MOAM series and lead to non-stationary ser-
0 and 4 for the various durations) do not satisfy indepen- ies, stationarity test confirms that this trend, if present,
dency, homogeneity or stationary criteria at the 5% confi- cannot be detected using 30-year MOAM series for most grid
dence level. boxes and durations even in future climate where the trend
Once parameter estimates have been obtained using the should presumably be more pronounced.
L-moment method, the v2 test was used to validate the To test the adequacy of the GEV or the GLO distributions
selection of the GEV distribution for each MOAM series. as a parametric model for the grid box MOAM values, the
According to this test, the hypothesis that the GEV distribu- parametric bootstrap technique was used. Results show that
tion fitted the data was accepted at the 5% or the 1% level individual grid box MOAM series can be fitted either by GEV
for all stations and durations with more than 20 years of or GLO distributions. This is a consequence of the short sim-
data. The GEV distribution was thus selected and fitted to ulation period available in control and future climates. The
all MOAM data sets using the L-moments. GEV distribution was however preferred. A comparison of
Regional analysis was then performed. The hypothesis MOAM estimates obtained using GEV or GLO distributions
that the region under study can be considered ‘‘accept- confirms that GEV and GLO estimates are very close for re-
ably’’ homogeneous was verified for 12-, 6- and 2-h rainfall turn periods shorter than 50 years.
durations (H values of 0.77, 0.99 and 0.19, respectively). A regional frequency analysis of CRCM simulation results
For 24-h rainfall durations, a regional analysis was done de- was also performed. Results of the homogeneity test for
spite the ‘‘possible’’ heterogeneity (H = 1.34) according to grid box MOAM values confirm that simulated values over
the Hosking and Wallis criteria. Southern Quebec for control and future climates can be re-
The selection of the parent distribution was achieved garded as ‘‘acceptably’’ homogeneous for all durations. It is
using the standard approach of Hosking and Wallis (1997). worth mentioning that almost all H values are negative (and
Among the usual distributions that can be considered, GEV some are smaller than 2; e.g. H = 2.05 for 24-h event in
and GLO distributions were retained as candidate parent future climate and H = 2.35 for 2-h event in future cli-
distributions. Results demonstrate that, except for the mate) suggesting a positive correlation between the data
24-h duration (for which only the GEV distribution was values at different grid boxes (Hosking and Wallis, 1997).
suitable), both distributions could reasonably fit the data. In fact, the most negative H values generally occur for long-
However, since it is the most frequently used distribution, er durations and increase as durations decrease. The only
202 A. Mailhot et al.

exception is the 2-h duration in future climate which shows and DeGaetano, 2002, 2005) is used to assess the perfor-
the smallest value. This result is not surprising considering mance of the CRCM in the simulation of extreme rainfall
that (1) the grid boxes completely cover the studied region, over Southern Quebec.
(2) the correlation between adjacent MOAM grid box series Fig. 2 shows scatter plots comparing MOAM estimates for
certainly exists and extends over many grid boxes and finally different return periods obtained from CRCM simulations (x-
(3) the inter-grid box variability of MOAM values is lower axis) and from station data (y-axis). Many similar y-values
than the inter-station variability. All this put together leads are expected for a given x-value since a grid box can include
to a regional homogeneity larger (smaller H) than would be many stations. Perfect agreement would result in points lo-
expected if grid box frequency distributions were indepen- cated on the line of slope one. Twenty four hour simulated
dent. Regional homogeneity was assumed despite this MOAM estimates slightly underestimate the observed MOAM
limitation. estimates while this underestimation is more pronounced
Here we should note that regional frequency analysis for 6-h events (this underestimation increases as duration
makes a few assumptions, among which is that the obser- decreases). This is consistent with the fact that spatial res-
vations at the different sites are independent. However, as olutions of observed and simulated data are very different.
stated by Hosking and Wallis (1997), meteorological events It was suggested that the grid box rainfall of GCM and RCM
(like intense rainfall) are almost always positively corre- has the spatial characteristics of areal averages (Osborn
lated for areas on which regional analysis is commonly per- and Hulme, 1997) and, accordingly, CRCM would operate
formed. Hosking and Wallis (1997) showed that their in a world drawn using 45 km · 45 km pixels (Fowler et al.,
procedure for regional analysis (applied here) yield suit- 2005).
ably robust and accurate quantile estimates for homoge- The concept of areal reduction factor (ARF) can be used
neous region even when correlations exist between sites. to relate the maximum areal average rainfall rate to the
Although we did not explicitly take into account the corre-
lation between MOAM grid box series for the calculation of
rainfall mean estimates for different durations and return
periods, this correlation was taken into consideration for
the calculation of confidence intervals on these estimates
(Section ‘‘Regional frequency analysis at the grid box
scale’’).
Results for the selection of the parent distribution reveal
that the GLO distribution gives acceptable close fit to the
regional average L-moments for control climate while it is
the GEV distribution that best fits regional average L-mo-
ments in future climate except for the 2-h event (for which
the GLO distribution is more suitable). This result suggests
that, at the grid box scale, the distribution that best suits
the data may change from control to future climate. This
means that non-stationarity induced by climate change
can translate not only in a modification of parameter values
but also in a changed form of the distribution itself. Regio-
nal parameter values were finally estimated using l1 , tR, tR3
for each climate and duration depending on the selected
parent distribution.

Comparison of observed and simulated MOAM

estimates

Validation of climate models is a difficult task. Two main

reasons explain this situation. One is the lack of observa-
tional data and the other is the low spatial resolution of cli-
mate models (Emori et al., 2005). In this section, the ability
of the CRCM to reproduce the statistical characteristics of
annual extreme rainfall is investigated.
Two different approaches have been applied to compare
heavy rainfall statistics based on observed and simulated
data. The first approach compares grid box values with val-
ues estimated from stations located within the grid box
(grid box analysis (GBA); Ekström et al., 2005). The second
is based on the regionalization frequency analysis (RFA)
where the comparison is achieved using regional estimates Figure 2 Comparison of 24-h (a) and 6-h (b) MOAM estimates
obtained from observed and simulated data. For both ap- obtained from simulated (grid boxes) and observed (stations)
proaches, the concept of areal reduction factor (ARF) (Allen values for the various return periods considered.
Assessment of future change in IDF curves for Southern Quebec using the CRCM 203

maximum rate estimated at a point. It has been shown that

ARFs vary with duration and size of the averaging area (Sri-
kanthan, 1995). ARFs will noticeably decrease when, for a
fixed area, annual maxima of shorter durations (less than
24 h) are considered, since events involved in shorter dura-
tion extreme rainfalls are spatially smaller than those in-
volved in longer duration extreme rainfalls (Srikanthan,
1995). The use of ARF derived from past records to relate
areal and point estimates in a future climate has also been
examined by Osborn (1997). Since ARF values depend on the
proportions of convective and stratiform rainfall and on the
spatial scale of events, these can change in future climate if
the spatial scale of rainfall events changes (Osborn, 1997).
However, since these eventual modifications cannot be con-
firmed or quantified at that time, analyses of future rainfall
extremes are usually based on the assumption that, for a gi-
ven area, duration and return period, ARF will not change in
Figure 3 Areal reduction factor (ARF) between regional
a future climate (Ekström et al., 2005; Fowler et al., 2005).
average rainfall depth values at the grid box and station scale
The observed differences between simulated and ob-
in control climate as a function of return period for different
served estimates can be related to the average sizes of
durations.
meteorological systems involved. Systems responsible for
2-h annual extremes are probably of convective nature
(summer thunderstorms) and therefore spatially localized
thunderstorms have spatial scales in the order of 10 km with reported ARF values. For example, using the values of
(Müller and von Storch, 2004) when compared to the size Allen and DeGaetano (2005) for 24-h extreme rainfalls on a
of a CRCM grid box. Water depths generated by those in- 2025-km2 basin in New Jersey and North Carolina, we ob-
tense but spatially limited systems are therefore much low- tained ARF values ranging respectively from 0.89 to 0.84
er once distributed over the whole grid box (as when for 2-year return period and from 0.89 to 0.88 for 50-year
parameterized in a regional climate model). On the other return periods. Although these values are comparable with
hand, meteorological systems responsible for 24-h annual those reported in Fig. 3, it is important to note that if we
extremes are mostly larger in scale (systems that last for observe increasing ARF values as a function of a return per-
longer periods and leads to bigger accumulations) and iod, Allen and DeGaetano (2005) reported lower ARF values
therefore can possibly completely cover one or more CRCM for higher return periods for 24-h events (this tendency is
grid boxes. The ‘smoothing’ effect will thus increase as the however very weak for North Carolina). One possible expla-
duration of annual extreme events decreases and this is in nation for this difference is related to the nature of meteo-
agreement with the model underestimation increases as rological systems involved in these different regions during
time window decreases. The fact that 24-h extreme precip- those extreme events. In the Southern Quebec region, one
itation estimates are close to observed estimates suggests can think that meteorological systems responsible for 24-h
that the average system size is larger than the grid box size MOAM rainfalls have a more important contribution from
in that case (small smoothing effect). This is consistent, stratiform precipitation for smaller return periods, and
although the local climatology of the two regions differs, therefore, their spatial extension increases as the return
with a result obtained by Osborn and Hulme (1997) that period gets longer.
showed that the characteristic scale of 24-h annual maxima
in Western Europe is in the order of a few hundred
kilometres. Comparison of MOAM estimates in control and
In order to further investigate the relationship between future climates
sites and grid box values, we have defined the average
ARF values as follows: Grid box analysis
ðgÞ
x p ðT; dÞ For the grid box analysis, the MOAM magnitude for 2-, 6-,
ARFðT; dÞ ¼ ðsÞ
ð1Þ
xp ðT; dÞ 12-, and 24-h and 2-, 5-, 10-, 25- and 50-year return periods
were computed and compared on a box-to-box basis be-
ðsÞ ðgÞ
where x p ðT; dÞ and x p ðT; dÞ are the average regional rain- tween control and future climates. Fig. 4a and b presents
fall depths associated to events of duration d and return respectively these results for 24-h and 6-h events. As can
periods T, respectively at the station scale (superscript s) be observed, estimates in future climate are predominantly
and the grid box scale (superscript g) in control climate higher than their corresponding values in control climate. It
(subscript p). ARF(T, d) is thus the average areal rainfall fac- is therefore expected that the region of interest will
tor between sites and grid box values obtained using region- experience increasing MOAM intensity in future climate for
alized values. Fig. 3 shows how ARF(T, d) values evolve as a all durations at the grid box scale (similar results are
function of return periods T for different durations d. To our obtained for 12- and 2-h events). The amplitude of this shift
knowledge, no ARF value has previously been estimated for between control and future periods will be estimated
the region under study. However, these values are coherent though a regional frequency analysis.
204 A. Mailhot et al.

performed and inter-site dependence was taken into ac-

count. The selected structure of inter-site dependence
assumes that if each site frequency distribution was trans-
formed to a normal distribution, then the joint distribution
for these N sites would be multivariate normal. The mean
correlation coefficient between grid boxes was used mean-
ing that the sites are assumed equicorrelated within the
simulated region. The observed sample L-moment ratios
were used as the population L-moment ratios of the simu-
lated region, even if Hosking and Wallis warn that this could
lead to simulated regions with higher heterogeneity than
observed. Some preliminary tests showed that this does
not happen in our case study.
Fig. 5 summarizes these results. Ninety percent confi-
dence intervals of the ratio between estimates in control
and future climates are presented (vertical lines) for the
different return periods and durations. Two main conclu-
sions can be drawn from this graph. The first one is that dif-
ferences between control and future estimates are
significant (at the 90% confidence level) for 2-h event up
to 25-year return period and for 6-h event up to 10-year re-
turn period while for 12- and 24-h event significant differ-
ences are observed only for the 2- and 5-year return
periods. The second conclusion is that, although 90% confi-
dence intervals often cross the unit value, all ratio mean
values are lower than one. Even if confidence intervals
can be large essentially because we have short series
(30 years) and since the spatial correlations between grid
boxes are rather large (see next section), the fact that fu-
ture values are systematically higher than their correspond-
ing values in control climate is certainly an indication
towards more intense heavy rainfalls in a future climate
over Southern Quebec.

Grid box correlation between simulated MOAM

series
Figure 4 Comparison of 24-h (a) and 6-h (b) MOAM estimates
obtained from CRCM simulations in control (x-axis) and future
The spatial correlation structure between the simulated
(y-axis) climates for the various return periods considered.
MOAM series at different grid boxes was investigated. It is

Regional frequency analysis at the grid box scale

Results of regional frequency analysis in future and control

climate have been compared. As mentioned previously, the
main motivation supporting this approach is that regional
estimates (integrating many grid boxes) of CRCM simulations
are probably more meaningful than those based on single
grid box values. Even though the comparison achieved in
the previous section (box-to-box comparison) is instructive,
we believe that reliable estimates of MOAM values in control
and future climates must be computed on a regional basis.
In order to assess if differences between control and future
climates are statistically significant, 90% bootstrap confi-
dence intervals of the ratio of MOAM regional estimates in
control and future climates (control/future) have been esti-
mated. These were estimated using the method proposed by
Hosking and Wallis (1997) which assumes that the sample
L-moments are normally distributed. According to this
method, rainfall series for the region under study are sto- Figure 5 Ratio of regionally averaged MOAM estimates in
chastically generated using regional values. Each realization control and future climates (control/future) at the grid box
considers the same number of sites (grid boxes) and record scale for the various durations and return periods. Vertical bars
lengths at each site. In our case, 10,000 realizations were are 90% bootstrap confidence intervals.
Assessment of future change in IDF curves for Southern Quebec using the CRCM 205

expected that these correlations could give information on

the spatial scale of events generating annual maximum rain-
fall depths. In fact, slow decaying spatial correlations be-
tween contiguous grid boxes may be an indication that
‘‘large’’ (covering many grid boxes) systems are involved
and that the same system is responsible for the annual max-
imum in adjacent grid boxes. However, it must be remem-
bered that a lower spatial cut-off exists (the grid box)
under which no information is available. The following cor-
relation between grid box MOAM series i and j was estimated
as
Pnt

k¼1 ðx ik  x i Þðx jk  x j Þ

rij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pnt P ffi ð2Þ
k¼1 ðx ik  x i Þ
 2 nk¼1 t
ðxjk  xj Þ2

where xik is the MOAM value at site i at year k, nt the number

of years for which both sites have data and
1 X nt
xi ¼ x ik ð3Þ
nt k¼1

is the time averaged MOAM at site i. The average correlation

between first, second, third, etc. neighbouring grid boxes
was then computed as
1 X
qm ¼ rhiji ð4Þ
nm
The sum is over all pairs of grid boxes separated by m grid
boxes (along the two axes of the spatial domain). Fig. 6a
and b presents the resulting correlations in control and fu-
ture climates, respectively. For control climate, spatial cor-
relation between neighbouring grid boxes is decaying more
rapidly as duration decreases, as expected under the
hypothesis that systems of smaller spatial extensions are in-
volved for extreme precipitation events of short duration. A
very different behaviour is observed in future climate where
spatial correlation for the various durations has approxi-
mately the same structure. This result is interesting as it Figure 6 Average correlation between grid box MOAM series
suggests that systems generating heavy precipitations for gi- as a function of distance (1 = neighbouring grid boxes; 2 = sec-
ven durations will be more localized in a future climate and ond neighbour grid boxes, etc.) for various durations in
presumably more convective in nature even for the 24-h (a) control and (b) future climates.
events. Although it is not possible to make a clear assess-
ment about the evolution of the spatial extension of systems
generating annual maximum rainfalls for sub-daily events,
and 2-h events. In these cases, return periods can be
this result suggests that they could be more localized (with-
approximately halved for events with return periods ranging
in grid boxes) and that the increases accounted for by the
between 20 and 50 years in present climate. However, it
CRCM at the grid box scale may be much more important
must be remembered that the confidence interval on MOAM
when transposed to the site scale.
estimates increases as a return period increases so that
changes for larger return periods are more uncertain.
IDF curves in control and future climates at the grid
box scale Estimation of regional MOAM rainfalls at the station
scale
Regional average estimates at the grid box scale have been
used to generate intensity–duration–frequency (IDF) IDF curves are usually developed from site-specific records.
curves in control and future climates (Fig. 7). A more conve- Therefore, some sets of hypotheses must be put forward in
nient way to visualize the expected changes is to express order to transpose the estimated regional average MOAM
these in terms of return periods. Fig. 8 presents the corre- rainfall intensities at the station scale from the estimated
sponding return period in control (x-axis) and future (y-axis) values at the grid box scale. Two approaches were consid-
climates for MOAM of a given duration and intensity. As pre- ered in the following. According to the first one, named
viously observed, smaller modifications of return periods the intensity-dependent ARF assumption, it is assumed that
are expected for 24-h and 12-h events, while larger and the ARF, for a given duration and a given rainfall intensity,
more significant changes (see Fig. 5) are expected for 6-h will remain unchanged in a future climate. In other words,
206 A. Mailhot et al.

Only the probability of occurrence of these events at the grid

box scale will change according to this hypothesis. However,
it is expected that the convective component of extreme
precipitation events will be more important in future climate
meaning that, for a given rainfall intensity at the grid box
scale, the meteorological systems could be more localized
(or have a more pronounced convective nature) in future cli-
mate and that accordingly, the ARF could be smaller. The ob-
served reduction of spatial correlation from control to future
climate (as seen in Fig. 6) may be an indication of this more
localized character of systems (see Section ‘‘Grid box corre-
lation between simulated MOAM series’’). If so, this suggests
that the following estimates of rainfall intensities at the sta-
tion scale in future climate would correspond to a lower
bound of the expected increase in rainfall values. More
important increases can be expected justifying the develop-
ment of higher RCMs resolution.
Figure 7 IDF curves at the grid box scale. Continuous curves Expressing the first hypothesis (intensity-dependent ARF)
refer to control climate while dotted curves refer to future in mathematical form, following the previously defined sets
climate. Circles, triangles, squares and diamonds are respec- of hypotheses, the ARF value applicable to the future MOAM
tively for 2-, 6-, 12-, and 24-h events. ðgÞ
estimates x f is given by the corresponding ARF for events
of same intensity in control climate and we have
ðgÞ ðgÞ
ARFf ðx f Þ ¼ ARFp ðx ðgÞ
p ¼ xf Þ ð5Þ
and therefore from Eq. (1):
ðgÞ ðgÞ
ðsÞ xf xf
xf ¼ ðgÞ
¼ ðgÞ
ð6Þ
ARFf ðx f Þ ARFp ðx f Þ
ðgÞ
Given xf (E value in Fig. 9), the corresponding probability
of non-exceedance in control climate at the grid box scale
ðgÞ
is F ðgÞ
p ðx f Þ (G value) and the corresponding rainfall depth
in control climate at the station scale is given by (F value)
ðsÞ ðgÞ
(brackets indicate that x p is a function of F ðgÞ
p ðx f ÞÞ:
h i
ðgÞ ðgÞ
xðsÞ
p F p ðx f Þ ð7Þ

And thus
ðgÞ
ðgÞ xf
ARFf ðx f Þ ¼ ðsÞ ðgÞ
ð8Þ
x p ½F ðgÞ
p ðx f Þ
Figure 8 Return periods in control climate and corresponding
return periods in future climate at the grid box scale.

the ARF in control climate corresponding to a given rainfall

depth is used in future climate to transpose grid box scale
value to the station scale. The second approach, named
the constant ARF assumption, assumes that the average rel-
ative change in extreme events between control and future
climates at the site scale is equal to the corresponding rel-
ative changes at the grid box scale for the same duration
and quantile. This is equivalent to saying that the expected
relative increases at the grid box scale will uniformly affect
the sites within the grid box.

Intensity-dependent ARF approach

It is important to realize that, according to the intensity-
dependent ARF hypothesis, the spatial extent of meteoro-
logical systems generating an intense rainfall of a given
intensity and duration at the grid box scale in future climate Figure 9 Construction of the IDF curves at the station scale in
will remain approximately unchanged when compared to sys- future climate from grid box curves and stations cumulative
tems with same intensity and duration in control climate. probability curves in control climate.
Assessment of future change in IDF curves for Southern Quebec using the CRCM 207

and finally and future climate (the delta) is applied to actual observa-
ðsÞ ðgÞ ðgÞ tions to get future values at the station. A comparison of
xf ¼ x ðsÞ
p ½F p ðx f Þ ð9Þ these two approaches indicates similar results, the average
The intensity-dependent ARF approach was applied to the absolute relative difference being estimated at 1.8% over all
grid box regional distributions for the various durations un- durations and return periods, a value much smaller than the
der study. Regional average IDF curves at the station scale uncertainties associated to these estimates.
were estimated and are presented in Fig. 10. It is important
to note that, according to the hypothesis used to build the
IDF curves at the station scale, the resulting relations be-
Summary and conclusions
tween return periods in control and future climates remain
CRCM-simulated Annual May to October maximum rainfall
identical to the ones obtained at the grid box scale and cor-
depth (MOAM) series (2-, 6-, 12- and 24-h durations) over
respond accordingly to those presented in Fig. 8.
Southern Quebec have been analyzed and compared to
available historical records. The CRCM simulations were dri-
Constant ARF approach
ven by the Canadian Coupled Global Climate Model (CGCM2)
Another approach (the constant ARF approach) was used to
following the SRES-A2 greenhouse gas emission scenario.
generate IDF curves at the station scale. In that case, in-
Two periods were simulated, 1961–1990 and 2041–2070,
creases in rainfall depths for a given duration and return
representative of control and future climates. A total of
period at the grid box scale were used to estimate the rain-
61 CRCM grid boxes (45 km · 45 km) and 51 rain gage sta-
fall depth at the station scale in future climate, namely
tions were considered in the statistical analysis achieved
ðsÞ
x f ¼ x ðsÞ
ðgÞ at individual grid boxes and stations. A regional frequency
p ½1 þ Dpf ðd; TÞ ð10Þ
analysis (both for stations and grid boxes) was performed
where mainly for two reasons: (1) it leads to more accurate statis-
tical estimates than a local analysis and (2) the CRCM results
ðgÞ ðgÞ
ðgÞ xf  xp should be interpreted at a regional scale (integrating many
Dpf ðd; TÞ ¼ ðgÞ
ð11Þ
xp grid boxes) since grid boxes represent the simulation ‘‘fun-
damental’’ spatial unit (‘‘pixels’’) (any interpretation of the
Once simplified, we see that this formulation is equivalent results at individual grid boxes is not recommended even
to though still common in practice).
ðsÞ
ðgÞ
xf The comparison of MOAM estimates based on CRCM sim-
x f ¼ x ðsÞ
p ðgÞ
ð12Þ ulations with the corresponding values estimated from
xp
available data must take into consideration the fact that
and means that, for a given duration and return period, we the spatial scale related to these estimates are different
suppose even when a regional climate model is used (areal vs. point
estimates). The areal reduction factor (ARF) was used to as-
ARFf ¼ ARFp ð13Þ sess the ability of CRCM to reproduce the observed statisti-
cal characteristics of MOAM. Regional ARF values (ratio of
This last approach is similar to the well-known delta ap-
regional grid box and stations MOAM estimates) were there-
proach where the model simulated change between present
fore computed. The resulting ARF values seem reasonable
and are consistent with what has been reported in the liter-
ature for other regions of the globe (to our knowledge, no
ARF values have been estimated for the area under study).
ARF dependence on return periods and durations can be re-
lated to the characteristic spatial scale of the systems gen-
erating extreme rainfalls. We believe that, although this
result is not a validation sensu stricto of the ability of the
model to simulate extreme precipitations, it is a good indi-
cation that CRCM simulations lead to results that are statis-
tically consistent with observations in terms of extreme
rainfall estimates.
In summary, the major findings of this study are:

• No clear indication of non-stationarity in MOAM series

(for all durations) is observed either at the stations (his-
torical records) or for the 30-year simulated series in
both control and future climate. Results for Southern
Quebec in control climate are therefore consistent since
Figure 10 IDF curves at the station scale using intensity- CRCM simulations suggest that, for a 30-year period in
dependent ARF. Continuous curves refer to control climate control climate, no trend could be detected even if such
while dotted curves refer to future climate. Circles, triangles, a trend is building up in the data, due to the change in
squares and diamonds correspond respectively to 2-, 6-, 12-, model GHG concentrations. Since the trend analysis
and 24-h events. (Mann-Kendall) was realized at individual stations, it is
208 A. Mailhot et al.

not impossible, as mentioned by Goswani et al. (2006) were not taken into consideration in this study since only
and Groisman et al. (1999), that a ‘‘regional trend anal- one RCM and one RCM experiment were analyzed. Clearly,
ysis’’ would reveal a trend otherwise difficult to detect multi-model ensemble systems need to be analyzed in order
using individual stations (or grid boxes). to quantify uncertainties. These multi-model ensemble sys-
• Increases in regional MOAM values at the grid box scale tems must include the use of the output of different GCMs,
are detected for Southern Quebec between CRCM simu- as well as different RCMs to estimate how model structure
lated control and future periods. Confidence intervals can modify extremes in future climate. Multi-member
on regional estimates show that uncertainties about ensembles should also be considered to investigate possible
these changes increase as return periods or durations sensitivity to initial conditions. Results obtained in this
increase. Thus, at the 90% confidence interval, increasing study remains model dependent since the use the output
MOAM rainfall depths are expected for 2-h MOAM of of different global climate models (GCM) might bring very
return periods between 2 and 25 years while for 24-h different results.
MOAM rainfall depths, statistically significant increases
are reported for 2- and 5-year return periods. Improve-
ment of these estimates would necessitate the analysis Acknowledgements
of larger samples. However, it must be noted that simu-
lated regional average MOAM rainfall depths are system- This work was funded by the Climate Change and Adaptation
atically larger in future climate. Program of Natural Resources Canada and by the Consor-
• The analysis of the spatial correlation of simulated MOAM tium Ouranos on Regional Climatology and Adaptation to
series in control and future climates for the region under Climate Change. The authors would like to thank the two
study shows that: (1) for a given duration, these correla- anonymous reviewers for their comments and suggestions.
tions will decrease in a future climate, and (2) the spatial
correlation in future climate is very similar for all dura-
tions. This correlation is thought to be related to the spa- Appendix A. Regional frequency analysis of
tial dimension of the meteorological systems generating rainfall extremes
the MOAM rainfall depth. A smaller correlation in a future
climate for a given duration means annual extreme rain- The regionalization technique supposes that frequency dis-
fall events resulting from more convective (and thus tributions for n sites are identical apart from a site-specific
more localized) weather systems. scaling factor which can be related to the average value or
• Return periods of MOAM rainfall depths for 2- and 6-h the median value at that site. The approach used here
events are approximately halved in future climate when encompasses three steps: (1) the statistical testing of the
compared to corresponding values in control climate. regional homogeneity, (2) the selection of the regional dis-
Reduction of return periods in future climate for 12- tribution and, (3) the estimation of regional quantiles.
and 24-h MOAM rainfall depths decreases as return peri-
ods increase. Statistical testing of the regional homogeneity

Estimation of future MOAM rainfall depths at the station In order to test for the regional homogeneity, weighted
scale for the area under study was achieved based on the average L-moments are used to estimate representative re-
hypothesis (intensity-dependent ARF) that the ratios of re- gional parameters. Average L-moments ratios ðtR ; tR3 ; tR4 Þ are
turn periods between control and future climates are the computed as follows:
same at the station scale and at the grid box scale. This
hypothesis on ARF values is equivalent to assuming that 1 Xn
tR ¼ nj tðjÞ
the spatial distribution of rainfalls inside grid boxes re- N j¼1
mains unchanged in future climate for a same total rainfall
1 Xn
ðjÞ
depth on the grid box. This may be an ‘‘optimistic’’ pro- tR3 ¼ nj t3 ðA:1Þ
N j¼1
jection since a wetter and more convective climate will
possibly lead to more localized thunderstorms and, conse- 1 Xn
ðjÞ
quently, lower ARF values. Therefore, once transposed to tR4 ¼ nj t4
N j¼1
the station (or site) scale, the increases of MOAM rainfalls
ðjÞ ðjÞ
at the grid box scale may result in even higher increases at where tðjÞ ; t3 ; t4 are the L-moments ratios at site j, nj is the
the site scale. This important issue could be further inves- length of the record at site j, n is the number of sites and N
tigated as higher-resolution RCM simulations become is the total number of available data:
available. Xn
Emori et al. (2005) showed that the simulation of ex- N¼ ni ðA:2Þ
treme daily precipitations can be highly dependent on their i¼1

parameterization, meaning that higher resolution models do Monte Carlo simulations have been used to generate sto-
not necessarily result in better simulations of climate than chastic samples at each site of the region under consider-
lower resolution models. This observation is probably also ation. These samples were drawn from the Kappa
valid for extreme sub-daily precipitations. Frei et al. distribution (Hosking, 1994) fitted to the weighted average
(2006) reach a similar conclusion and suggest that the for- L-moments 1; tR ; tR3 ; tR4 . A large number of samples Ns are
mulation of RCMs contributes significantly to uncertainties generated and the following values are estimated for each
in extreme precipitation assessment. Such uncertainties sample:
Assessment of future change in IDF curves for Southern Quebec using the CRCM 209
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi
ðiÞ R 2 Meteorology, May 13–14. American Meteorological Society,
i¼1 n i ðt  t Þ
V¼ Pn ðA:3Þ Oregon.
i¼1 ni Allen, R.J., DeGaetano, A.T., 2005. Areal reduction factors for two
eastern United States regions with high rain-gauge density. J.
This value corresponds to the weighted standard deviation
Hydraul. Eng. 10 (4), 327–335.
of t. Defining V(obs), the observed network values, the homo- Allen, M.R., Ingram, W.J., 2002. Constraints on future changes in
geneity measure is given by climate and the hydrologic cycle. Nature 419, 224–232.
ðV ðobsÞ  lV Þ Arnbjerg-Nielsen, K., 2006. Significant climate change of extreme
H¼ ðA:4Þ rainfall in Denmark. Water Sci. Technol. 54 (6–7), 1–8.
rV Boer, G.J., McFarlane, N.A., Lazare, M., 1992. Greenhouse
where lV and rV are the mean and standard deviation of the gas-induced climate change simulated with the CCC second-
Ns values synthetically generated. According to Hosking and generation general circulation model. J. Climate 5, 1045–
Wallis (1997), a region may be regarded as ‘‘acceptably’’ 1077.
Boo, K.-O., Kwon, W.-T., Baek, H.-J., 2006. Change of extreme
homogeneous if H < 1, ‘‘possibly’’ heterogeneous if
events of temperature and precipitation over Korea using
1 6 H < 2 and ‘‘definitively’’ heterogeneous if 2 6 H. regional projection of future climate change. Geophys. Res.
Lett. 33, L01701. doi:10.1029/2005GL023378.
Identification of the regional distribution Caya, D., Laprise, R., 1999. A semi-implicit semi-Lagrangian
regional climate model: the Canadian RCM. Mon. Weather Rev.
Once the homogeneity test confirms that weighed average 127 (3), 341–362.
values are representative of regional values, a goodness- Easterling, D.R., Evans, J.L., Groisman, P.Ya., Karl, T.R., Kunkel,
of-fit test must be defined to select the appropriate regional K.E., Ambenje, P., 2000. Observed variability and trends in
distribution. A candidate distribution is first selected (in our extreme climate events: a brief review. Bull. Am. Meteorol. Soc.
81 (3), 417–425.
case the GEV and GLO distributions were considered) and
Ekström, M., Fowler, H.J., Kilsby, C.G., Jones, P.D., 2005. New
the L-kurtosis of the fitted distributions is denoted by estimates of future changes in extreme rainfall across the UK
tDIST
4 . A large number of realizations for a region is then sim- using regional climate model integrations. 2. Future estimates
ulated using the Kappa distribution fitted to the regional and use in impact studies. J. Hydrol. 300, 234–251.
average L-moment ratios (the Kappa distribution includes Emori, S., Hasegawa, A., Suzuki, T., Dairaku, K., 2005. Validation,
the GEV and GLO distributions as special cases; a descrip- parametrization dependence, and future projection of daily pre-
tion of the Kappa distribution can be found in Hosking, cipitation simulated using a high-resolution atmospheric GCM.
1994). The following statistic is used (Hosking and Wallis, Geophys. Res. Lett. 32, L06708. doi:10.1029/2004GL022306.
1997) to identify the regional distribution: Flato, G.M., Boer, G.J., 2001. Warming asymmetry in climate
change simulations. Geophys. Res. Lett. 28, 195–198.
1 DIST Flato, G.M., Boer, G.J., Lee, W.G., McFarlane, N.A., Ramsden, D.,
ZDIST ¼ ðt  tR4 þ B4 Þ ðA:5Þ
r4 4 Reader, M.C., Weaver, A.J., 2000. The Canadian Centre for
climate modelling and analysis global coupled model and its
where tR4 is the weighted regional average L-kurtosis, and B4 climate. Climate Dyn. 16, 451–467.
and r4 are given by Fowler, H.J., Kilsby, C.G., 2003. A regional frequency analysis of
United Kingdom extreme rainfall from 1961 to 2000. Int. J.
1 X
Nsim
ðmÞ
B4 ¼ ðt  tR4 Þ Climatol. 23 (11), 1313–1334.
Nsim m¼1 4 Fowler, H.J., Ekström, M., Kilsby, C.G., Jones, P.D., 2005. New
( " #)1=2 ðA:6Þ estimates of future changes in extreme rainfall across the UK
1 X
Nsim
r4 ¼
ðmÞ R 2 2
ðt  t4 Þ  Nsim B4 using regional climate model integrations – 1. Assessment of
ðNsim  1Þ m¼1 4 control climate. J. Hydrol. 300, 212–233.
Frei, C., Schöll, R., Fukutome, S., Schidli, J., Vidale, P.L., 2006.
ðmÞ
t4 corresponds to the regional average L-kurtosis for the Future change of precipitation extremes in Europe: intercom-
mth simulated region. The goodness-of-fit is adequate if parison of scenarios from regional climate models. J. Geophys.
jZDISTj 6 1.64. Res. 111, D06105. doi:10.1029/2005JD005965.
Gal-Chen, T., Sommerville, R.C., 1975. On the use of a coordinate
transformation for the solution of Navier–Stokes. J. Comput.
References Phys. 17, 209–228.
Goswani, B.N., Venugopal, V., Sengupta, D., Madhusoodanan, M.S.,
Alexander, L.V., Zhang, X., Peterson, T.C., Ceasar, J., Gleason, B., Xavier, P.N., 2006. Increasing trend of extreme rain events over
Klein Tank, A.M.G., Haylock, M., Collins, D., Trewin, B., India in a warming environment. Science 314, 1442–1445.
Rahimzadeh, F., Tagipour, A., Rupa Kumar, K., Revadekar, J., Goyette, S., McFarlane, N.A., Flato, G.M., 2000. Application of the
Griffiths, G., Vincent, L., Stepenson, D.B., Burn, J., Aguilar, E., Canadian regional climate model to the Laurentian Great Lakes
Brunet, M., Taylor, M., New, M., Zhai, P., Rusticucci, M., region: implementation of a lake model. Atmos. Ocean 38, 481–
Vasquez-Aguirre, J.L., 2006. Global observed changes in daily 503.
climate extremes of temperature and precipitation. J. Geophys. Groisman, P.Y., Karl, T.R., Easterling, D.B., Knight, R.W., Jamason,
Res. 111, DO5109. doi:10.1029/2005JD006290. P.F., Hennessy, K.J., Suppiah, R., Page, C.M., Wibig, J.,
Alila, Y., 1999. A hierarchical approach for the regionalization of Fortuniak, K., Razuvaev, V.N., Dougals, A., Førland, E., Zhai,
precipitation annual maxima in Canada. J. Geophys. Res. 104 P.-M., 1999. Changes in the probability of heavy precipitation:
(D24), 31645–31655. important indicators of climatic change. Clim. Change 42, 243–
Alila, Y., 2000. Regional rainfall depth-duration-frequency equa- 283.
tions for Canada. Water Resour. Res. 36 (7), 1767–1778. Groisman, P.Y., Knight, R.W., Easterling, D.R., Karl, T.R., Hegerl,
Allen, R.J., DeGaetano, A.T., 2002. Re-evaluation of extreme G.C., Razuvaev, V.N., 2005. Trends in intense precipitation in
rainfall areal reduction factors. In: 13th Conference on Applied the climate record. J. Climate 18, 1326–1350.
210 A. Mailhot et al.

Grum, M., Jørgensen, A.T., Johansen, R.M., Linde, J.J., 2006. The langa, E., Mbambalala, E., Adesina, F., Saleh, H., Kanyanga, J.,
effect of climate change on urban drainage: an evaluation based Adosi, J., Bulane, L., Fortunata, L., Mdoka, M.L., Lajoie, R.,
on regional climate model simulations. Water Sci. Technol. 54 2006. Evidence of trends in daily climate extremes over southern
(6–7), 9–15. and West Africa. J. Geophys. Res. 111, D14102. doi:10.1029/
He, J., Valeo, C., Bouchart, F.J.-C., 2006. Enhancing urban 2005JD006289.
infrastructure investment planning practices for a changing Onibon, H., Ouarda, T.B.M.J., Barbet, M., St-Hilaire, A., Bobée, B.,
climate. Water Sci. Technol. 53 (10), 13–20. Bruneau, P., 2004. Anlayse fréquentielle régionale des précip-
Hosking, J.R.M., 1994. The four-parameter kappa distribution. IBM itations journalières maximales annuelles au Québec, Canada. J.
J. Res. Develop. 38 (3), 251–258. Sci. Hydrol. 49 (4), 717–735.
Hosking, J.R.M., Wallis, J.R., 1997. Regional Frequency Analysis: An Osborn, T.J., 1997. Areal and point precipitation intensity changes:
Approach Based on L-Moments. Cambridge University Press, implications for the application of climate models. Geophys.
Cambridge, UK, 224p. Rev. Lett. 24, 2829–2832.
Hosking, J.R.M., Wallis, J.R., Wood, J.R., 1985. Estimation of the Osborn, T.J., Hulme, M., 1997. Development of a relationship
generalized extreme-value distribution by the method of prob- between station and grid box rainday frequencies for climate
ability-weighted moments. Technometrics 27, 251–261. model validation. J. Climate 10, 1885–1908.
Houghton, J.T., Ding, Y., Griggs, D.J., Noguer, N., van der Linden, Pacanowski, R.C., Dixon, K., Rosati, A., 1993. The GFDL
P.J., Dai, X., Maskell, K., Johnsons, C.A., 2001. Climate Change Modular Ocean Model User’s Guide. GFDL Ocean Group
2001: The Scientific Basis. Contribution of Working Group I to the Tech. Rep. 2, Geophysical Fluid Dynamics Laboratory,
Third Assessment Report of the Intergovernmental Panel on Princeton, NJ, 46p.
Climate Change. Cambridge University Press, Cambridge, UK, Papa, F., Guo, Y., Thoman, G.W., 2004. Urban drainage infra-
881p. structure planning and management with a changing climate.
Hundecha, Y., Bárdossy, A., 2005. Trends in daily precipitation and In: Proceedings of the 57th Canadian Water Resources Asso-
temperature extremes across western Germany in the second ciation Annual Congress – Water and Climate Change: Knowl-
half of the 20th century. Int. J. Climatol. 25, 1189–1202. edge for Better Adaptation, Montréal, Québec, June 16–18,
IPCC, 2007. Climate Change 2007: The Physical Science Basis – 2004, 6p.
Summary for Policy Makers. Contribution of Working Group I to Plummer, N., Salinger, M.J., Nicholls, N., Suppiah, R., Hennessy,
the Fourth Assessment Report of the Intergovernmental Panel on K.J., Leighton, R.M., Trevin, B., Page, C.M., Lough, J.M., 1999.
Climate Change. IPCC, Geneva, Switzerland, 18p. Changes in climate extremes over the Australian region and New
Iwashima, T., Yamamoto, R., 1993. A statistical analysis of extreme Zealand during the twentieth century. Clim. Change 42 (11),
events: long-term trend of heavy daily precipitation. J. Mete- 183–202.
orol. Soc. Jpn. 71, 637–640. Plummer, D.A., Caya, D., Frigon, A., Côté, H., Giguère, M., Paquin,
Jones, P.D., Reid, P.A., 2001. Assessing future changes in extreme D., Biner, S., Harvey, R., de Elia, R., 2006. Climate and climate
precipitation over Britain using regional climate model integra- change over North America as simulated by the Canadian RCM. J.
tions. Int. J. Climatol. 21, 1337–1356. Climate 19 (13), 3112–3132.
Karl, T.R., Knight, R.W., 1998. Secular trends of precipitation Schmidli, J., Frei, C., 2005. Trends of heavy precipitation and wet
amount, frequency, and intensity in the United States. Bull. Am. and dry spells in Switzerland during the 20th century. Int. J.
Meteorol. Soc. 79 (2), 231–241. Climatol. 25, 753–771.
Karl, T.R., Knight, R.W., Plummer, N., 1995. Trends in high- Semmler, T., Jacob, D., 2004. Modeling extreme precipitation
frequency climate variability in the twentieth century. Nature events – a climate change simulation for Europe. Global Planet.
377, 217–220. Change 44, 119–127.
Kunkel, K.E., Pielke Jr., R.A., Changnon, S.A., 1999. Temporal Srikanthan, R.A., 1995. Review of the methods for estimating areal
fluctuations in weather and climate extremes that cause reduction factors for design rainfalls. Report 95/3, Cooperative
economic and human health impacts: a review. Bull. Am. Research Centre for Catchment Hydrology, Melbourne, Austra-
Meteorol. Soc. 80 (6), 1077–1099. lia, 36p.
Laprise, R., Caya, D., Giguère, M., Bergeron, G., Côté, H., Stone, D.A., Weaver, A.J., Zwiers, F.W., 2000. Trend in Canadian
Blanchet, J.-P., Boer, G.J., McFarlane, N.A., 1998. Climate precipitation intensity. Atmos. Ocean 38 (2), 321–347.
and climate change in western Canada as simulated by the Tebaldi, C., Hayhoe, K., Arblaster, J.M., Meehl, G.A., 2006. Going
Canadian regional climate model. Atmos. Ocean 36, 119–167. to the extremes – an intercomparison of model-simulated
Laprise, R., Caya, D., Frigon, A., Paquin, D., 2003. Current and historical and future changes in extreme events. Clim. Change
perturbed climate as simulated by the second-generation Cana- 79 (3-4), 181–211.
dian regional climate model (CRCM-II) over northwestern North Trenberth, K.E., 1999. Conceptual framework for changes of
America. Climate Dyn. 21, 405–421. extremes of the hydrological cycle with climate change. Clim.
Mailhot, A., Duchesne, S., Rivard, G., Nantel, E., Caya, D., Change 42, 327–339.
Villeneuve, J.-P., 2006. Climate change impacts on the perfor- Trenberth, K.E., Dai, A., Rasmussen, R.M., Parsons, D.B., 2003. The
mance of urban drainage systems for Southern Quebec. EIC changing character of precipitation. Bull. Am. Meteorol. Soc. 84,
Climate Change Technology Conference, May 9–12, 2006, 1205–1217.
Ottawa, Ontario, Canada. Vincent, L.A., Mekis, É., 2005. Changes in daily and extreme
McFarlane, N.A., Boer, G.J., Blanchet, J.-P., Lazare, M., 1992. The temperature and precipitation indices for Canada over the 20th
Canadian Climate Centre second-generation general circulation century. Atmos. Ocean 44 (2), 177–193.
model and its equilibrium climate. J. Climate 5, 1013–1044. Zhai, P., Zhang, X., Wan, H., Pan, X., 2005. Trends in total
Müller, P., von Storch, H., 2004. Computer Modeling in Atmospheric precipitation and frequency of daily precipitation extremes over
and Oceanic Sciences – On the Building of Knowledge. Springer- China. J. Climate 18, 1096–1108.
Verlag, Heidelberg, pp. 304. ISBN 1437-028X. Zwiers, F.W., Kharin, V.V., 1998. Changes in the extremes of the
New, M., Hewitson, B., Stephenson, D.B., Tsiga, A., Kruger, A., climate simulated by CCC GCM2 under CO2 doubling. J. Climate
Manhique, A., Gomez, B., Coelho, C.A.S., Masisi, D.N., Kulu- 11, 2200–2222.