You are on page 1of 20

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 115, D10101, doi:10.

1029/2009JD012882, 2010
Click
Here
for
Full
Article

Bias correction of monthly precipitation and temperature fields


from Intergovernmental Panel on Climate Change AR4 models
using equidistant quantile matching
Haibin Li,1 Justin Sheffield,1 and Eric F. Wood1
Received 23 July 2009; revised 24 November 2009; accepted 23 December 2009; published 18 May 2010.

[1] A new quantile‐based mapping method is developed for the bias correction of
monthly global circulation model outputs. Compared to the widely used quantile‐based
matching method that assumes stationarity and only uses the cumulative distribution
functions (CDFs) of the model and observations for the baseline period, the proposed
method incorporates and adjusts the model CDF for the projection period on the basis
of the difference between the model and observation CDFs for the training (baseline) period.
Thus, the method explicitly accounts for distribution changes for a given model between
the projection and baseline periods. We demonstrate the use of the new method over
northern Eurasia. We fit a four‐parameter beta distribution to monthly temperature fields
and discuss the sensitivity of the results to the choice of distribution range parameters. For
monthly precipitation data, a mixed gamma distribution is used that accounts for the
intermittent nature of rainfall. To test the fidelity of the proposed method, we choose
1970–1999 as the baseline training period and then randomly select 30 years from 1901–
1999 as the projection test period. The bootstrapping is repeated 30 times to mimic
different climate conditions that may occur, and the results suggest that both methods are
comparable when applied to the 20th century for both temperature and precipitation for
the examined quartiles. We also discuss the dependence of the bias correction results on
the choice of time period for training. This indicates that the remaining biases in the bias‐
corrected time series are directly tied to the model’s performance during the training
period, and therefore care should be taken when using a particular training time period.
When applied to the Intergovernmental Panel on Climate Change fourth assessment report
(AR4) A2 climate scenario projection, the data time series after bias correction from both
methods exhibit similar spatial patterns. However, over regions where the climate
model shows large changes in projected variability, there are discernable differences
between the methods. The proposed method is more sensitive to a reduction in variability,
exemplified by wintertime temperature. Further synthetic experiments using the lower 33%
and upper 33% of the full data set as the validation data suggest that the proposed
equidistance quantile‐matching method is more efficient in reducing biases than the
traditional CDF mapping method for changing climates, especially for the tails of the
distribution. This has important consequences for the occurrence and intensity of future
projected extreme events such as heat waves, floods, and droughts. As the new method is
simple to implement and does not require substantial computational time, it can be used to
produce auxiliary ensemble scenarios for various climate impact‐oriented applications.
Citation: Li, H., J. Sheffield, and E. F. Wood (2010), Bias correction of monthly precipitation and temperature fields from
Intergovernmental Panel on Climate Change AR4 models using equidistant quantile matching, J. Geophys. Res., 115, D10101,
doi:10.1029/2009JD012882.

1. Introduction making to cope with the consequences of changing climate


(World Modeling Summit for Climate Prediction, 2008;
[2] Global circulation models (GCMs) are an important
http://wcrp.ipsl.jussieu.fr/Workshops/ModellingSummit/
tool in assessing climate change and in helping decision
Documents/FinalSummitStat_6_6.pdf). These numerical cou-
1
Department of Civil and Environmental Engineering, Princeton
pled models, based on the governing physical laws and
University, Princeton, New Jersey, USA. expressed in forms of primitive equations, represent large‐
scale (generally hundreds of kilometers) flow patterns and
Copyright 2010 by the American Geophysical Union. dynamics for Earth system components, including the atmo-
0148‐0227/10/2009JD012882

D10101 1 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

sphere, ocean, land surface, and sea ice. It is on such spatial higher‐order moments, may result in underestimation of
scales that GCMs exhibit encouraging skill in simulating future climate extremes such as heat waves related to tem-
climate responses [Randall et al., 2007]. However, their perature or droughts and floods associated with precipitation
ability to capture local‐scale (usually tens of kilometers) or extremes. In this study, we focus on statistical bias correction
even regional‐scale patterns that are directly relevant to end and propose a simple methodology called equal distance‐
users for decision making and mitigation strategy planning based quantile‐to‐quantile matching to correct monthly pre-
is less promising, especially for precipitation [e.g., Wood et cipitation and temperature fields from the World Climate
al., 2004]. Thus, there is a mismatch between the capacity of Research Program’s Coupled Model Intercomparison Proj-
current GCMs and desired details at local scales. What is ect phase 3 (CMIP3) GCM simulations that formed the basis
more, imposed on this is the inevitable model bias due to for the fourth assessment report (AR4) of the IPCC. The
inadequate knowledge of key physical processes (e.g., cloud proposed method has the advantage that it explicitly in-
physics) and simplification of the natural heterogeneity of corporates changes in the distribution in the future climate.
the climate system that exist at finer spatial scales. Although We will show that changes in the tails of the distribution
systematic improvements have been reported in the latest can be better reflected with our method when compared to
generation of climate models [Reichler and Kim, 2008], a the traditional quantile‐based mapping method (see section 4
general cold bias is still evident for temperature, and sub- for details).
stantial precipitation biases, especially in the tropics, are not [5] The paper is structured as follows. In section 2, we
uncommon in many models [Randall et al., 2007]. describe the study domain, followed by the data and meth-
[3] Downscaling approaches, either physical process‐ odology in sections 3 and 4, respectively. Section 5 presents
based dynamic downscaling or statistically based ones, are results of the methodology validation. Section 6 details the
required to remove systematic biases in models and trans- application of the methodology to future projections of
form simulated climate patterns at coarse grid to a finer precipitation and temperature. Finally, discussion and con-
spatial resolution of local interest [Maurer and Hidalgo, clusions are given in section 7.
2008]. The dynamic approach uses limited area models or
high‐resolution GCMs to simulate physical processes at
fine scales with boundary conditions given by the coarse‐
2. Study Domain
resolution GCMs. The statistical approach transforms coarse‐ [6] We select the Northern Eurasian Earth Science Part-
scale climate projections to a finer scale through trained nership Initiative (NEESPI) domain as the study area, which
transfer functions that connect the climate at the two spatial is defined as being between 15°E in the west and the Pacific
resolutions. To capture the anthropogenic climate change Coast in the east and between 40°N in the south and north to
signal, the choice of predictor variables is a critical step the Arctic Ocean coastal zone. In this study, we extend the
[Hewitson and Crane, 2006]. Two important considerations NEESPI domain to 34°N. The NEESPI area accounts for
are that (1) the selected predictors should reflect the primary about 19% of the global land surface, covering diverse cli-
circulation dynamics of the atmosphere reasonable well and mate zones from Mediterranean climate in the southwest
(2) there is a physical connection to the predictant. There coast to subarctic climate in Siberia in the far northeast. This
are also statistical downscaling methods primarily for the region is also characterized by a carbon‐rich, cold region
purpose of bias correction which involve some form of component of the Earth system with potentially large im-
transfer function derived from cumulative distribution plications for the global climate system [Groisman et al.,
functions (CDFs) of observations and model simulations 2009]. The NEESPI region experiences an amplified re-
[e.g., Ines and Hansen, 2006; Piani et al., 2010; Wood et sponse to climate forcing due to positive feedbacks [Moritz
al., 2004]. The advantages and disadvantages of both et al., 2002] related to, for example, albedo reduction,
approaches have been thoroughly documented [e.g., Fowler glacial and sea ice retreat, and vegetation expansion pro-
et al., 2007; Wilby et al., 2009]. The key advantage of the cesses. Climatic changes have already been reported in this
statistical approach is the lower computational requirement region in recent decades, including increased winter air
compared to the dynamical model–based alternative, and temperature [Rawlins and Willmott, 2003; Robeson, 2004],
thus, statistical downscaling approaches are widely used in reduced snow cover [Serreze et al., 2000], reduced sea ice
climate impact–related research work. retreat [Serreze et al., 2003; Serreze and Francis, 2006],
[4] An underlying assumption for statistical approaches is and increased runoff and river discharge [Peterson et al.,
that the statistical model which connects the large‐scale 2002]. These rapid changes have raised concerns of
circulation and the local climate (or the transfer function abrupt climate change. For regions north of 60°N for the 21st
relating observations to model simulations) remains un- century, regional model simulations project continuous
changed in an altered climate, which may not hold. How- warming (particularly large in autumn and early winter,
ever, if the time series used to train the statistical model is being 2°C–4°C higher than spring and summer) and an
long enough, many different situations, including the altered overall wetter climate. The associated impacts include
climate, may be included [Zorita and von Storch, 1999]. In widespread reductions in snow cover and a shortened snow
applying transfer functions to future climate projections, cover season, accompanied by more frequent climate ex-
many existing statistical bias correction methods simply tremes [Christensen et al., 2007] that may have devastating
assume the higher‐order moments do not change much in consequences. A recent study by Solomon et al. [2009] points
the future climate scenario, yet the latest report from the out that greenhouse gas–induced changes may be irreversible
IPCC suggests otherwise [Meehl et al., 2007]. The non- in the time frame of a million years. Thus, there is an urgent
stationarity issue has a profound influence on climate impact need to investigate potential changes, especially those be-
studies. Neglecting changes in the distribution, particularly yond the climatological change in this region, to form the

2 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

scientific basis for possible disaster mitigation and design tating. This supports our intention to develop the new
strategies for adaptation. quantile‐mapping technique that deals with nonstationarity
in climate time series and particularly in higher moments.
On an annual scale, a general cold bias exists in the multi-
3. Data Sets model averages, usually less than 2°C outside of the polar
[7] The observations used in this study are gridded regions. Individual CMIP3 models can exhibit larger errors.
monthly temperature and precipitation data prepared by the For PCM1, the errors are as large as 5°C in the Sahara desert
Climate Research Unit (CRU) of University of East Anglia and coastal areas of midlatitudes to high latitudes [Randall
at 0.5° resolution for 1901–2000 [New et al., 2002]. The et al., 2007]. In terms of seasonal patterns, the cold bias
CRU data set is one of the best available consistent long‐ in summer (July) is more pronounced than that in winter
term gridded observational records. These data are averaged (January). For 1970–1999, the spatially averaged mean bias
to 1.0° resolution. It should be noted that the CRU data set is about −1.2°C in January and up to −3.9°C in July.
does not make adjustments for gauge undercatch of snow- [9] For precipitation (Figure 2), the NEESPI region has a
fall. For high‐latitude basins, this means wintertime snow- very diverse climate, characterized by low precipitation over
falls may be underestimated by the CRU data set, and some the majority of inland areas in wintertime as a result of the
modeling has shown that this may be important [e.g., Tian et strong influence from cold and dry polar air masses from
al., 2007]. However, other studies have shown that gauge high latitudes. The exception is the Mediterranean region,
undercatch adjustments overpredict regional observations of which is characterized by a seasonal peak in winter. In
precipitation and snow water equivalent [Troy et al., 2008]. summertime, the Asian monsoon brings moist air from the
For illustration purposes, we use the model outputs from the tropical Pacific Ocean, resulting in high precipitation in East
Parallel Climate Model (PCM1) [Meehl et al., 2001, 2004; Asia. Extremely dry summers for the Mediterranean coasts
Washington et al., 2000] developed at the National Center are caused by the sinking air of the subtropical high. The
for Atmospheric Research. We choose to use one ensemble large‐scale spatial structure of modeled precipitation agrees
member for simplicity and to illustrate the method. How- with observations for the period 1970–1999, suggesting that
ever, the method can be readily applied to an ensemble of the model captures the large‐scale dynamics reasonably
realizations to form an envelope to reflect the uncertainty in well. The precipitation brought by extratropical cyclones
future projections. We use the model data sets for the 20th along midlatitude storm tracks through western Europe and
century (20C3M) and the Special Report on Emissions into Eurasia [Money, 2000] is overestimated by the model.
Scenarios A2 future scenarios for 2001–2099 [Nakićenović This has also been noted by Randall et al. [2007], who
and Swart, 2000], which were run as part of CMIP3. The found a positive bias in the PCM1 compared to the Climate
A2 scenario is generally regarded as a worst‐case scenario Prediction Center Merged Analysis of Precipitation [Xie and
that sees a fourfold to fivefold increase in CO2 emissions Arkin, 1997] climatology over Eurasia. This may be an in-
over 2000–2099, during which CO2 concentrations increase dication that the model‐simulated westerlies are too strong
from about 350 to 850 ppm. The model data are regridded to or the model topography fields are oversimplified (too flat),
the same resolution as the observations (1.0°) using bilinear allowing depressions to penetrate farther eastward into
interpolation. Thus, the preprocessed model outputs are a central Eurasia. Also, the precipitation variability is under-
combination of inherent model errors and uncertainties in- estimated during the monsoon season for coastal areas of
troduced by the interpolation scheme. If a much finer res- East Asia. The same can be seen for the Mediterranean region
olution is desired, a more sophisticated spatial downscaling during wintertime. Regions with high precipitation are usu-
scheme utilizing auxiliary information, for example, eleva- ally characterized by relatively high variability. For CMIP3
tion and pressure level, may be used. The CRU data are used models in general, they adequately capture most climatic
to bias correct the 20C3M model data, and these corrections features, for example, lower precipitation at higher latitudes.
are then applied to the A2 future projections to provide bias‐ The models also represent well a local minimum in precipi-
corrected future climate data. tation near the equatorial Pacific and local maxima at mid-
[8] Figure 1 compares the basic statistics of temperature latitudes due to frequent storms. Despite the apparent skill
for the observations and 20C3M data for 1970–1999 and the exhibited by the CMIP3 multimodel mean, models individ-
A2 scenario data for 2070–2099. The model has a cool bias, ually still display substantial biases that may exceed the
most noticeable in higher latitudes, which is typical of most magnitude of the mean observed precipitation climatology in
climate models [Randall et al., 2007]. High latitudes gen- the tropics in particular [Randall et al., 2007]. Under the A2
erally show stronger variability regardless of the season. For scenario, July precipitation is projected to increase more
the baseline period of 1970–1999, PCM1 captures the spatial than 50% for the Tibetan Plateau, which is picked up by
distribution of regions with strong variability but exaggerates most CMIP3 models, and the higher the elevation, the more
wintertime variability. For July, there is less agreement for the increase. A reduction in precipitation is projected for
high‐variability regions between the PCM1 and CRU data. In central Asia and southern Europe, due to more easterly and
addition to an increase in mean for the future projection, the anticyclonic flow. However, there is less consensus among
model predicts a decrease in wintertime variability but an GCMs for the summertime precipitation decrease for these
increase in summertime. The higher‐order moment skew also regions [Christensen et al., 2007].
changes in the future. It is clear that change in the distribution
of the future climate is not limited to shifts in the mean;
there are also changes in higher‐order moments. These
4. Methodology
changes are more important for resources managers and [10] The goal of this study is to develop a simple and
stakeholders, as the consequences are usually more devas- effective statistical bias correction method. The basic prin-

3 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 1. (top) January and (bottom) July temperature statistics. (left) Observations (1970–1999).
(middle) PCM1 (1970–1999) for the 20C3M scenario. (right) PCM1 (2070–2099) for the A2 scenario.
For each month, the maps show the statistics for mean (m), standard deviation (s), and skew (g).

4 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 2. Same as Figure 1 but for precipitation (units: mm/month for m and s).

5 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 3. Illustration of methodology of (a) CDFm and (b) EDCDFm using synthetic generated winter-
time temperature data for a grid point. Solid black line shows observations (OBS). Thick dashed line
shows model simulation for current climate (MODc). Thin dashed line shows model simulation for future
projection (MODp). Dotted‐dashed line shows CDFm in Figure 3a and EDCDFm in Figure 3b. See text
for details.

ciple, regardless of the complexity of the statistical model, is solid circle). The major advantage of the method is that it
to establish a statistical relationship or transfer function adjusts all moments (i.e., the entire distribution matches that
between model outputs and observations based on available of the observations for the training period) while maintain-
historical data sets and then apply the established transfer ing the rank correlation between models and observations.
function to future model projections to infer the possible However, an underlying assumption of the method is that
trajectory of future observations. The quantile‐based map- the climate distribution does not change much over time,
ping method (CDF matching (CDFm hereafter)) [Panofsky that is, it is stationary in the variance and skew of the dis-
and Brier, 1968] maps the distribution of monthly GCM tribution, and only the mean changes. This, however, may
variables (precipitation and temperature) onto that of gridded not hold [Meehl et al., 2007; Milly et al., 2008] with the
observed data. The method is a relatively simple approach possibility that the model projects changes in the higher
that has been successfully used in hydrologic and many moments as well, as shown in Figures 1 and 2. One way to
other climate impact studies [e.g., Cayan et al., 2008; allow for this is to incorporate information from the CDF of
Hayhoe et al., 2004; Maurer and Hidalgo, 2008]. For a the model projection instead of assuming that the historic
climate variable x, the method can be written as model distribution applies to the future period. For a given
 percentile, we assume that the difference between the model
~xm-p:adjst: ¼ Fo1
-c Fm-c xm-p ; ð1Þ and observed value during the training period also applies to
the future period, which means the adjustment function
where F is the CDF of either the observations (o) or model remains the same. However, the difference between the CDFs
(m) for a historic training period or current climate (c) or for the future and historic periods is also taken into account.
future projection period (p). Basically, to bias correct model Accordingly, the method can be written mathematically as
values for a future period, we first find the corresponding  
percentile values for these future projection points in the ~xm-p:adjst: ¼ xm-p þ Fo1
-c Fm-p xm-p  Fm1-c Fm-p xm-p : ð2Þ
CDF of the model for the training period and then locate the
observed values for the same CDF values of the observa- This method is termed the equidistant CDF matching
tions. These are the model values after bias correction. method (EDCDFm). Figure 3b illustrates how the method
Figure 3a illustrates how the method works. The CDFs for works. A value of 25 corresponds to 0.05 in the CDF of the
the observations and model are constructed from the January future model projection. For the 5th percentile, the differ-
temperature field for a point near 60°N, 150°E. According to ence between the model and observations under current
equation (1), a value of 25 in a future projection time series climate is subtracted from the model projection to get the
corresponds to a Fm‐train value of 0.2 under the current bias‐corrected value (the solid circle in Figure 3b, which, in
climate, which can be transferred to the observed value this example, is different from that estimated by the CDFm
according to the quantile function of the observations (the method). The same procedure is repeated for every point for

6 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

the future projection time series, which can then be used to for the summer dry regions but totally misses the dry region
construct the CDF for the bias‐corrected future projection in wintertime, and the implications of this in the bias cor-
time series (the dotted‐dashed line). If the distribution for rection are discussed in section 5.1.
the future climate is the same as the current one (training
period), the results from the CDFm and EDCDFm methods 5. Methodology Validation
will be identical. Moreover, if the changes in variability
are small, results from both methods will be close. In this 5.1. Validation Using a Bootstrapping Approach
example, however, the distribution for the model has [13] The validation of a method or model is done tradi-
changed from the training period to the future period. As a tionally by testing its performance for a specific historical
comparison, the distribution function using the method of period, say, 30 years, for which observations are available, a
EDCDFm is different from that using CDFm, especially near so‐called “pseudoprojection.” However, the performance
the tails. Figure 3 is solely for the purpose of illustration; the statistics calculated in this way may not be representative, as
method is validated comprehensively in sections 5 and 7. the results may be different if we choose a different vali-
[11] If empirical CDFs are used for the quantile mapping, dation period. This may arise since the model performance
frequent interpolation or extrapolation is required, which is time dependent, particularly for areas with pronounced
is unsatisfactory. We therefore fit parametric distributions seasonality and large variability, for example, wintertime
to both the temperature and precipitation fields. A four‐ temperature for continental interior and summertime pre-
parameter beta distribution is fitted to temperature fields: cipitation for high latitudes. Ideally, we should use as many
different periods as possible for the validation, and some
1
f ð x; a; b; p; qÞ ¼ ð x  aÞp1 statistics can then be derived to assess the performance of
Bð p; qÞðb  aÞpþq1 ð3Þ the methodology. Limited by the available observational
 ðb  xÞq1 a  x  b; p; q > 0; records (100 years), we use a bootstrapping approach. We
also sampled continuous overlapping 30 year periods (e.g.,
where B is the beta function. We first approximate the range 1901–1930, 1911–1940), but these gave essentially similar
parameters (a and b) as the extreme values from the data, results to those of the bootstrapping approach. Specifically,
extended by a certain percentage of the standard deviation we choose 1970–1999 as the training period, in accordance
(s), similar to the approach by Watterson [2008]. Once the with many climate impact studies. Then we randomly select
range parameters are determined, the shape parameters p and 30 years out of 100 (1901–2000) as the test data. This
q can be estimated by the method of maximum likelihood procedure is repeated 30 times. The model performance
estimation. To investigate the sensitivity of the fitted dis- is then evaluated in terms of biases in reproducing the ob-
tribution to the choice of range parameters, we extended the servations for selected quantiles. The results collectively tell
extreme values from 5% to 200% of s. The differences how well the method performs using 1970–1999 as the
between the constructed CDFs are small once the selected baseline climatology.
percentage value is above 30%. We therefore set range [14] Figure 4 shows the spatial patterns for median values
parameters for every grid point as the extreme values averaged from the 30 bootstrap experiments using the
extended by half of the s values. Our analysis suggests CDFm and EDCDFm methods. For temperature, the spatial
that this value is adequate to enclose possible extreme values patterns of the bias‐corrected fields resemble the observa-
in future projections. tions for both the cold (January) and warm (July) seasons
[12] Precipitation is intermittent in nature, and even at (Figure 4). Similar conclusions can be drawn for other
monthly time scales a mixture of months with no rain and quartiles as well. There are distinct differences between the
months with rain can occur, especially for dry regions. We January and July model biases. For January, PCM1 is too
construct a mixed gamma distribution to account for this warm for a large area of continental interior and northern
feature. The CDF for the mixed distribution can be written as Asia north of 60°N but is too cold for the northwest and
southeast portions (Figures 4a and 4b). Overall, PCM1 ex-
Gð xÞ ¼ ð1  PÞH ð xÞ þ pF ð xÞ; ð4Þ hibits a moderate cold bias when averaged spatially but with
a large spread in winter (Figure 5). The results after bias
where P is the percentage of months with rain and H(x) is a correction show a dramatic reduction in the mean and range
step function having a value of zero when there is no rain and of the biases. The results for the EDCDFm and CDFm
a value of 1 when there is rain. For the portion of a given methods are very similar (Figures 4c, 4d, and 5). For July,
time series with rain, a two‐parameter gamma distribution is the model shows a strong and systematic cold bias for nearly
used: the entire NEESPI region (Figures 4e and 4f) with a mean
bias of over 4° for all quartiles (Figure 6), though there is a
ex=
f ð x; k; Þ ¼ xk1 for x > 0 and k;  > 0: ð5Þ much smaller variability compared to the results for January.
k ðk Þ Both the EDCDFm and CDFm methods are very effective in
terms of reducing the original model biases (Figures 4g, 4h,
At the monthly scale, the majority of the NEESPI region has and 6). The remaining biases, after bias correction, also tend
a value of 1 for P. Regions with values of <1 are limited to to cluster more tightly around zero compared to the raw
small dry areas: centered on central Asian countries east of PCM1 data, indicating a much smaller mean and greatly
the Caspian Sea (including northern Iran, Afghanistan, reduced variability (Figure 6). It is worthwhile to note that
Turkmenistan, and Uzbekistan) in summer and western the EDCDFm and CDFm methods exhibit comparable skill
China and the Tibetan Plateau in winter. Note that the with no one method being superior to the other, at least for
PCM1 model roughly captures the geographical distribution the 20th century data and for the examined quartile values.

7 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 4. Spatial distribution of temperature (median values for (a–d) January and (e–h) July) for the ob-
servations, model, EDCDFm, and CDFm bias‐corrected data sets averaged over the 30 bootstrap samples.

8 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 5. (top) Mean biases for the 25th, 50th, and 75th percentiles (temperature) for January. Histograms
of (bottom left) model biases, (bottom middle) results from the EDCDFm, and (bottom right) results from
the CDFm for the 25th, 50th, and 75th percentiles.

9 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

[15] For precipitation, an issue arises in some dry regions lower 33% and upper 33% of the full 20th century data for
where the model has essentially no skill at all: either (1) the testing, thus representing the two extreme situations. We
observations have no or very little rainfall in a few months, then calculated the mean absolute biases, root‐mean‐square
but the model indicates that it rains every month and total error (RMSE), and reduction of RMSE (one minus the ratio
rainfall can be of the order of 1000 mm, or (2) the ob- between the RMSE from the bias correction method and
servations indicate significant rainfall (hundreds of milli- RMSE from PCM1) for the 5th (95th) percentiles of the
meters), but the model is very dry. This issue arises for lower 33% and upper 33% of the full data record. The results
about 0.4% of grid points in January and 0.2% in July. For are presented in Table 1. First, both methods show significant
these grid points, we use the ratio of the future to the current reduction in all error statistics analyzed, in line with the
climate in the model instead of the difference as the transfer findings reported in section 5.1. The reduction of RMSE
function to adjust the observed CDFs. The method works ranges from 70% in January to about 95% in July. Both
well for those problematic grid points. As the model shows methods show smaller mean biases and RMSE in both
such poor skill for these grid points, it may even be more wintertime and summertime, with greater skill in summer-
appropriate to directly use the observed climatology (perhaps time as a result of the smaller variability. From Table 1, we
with the addition of some white noise) instead of attempting may say that the EDCDFm method is superior to the CDFm
to bias correct the model data. method. The former shows, on average, a 4–5% (10%)
[16] Figure 7 shows the median values for the precipitation further reduction in RMSE in summer (winter). Thus, the
field (January in Figures 7a–7d and July in Figures 7e–7h). EDCDFm developed in this study seems more skillful in
Again, both the EDCDFm and CDFm methods show im- dealing with changed variability.
provements compared to the original model outputs with
5.3. Choice of Training Period
more realistic spatial patterns. In January, the negative bias in
the PCM1 for the Mediterranean regions and positive bias for [19] Another question that has not been addressed in the
the East Asian Peninsula (Figures 7a and 7b) are substan- literature is how the results would differ when using another
tially reduced by both methods (Figures 7c and 7d). The time period for training. To shed light on this, we carried out
underestimation of summertime Asian monsoon rainfall for another bootstrapping test for the January temperature field
East Asia is also dramatically improved after bias correction using 1901–1930 as the training period. The spatial patterns
(Figures 7e–7h). In terms of the remaining biases, the spatial are similar to those shown in Figure 4a. Both statistical bias
patterns for the different quartiles are very similar for the two correction methods produce data in better agreement with
methods (Figures 8 and 9). Over regions with relatively high observations. However, it is also interesting to note that the
precipitation and high variability (Figure 2) the biases tend to spatial patterns for the presented mean biases differ from
also be relatively large. This is understandable, as such those depicted in Figure 5 (top) in terms of the sign of the
regions are usually characterized by strong dynamics, mean biases: the slight positive biases for the southeast
which can be difficult to simulate. Both the EDCDFm and when using 1970–1999 switch to negative biases when
CDFm are able to remove biases considerably, but less using 1901–1930, and the northwest of Russia now exhibits
skillfully in July compared to January. Relatively large positive biases (not shown). The analysis suggests that the
biases are still apparent for regions with high precipitation bias‐corrected time series are directly tied to the model’s
and large variability (up to 30 mm/month), especially for performance during the training period, supporting the ar-
July, though of a much smaller magnitude than the model gument that model biases are time dependent, and agrees
biases. with the finding by Reifen and Toumi [2009]. They dem-
[17] The reason for the similar performance from both onstrated that GCMs that perform well in one period may
statistical bias correction methods may be explained by the not necessarily perform equally well for another period,
way the sampling is done in the bootstrapping experiment possibly a result of nonstationarity of climate feedback
and the very limited length of the climate records (30 out of strength or a model’s representation of external forcings
100 points for each sample). Given long enough climate (e.g., CO2 concentration, aerosols). The results indicate that
records and substantially more samples, we may pick up care should be taken when using a specified time period for
extreme distributions, which can be used to investigate the training. However, if the temporal variation of the bias has a
sensitivity of methods to data having quite different dis- much smaller magnitude than the bias itself, we may simply
tributions. As these conditions, especially the former, are choose a period we feel most confident with if the main
hard to satisfy, we have designed another synthetic ex- objective is bias removal. The choice of the past few dec-
periment specifically for the purpose of examining climate ades (1970–1999) as the training period for bias correction
extremes. of future projections in many climate downscaling studies
may be justified for the following reasons: (1) recent
5.2. A Synthetic Experiment for Bias Correction observations are more reliable, and (2) recent decades have
of Climate Extremes experienced warming and thus are more likely to resemble
[18] Our main goal is to develop a method that can better future projections.
handle a changing climate, including changes in variability
and extreme monthly values. Therefore, a key question is
whether the EDCDFm is more efficient in dealing with 6. Future Projection
changed variability than the CDFm. To answer this ques- [20] In this section, we use the EDCDFm method to
tion, we conducted a synthetic experiment using the 20th downscale a future model run of the A2 scenario, a high‐
century temperature data. First, we used the entire 20th emission scenario that is associated with high population
century data (1901–1999) for training and then chose the growth, regionally oriented economic development, and

10 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 6. Same as Figure 5 but for July.

11 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 7. Spatial distribution of precipitation (median values for (a–d) January and (e–h) July) for the
observations, model, EDCDFm, and CDFm bias‐corrected data sets averaged over the 30 bootstrap
samples.

12 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 8. (top) January mean biases for the 25th, 50th, and 75th percentiles for precipitation. Histograms
of bias for (bottom left) the model, (bottom middle) results from the EDCDFm method, and (bottom right)
results from the CDFm method for the 25th, 50th, and 75th percentiles.

13 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 9. Same as Figure 8 but for July.

14 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Table 1. Summary of Statistics for the Synthetic Bias Correction Experiments Using the Lower 33% and Upper 33% of the Full 20th
Century Data Sets for Temperaturea
Januaryb Julyb
5% of Lower 33% 95% of Upper 33% 5% of Lower 33% 95% of Upper 33%
I II III I II III I II III I II III
MAB 4.26 0.95 1.70 0.52 0.12 −0.47 4.51 0.04 0.13 4.32 0.06 0.16
RMSE 7.10 1.30 2.08 4.56 0.68 1.28 5.24 0.25 0.50 4.87 0.29 0.48
R‐RMSE (%) 81.71 70.63 85.05 71.90 95.18 90.49 93.99 90.09
a
See text for details about the experiment design. MAB, mean absolute bias; RMSE, root mean square error; R‐ RMSE, reduction, defined as 1‐RMSEBC/
RMSEPCM1, where RMSEBC represents RMSE model outputs after bias correction. The statistics represent spatial averages.
b
I, PCM1; II, EDCDFm; III, CDFm.

slow technological change [Nakićenović and Swart, 2000]. wintertime, the bias‐corrected precipitation is higher for
The objectives of this are twofold: first, to see how the bias‐ the Mediterranean region and the northwest of Russia. A
corrected future projection compares to the original PCM1 reduced precipitation amount is observed for Japan. In
model output and, second, for such a high‐emission scenario summertime, both downscaling methods suggest higher
with associated changes in temperature and precipitation, to precipitation for East Asia, particularly for east coastal
see how different the results are when using the EDCDFm China, the Korean Peninsula, and Japan. These changes
or CDFm. We present the results for the period 2070–2099. seem reasonable considering the biases in the model‐simulated
[21] For temperature, the two methods give very similar precipitation for the training period. The differences between
spatial patterns for the mean and variability (Figure 10). these two methods are more subtle and isolated. With respect
Compared to the original model projections, both methods to differences for given percentiles, differences are small for
show higher temperatures in July, suggesting a much lower percentile values, which comes as no surprise because
stronger warming will occur (because of the cold 20th the lower percentiles correspond to low precipitation
century bias in the model, which the methods correct for). In amounts; thus, the absolute differences are small. On the other
January, a colder northern Asia and warmer northwest and hand, higher percentiles correspond to higher precipitation
southeast are expected at the end of the 21st century under amounts; thus, the absolute differences can be relatively large
the A2 scenario. These features are consistent with the (up to 10 mm/month), especially for regions where the model
validation results discussed in section 5. The bias‐corrected shows marked changes in variability.
projections show a dramatic reduction in variability for large
areas in January. On the other hand, enhanced variability is
evident for Barents Sea coastal areas from both downscaling 7. Discussion and Concluding Remarks
methods in July, a result of a combination of underestimated [23] We have developed a new quantile‐based mapping
variability for the baseline period and increased variability method, called EDCDFm, for the purpose of bias correction
in the future projection. However, there are some subtle of GCMs near surface meteorological fields and compared
differences, primarily related to the variability in January: it to the traditional CDFm. Different from most previous
the EDCDFm method shows smaller variability for the studies, which usually only select one time period for vali-
eastern coast and the central and eastern parts of northern dation, a bootstrapping method is used that samples the full
Asia compared to that from the CDFm method (Figure 10). historic record to test the fidelity of the downscaling. This
The reduced variability for these regions is a reflection gives an indication of the robustness of the method and
of the method’s adjustment to the exaggerated variability for highlights regions of high confidence or large uncertainty.
the current climate and a reduced variability in future pro- Both methods are able to reduce biases in the PCM1 model
jection by PCM1. For these regions, EDCDFm tends to have temperature and precipitation fields significantly. The vali-
a distribution with higher values for lower percentiles and dation results from the bootstrapping experiment suggest
lower values for higher percentiles. The further away from that the EDCDFm is comparable to the traditional quantile‐
the 50th percentile, the larger the differences between these based mapping method if we use 1970–1999 as the training
two methods, exceeding 1°C near the tails in January period.
(Figure 11, top). The differences in July are relatively smaller [24] We further designed a synthetic experiment to in-
in both magnitude and geographical coverage (Figure 11, vestigate the efficiency of both methods in reducing climate
bottom). The results also suggest that EDCDFm is more model bias in a synthetically generated changing climate.
sensitive to reduced variability than the CDFm counterpart. The analysis for selected statistical metrics (mean absolute
Comparatively, July differences between these two methods bias, RMSE, reduction of RMSE) suggests that the
are less discernable, though results from EDCDFm are about EDCDFm method is superior to the CDFm method. Thus,
0.5°C–1°C higher than those from CDFm for the coastal the EDCDFm developed in this study seems more skillful in
areas of the Kara Sea and Laptev Sea. The values between dealing with changed variability. This may have profound
the 25th to 75th percentiles are comparable. The differences implications for climate impact studies that are concerned
are more apparent for percentiles below the 5th and above with changing climate variability.
95th percentiles for January. [25] To investigate the sensitivity of bias correction results
[22] The spatial structure of precipitation after bias cor- to the choice of different baseline periods, we performed the
rection is quite similar for the two methods (Figure 12). In same bootstrapping experiments using the period of 1901–

15 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 10. Mean and variability of the original model temperature projection (2070–2099) and after bias
correction using the EDCDFm and CDFm methods. (top) January and (bottom) July.

16 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 11. Differences for selected percentiles for bias‐corrected temperature projections using
EDCDFm and CDFm. (top) January and (bottom) July.

17 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Figure 12. Same as Figure 10 but for precipitation.

18 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

1930 as the training period. We found that the choice of random sampling method [e.g., Hirabayashi et al., 2005],
baseline period does have an impact on the bias correction gridded time series at daily or even finer temporal scales
results: the bias‐corrected results are tightly connected to the can be derived for, but not limited to, hydrologic impact
model biases for the training period. The results indicate that studies. This is the subject of an accompanying study in
care should be taken when using a specified time period for which our approach is to sample the daily and subdaily
training. variation from the historic observations (thus maintaining
[26] In bias correcting future scenario projections, the realistic temporal and spatial structure in the data) but
differences between the two methods are subtle for the looking at improved methods that take into account projected
selected quartiles. However, for regions where the model changes in the daily and subdaily statistics (e.g., storm fre-
exhibits changes in variability, the differences are more quencies and intensities and diurnal cycles of temperature and
apparent at the tails of the distributions. The EDCDFm precipitation).
method is more sensitive to reduced variability in the future [29] Changes in the hydrological cycle, especially more
projection and tends to produce a probability density func- frequent extreme events, can have the most devastating
tion that has a much narrower range than the CDFm method. impacts on human society and result in losses exceeding
For these regions, although the mean values are very close billions of dollars. Meanwhile, there is increasing evidence
for both methods, the differences can be as large as 1°C of the intensification of the global hydrological cycle for the
when approaching the tails (e.g., January temperature pro- past century [Huntington, 2006; Milly et al., 2002]. The
jection for the 5th or 95th percentile). This may have pro- proposed method offers decision makers another possible
found implications for the hydrology in high‐latitude regions scenario of the future climate. It is especially useful for
where, for example, temperature‐dependent snowmelt in the regions experiencing changes at a relatively fast pace, such
spring is the predominant source of water. as high latitudes. These scenario forcings will be invalu-
[27] Although statistical downscaling approaches do not able in assisting, for example, water resources planners to
provide a physical explanation for biases, they have a assess future flood risk, update reservoir operation rules,
computational advantage to dynamic downscaling and have and initiate talks for sharing cross‐border water resources,
skill comparable to limited area climate models, at least for especially for water‐scarce regions.
present climate conditions [Hay and Clark, 2003; Wilby and
Wigley, 2000; Wood et al., 2004]. This makes statistical
downscaling very attractive to climate impact and assess- [30] Acknowledgments. The study is supported by NSF grant ARC‐
0629471 and NASA grant NNG06GE62G. Their support is gratefully
ment applications because large ensemble members can be acknowledged. We acknowledge the modeling groups, the Program
generated relatively easily. On the other hand, the EDCDFm for Climate Model Diagnosis and Intercomparison (PCMDI) and the
assumes a time‐invariant bias (or transfer function) like WCRP’s Working Group on Coupled Modeling (WGCM), for their roles
in making available the WCRP CMIP3 multimodel data set. Support of this
other statistical downscaling techniques. This assumption data set is provided by the Office of Science, U.S. Department of Energy.
may not necessarily hold. For example, if the region of
interest experiences a different rainfall regime in the future
as a result of changes in large‐scale circulation, then the References
future bias would be controlled by different processes, and
Benestad, R. E. (2001), A comparison between two empirical downscaling
the behavior of the bias may change. This is a limitation of strategies, Int. J. Climatol., 21(13), 1645–1668, doi:10.1002/joc.703.
the downscaling method, but one that cannot be tested, as Cayan, D. R., E. P. Maurer, M. D. Dettinger, M. Tyree, and K. Hayhoe
we cannot validate future projections. Compared to other (2008), Climate change scenarios for the California region, Clim.
Change, 87, Suppl. 1, 21–42, doi:10.1007/s10584-007-9377-6.
statistical downscaling approaches, the EDCDFm has two Christensen, J., et al. (2007), Regional climate projections, in Climate
attractive features. First, the method explicitly considers Change 2007: The Physical Science Basis. Contribution of Working
changes in the distribution of the future climate, including Group I to the Fourth Assessment Report of the Intergovernmental Panel
the tails of the distribution, which are most pertinent for on Climate Change, edited by S. Solomon et al., pp. 874–940, Cambridge
Univ. Press, New York.
climate impact and assessment studies. Second, the pro- Fowler, H. J., S. Blenkinsop, and C. Tebaldi (2007), Linking climate
posed method is very straightforward and simple to use. It change modelling to impacts studies: Recent advances in downscaling
is particularly useful for a large spatial domain such as the techniques for hydrological modelling, Int. J. Climatol., 27(12), 1547–
1578, doi:10.1002/joc.1556.
Eurasian continent or the entire globe. In contrast, many Groisman, P. Y., et al. (2009), The Northern Eurasia Earth Science Partner-
statistical downscaling approaches are based on regression‐ ship: An example of science applied to societal needs, Bull. Am. Meteorol.
type models or weather classification schemes developed Soc., 90, 671–688, doi:10.1175/2008BAMS2556.1.
for small regions or specific sites and therefore are inap- Hay, L. E., and M. P. Clark (2003), Use of statistically and dynamically
downscaled atmospheric model output for hydrologic simulations in
propriate for large‐scale applications. Furthermore, nearly three mountainous basins in the western United States, J. Hydrol., 282,
every statistical downscaling approach is developed for a 56–75, doi:10.1016/S0022-1694(03)00252-X.
specific purpose or location, and so there is no panacea, Hayhoe, H., et al. (2004), Emissions pathways, climate change, and impacts
on California, Proc. Natl. Acad. Sci. U. S. A., 101, 12,422–12,427,
although general guidelines on the choice of statistical doi:10.1073/pnas.0404500101.
method have been developed [Wilby et al., 2004]. Hewitson, B. C., and R. G. Crane (2006), Consensus between GCM climate
[28] The EDCDFm method can be readily applied to other change projections with empirical downscaling: Precipitation downscaling
over South Africa, Int. J. Climatol., 26(10), 1315–1337, doi:10.1002/
climate model outputs in producing ensemble climate pro- joc.1314.
jections for a variety of scenarios. We have successfully Hirabayashi, Y., S. Kanae, I. Struthers, and T. Oki (2005), A 100‐year
applied the method to several other coupled climate model (1901–2000) global retrospective estimation of the terrestrial water cycle,
projections. Combined with a temporal disaggregation J. Geophys. Res., 110, D19101, doi:10.1029/2004JD005492.
Huntington, T. G. (2006), Evidence for intensification of the global water
approach such as a climate analog approach [e.g., Benestad, cycle: Review and synthesis, J. Hydrol., 319, 83–95, doi:10.1016/
2001; Zorita and von Storch, 1999] or a more traditional j.jhydrol.2005.07.003.

19 of 20
D10101 LI ET AL.: BIAS CORRECTION OF IPCC MODEL OUTPUTS D10101

Ines, A. V. M., and J. W. Hansen (2006), Bias correction of daily GCM Robeson, S. M. (2004), Trends in time‐varying percentiles of daily mini-
rainfall for crop simulation studies, Agric. For. Meteorol., 138(1–4), mum and maximum air temperature over North America, Geophys.
44–53, doi:10.1016/j.agrformet.2006.03.009. Res. Lett., 31, L04203, doi:10.1029/2003GL019019.
Maurer, E. P., and H. G. Hidalgo (2008), Utility of daily vs. monthly large‐ Serreze, M. C., and J. A. Francis (2006), The Arctic amplification debate,
scale climate data: An intercomparison of two statistical downscaling Clim. Change, 76(3–4), 241–264, doi:10.1007/s10584-005-9017-y.
methods, Hydrol. Earth Syst. Sci., 12, 551–563. Serreze, M. C., J. E. Walsh, F. S. Chapin III, T. Osterkamp, M. Dyurgerov,
Meehl, G. A., P. R. Gent, J. M. Arblaster, B. L. Otto‐Bliesner, E. C. Brady, V. Romanovsky, W. C. Oechel, J. Morison, T. Zhang, and R. G. Barry
and A. Craig (2001), Factors that affect amplitude of El Nino in global (2000), Observational evidence of recent change in the northern high‐
coupled climate models, Clim. Dyn., 17, 515–526, doi:10.1007/ latitude environment, Clim. Change, 46(1–2), 159–207, doi:10.1023/
PL00007929. A:1005504031923.
Meehl, G. A., W. M. Washington, C. M. Ammann, J. M. Arblaster, T. M. Serreze, M. C., J. A. Maslanik, T. A. Scambos, F. Fetterer, J. Stroeve,
L. Wigley, and C. Tebaldi (2004), Combinations of natural and anthro- K. Knowles, C. Fowler, S. Drobot, R. G. Barry, and T. M. Haran (2003), A
pogenic forcings and 20th century climate, J. Clim., 17, 3721–3727, new record minimum Arctic sea ice and extent in 2002, Geophys. Res.
doi:10.1175/1520-0442(2004)017<3721:CONAAF>2.0.CO;2. Lett., 30(3), 1110, doi:10.1029/2002GL016406.
Meehl, G. A., et al. (2007), Global climate projections, in Climate Change Solomon, S., G.‐K. Plattnerb, R. Knuttic, and P. Friedlingsteind (2009),
2007: The Physical Science Basis. Contribution of Working Group I to Irreversible climate change due to carbon dioxide emissions, Proc. Natl.
the Fourth Assessment Report of the Intergovernmental Panel on Climate Acad. Sci. U. S. A., 106, 1704–1709, doi:10.1073/pnas.0812721106.
Change, edited by S. Solomon et al., pp. 749–845, Cambridge Univ. Tian, X., A. Dai, D. Yang, and Z. Xie (2007), Effects of precipitation‐bias
Press, New York. corrections on surface hydrology over northern latitudes, J. Geophys.
Milly, P. C. D., R. T. Wetherald, K. A. Dunne, and T. L. Delworth (2002), Res., 112, D14101, doi:10.1029/2007JD008420.
Increasing risk of great floods in a changing climate, Nature, 415, 514– Troy, T. J., E. F. Wood, and J. Sheffield (2008), Quantifying the mean, var-
517, doi:10.1038/415514a. iability, and trends in the water and energy cycles across NEESPI
Milly, P. C. D., J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W. through modeling and observations, Eos Trans. AGU, 89(53), Fall Meet.
Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer (2008), Stationarity is Suppl., Abstract GC53C‐05.
dead: Whither water management?, Science, 319, 573–574, doi:10.1126/ Washington, W. M., et al. (2000), Parallel climate model (PCM) control
science.1151915. and transient simulations, Clim. Dyn., 16, 755–774, doi:10.1007/
Money, D. (2000), Weather and Climate, Nelson Thornes, Cheltenham, s003820000079.
U. K. Watterson, I. G. (2008), Calculation of probability density functions for
Moritz, R. E., C. M. Bitz, and E. J. Steig (2002), Dynamics of recent temperature and precipitation change under global warming, J. Geophys.
climate change in the Arctic, Science, 297, 1497–1502, doi:10.1126/ Res., 113, D12106, doi:10.1029/2007JD009254.
science.1076522. Wilby, R. L., and T. M. L. Wigley (2000), Precipitation predictors for
Nakićenović, N., and R. Swart (Eds.) (2000), Special Report on Emissions downscaling: Observed and general circulation model relationships, Int.
Scenarios: A Special Report of Working Group III of the Intergovern- J. Climatol., 20(6), 641–661, doi:10.1002/(SICI)1097-0088(200005)
mental Panel on Climate Change, 599 pp., Cambridge Univ. Press, 20:6<641::AID-JOC501>3.0.CO;2-1.
New York. Wilby, R. L., S. P. Charles, E. Zorita, B. Timbal, P. Whetton, and L. O.
New, M., D. Lister, M. Hulme, and I. Makin (2002), A high‐resolution data Mearns (2004), Guidelines for use of climate scenarios developed from
set of surface climate over global land areas, Clim. Res., 21, 1–25, statistical downscaling methods, report, 27 pp., Intergov. Panel on Clim.
doi:10.3354/cr021001. Change, Geneva, Switzerland. (Available at http://www.ipcc‐data.org/
Panofsky, H. A., and G. W. Brier (1968), Some Application of Statistics to guidelines/.)
Meteorology, 224 pp., Pa. State Univ., University Park, Pa. Wilby, R. L., J. Troni, Y. Biot, L. Tedd, B. C. Hewitson, D. M. Smith, and
Peterson, B. J., R. M. Holmes, J. W. McClelland, C. J. Vörösmarty, R. B. R. T. Sutton (2009), A review of climate risk information for adaptation
Lammers, A. I. Shiklomanov, I. A. Shiklomanov, and S. Rahmstorf and development planning, Int. J. Climatol., 29(9), 1193–1215,
(2002), Increasing river discharge to the Arctic Ocean, Science, 298, doi:10.1002/joc.1839.
2171–2173, doi:10.1126/science.1077445. Wood, A. W., L. R. Leung, V. Sridhar, and D. P. Lettenmaier (2004), Hydro-
Piani, C., J. O. Haerter, and E. Coppola (2010), Statistical bias correction logic implications of dynamical and statistical approaches to downscaling
for daily precipitation in regional climate models over Europe, Theor. climate model outputs, Clim. Change, 62(1–3), 189–216, doi:10.1023/B:
Appl. Climatol., 99, 187–192, doi:10.1007/s00704-009-0134-9. CLIM.0000013685.99609.9e.
Randall, D. A., et al. (2007), Climate models and their evaluation, in Cli- Xie, P., and P. A. Arkin (1997), Global precipitation: A 17‐year monthly
mate Change 2007: The Physical Science Basis. Contribution of Working analysis based on gauge observations, satellite estimates, and numerical
Group I to the Fourth Assessment Report of the Intergovernmental Panel model outputs, Bull. Am. Meteorol. Soc., 78, 2539–2558, doi:10.1175/
on Climate Change, edited by S. Solomon et al., pp. 591–648, Cambridge 1520-0477(1997)078<2539:GPAYMA>2.0.CO;2.
Univ. Press, New York. Zorita, E., and H. von Storch (1999), The analog method as a simple
Rawlins, M. A., and C. J. Willmott (2003), Winter air‐temperature change statistical downscaling technique: Comparison with more complicated
across the terrestrial Arctic, 1961–1990, Arct. Antarct. Alp. Res., 35, methods, J. Clim., 12, 2474–2489, doi:10.1175/1520-0442(1999)
530–537, doi:10.1657/1523-0430(2003)035[0530:WATCOT]2.0.CO;2. 012<2474:TAMAAS>2.0.CO;2.
Reichler, T., and J. Kim (2008), How well do coupled models simulate
today’s climate?, Bull. Am. Meteorol. Soc., 89, 303–311, doi:10.1175/ H. Li, J. Sheffield, and E. F. Wood, Department of Civil and
BAMS-89-3-303. Environmental Engineering, Princeton University, E‐209A Engineering
Reifen, C., and R. Toumi (2009), Climate projections: Past performance no
Quad, Princeton, NJ 08544, USA. (haibinli@princeton.edu)
guarantee of future skill?, Geophys. Res. Lett., 36, L13704, doi:10.1029/
2009GL038082.

20 of 20

You might also like