You are on page 1of 21

Climate Dynamics (2019) 53:4629–4649

https://doi.org/10.1007/s00382-019-04809-x

An intercomparison of multiple statistical downscaling methods


for daily precipitation and temperature over China: present climate
evaluations
Yi Yang1 · Jianping Tang1   · Zhe Xiong2 · Shuyu Wang1 · Jian Yuan3

Received: 4 February 2018 / Accepted: 8 May 2019 / Published online: 15 May 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
Four statistical downscaling methods, that is, three quantile mapping based techniques including bias-correction and spatial
downscaling (BCSD), bias-correction and climate imprint (BCCI), and bias correction constructed analogues with quantile
mapping reordering (BCCAQ), and the cumulative distribution function transform (CDF-t) method, are evaluated with
daily observed precipitation and surface temperature for 1961–2005 over China. The four downscaling methods improve the
accuracy over the driving general climate models (GCMs) significantly in terms of spatial variability, bias, seasonal cycle,
and probability density functions of daily series and extreme events. Overall, BCSD outperforms other methods in frequency
distributions of daily temperature, precipitation, and extreme precipitation indices such as wet and dry spell lengths. But it
comparably has larger biases in temperature-related extremes. When downscaling the seasonal and extreme precipitation,
the three quantile mapping based techniques exhibit better capacity than CDF-t in terms of the spatial correlation and bias
over all subregions. Whereas CDF-t methods overestimate consecutive wet days and extreme wet indices significantly, as it
displays limited improvement over the driving GCMs by producing too many drizzle days using either absolute or relative
threshold. All methods are equally skillful in downscaling monthly and seasonal temperature, and the temperature extremes
are better reproduced by BCCI, BCCAQ and CDF-t. However, the statistical downscaling methods show limited capacity
in improving the interannual variability of temperature and precipitation extremes.

Keywords  Statistical downscaling · Intercomparison · China · Extreme

1 Introduction to the more frequently occurring disasters, which impose


profound impacts on human safety and local economy.
Because of the vast territory, complex topography, and the Many studies have been conducted to simulate and project
strong influence of East Asian monsoon, China is susceptible climate changes in China based on general climate models
to extreme weather and climate events such as floods and (GCMs) (Jiang et al. 2004; Li and Zhou 2010; Chen et al.
droughts (Huang et al. 2007). Regional climate change under 2011; Chong-Hai and Ying 2012; Zhou et al. 2014). Despite
the influence of global warming is of increasing concern due continuous improvements, GCMs are restricted in their abil-
ity to provide detailed information for regional to local scale
climate impact studies (such as hydrologic modeling and
* Jianping Tang
jptang@nju.edu.cn water resources assessment). To address the scale mismatch
issues, downscaling techniques of dynamical and statistical
1
CMA‑NJU Joint Laboratory for Climate Prediction Studies, downscalings have been developed and widely applied over
Institute for Climate and Global Change Research, School the past 2 decades (Wilby and Wigley 1997). Dynamical
of Atmospheric Sciences, Nanjing University, 163 Xianlin
Road, Nanjing, China downscaling nests a regional climate model (RCM) in the
2 GCM over the domain of interest to provide better repre-
Key Laboratory of Regional Climate‑Environment
for Temperate East Asia, Institute of Atmospheric Physics, sentation of physical processes consistent with large-scale
Chinese Academy of Sciences, Beijing, China dynamics; however, it is computationally demanding and
3
Institute for Climate and Global Change Research, School inherits bias from GCM boundary conditions. On the other
of Atmospheric Sciences, Nanjing University, Nanjing, China hands, statistical downscaling is computationally efficient,

13
Vol.:(0123456789)

4630 Y. Yang et al.

and able to directly incorporate observations for bias correc- is an effective tool to remove biases of models. They then
tion. It can be easily applied to multiple regions and GCMs used this approach to predict future variations of tempera-
to generate large scenario ensembles (Wilby et al. 2004; ture, precipitation and climatological drought under differ-
Fowler et al. 2007), and has been widely used in climate ent emission scenarios (Wang and Chen 2014). In addition,
investigation and climate impact assessments. the intercomparison of different approaches is relatively
Statistical downscaling approaches establish statistical scarce over China (Chu et al. 2010; Hu et al. 2013; Liu et al.
relationships relating the large-scale predictors to regional 2013), and BCSD has not been fully evaluated and exam-
and local predictands (Trigo and Palutikof 1999; Wilby and ined against other statistical downscaling methods. The key
Wigley 2000; Jakob Themeßl et al. 2011; Tripathi et al. elements such as dry spell length, are barely evaluated thor-
2006), and typically assume that the relationships estab- oughly, though they are important in hydrologic and agri-
lished under present climate remain valid under changed cultural applications.
future climate (stationarity assumption). There are many This paper analyzes the application of four commonly
categories of statistical methods. With the growing avail- used statistical downscaling methods, namely, BCSD, BCCI
ability of climate model ensembles such as those from COR- (bias-correction and climate imprint), BCCAQ (bias cor-
DEX (Coordinated Regional Climate Downscaling Experi- rection constructed analogues with quantile mapping reor-
ment, Giorgi et al. 2009), model output statistics (MOS) dering), and CDF-t (cumulative distribution function trans-
that directly applied to RCM or GCM outputs has become form). We assess their capabilities to reproduce the daily
popular (Maraun et al. 2010; Hanson et al. 2012; Hwang surface climatology and extreme events in present climate
and Graham 2013; Stoner et al. 2013; Gutiérrez et al. 2018). over China. The paper is organized as follows. Section 2
Under the MOS approach, global and regional climate model describes the datasets, statistical downscaling methods
simulations are statistically corrected against observations, and the metrics for evaluation. The results are presented in
thus accounting for systematic model biases. Previous stud- Sect. 3. Concluding remarks are given in Sect. 4.
ies show the MOS methods (sometimes called bias correc-
tion methods) are easy to implement with high skill (Sala-
thé 2003; Widmann et al. 2003; Maurer and Hidalgo 2008),
and have been widely applied in regional climate change 2 Data and methods
and adaptation studies especially over the United States and
European countries (Payne et al. 2004; Iizumi et al. 2011; 2.1 Data
Abatzoglou and Brown 2012; Sunyer Pinya et al. 2015).
Among these approaches, bias-correction and spatial downs- The observed daily gridded precipitation, surface mean,
caling (BCSD, Wood et al. 2004) and constructed analogues maximum and minimum temperature data at 0.25° spa-
(CA, Hidalgo et al. 2008) are regularly used, owing to their tial resolution over China for the period of 1961–2005 are
good performances in the hydrologic assessments. The inter- obtained from the National Climate Center of China Mete-
comparison studies of statistical downscaling methods for orological Administration (CN05.1; Wu and Gao 2013) and
climate impact assessments (Maurer et al. 2010; Bürger et al. are used in the calibration and assessment of the downscal-
2013; Pierce et al. 2013; Mizukami et al. 2016) demonstrate ing methods. The observation datasets are based on interpo-
that these methods are effective to simulate both mean cli- lations from 2416 meteorological stations in China and have
mate and extreme events, with relatively better performance been widely used for model assessment (Gao et al. 2013;
for temperature-related indices. Bao et al. 2015).
Meanwhile, much attention has been paid to statistical The historical daily outputs of precipitation, surface
downscaling methods in China (Wetterhall et al. 2006; Chu mean, maximum and minimum temperature data from
et al. 2010; Huang et al. 2011; Liu et al. 2011, 2012; Chen seven CMIP5 models developed by different institutes are
et al. 2012; Tang et al. 2016). However, focusing on the employed for downscaling based on their performance on
monthly to annual time-scales and mostly at the station large-scale simulations of climate extreme indices (Sill-
level, very few studies have been conducted to investigate mann et al. 2013). Table 1 lists the basic information on the
the extremes, especially over the arid and semiarid regions names, institution, and resolution for these models. Since the
of China. The BCSD method, shown to be robust over the focus of this study lies on the evaluation of current climate,
continental US (Wood et al. 2004), has recently been used to only the GCM outputs of historical scenario run are used in
obtain local meteorological variables (e.g., temperature and daily surface climate downscaling. All downscaling methods
evaporation) and to assess climate change impacts in China used to generate daily precipitation and temperature data are
(Fang et al. 2016; Wen et al. 2015, 2016). Wang and Chen calibrated for the 30 years of 1971–2000, and the validation
(2013) evaluated the performance of BCSD for downscal- is conducted for the remaining 15 years (1961–1970 and
ing monthly precipitation in China, and found that BCSD 2001–2005).

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4631

Table 1  Descriptions of the GCM name Institution (country) Resolution (lon-


7 CMIP5 models used in this gitude × latitude)
study
CCSM4 National Center for Atmospheric Research (USA) 288 × 192
CESM1-BGC Community Earth System Model Contributors (USA) 288 × 192
CMCC-CM Centro Euro-Mediterraneo per I Cambiamenti Climatici (Italy) 480 × 240
CNRM-CM5 Centre National de Recherches Météorologiques/Centre Européen 256 × 128
de Recherche et Formation Avancée en Calcul Scientifique
(France)
EC-EARTH EC-Earth consortium (Europe) 320 × 160
MIROC5 Atmosphere and Ocean Research Institute (The University of 256 × 128
Tokyo), National Institute for Environmental Studies, and Japan
Agency for Marine-Earth Science and Technology (Japan)
MRI-CGCM3 Meteorological Research Institute (Japan) 320 × 160

2.2 Statistical downscaling methods means. Daily data can be obtained by randomly resampling a
month from the historical record (1971–2000), and rescaling
All downscaling methods in this study use the same cli- the observed daily values to match the downscaled monthly
mate variables for the coarse- and fine- resolution fields, values. Following Maurer and Hidalgo (2008), this temporal
and belong to the MOS approach. These methods are imple- disaggregation is constrained by a check on precipitation
mented for each month separately. Table 2 gives some infor- scaling factor (no more than three) and number of wet days
mation about the structural properties of the methods. to avoid unrealistic precipitation values.

2.2.1 Bias‑correction and spatial downscaling (BCSD) 2.2.2 Bias‑correction and climate imprint (BCCI)

The BCSD approach developed by Wood et al. (2002, 2004) The BCCI method, described in detail by Hunter and
has been widely used to downscale GCM simulations for Meentemeyer (2005), is a statistical technique using monthly
impact assessments. This two-step method first corrects the observed climatologies at the fine scale to represent envi-
biases in monthly GCM outputs against gridded observations ronmental gradients. Daily anomalies of GCM outputs rela-
aggregated to the coarse model resolution using detrended tive to the coarse-scale model monthly climatology over the
quantile mapping (DQM) with the delta method extrapola- calibration period are calculated and then interpolated to
tion (Werner 2011; Bürger et al. 2013). Second, the bias- the high-resolution observational grid. These interpolated
corrected monthly GCM fields are spatially disaggregated to daily anomalies are multiplied (for precipitation) or added
the fine-scale grid using high-resolution monthly anomalies. (for temperature) to the fine-scale observational climatol-
The anomalies are calculated as the ratio (for precipitation) ogy for each month to generate the daily fine-scale data.
or difference (for temperature) between large-scale bias- Finally, BCCI applies bias-correction as a post-processing
corrected GCM values and coarse-scale observed monthly step by the same quantile mapping as in BCSD but for daily

Table 2  The downscaling methods used in this study

Method TS BCA Characteristic CCDF

BCSD Monthly DQM Perform bias correction at the large scale No


Spatial disaggregation
Temporal disaggregation
BCCI Daily DQM Interpolate the raw model output to fine resolution with the CI delta method No
Perform bias correction at the fine scale
BCCAQ Daily DQM Daily BCCI outputs at each fine-scale grid point are reordered within a given month No
according to the daily BCCA ranks
CDF-t Daily CDF-t Perform a bilinear interpolation of raw model output to fine resolution Yes
CDF transformation between the future large- and local-scale CDFs

All methods are used to generate daily data. TS indicates the time scale the methods applied to. BCA indicates the bias correction approach used
in each method. CCDF indicates whether the methods account for changes in the large-scale CDF in the future. All methods directly downscale
temperature from the large-scale models, except BCSD, which indirectly derives minimum and maximum temperature from mean temperature

13

4632 Y. Yang et al.

series within each month. This method preserves the daily CDF-t2), while the non-zero values are retained for the
sequence of GCM variability but suffers from poor spatial observations. The relative threshold is determined such
covariability due to the smooth interpolation from coarse to that the frequency of large-scale precipitation occurrences
fine resolution (Sobie and Murdock 2017). matches the observed local-scale one.

2.2.3 Bias correction constructed analogues with quantile


mapping reordering (BCCAQ) 2.3 Evaluation metrics

BCCAQ (Werner and Cannon 2016) is a hybrid downscaling The assessment of the statistical downscaling methods
method incorporating results from BCCI and bias correc- is performed against the observations over the period of
tion constructed analogs (BCCA; Maurer et al. 2010) by 1961–2005, setting the validation period of 1961–1970
reordering the daily BCCI outputs at each grid cell within a and 2001–2005. The results over the calibration period of
given month based on the BCCA ranks. BCCA uses similar 1971–2000 are briefly discussed as well. The skill of sta-
quantile mapping step as BCSD to bias correct the daily tistical methods in downscaling the monthly temperature
coarse-scale GCM outputs, and obtains spatial informa- and precipitation from GCMs is examined by presenting
tion from a linear combination of historical coarse-reso- the root-mean-square error (RMSE), mean absolute error
lution observations that best match the large-scale GCM (MAE), and correlation coefficient (CORR) at each grid cell
temperature/precipitation pattern on a given day (Hidalgo over China. Besides the grid-to-grid evaluation, six subre-
et al. 2008). BCCA has the advantage of capturing the topo- gions are selected to better understand the performance of
graphic effects that are not depicted by interpolation based downscaling methods on regional scales (Fig. 1). Taylor
methods, however, it tends to produce too much low-level diagram (Taylor 2001) for both boreal summer (June, July
precipitation (drizzle) and underestimate the large precipi- and August) and winter (December, January and February)
tation events when downscaling precipitation (Abatzoglou is produced for seasonal skill assessment.
and Brown 2012; Sobie and Murdock 2017). The BCCAQ On a daily basis, the degree of fit on the distributions
method, which combines BCCI and BCCA, retains the of precipitation and temperature is qualitatively examined
advantages of each method and produces more accurate in different subregions using quantile–quantile (Q–Q) plot
downscaled values than the component methods (Werner (Wilk and Gnanadesikan 1968), a graphical technique for
and Cannon 2016). comparing the distribution of downscaled data against that
of the observations at each quantile. In addition, raw GCM
2.2.4 Cumulative distribution function transform (CDF‑t) outputs bilinearly interpolated to the observational grid
are plotted as a reference to evaluate the value added from
The CDF-t method, originally developed by Michelangeli downscaling. Nine indices recommended by the ETCCDI
et al. (2009) for wind downscaling and recently applied (Expert Team on Climate Change Detection and Indices) are
to downscale temperature and precipitation (Vigaud et al. adopted to represent precipitation and temperature extremes
2013), is similar to quantile mapping but takes into account in this study (Sillmann et al. 2013). These indices derived
the change in the large-scale CDF (cumulative distribution from downscaled daily temperature and precipitation data
function) from the calibration to the future (or validation) are calculated per grid cell for each year by the R software
period. This method is applied to the GCMs bilinearly inter- package “climdex.pcic”. Definitions of these indices are
polated to the fine-scale grid. A mathematical transformation summarized in Table 3 and more details are available at the
is determined to transform the CDF of large-scale predic- ETCCDI website (http://etccd​i.pacif​i ccli​mate.org/list_27_
tors (i.e., GCM outputs) as close as possible to the CDF indic​es.shtml​). The significance score (­ Sscore) (Perkins et al.
representing local-scale predictands (i.e., observations) at 2007) is used to evaluate the consistency between the down-
each grid point during the calibration period, and the same scaled and observed probability density functions (PDFs)
transformation is applied to the model projections. of extreme indices. The ideal value is 1 for S ­ score, indicating
Although CDF-t is directly applied to temperature, two an unbiased simulation where the PDF generated by down-
different processings following Lavaysse et al. (2012) are scaled data is exactly the same as the observed conditions.
used for precipitation as it is highly discontinuous. Both
m
absolute and relative thresholds are applied to define rainy ∑
Sscore = Minimum (Pdi , Poi ),
days. In the first processing (hereafter CDF-t1), only the i=1
non-zero precipitation data in both model and observations
are supplied so as to downscale the CDF of strictly posi- where m is the number of bins, and Pdi and Poi are the
tive rainfall. A relative threshold for rainy days is used for downscaled and observed probability values at certain bin,
GCM at each grid point in the second processing (henceforth respectively.

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4633

Fig. 1  Spatial distributions of root-mean-square error (RMSE), mean in a illustrate the six subregions in China (NEC Northeast China, NC
absolute error (MAE), and correlation coefficient (CORR) between North China, YR Yangtze River basin, SC South China, NWC North-
the downscaled monthly precipitation (BCSD, BCCI, BCCAQ, CDF- west China, SWC Southwest China)
t1, and CDF-t2) and observations during the validation period. Boxes

It should be noted that outputs from each of the seven ensemble is built for each method through the arithmetic
GCMs are downscaled by the four downscaling methods mean in this study. Preliminary analyses with individual
in both the calibration and validation periods. Because models did not show significant differences in the results
this paper focuses on the evaluation and comparison of presented later (not shown).
downscaling methods in the present climate, multimodel

13

4634 Y. Yang et al.

Table 3  Definitions of the climate extreme indices


Index Definitions Units

Precipitation indices CDD Consecutive dry days: maximum number of consecutive days with RR (daily precipitation) < 1 mm days
CWD Consecutive wet days: maximum number of consecutive days with RR ≥ 1 mm days
R10mm Number of heavy precipitation days: annual number of days when RR ≥ 10 mm days
Rx5day Maximum 5-day precipitation amount: annual maximum consecutive 5-day precipitation mm
R95p Very wet day precipitation: annual total precipitation when RR > 95th percentile of precipitation mm
Temperature indices GSL Growing season length: annual number of days between the first occurrence of 6 consecutive days with days
mean temperature > 5 °C and the first occurrence of 6 consecutive days with mean temperature < 5 °C
SU Summer days: number of days with daily maximum temperature above 25 °C days
FD Frost days: number of days with daily minimum temperature below 0 °C days
DTR Diurnal temperature range: annual mean difference between daily maximum and minimum temperature °C

3 Results the East Asian summer monsoon has been partly corrected
by the downscaling approaches (Fig. 2c). Over South China,
3.1 Precipitation all downscaling methods can reproduce the bimodal curve
of precipitation with peaks in June and August, and reduce
3.1.1 Monthly and seasonal skill the dry bias of models in the highest peak (Fig. 2d). The
CDF-t methods consistently overestimate the amplitude of
Figure 1 shows the spatial distributions of RMSE, MAE and the seasonal cycle, even larger than that of the GCMs in
correlation of downscaled monthly averaged precipitation some cases. This overestimation of precipitation intensity by
against the observations for the validation period over China. CDF-t has also been reported by Gutiérrez et al. (2018) for
The RMSE and MAE for all downscaling methods show a Europe. Generally, these downscaling methods contribute to
similar decreasing pattern from southeast to northwest of the satisfactory precision in reproducing seasonal cycle of
China. Compared with the other three approaches, the two precipitation over different subregions of China.
CDF-t methods generally show larger RMSE values over the As is well reflected in the Taylor diagram (Fig. 3), for
entire China. The bias of downscaled precipitation quantified both calibration and validation periods, downscaling meth-
by the MAE is higher over southeastern China than over the ods are more consistent with the observed precipitation
arid and semi-arid areas of Northwest China, with a rela- pattern than the driving models. The three quantile map-
tively small bias of about 1.8 mm/day versus 0.2 mm/day ping based methods reproduce satisfactorily the summer
for the methods using quantile mapping for bias correction and winter spatial distributions, and are located closer to
(BCSD, BCCI and BCCAQ). Both CDF-t methods yield a the observations [represented by the reference point (REF)]
much larger MAE value, which is notable mainly along the than CDF-t, with lower RMSE values and spatial correla-
southeast coast and over the Yangtze River basin with MAE tions above 0.95 in most cases. The downscaled standard
up to 2.6 mm/day. All downscaling methods are generally deviation is close to observations. The results are very simi-
comparable and present correlation coefficients above 0.6 lar whatever the period considered, with slightly better per-
over almost the whole domain. formance of downscaling, as expected, during calibration
The seasonal cycles of precipitation over six subregions period. Comparing statistics calculated by the two CDF-t
and China based on various downscaled data are demon- processings, they perform similarly with the CDF-t1 hav-
strated in Fig. 2. For China as a whole, BCSD, BCCI and ing slightly higher correlation and lower RMSE, implying
BCCAQ can perfectly portray the unimodal characteristics better performance of absolute threshold processing applied
of seasonal cycle of the domain-averaged precipitation, in CDF-t1 to produce realistic precipitation patterns. How-
whereas both CDF-t methods tend to overestimate precipi- ever, both CDF-t methods tend to overestimate the spatial
tation amounts by 0.1–0.8 mm/day for all months, which variability of seasonal precipitation, particularly in winter,
is consistent with the driving GCMs, although these two although they clearly improve the spatial correlation of pre-
methods can reproduce the peak that occurs in July (Fig. 2g). cipitation distributions compared to the raw GCMs.
Similar conclusions could be drawn at different subregions.
Over Yangtze River basin, the GCMs tend to show wet 3.1.2 Daily skill
biases from January to May and dry biases from July to
October, indicating an early onset of monsoon compared to The daily precipitation amounts of the raw and downscaled
observations, and this bias in the timing and magnitude of GCM data are compared to observations for different

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4635

Fig. 2  Seasonal cycles of
precipitation over different
subregions and China during
validation period. The shaded
area indicates the standard
deviation of observation on the
basis of data for 15 years

Fig. 3  Taylor diagram of sea-


sonal mean precipitation for the
driving GCMs and downscaling
methods over China. The radial
distance represents the ratio of
simulated and observed stand-
ard deviation and the angle in
the polar plot stands for spatial
correlation

13

4636 Y. Yang et al.

subregions by Q–Q plot as shown in Fig. 4. Generally, produce similar distributions of daily precipitation, and
all downscaling methods improve the frequency dis- show better performances for rain rates at high intensities
tributions over the raw climate models by reducing the than BCCI and BCCAQ. However, both methods have too
underestimation of upper quantiles of daily precipitation, many drizzle days and show no clear improvement over
which is noticeable over areas such as South China and the driving fields, which could be partially responsible
Yangtze River basin, and correcting the drizzle effect to for the wet biases at monthly and longer time scales. Both
some extent over China. Directly rescaling monthly coarse BCCI and BCCAQ well reproduce the lower tail of the dis-
model precipitation, the BCSD significantly outperforms tribution but producing fewer high rainfall events against
other methods at all percentiles and exhibits good agree- observed heavy precipitation, with a slightly superior per-
ment with observations, except for high extreme values, formance for BCCAQ, partly owing to the combination of
which is consistent with the conclusions from Bürger BCCA and BCCI.
et al. (2012) and Eum et al. (2017). Two CDF-t methods

Fig. 4  The Q–Q plots for the spatially averaged daily precipitation of original and downscaled GCMs (y axis) with respect to observations (x
axis) during validation period

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4637

3.1.3 Extreme precipitation both CDF-t can reasonably reproduce large centers of CDD
over northwestern regions, they, consistent with the driving
Five precipitation extreme indices (i.e., CDD, CWD, GCMs, fail to capture the main spatial characteristics over
R10mm, Rx5day, and R95p) based on daily precipitation South China with low spatial correlations. For the wet spell
data are used to assess the ability of downscaling meth- length represented by CWD, BCSD is relatively unbiased
ods for simulating the intensity, frequency and duration of and presents spatial correlations above 0.8. All the other
precipitation events. Figure 5 displays the spatial distribu- methods, especially two CDF-t processings, are positively
tions of annual mean extreme precipitation indices over biased over each subregion, indicating that these methods
China. Spatial statistics are reported in Fig. 6, where the produce little rainfall amount too often, which can also be
downscaled values are compared to those of observations. found in the results of Q–Q plot (Fig. 4). Although both
Overall, all downscaling methods present a spatial pattern CDF-t approaches replicate the observed large CWD cent-
close to what is observed for all extreme indices, with high ers over the Sichuan Basin (28°N–32°N, 100°E–105°E) and
correlations in most cases. For CDD, methods employing the southern part of the Tibetan Plateau, they are biased
quantile mapping perform well. They show high correla- high (84–93%) over the southeastern domain and, to a lesser
tions and biases close to zero over all subregions. The two extent, in Southwest China (8–24%). The shorter dry spell
CDF-t approaches show relatively high underestimation length along with longer wet spell length for both CDF-t
over entire China, in particular over Northwest China with methods with respect to observations confirms the limited
bias up to − 28% for CDF-t1 and − 24% for CDF-t2, indicat- improvement of CDF transform technique in correcting the
ing shorter dry spell lengths for these methods. Although drizzle effect over the driving GCMs.

Fig. 5  Spatial distributions of the extreme precipitation indices produced by the observations and downscaling methods during validation period.
From left to right column, indices of CDD, CWD, R10mm, Rx5day, and R95p are shown, respectively

13

4638 Y. Yang et al.

Fig. 6  Spatial correlations and relative errors (units: %) of downscaled extreme climate indices at each subregion. Numbers inside the shaded
boxes indicate the corresponding values

For the indices of R10mm, Rx5day, and R95p, they show with previous studies in Europe (Gutiérrez et al. 2018; Her-
a similar spatial distribution as the seasonal precipitation in tig et al. 2018). Overall, BCSD is closest to observed spa-
both observations and downscaled GCMs, with highest val- tial pattern of extreme precipitation indices except for R95p
ues over southeastern China and decreasing northwestward. among all downscaling methods in terms of correlation and
BCSD, BCCI, and BCCAQ generally generate more reliable RE, possibly due to the resampling of historical records and
spatial distributions of these extreme precipitation indices thus obtaining realistic pattern of spatial variability and the
in China and six subregions with higher spatial correlations best performance in the distribution of daily sequences.
and closer magnitude, although slight overestimation exists The observed and downscaled PDFs for extreme precipi-
over most areas (Fig. 6). The two CDF-t approaches present tation indices are illustrated in Fig. 7. Substantial improve-
the highest degree of overestimation among all methods in ment for downscaled data compared to large-scale outputs
both periods, especially for R95p with RE above 68.4% over is detected for most indices, especially for wet and dry
Northwest China, which is in agreement with their overesti- spell lengths, by increasing the S ­ score of CWD from 0.79
mation of monthly and seasonal precipitation and consistent to exceeding 0.9 for the quantile mapping based methods.

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4639

Fig. 7  PDFs of extreme precipitation indices generated from observations, driving GCMs, and downscaling methods over China in the valida-
tion period. The PDF-based statistic ­Sscore with a perfect value of 1 is calculated for each dataset

In general, all downscaling methods present reasonable R10mm and Rx5day, suggesting overestimation for extreme
PDFs of extreme precipitation indices compared to those precipitation events.
derived from observations, although there exists a slight
horizontal shift in the PDF spectrums and the magnitudes 3.2 Temperature
of peak values are slightly biased. The PDFs generated by
BCSD, BCCI and BCCAQ are typically quite tightly clus- 3.2.1 Monthly and seasonal skill
tered around the observed ones for the five indices, with
skill scores above 0.8, whereas both CDF-t methods yield Seasonal cycles of surface temperature during the valida-
relatively lower S­ score for all indices and overestimate the tion period are shown in Fig. 8. All downscaling meth-
probability of longer wet spell length and upper tails of ods can accurately reproduce the seasonal cycle in the

13

4640 Y. Yang et al.

Fig. 8  The same as Fig. 2, but


for surface temperature

domain-averaged surface temperature at each subregion Xinjiang province, and a slight cold bias (− 0.1 °C) over
with dry or more humid conditions, and present an obvi- Qinghai province, Sichuan Basin, and the south edge of
ous improvement over the driving GCMs with a decreasing China. The interannual variations in RMSE and spatial
RMSE from 1.36 to 0.24 °C over China as a whole. correlation are more or less the same for all downscaling
Figure 9 shows the biases of downscaled seasonal sur- methods, with high spatial correlation coefficients (greater
face temperature over China in the validation period. The than 0.98) and relatively small RMSE values (about 0.5 °C
associated metrics (bias, RMSE, and spatial correlations) in JJA and 1.1 °C in DJF), indicating a good performance
are shown in Fig. 10. All four downscaling methods pro- of statistical downscaling methods in reproducing the vari-
duce similar results for both JJA and DJF. In summer, all ability of seasonal mean temperature. The relatively high
methods present a slight cold bias around − 0.3 °C over RMSE values in the years such as 1967 and 1976 may be
most areas, whereas for winter season they show a marked related to the extreme cold winters in these years (Zheng
warm bias over Northeast China (regional bias of approxi- et  al. 2012). Overall, there is little difference between
mately 1 °C), Inner Mongolia, and the northern part of methods for monthly and seasonal temperature.

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4641

Fig. 9  Biases (°C) in seasonal


surface air temperature by four
downscaling methods of BCSD,
BCCI, BCCAQ, and CDF-t for
the validation period

3.2.2 Daily skill Systematic warm biases are generated at the lower quantiles
of surface temperature by BCCI, BCCAQ and CDF-t, with
To analyze the fit of the downscaled distributions to the little discrepancy among the different results. Similar con-
observed ones, the Q–Q plot for daily surface mean tem- clusions can be drawn for the daily maximum and minimum
perature at different subregions is shown in Fig. 11. Gener- temperatures (not shown).
ally, GCMs tend to overestimate the lower end of extreme
values and underestimate the upper tails over North China 3.2.3 Extreme temperature
and Yangtze River basin, whereas underestimation at all
quantiles is detected over Southwest China. The downscal- Figure 12 displays spatial patterns of extreme temperature
ing methods vastly improve the frequency distribution by indices during the validation period for observations, driv-
reducing the biases of GCM simulated percentiles over the ing GCMs and the four statistical downscaling methods. In
whole China, especially over Southwest China. Overall, all general, all downscaling methods well capture the spatial
downscaling methods can well replicate the statistical distri- characteristics of observed extreme temperature indices,
bution of the observed daily temperature and the extremes. with correlation coefficients up to 0.99. The improvement

13

4642 Y. Yang et al.

Fig. 10  The first column dem-


onstrates the bar plots of bias
(°C) for downscaled seasonal
mean temperature at each subre-
gion in a summer and b winter
during the validation period.
The second column is the inter-
annual variation in RMSE (°C)
(scatters) and spatial correla-
tion coefficients (line) between
downscaled c summer and d
winter temperature and observa-
tions over China

over GCMs is substantial, as four methods display more Figure 14 presents the time series of extreme indices
skills in reproducing both the spatial distributions and mag- spatially averaged over China. As expected, downscaling
nitudes of extreme events. The growing season length (GSL) methods are (obviously) more reliable in reproducing the
is slightly underestimated by all methods over most areas, magnitudes and ranges of the interannual variability of
whereas BCSD presents relatively high overestimation of extreme temperature indices compared to GCMs, as down-
about 15 days along Yangtze River. The biases range from scaling approach presents more reasonable regional aver-
− 12 to 4 days for summer days (SU) with distinctive spatial ages of extreme indices. However, downscaling certainly
characteristic, that is, underestimation in most of eastern can not reproduce the observed year-to-year variances of
and northern China, while overestimation of about 4 days the extremes, as they are strongly influenced by the skill
over Northeast China. For frost days (FD), the magnitudes of driving GCMs and therefore because the bias-correction
of differences between the downscaled and observed values methods using CDF mapping do not explicitly change the
are really small (within ± 4 days) for all methods, except temporal pattern or timing of temperature events (Hwang
for BCSD, which produces large underestimation up to and Graham 2013). The standard deviation of annual mean
− 10 days over the northern part of Yangtze River basin, values is adopted as a measure of interannual variability
consistent with the warm bias in the lower tails of daily of extreme indices. Compared to observations, the GCMs
minimum temperature. For DTR, all methods present slight tend to underestimate interannual variability of temperature
underestimation with a regional bias of around − 0.55 °C for extremes, especially for FD and DTR (Fig. 14, right panel).
BCSD and − 0.27 °C for the other three methods. This performance is generally unchanged by the downscal-
The PDFs for extreme temperature indices during the ing methods. An exception is the BCSD for DTR (Fig. 14h),
validation period are shown in Fig. 13. It is evident that which shows a much smaller spread across models than the
GCMs display shift toward either the lower or higher ends raw GCMs and other methods, and the spatial averages
of spectrums. Such biases are generally corrected by the are broadly uniform over time (Fig. 14d). Note that BCSD
application of statistical downscaling methods. Overall, indirectly derives minimum and maximum temperature
all downscaling methods reproduce the PDF of observed from mean temperature by assuming a fixed climatological
extreme temperature indices, and generate a skill score in diurnal temperature range (Bürger et al. 2012; Werner and
excess of 0.76 for GSL and 0.9 for SU, FD and DTR. The Cannon 2016), thus this method is not suitable to model
spectrum of diurnal temperature range is shifted towards low the possible changes in DTR. Similar results are found for
values by all methods especially BCSD, implying smaller precipitation extremes: downscaling methods add substantial
range of diurnal temperature. Comparably, BCCI, BCCAQ value by simulating more reliable regional averages, even
and CDF-t perform slightly better than BCSD for DTR as for the spell-length-related extremes, while generally retain-
far as the commonly used ­Sscore is concerned. ing the interannual correlation of the driving models (not

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4643

Fig. 11  The same as Fig. 4, but for the daily surface temperature

shown). However, in contrast to temperature, downscaling and extreme events. The comparison is conducted for daily,
methods especially BCSD can modify the interannual vari- monthly and seasonal data across spatial scales ranging from
ability of precipitation extremes, in particular for the dry and six subregions to the whole domain. The historical record
wet spells. Such improvement of the MOS methods in tem- is divided into two parts, a calibration period of 1971–2000
poral variability may be indirectly achieved via modification and the remaining period for independent validation.
of the marginal distribution (Gutmann et al. 2014; Maraun The quantile mapping based methods (i.e., BCSD, BCCI
et al. 2017). and BCCAQ) can accurately reproduce the spatial patterns
and magnitudes of seasonal precipitation and exhibit compa-
rable skills at all subregions with different climate character-
4 Summary and conclusions istics. However, the two CDF-t processings tend to overesti-
mate the observed precipitation regardless of the season over
In this study, four widely used statistical downscaling meth- the whole domain. Considering spatial structure, BCSD,
ods, BCSD, BCCI, BCCAQ and CDF-t, are evaluated via a BCCI and BCCAQ present excellent capacity to downscale
comparison with gridded observations over China for the precipitation from large-scale fields over China, whereas
present climate (1961–2005), as applied to seven GCMs the CDF-t methods present poor performance as compared
from CMIP5. The skill of each downscaling method is to other methods with relatively low spatial correlations.
evaluated based on the ensemble mean of the seven down- This is expected since the CDF-t approach applied at each
scaled models in terms of climatology, temporal variability individual grid point is not designed to improve the spatial

13

4644 Y. Yang et al.

Fig. 12  The distributions of extreme temperature indices from observations (the top panel) for the validation period and bias for the raw and
downscaled GCMs (from second to bottom panel)

variability of initial driving fields (Vrac et al. 2012; Flaounas and little difference is found among the four downscaling
et al. 2013). In terms of seasonal cycles of precipitation, all methods.
downscaling methods present substantial improvements over At the daily time scale, all downscaling methods gener-
the driving GCMs in China, especially over Yangtze River ally display obvious advantages over the driving GCMs in
basin, where the biases in both magnitude and the begin- simulating the frequency distribution of daily precipitation.
ning of monsoon system are generally corrected, though the The BCSD method shows the highest skill in producing the
two CDF-t methods overestimate precipitation throughout distribution of observed daily sequences and extremes, and
the year. As expected, performance is better for temperature effectively corrects the drizzle effect of GCMs. This may be
than precipitation at both monthly and seasonal time scales, attributed to its random sampling of historical months when

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4645

Fig. 13  The same as Fig. 7, but for the extreme temperature indices

temporally disaggregating to a daily time scale, however, the simulating large-scale temperature than precipitation, which
temporal disaggregation in BCSD may lead to discontinu- makes the correction by CDF transform easier. The Q–Q plot
ity between months. In contrast, although the two CDF-t reveals that, BCSD demonstrates a superior ability in repro-
approaches produce some high rainfall events matching ducing the frequency distribution of surface mean tempera-
observed heavy precipitation, they still overestimate the ture, whereas the other methods tend to slightly overestimate
probability of light rainfall as observed in the driving GCMs, the lower tails of the quantiles, indicating a warm bias in the
which may be responsible for the shorter dry spell length extreme cold temperature. All methods can accurately cap-
and longer wet spell length generated by these methods. All ture the main spatial characteristics of extreme temperature
downscaling methods can capture well the spatial features indices and reduce biases in the driving GCMs. Although
of precipitation extremes and exhibit high spatial correla- BCSD is generally unbiased for precipitation extremes, the
tions with observations, although the bias of extreme indices bias in BCSD-downscaled temperature extremes is slightly
focusing on frequency and intensity (R10mm, Rx5day, and higher than the other three methods, which may be due to the
R95p) is relatively larger compared to the indices related to fact that minimum and maximum temperature are indirectly
persistence (CDD and CWD). BCSD, BCCI and BCCAQ downscaled, and the effect of this implementation is most
outperform CDF-t for all extreme precipitation indices with clear for DTR. The overall shape of the observed PDFs for
smaller bias and higher correlations. And the two CDF-t temperature extremes is well replicated, although the spec-
processings for precipitation do not appear to differ greatly trum of DTR is shifted towards lower ends by all methods
based on all performance metrics. especially BCSD, which is in consensus with the relatively
The skill in downscaling daily temperature is better as high underestimation of diurnal temperature range for
compared to that for precipitation, and the discrepancy BCSD, possibly due to the assumption of a uniform DTR in
among the four downscaling methods is not evident. The this method (Bürger et al. 2012). Regarding interannual vari-
satisfactory performance of CDF-t to downscale temperature ability, consistent with previous studies (Hwang and Graham
may be partially attributed to the better skill of GCMs in 2013; Flaounas et al. 2013; Maraun et al. 2017), none of

13

4646 Y. Yang et al.

Fig. 14  a–d Time series of temperature indices from 1961 to 2005 calculated over 1961–2005. For each method, boxes denote the 25th
of the observations (black), GCM results (dashed) and downscal- and 75th percentiles of model spread with the horizontal line indicat-
ing methods (BCSD: blue, BCCI: red, BCCAQ: green, and CDF-t: ing the median and whiskers representing the maximum values within
orange). e–h Bias of the individual methods in interannual variability 1.5 times the interquartile range, values outside that range are plotted
represented as the ratio of simulated and observed standard deviation individually

these methods brings a gain over the driving GCMs in terms precipitation indices. For temperature, interannual standard
of temporal correlation, although the improvement in terms deviation is generally inherited from the driving models, but
of magnitude is clear for both extreme temperature and for precipitation, downscaling especially BCSD can improve

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4647

the interannual variances of dry and wet spells, which are Bürger G, Sobie SR, Cannon AJ, Werner AT, Murdock TQ (2013)
generally underestimated and overestimated by the GCMs, Downscaling extremes: an intercomparison of multiple methods
for future climate. J Clim 26(10):3429–3449
respectively. This can be partly explained by a side effect Chen W, Jiang Z, Li L (2011) Probabilistic projections of climate
of adjusting the marginal distribution by the downscaling. change over China under the SRES A1B scenario using 28
This study investigates the performance of statistical AOGCMs. J Clim 24(17):4741–4756
downscaling methods that directly use GCM daily precipi- Chen H, Xu CY, Guo S (2012) Comparison and evaluation of mul-
tiple GCMs, statistical downscaling and hydrological models
tation and temperature as predictors for the corresponding in the study of climate change impacts on runoff. J Hydrol
fine-scale fields, and provides water resource managers 434:36–45
with information on the fidelity of different downscaling Chong-Hai XU, Ying X (2012) The projection of temperature and pre-
methods over China. The methods used in this study can cipitation over China under RCP scenarios using a CMIP5 multi-
model ensemble. Atmos Ocean Sci Lett 5(6):527–533
also be applied to RCM outputs. All downscaling methods Chu JT, Xia J, Xu CY, Singh VP (2010) Statistical downscaling of daily
exhibit good stabilities and generate results with little skill mean temperature, pan evaporation and precipitation for climate
distinction between the calibration and validation periods. change scenarios in Haihe River, China. Theor Appl Climatol
However, a common limitation of statistical downscaling 99(1–2):149–161
Dixon KW, Lanzante JR, Nath MJ, Hayhoe K, Stoner A, Radhakrishnan
is the stationarity assumption (Fowler et al. 2007), which A et al (2016) Evaluating the stationarity assumption in statisti-
is difficult to verify, and only a few studies have focused on cally downscaled climate projections: is past performance an indi-
this problem (Vrac et al. 2007; Hertig and Jacobeit, 2013; cator of future results? Clim Change 135(3–4):395–408
Dixon et al. 2016). The CDF-t method seems to be very Eum HI, Cannon AJ, Murdock TQ (2017) Intercomparison of multiple
statistical downscaling methods: multi-criteria model selection
promising due to its consideration of change in large-scale for South Korea. Stoch Environ Res Risk Assess 31(3):683–703
fields, in spite of its wet bias for precipitation, which has Fang GH, Qi HS, Wen X, Zhou L (2016) Analysis of spatiotempo-
also been indicated in other studies (Gutiérrez et al. 2018; ral evolution of extreme monthly precipitation in the nine major
Hertig et al. 2018). Although the good skill of downscal- basins of China in 21st century under climate change (in Chinese).
J Nat Disasters 25(2):15–25
ing methods in reproducing accurate present climate and Flaounas E, Drobinski P, Vrac M, Bastin S, Lebeaupin-Brossier C,
extreme events do not guarantee a good representation of Stéfanon M et al (2013) Precipitation and temperature space-time
altered climate conditions, a natural follow-up of this study variability and extremes in the Mediterranean region: evaluation
is to apply downscaling methods to future scenarios for cli- of dynamical and statistical downscaling methods. Clim Dyn
40(11–12):2687–2705
mate change impact studies and perform consistent future Fowler HJ, Blenkinsop S, Tebaldi C (2007) Linking climate change
intercomparison studies over China. modelling to impacts studies: recent advances in downscal-
ing techniques for hydrological modelling. Int J Climatol
Acknowledgements  The work is jointly funded by the National Key 27(12):1547–1578
Research and Development Program of China (2018YFA0606003 and Gao XJ, Wang ML, Giorgi F (2013) Climate change over China in the
2016YFC0202000: Task 2) and the National Natural Science Founda- 21st century as simulated by BCC_CSM1.1-RegCM4.0. Atmos
tion of China (41875124, 91425304, 41575099 and 41275004). This Ocean Sci Lett 6:381–386
work is also supported by the Chinese Jiangsu Collaborative Innovation Giorgi F, Jones C, Asrar GR (2009) Addressing climate information
Center for Climate Change. The authors would like to thank National needs at the regional level: the CORDEX framework. Bull World
Climate Center of China Meteorological Administration for the pro- Meteorol Organ 58(3):175–183
vision of high-resolution gridded observations (CN05.1), the World Gutiérrez JM, Maraun D, Widmann M, Huth R, Hertig E, Benestad R
Climate Research Programme (https:​ //esgf-node.llnl.gov/search​ /cmip5​ et al (2018) An intercomparison of a large ensemble of statisti-
/) for providing the CMIP5 data, and Michelangeli et al. (2009) for cal downscaling methods over Europe: results from the VALUE
developing and making available the R-package “CDFt”. The statistical perfect predictor cross-validation experiment. Int J Climatol. https​
downscaling in this paper has been performed on the computing facili- ://doi.org/10.1002/joc.5462
ties in the High Performance Computing Center (HPCC) of Nanjing Gutmann E, Pruitt T, Clark MP, Brekke L, Arnold JR, Raff DA, Ras-
University. mussen RM (2014) An intercomparison of statistical downscaling
methods used for water resource assessments in the United States.
Water Resour Res 50(9):7167–7186
Hanson RT, Flint LE, Flint AL, Dettinger MD, Faunt CC, Cayan D,
Schmid W (2012) A method for physically based model analysis
References of conjunctive use in response to potential climate changes. Water
Resour Res 48(6):W00L08
Abatzoglou JT, Brown TJ (2012) A comparison of statistical down- Hertig E, Jacobeit J (2013) A novel approach to statistical downscaling
scaling methods suited for wildfire applications. Int J Climatol considering nonstationarities: application to daily precipitation in
32(5):772–780 the Mediterranean area. J Geophys Res Atmos 118(2):520–533
Bao J, Feng J, Wang Y (2015) Dynamical downscaling simulation Hertig E, Maraun D, Bartholy J, Pongracz R, Vrac M, Mares I et al
and future projection of precipitation over China. J Geophys Res (2018) Comparison of statistical downscaling methods with
Atmos 120(16):8227–8243 respect to extreme events over Europe: Validation results from
Bürger G, Murdock TQ, Werner AT, Sobie SR, Cannon AJ (2012) the perfect predictor experiment of the COST Action VALUE.
Downscaling extremes—an intercomparison of multiple statistical Int J Climatol. https​://doi.org/10.1002/joc.5469
methods for present climate. J Clim 25(12):4366–4388 Hidalgo HG, Dettinger MD, Cayan DR (2008) Downscaling with con-
structed analogues: Daily precipitation and temperature fields over

13

4648 Y. Yang et al.

the United States. California Energy Commission PIER Final Pro- choices for hydrologic portrayals of climate change over the
ject Report CEC-500-2007-123 contiguous United States: statistically downscaled forcing data
Hu Y, Maskey S, Uhlenbrook S (2013) Downscaling daily precipita- and hydrologic models. J Hydrometeorol 17(1):73–98
tion over the Yellow River source region in China: a comparison Payne JT, Wood AW, Hamlet AF, Palmer RN, Lettenmaier DP (2004)
of three statistical downscaling methods. Theor Appl Climatol Mitigating the effects of climate change on the water resources
112(3–4):447–460 of the Columbia River basin. Clim Change 62(1):233–256
Huang R, Chen J, Huang G (2007) Characteristics and variations of the Perkins SE, Pitman AJ, Holbrook NJ, McAneney J (2007) Evaluation
East Asian monsoon system and its impacts on climate disasters of the AR4 climate models’ simulated daily maximum tempera-
in China. Adv Atmos Sci 24(6):993–1023 ture, minimum temperature, and precipitation over Australia
Huang J, Zhang J, Zhang Z, Xu C, Wang B, Yao J (2011) Estimation using probability density functions. J Clim 20:4356–4376
of future precipitation change in the Yangtze River basin by using Pierce DW, Cayan DR, Das T, Maurer EP, Miller NL, Bao Y et al
statistical downscaling method. Stoch Environ Res Risk Assess (2013) The key role of heavy precipitation events in climate
25(6):781–792 model disagreements of future annual precipitation changes in
Hunter RD, Meentemeyer RK (2005) Climatologically aided map- California. J Clim 26(16):5879–5896
ping of daily precipitation and temperature. J Appl Meteorol Salathé EP (2003) Comparison of various precipitation downscaling
44(10):1501–1510 methods for the simulation of streamflow in a rainshadow river
Hwang S, Graham WD (2013) Development and comparative evalua- basin. Int J Climatol 23(8):887–901
tion of a stochastic analog method to downscale daily GCM pre- Sillmann J, Kharin VV, Zhang X, Zwiers FW, Bronaugh D (2013)
cipitation. Hydrol Earth Syst Sci 17(11):4481–4502 Climate extremes indices in the CMIP5 multimodel ensemble:
Iizumi T, Nishimori M, Dairaku K, Adachi SA, Yokozawa M (2011) part 1. Model evaluation in the present climate. J Geophys Res
Evaluation and intercomparison of downscaled daily precipitation Atmos 118(4):1716–1733
indices over Japan in present-day climate: strengths and weak- Sobie SR, Murdock TQ (2017) High-resolution statistical downscal-
nesses of dynamical and bias correction-type statistical downscal- ing in southwestern British Columbia. J Appl Meteorol Clima-
ing methods. J Geophys Res Atmos 116:D1 tol. https​://doi.org/10.1175/JAMC-D-16-0287.1
Jakob Themeßl M, Gobiet A, Leuprecht A (2011) Empirical-statisti- Stoner AM, Hayhoe K, Yang X, Wuebbles DJ (2013) An asynchro-
cal downscaling and error correction of daily precipitation from nous regional regression model for statistical downscaling of
regional climate models. Int J Climatol 31(10):1530–1544 daily climate variables. Int J Climatol 33(11):2473–2494
Jiang DB, Wang HJ, Lang XM (2004) Multimodel ensemble prediction Sunyer Pinya MA, Hundecha Y, Lawrence D, Madsen H, Willems
for climate change trend of China under SRES A2 scenario. Chin P, Martinkova M et al (2015) Inter-comparison of statistical
J Geophys 47(5):878–886 downscaling methods for projection of extreme precipitation in
Lavaysse C, Vrac M, Drobinski P, Lengaigne M, Vischel T (2012) Sta- Europe. Hydrol Earth Syst Sci 19(4):1827–1847
tistical downscaling of the French Mediterranean climate: assess- Tang JP, Niu XR, Wang SY, Gao HX, Wang XY, Wu J (2016) Statis-
ment for present and projection in an anthropogenic scenario. Nat tical downscaling and dynamical downscaling of regional cli-
Hazards Earth Syst Sci 12(3):651–670 mate in China: present climate evaluations and future climate
Li B, Zhou TJ (2010) Projected climate change over China under SRES projections. J Geophys Res Atmos 121(5):2110–2129
A1B scenario: multi-model ensemble and uncertainties. Adv Clim Taylor KE (2001) Summarizing multiple aspects of model
Change Res 6(4):270–276 performance in a single diagram. J Geophys Res Atmos
Liu Z, Xu Z, Charles SP, Fu G, Liu L (2011) Evaluation of two statisti- 106(D7):7183–7192
cal downscaling models for daily precipitation over an arid basin Trigo RM, Palutikof JP (1999) Simulation of daily temperatures for
in China. Int J Climatol 31(13):2006–2020 climate change scenarios over Portugal: a neural network model
Liu CM, Liu WB, Fu GB, Ouyang RL (2012) A discussion of some approach. Clim Res 13(1):45–59
aspects of statistical downscaling in climate impacts assessment Tripathi S, Srinivas VV, Nanjundiah RS (2006) Downscaling of pre-
(in Chinese). Adv water Sci 23(3):427–437 cipitation for climate change scenarios: a support vector machine
Liu W, Fu G, Liu C, Charles SP (2013) A comparison of three multi- approach. J Hydrol 330(3):621–640
site statistical downscaling models for daily rainfall in the North Vigaud N, Vrac M, Caballero Y (2013) Probabilistic downscal-
China Plain. Theor Appl Climatol 111(3–4):585–600 ing of GCM scenarios over southern India. Int J Climatol
Maraun D, Wetterhall F, Ireson AM, Chandler RE, Kendon EJ, Wid- 33(5):1248–1263
mann M et al (2010) Precipitation downscaling under climate Vrac M, Stein ML, Hayhoe K, Liang XZ (2007) A general method for
change: recent developments to bridge the gap between dynamical validating statistical downscaling methods under future climate
models and the end user. Rev Geophys 48(3):RG3003 change. Geophys Res Lett 34(18):L18701
Maraun D, Huth R, Gutiérrez JM, Martín DS, Dubrovsky M, Fischer A Vrac M, Drobinski P, Merlo A, Herrmann M, Lavaysse C, Li L, Somot
et al (2017) The VALUE perfect predictor experiment: evaluation S (2012) Dynamical and statistical downscaling of the French
of temporal variability. Int J Climatol 1:1. https:​ //doi.org/10.1002/ Mediterranean climate: uncertainty assessment. Nat Hazards
joc.5222 Earth Syst Sci 12(9):2769–2784
Maurer EP, Hidalgo HG (2008) Utility of daily vs. monthly large-scale Wang L, Chen W (2013) Application of bias correction and spatial
climate data: an intercomparison of two statistical downscaling disaggregation in removing model biases and downscaling over
methods. Hydrol Earth Syst Sci Discuss 4(5):3413–3440 China (in Chinese). Adv Earth Sci 28(10):1144–1153
Maurer EP, Hidalgo HG, Das T, Dettinger MD, Cayan DR (2010) The Wang L, Chen W (2014) A CMIP5 multimodel projection of future
utility of daily large-scale climate data in the assessment of cli- temperature, precipitation, and climatological drought in China.
mate change impacts on daily streamflow in California. Hydrol Int J Climatol 34(6):2059–2078
Earth Syst Sci 14:1125–1138 Wen X, Fang GH, Zhou L, Qi HS (2015) Regional climate change and
Michelangeli PA, Vrac M, Loukos H (2009) Probabilistic downscaling its possible effects on river runoff in Qiantang River Basin—past
approaches: application to wind cumulative distribution functions. and future. Fresenius Environ Bull 24(11B):3880–3894
Geophys Res Lett 36(11):L11708 Wen X, Fang GH, Qi HS, Zhou L, Gao YQ (2016) Changes of tempera-
Mizukami N, Clark MP, Gutmann ED, Mendoza PA, Newman AJ, ture and precipitation extremes in China: past and future. Theor
Nijssen B et  al (2016) Implications of the methodological Appl Climatol 126(1–2):369–383

13
An intercomparison of multiple statistical downscaling methods for daily precipitation and… 4649

Werner AT (2011) BCSD downscaled transient climate projections for Wilk MB, Gnanadesikan R (1968) Probability plotting methods for the
eight select GCMs over British Columbia. Pacific Climate Impacts analysis of data. Biometrika 55(1):1–17
Consortium, University of Victoria, Victoria, p 63 Wood AW, Maurer EP, Kumar A, Lettenmaier DP (2002) Long-range
Werner AT, Cannon AJ (2016) Hydrologic extremes—an intercompari- experimental hydrologic forecasting for the eastern United States.
son of multiple gridded statistical downscaling methods. Hydrol J Geophys Res Atmos 107:D20
Earth Syst Sci 20(4):1483 Wood AW, Leung LR, Sridhar V, Lettenmaier DP (2004) Hydrologic
Wetterhall F, Bárdossy A, Chen D, Halldin S, Xu CY (2006) Daily implications of dynamical and statistical approaches to downscal-
precipitation-downscaling techniques in three Chinese regions. ing climate model outputs. Clim Change 62(1):189–216
Water Resour Res 42(11):116 Wu J, Gao XJ (2013) A gridded daily observation dataset over China
Widmann M, Bretherton CS, Salathé EP Jr (2003) Statistical pre- region and comparison with the other datasets (in Chinese). Chin
cipitation downscaling over the northwestern United States J Geophys 56:1102–1111
using numerically simulated precipitation as a predictor. J Clim Zheng J, Ding L, Hao Z, Ge Q (2012) Extreme cold winter events in
16(5):799–816 southern China during AD 1650–2000. Boreas 41(1):1–12
Wilby RL, Wigley TML (1997) Downscaling general circulation model Zhou B, Wen QH, Xu Y, Song L, Zhang X (2014) Projected changes
output: a review of methods and limitations. Prog Phys Geogr in temperature and precipitation extremes in China by the CMIP5
21(4):530–548 multimodel ensembles. J Clim 27(17):6591–6611
Wilby RL, Wigley TML (2000) Precipitation predictors for downscal-
ing: observed and general circulation model relationships. Int J Publisher’s Note Springer Nature remains neutral with regard to
Climatol 20(6):641–661 jurisdictional claims in published maps and institutional affiliations.
Wilby RL, Charles SP, Zorita E, Timbal B, Whetton P, Mearns LO
(2004) Guidelines for use of climate scenarios developed from
statistical downscaling methods. Supporting material of the Inter-
governmental Panel on Climate Change, available from the DDC
of IPCC TGCIA. Available from: IPCC-DDC: http://www.ipcc-
data.org/

13

You might also like