J. Meteor. Res.  2016, Vol. 30 Issue (6): 867-880   PDF    
http://dx.doi.org/10.1007/s13351-016-6030-1
The Chinese Meteorological Society
0

Article Information

WU Dan, JIANG Zhihong, MA Tingting . 2016.
Projection of Summer Precipitation over the Yangtze-Huaihe River Basin Using Multimodel Statistical Downscaling Based on Canonical Correlation Analysis. 2016.
J. Meteor. Res., 30(6): 867-880
http://dx.doi.org/10.1007/s13351-016-6030-1

Article History

Received March 18, 2016
in final form May 16, 2016
Projection of Summer Precipitation over the Yangtze-Huaihe River Basin Using Multimodel Statistical Downscaling Based on Canonical Correlation Analysis
WU Dan(吴丹), JIANG Zhihong(江志红), MA Tingting(马婷婷)     
1. (Nansen-Zhu International Research Centre, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029);
2. (Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science & Technology, Nanjing 210044);
3. (University of Chinese Academy of Sciences, Beijing 100049);
ABSTRACT: By using observational daily precipitation data over the Yangtze-Huaihe River basin, ERA-40 data, and the data from eight CMIP5 climate models, statistical downscaling models are constructed based on BP-CCA (combination of empirical orthogonal function and canonical correlation analysis) to project future changes of precipitation. The results show that the absolute values of domain-averaged precipitation relative errors of most models are reduced from 8%-46% to 1%-7% after statistical downscaling. The spatial correlations are all improved from less than 0.40 to more than 0.60. As a result of the statistical downscaling multi-model ensemble (SDMME), the relative error is improved from -15.8% to -1.3%, and the spatial correlation increases significantly from 0.46 to 0.88. These results demonstrate that the simulation skill of SDMME is relatively better than that of the multimodel ensemble (MME) and the downscaling of most individual models. The projections of SDMME reveal that under the RCP (Representative Concentration Pathway) 4.5 scenario, the projected domain-averaged precipitation changes for the early (2016-2035), middle (2046-2065), and late (2081-2100) 21st century are -1.8%, 6.1%, and 9.9%, respectively. For the early period, the increasing trends of precipitation in the western region are relatively weak, while the precipitation in the east shows a decreasing trend. Furthermore, the reliability of the projected changes over the area east of 115°E is higher than that in the west. The stations with significant increasing trends are primarily located over the western region in both the middle and late periods, with larger magnitude for the latter. Stations with high reliability mainly appear in the region north of 28.5°N for both periods.
Key words: summer precipitation     BP-CCA     statistical downscaling     multimodel ensemble     projection    
1Introduction

The projection of regional climate change under global warming has been a hot topic within the global scientific community in recent years, with a considerable body of research carried out (Chen et al., 2011b; Zhou et al., 2014). With economic development, the progress of society, and the change of the climate system, economic losses caused by meteorological disasters are increasing substantially, becoming a potential limiting factor for socioeconomic development. The Yangtze-Huaihe River basin, located in the climatic transition zone of the subtropical zone and warm temperate belt, with its complex climatic conditions and frequent meteorological disasters, is highly sensitive to climate change. In addition, it is a densely populated region with a highly developed economy. Therefore, it is very important to carry out projections of summer precipitation for this region, at as higher resolutions as possible, against the background of global warming.

General circulation models (GCMs) have become an important tool for climate simulation and projection (Zhao, 2006; Chen et al., 2011a; Jiang et al., 2012). Currently, CMIP5 provides simulation results of more than 50 models. These models produce better simulations of large-scale climatic features than does the CMIP3 models (IPCC, 2013). However, in terms of simulating present local climate change over China, the improvements in this new generation of climate models in CMIP5 are not obvious, as compared with analyses from CMIP3 multimodel databases (Xu and Xu, 2012). Therefore, downscaling techniques need to be introduced to provide higher-resolution regional climate scenarios (Chen et al., 2012; IPCC, 2013). Statistical downscaling has come to be widely used because of its lesser computational requirements (Wilby and Wigley, 1997; Chu et al., 2008). There are different types of statistical downscaling, which can be grouped in three categories: regression model, weather classification, and weather generator (Busuioc and von Storch, 2003; Katz et al., 2003; Buishand et al., 2004; Fan et al., 2005). Linear regression models are most popular (Palutikof et al., 2002). For example, Han et al. (2014) and Liu and Li (2014) developed a statistical scheme based on multivariate linear regression and timescale decomposition to predict summer rainfall; the results suggested that this scheme exhibits a considerable advantage relative to the conventional without timescale decomposition method, as well as the models’ raw predictions, and may be appropriate for operational short-term climate predictions. Additionally, the BP-CCA method (a combination of empirical orthogonal function and canonical correlation analysis) has also been widely used in statistical downscaling (Benestad, 2002a, b; Tomozeiu et al., 2007, 2014). Compared with conventional CCA (canonical correlation analysis), BP-CCA can not only effectively reduce the dimensions of original climate variables and extract the main large-scale information, but also reduce noise (Xoplaki et al., 2000) and obtain stable relationships. Cui et al. (2010) showed that the simulation capacity of HadCM3 with respect to regional climate characteristics can be effectively improved by BP-CCA, with the relative errors of the climatological average state of extreme precipitation indices decreasing by 30%-100%, as well as the relative errors at all stations reducing to less than 10%. Thus, it is apparent that a statistical downscaling model based on BP-CCA exhibits good simulation ability.

On the other hand, considerable uncertainties still exist in the projection of climate change. The multimodel ensemble (MME) is supposed to be an effective method to reduce uncertainties (Palmer et al., 2005; Thomson et al., 2006; Semenov and Stratonovitch, 2010), and generally speaking, the projected results of the MME are indeed more reliable than those of individual models (Barnston et al., 2003; Raftery et al., 2003; Georgakakos et al., 2004). For instance, Jiang et al. (2009) studied the capabilities of seven IPCC-AR4 (Fourth Assessment Report of the Intergovernmental Panel on Climate Change) global models in simulating extreme precipitation in China, and concluded that the MME performed better than individual models. For example, the relative error of R95t (fraction of total rainfall from extreme events) simulated by the MME was 7%, while that of most individual models was more than 10%. Dong et al. (2015) assessed the overall performance of CMIP5 models based on relative root-mean-square error (RMSE) (the RMSE of the MME relative to the median RMSE of all models). The results showed that the relative errors of the MME for most extreme temperature indices were in the range of -30% to -12%, demonstrating that the results of the MME were closer to observed values than those of most individual models. On regional scales, Benestad (2002a, b) applied a statistical downscaling technique based on common EOF (empirical orthogonal function) analysis and CCA to different global climate scenarios, and projected local climate scenarios at various locations in Norway and northern Europe, respectively. Kang et al. (2007) showed significant skill improvement by applying the statistical downscaling MME (SDMME) to the Philippines and Thailand. They found that the correlation coefficients between the area-averaged precipitation series and the observations for both areas were improved from -0.39 and -0.47 to 0.62 and 0.75, respectively. Zhu et al. (2008) found significantly increased skill in predicting summer precipitation over East Asia. The anomaly correlation coefficient was increased by 0.14 and the corresponding RMSE was reduced by 10.4% for the statistical MME forecast. Sun and Chen (2012) developed a statistical downscaling method to improve the global precipitation forecasting of GCMs and demonstrated that the 7-GCM ensemble mean after applying the statistical downscaling showed better performance, with spatial anomaly correlation coefficients of 0.86, 0.75, 0.80, and 0.84 for 1998, 1999, 2000, and 2001, respectively. However, the correlation coefficients for the downscaling of individual models were in the range of 0.58-0.75. The above-mentioned studies indicate that SDMME is an effective method to improve the reliability of projections. However, there has been relatively little research on the SDMME in China compared to other countries and regions of the world.

This study uses observational daily precipitation data over the Yangtze-Huaihe River basin, ECMWF 40-yr reanalysis (ERA-40) data, and data from eight CMIP5 climate models, to construct statistical downscaling models to project the future changes of precipitation for the early (2016-2035), middle (2046-2065), and late (2081-2100) 21st century under the RCP4.5 (+ 4.5 W m-2 Representative Concentration Pathway) scenario, based on CCA.

The paper is organized as follows: Section 2 introduces the models, data, and statistical downscaling technique used in this study. Section 3 provides a brief description of the selection of predictors, set-up and validation of the statistical downscaling models. The simulation abilities of the statistical downscaling models for the current GCMs, including the SDMME, are discussed in Section 4. In Section 5, scenarios of changes in summer precipitation over the Yangtze-Huaihe River basin and reliability analysis are presented. Finally, a summary and conclusions are given in Section 6.

2Data and methods 2.1Datasets

The main datasets employed in this study include: (1) summer (June-August) daily precipitation at 56 stations over the Yangtze-Huaihe River basin (27.5°-32°N, 110°-122.5°E) from the National Meteorological Information Center, covering the period 1961 -2005. From Fig. 1, it can be seen that the stations have a rather uniform spatial distribution. (2) The monthly mean circulation data of ERA-40 from 1961 to 2002, gridded at a resolution of 2.5° × 2.5°, including geopotential height fields, temperature fields, specific humidity fields, wind fields, and so on. (3) Eight CMIP5 climate models that have certain abilities to simulate summer circulations over East Asia, and extreme precipitation indices in China (Huang et al., 2015; Jiang et al., 2015) are used. The variables used are monthly geopotential height, temperature, specific humidity, zonal wind, and precipitation, for summers during 1961-2005. The RCP4.5 scenario experiments for the early, middle, and late 21st century are also used. Table 1 illustrates the eight models utilized in this study. The above-mentioned large-scale circulations of reanalysis and model datasets will be referred to as “predictors” hereafter. Before carrying out the statistical downscaling and any analysis, we first use a bilinear scheme to interpolate all model data onto the same grid as the reanalysis data. Both precipitation and predictor series are standardized based on the climate state of 1961-1990, to avoid systematic errors.

Figure 1 Station locations over the Yangtze-Huaihe River basin.
Table 1 Model institution identification (ID), modeling center and country, model name, and atmospheric resolution of the eight CMIP5 global climate models used in this study
2.2Methods

In this study, the BP-CCA method is employed to obtain stable statistical relationships between largescale circulations and regional precipitation. The following key steps are followed in order to construct the future projections:

(1) Select the optimal large-scale predictors and corresponding regions by calculating the correlation coefficient between the first principal component (PC) of summer precipitation and multiple ERA-40 largescale fields.

(2) The common EOF approach (Benestad, 2001; Cui et al., 2010), used to combine the ERA-40 largescale fields and identical variables of individual GCMs along the time axis, is first applied. Then, the leading modes of precipitation and the ERA-40 largescale fields during 1961-1990 are used for the CCA to establish the statistical downscaling models, and an independent-sample validation based on observations is carried out for the 12-yr period from 1991 to 2002. The skill of the statistical downscaling models is quantified at station level in terms of the following: correlation coefficient (COR), RMSE, and relative errors computed between observed and simulated values, which are respectively defined as

(1)

and

(2)

where N is the number of years in the validation period, Mi represents the simulation results, and Oi represents the observations.

(3) The statistical downscaling models established in (2) are applied to the historical outputs of the CMIP5 models (1986-2005) to examine the GCMs’ simulation skill before projecting the climate change scenarios. The predictors used for the individual models are the same as observed. A Taylor diagram (Taylor, 2001) is introduced to assess the performance of the GCMs in simulating the spatial pattern of precipitation before and after downscaling. This provides a statistical summary of the comparisons between simulations and observations in terms of spatial correlation coefficients, centered pattern root-mean-square (RMS) difference, and the ratio of spatial standard deviations of the model and observations.

(4) The statistical downscaling models established in (2) are applied to the future predictors simulated by the GCMs. To improve the reliability of the projections, we calculate the SDMME by using the arithmetic average of the eight GCMs’ downscaled results.

In addition, we use the signal-to-noise ratio (SNR) to discuss the reliability of the SDMME’s projections, which is defined as follows:

(3)

where is the absolute value of the projected change rate of the SDMME (signal), and δΔP is the mean square deviation of the projected change rates of all individual models after downscaling (noise). The area where the signal is smaller than the noise (i.e., SNR < 1) indicates that the reliability of the projections is relatively low; on the contrary, the area where the signal is larger than the noise (i.e., SNR > 1) indicates that the projected results are credible, and a larger SNR indicates higher reliability.

It is notable that the downscaled results have smaller variance than the observed series. Therefore, the method of variance inflation (von Storch, 1999) is adopted in this study.

3The statistical downscaling model 3.1Choice of predictors and predictor domains for statistical downscaling

It is very important to select suitable predictors for statistical downscaling. In our research, the predictors and their corresponding domains are selected based on correlation analysis between the first PC of summer precipitation from 1961 to 1990 and multiple large-scale circulations (Z500, geopotential height at 500 hPa; Z850, geopotential height at 850 hPa; T850, temperature at 850 hPa; Q850, specific humidity at 850 hPa; Q500, specific humidity at 500 hPa; U850, zonal wind at 850 hPa; U500, zonal wind at 500 hPa, etc.). The first mode of summer precipitation accounts for 39.14% of the total variance, and it correlates closely (0.998) with the domain-averaged summer precipitation series. Thus, we use the first PC, which represents the domain-averaged summer precipitation series, in investigating the correlation. The variables that have a feasible physical relationship and high degree of correlation with the precipitation are selected as predictors, as shown in Fig. 2. The relationship between the western Pacific subtropical high and summer precipitation over the Yangtze-Huaihe River basin is presented in Figs. 2a and 2d. According to Fig. 2a, the area with a positive relationship is primarily south of 30°N, indicating a southward subtropical high associated with a significant westerly anomaly prevailing over the Yangtze-Huaihe River basin. Furthermore, the northwesterly of the subtropical high can bring large quantities of water vapor to this region, resulting in much more precipitation. On the contrary, when the subtropical high is situated farther north than usual, the basin will be under the control of the subtropical high or a significant easterly anomaly prevails, leading to less precipitation. Figure 2b presents an area with low temperature at 850 hPa located north of the Yangtze-Huaihe River basin, which is primarily controlled by cold air. The confluence of the cold air north of the basin and the warm, moist air from the south can bring more precipitation to the region. Sufficient water vapor is one of the necessary conditions for precipitation, which can be explained by Fig. 2c. According to these results, four predictors and corresponding regions are preliminarily selected: Z500 (10°-25°N, 95°-170°E); T850 (27.5°-37.5°N, 107.5°-125°E); Q500 (27.5°-32.5°N, 105°-125°E); and U500 (25°-32.5°N, 95°-140°E).

Figure 2 Correlation coefficients between the first PC of summer precipitation and (a) Z500, (b) T850, (c) Q500, and (d) U500. Regions that are statistically significant at the 90% confidence level are shaded. Predictor domains are shown by black boxes, and the black dots indicate the stations shown in Fig. 1.

Given the multicollinearity problem of predictors, the correlation between various large-scale fields (Z500, T850, Q500, and U500) during 1961-1990 is calculated (Table 2). The indication is that U500 is highly correlated with Z500 and T850, and the correlation between the other two predictors is relatively weak. Thus, we consider that the removal of U500 may improve the results of statistical downscaling, which is further proven by several statistical downscaling experiments (Table 3). We can see that the correlation coefficients of both the domain-averaged precipitation interannual change and the climate state between the observed and simulated values are all equal to or greater than 0.75 (exceeding the 99% confidence level). However, in terms of relative error, a weaker performance is exhibited in the experiments including U500, with the absolute values of the relative errors exceeding 5%. A similar situation can be found for the RMSE, which is equal to or greater than 2.35 mm day-1. In the case of the combination of Q500, Z500, and T850, the RMSE has a minimum value. Therefore, the results verify that the predictor combination of Q500, Z500, and T850 is the best for the statistical downscaling of summer precipitation over the Yangtze-Huaihe River basin.

Table 2 Correlation coefficients between each pair of predictors (1961-1990)
Table 3 Relative error, RMSE, interannual correlation, and spatial correlation between the downscaled summer precipitation and observations in the validation period of 1991-2002 (Q, Z, T, and U indicate Q500, Z500, T850, and U500, respectively)
3.2Selection of downscaling model parameters

It is necessary to change the number of PCs of the local precipitation and large-scale fields, as well as the CCA modes, to construct the optimal statistical downscaling model. Figure 3 depicts the variance of RMSE with the PCs of the local precipitation/largescale fields truncating at 7/8, 13/15, and 18/18, respectively, associated with different CCA modes. We can infer that, in CCA, the increase in the number of canonical modes first leads to an improvement. However, after attaining a “saturation level” (e.g., 10 and 12 modes), adding more modes into the statistical downscaling model deteriorates the results, which is consistent with Huth (1999). The number of PCs of the local precipitation and large-scale fields and CCA modes most suitable to retain are 13, 15, and 10, respectively, with the minimal spatially averaged value of RMSE being 2.27 mm day-1.

Figure 3 Accuracy (in terms of spatially averaged RMSE; mm day-1) for different numbers of PCs of precipitation/large-scale fields and CCA modes in the validation period of 1991-2002.

A byproduct of BP-CCA is that it automatically identifies the link between large-scale variability and local climate. In the following, the precipitation-predictors connection is presented by the first CCA pairs obtained by retaining the first 15 PCs of combined predictors (Q500, T850, and Z500) and 13 PCs of precipitation (Fig. 4). The first canonical correlation coefficient is 0.62, indicating strong links between precipitation and predictors. The pattern of Z500 presents a negative anomaly center in the area north of 25°N, with small values over the large-area region, suggesting that the subtropical high is situated farther south than usual. Therefore, the water vapor from the south cannot be transported to the Yangtze-Huaihe River basin, which can explain the negative precipitation anomalies at most stations, associated with the same sign in Q500.

Figure 4 The first CCA patterns of (a) precipitation, (b) T850, (c) Q500, and (d) Z500 at the 13th and 15th PC truncation of precipitation and combined predictors, respectively.
3.3Validation of the results

In this section, the statistical downscaling models are applied to the large-scale fields during 1961-2002 to obtain the downscaled precipitation at 56 stations. The results for the period 1961-1990 are hindcasted by the statistical downscaling models. Additionally, we assess the skill of the statistical downscaling models for summer precipitation in terms of temporal variation and spatial distribution.

3.3.1Evaluation of temporal variation

Figure 5 displays the spatial distribution of the temporal correlation coefficients (TCCs) between the downscaled and observed summer precipitation during 1991-2002. The statistical downscaling model performs well in simulating the temporal variation of summer precipitation at most stations. The stations with correlation coefficients exceeding 0.50 account for 61% (TCC of 0.50 indicates the 90% confidence level), and the performance for the western region is better than that for the eastern region. The stations with higher correlation coefficients are mainly distributed over the region west of 119°E. The stations with low correlation coefficients are primarily distributed along the eastern coastal zone, which could be explained by the greater strength of local processes there (such as convection and mesoscale synoptic processes).

Figure 5 Spatial distribution of TCCs between the downscaled and observed summer precipitation during 1991-2002. TCCs of 0.50, 0.58, and 0.71 indicate the 90%, 95%, and 99% confidence levels, respectively.

Figure 6 presents the interannual change in downscaled and observed area-averaged summer precipitation over the period 1961-2002. The correlation coefficients between the downscaled and observed areaav eraged summer precipitation series are 0.86 and 0.83 over the periods 1961-1990 and 1991-2002, respectively. Although there are some years when the intensity of precipitation is under-or over-estimated, the statistical downscaling model is able to capture the temporal variability of the area-averaged precipitation.

Figure 6 Downscaled (solid lines) and observed (dashed lines) area-averaged summer precipitation during 1961-2002. R1 and R2 are the correlation coefficients for the periods 1961-1990 and 1991-2002, respectively.
3.3.2Evaluation of spatial distribution

In terms of the spatial climatology (1991-2002), the correlation coefficient between the observed and downscaled results is 0.80, indicating that the statistical downscaling model has good skill in reproducing the spatial pattern of the climate state. In addition, the correlation coefficients between the downscaled and observed precipitation are also calculated on the spatial scale for each year (Fig. 7). The statistical downscaling model can simulate the spatial pattern for each year relatively well. The majority of the correlation coefficients are greater than 0.34 (correlation coefficient of 0.34 indicates the 99% confidence level), except for 2001. It is found that the statistical downscaling model performs better in the cases of 1980, with the correlation coefficient reaching 0.93.

Figure 7 Spatial correlation between the downscaled and observed summer precipitation during 1961-2002.

The above results demonstrate that the statistical downscaling model is able to simulate the characteristics of summer precipitation well over the Yangtze-Huaihe River basin. As a result, it can be further applied to the GCMs.

4Downscaling of summer precipitation for the present climate

GCMs are less capable when it comes to simulating local precipitation but better when simulating large-scale atmospheric variables, which can be further manipulated to provide a reasonable prediction of precipitation at local stations. Therefore, the statistical downscaling models are first applied to the predictors (1986-2005) simulated by GCMs to examine their simulation skill, before projecting the climate change scenarios. The precipitation of the GCMs is interpolated to the station coordinates for the convenience of comparing with the downscaled results.

4.1Domain averages

We present in Fig. 8a quantitative evaluation of the statistical downscaling models’ performance in terms of the domain average. We can see that, before downscaling, there are dry biases in all the GCMs except ACCESS1.3. Furthermore, the absolute values of relative error in most models are between 8% and 46%, which are reduced by 15%-41% after downscaling, reaching 1%-7%. Among the eight models, wet biases in precipitation are found in IPSL-CM5AMR, CMCC-CM, ACCESS1.3, and EC-EARTH after downscaling, associated with the absolute values of relative error decreasing by more than 15%. In addition, the amplitudes of the dry biases decrease in all the other models except CESM1-CAM5. BCC-CSM1.1(m) and EC-EARTH perform best after downscaling, with their relative errors improving from -44.6% to -6.6% and from -45.7% to 4.8%, respectively. In addition, CMCC-CM shows the lowest deviation after downscaling, with the relative error reducing to 0.9%. It is clear that the simulation ability of most GCMs is significantly improved after downscaling. The contrast between the SDMME and MME suggests that there are dry biases in both results, but the SDMME shows better performance. The relative error is improved from -15.8% to -1.3% by the SDMME. In summary, the downscaled result of the SDMME is better than that of most individual models.

Figure 8 Relative error [(modeled-observed) / observed × 100; %] in the raw GCM outputs, GCM downscaled results, MME, and SDMME.
4.2Evaluation of spatial variation by Taylor diagram

A Taylor diagram is used to comprehensively assess the performance of the GCMs, before and after statistical downscaling, in simulating the spatial pattern of summer precipitation. Figure 9 shows the Taylor diagram of the model simulations and model downscaled results against observations, with blue for the GCMs and MME and red for the models’ downscaling and SDMME. Where the distance between a square and the circle on the bottom axis is shorter, the simulation capability is better. Most GCMs have spatial correlations of less than 0.40 with the reference data, while the correlations are all improved to above 0.60 after downscaling. HadGEM2-ES has the highest correlation value of 0.91. This indicates that the performance of coupled models in simulating the spatial distribution of summer precipitation is significantly improved by statistical downscaling. The pattern of precipitation is clearly separated by the ratio of variance (solid black lines): most models have a ratio of variance between 0.30 and 0.91, but the majority of models have a ratio of variance larger than 1.0 after downscaling, indicating that the simulated spatial variation is smaller/larger than observed before/after downscaling for most models. Almost all models have a centered normalized RMS difference-expressed by the blue solid line-of more than 1.0. However, most of them are reduced to 0.40-0.80 after downscaling, indicating that the amplitude of bias is decreased by statistical downscaling. As for the SDMME, it can be clearly seen that the spatial correlation increases significantly from 0.46 to 0.88, associated with the centered normalized RMS difference reducing by 0.38 mm day-1 compared to the MME. Further, the ratio of variance is closer to 1.0 than the MME. All these results reveal that the simulation skill of the SDMME is relatively better than that of the MME, as well as the downscaling of most individual models.

Figure 9 Taylor diagram of summer precipitation between observations and GCMs/MME (blue), and multimodel downscaling/SDMME (red), over the Yangtze-Huaihe River basin. In the diagram, angular axes show the spatial correlation between simulated/downscaled and observed fields; and radial axes show the spatial standard deviation (RMS deviation), normalized against that of the observations. Each square represents a model, identified by its number on the right. For the squares located between the two dotted lines, correlations are between 0.60 and 0.90.

In summary, the ability of each coupled model to simulate local precipitation can be significantly improved by statistical downscaling. The SDMME performs better than the MME, as well as the downscaling of most individual models. This gives us confidence in using the SDMME to project summer precipitation over the Yangtze-Huaihe River basin for the 21st century under the RCP4.5 scenario.

5Projections of summer precipitation and reliability analysis

SDMME scenarios of summer precipitation at 56 stations are obtained by applying the statistical downscaling models to the outputs of GCMs under the RCP4.5 scenario. Future changes in precipitation over the Yangtze-Huaihe River basin for the early (2016-2035), middle (2046-2065), and late (2081-2100) 21st century, relative to 1986-2005, are presented in this section.

Figure 10a presents the climatology of observed summer precipitation over the Yangtze-Huaihe River basin during 1986-2005. The summer precipitation at all stations is in the range of 4-12 mm day-1, and the precipitation at most stations in the area east of Hubei Province, south of Anhui Province, and north of Jiangxi Province is above 6 mm day-1. A common feature exists in the different steps of future climate projection that the increasing trend in the west is much greater than that in the east (Figs. 10b-d). With respect to the period 1986-2005, the area-averaged precipitation relative change for the early period of the 21st century is -1.8%. The increasing trends of precipitation in the western region are relatively weak, with a precipitation anomaly increase of 0-15%; the stations in Hubei Province show the most obvious increasing trends; while the precipitation in the eastern part shows a decreasing trend. These findings are consistent with Li (2008), who reported a drying trend in southern and eastern coastal areas of China for the first half of the 21st century, with a precipitation anomaly decrease of 10%-20%. For the middle and end of the century, the precipitation at most stations is projected to increase. For the middle of the 21st century, the domain-averaged precipitation increases by 6.1%, with an increase of 7.9% compared with the early period. The spatial distribution of precipitation changes is similar to the early period. However, more intense precipitation is projected at most stations in the west, with a range of 15%-30%. Meanwhile, the decreasing trends at some stations in the east turn to opposite signs, such as at Hangzhou, Cixi, Hongjia, and Yuhuan stations. The end of the 21st century is expected to feature more intense precipitation. The domain-averaged precipitation is projected to increase by 9.9%, with an increase of 11.7% and 3.8% compared with the early and middle periods, respectively. The relative change at most stations is in the range of 15%-45%, with obvious increasing trends occurring in Hunan, Hubei, and Zhejiang provinces. In addition, weaker decreases are projected to occur in Anhui and Jiangsu provinces compared with the former two periods.

Figure 10 (a) Climatology of observed summer precipitation (mm day-1) over the Yangtze-Huaihe River basin for the period 1986-2005, and spatial distributions of projected relative changes of the SDMME (%) for the (b) early, (c) middle, and (d) late 21st century, relative to 1986-2005.

In order to evaluate quantitatively the reliability of the projected results of the SDMME, we calculate the SNR. The greater the sign in Fig. 11, the higher the reliability of the projection. In the spatial patterns, the SNR exhibits somewhat different features with time. For the early 21st century, the stations where the signal is larger than the noise (i.e., SNR > 1) are mainly distributed in the area east of 115°E, associated with stations having higher reliability being distributed in the area east of 119°E. The reliability of the western part of the region is relatively low, demonstrating that the SDMME performs better over the east of the basin than over the west. For the middle 21st century, the spatial distribution of SNR is different to that of the early period. Notable signals of precipitation change relative to the intermodal noise appear only in the area north of 28.5°N, demonstrating a low level of uncertainty and a high level of reliability. The reliability of the stations in the west increases, especially in Hubei Province, while the number of stations with high reliability reduces in the eastern part. For the late 21st century, the spatial distribution of SNR is not largely different from that of the middle period; high reliability also appears in the region north of 28.5°N, and the reliability in the west increases further, especially in Hubei and Hunan provinces. It is also apparent that the number of stations with credible projections in the western part of this region increases with time.

Figure 11 Spatial distributions of SNR over the Yangtze-Huaihe River basin for the (a) early, (b) middle, and (c) late 21st century. The bigger the sign, the higher the reliability of the projection.
6Summary and conclusions

In the present study, we project future changes in summer precipitation over the Yangtze-Huaihe River basin by using a statistical downscaling technique based on BP-CCA. We first select the optimal largescale predictors and corresponding regions. Then, statistical downscaling models are constructed for different GCMs and validated by using an independent dataset. Finally, statistical downscaling models are applied to the current and future predictors simulated by the GCMs. The main findings can be summarized as follows:

(1) The predictors and their corresponding domains are first selected based on correlation analysis between the first PC of summer precipitation and multiple large-scale fields. Then, the combination of geopotential height at 500 hPa, temperature at 850 hPa, and specific humidity at 500 hPa is confirmed as the optimal set of predictors according to several statistical downscaling experiments.

(2) Validation with an independent sample demonstrates that the statistical downscaling models are capable of simulating the temporal variation and spatial distribution of summer precipitation well. For instance, the stations with correlation coefficients between the downscaled and observed series exceeding 0.50 account for 61%. Most annual spatial correlations between the downscaled and observed climatological average state of precipitation are greater than 0.34.

(3) The absolute values of domain-averaged precipitation relative errors of most models are reduced by 15%-41% after downscaling, reaching 1%-7%, associated with the spatial correlation increasing to greater than 0.60. The contrast between the SDMME and MME suggests that the relative error is improved from -15.8% to -1.3% by the SDMME; the spatial correlation also increases significantly from 0.46 to 0.88, associated with a decrease of 0.38 mm day-1 in the centered normalized RMS difference. All these results reveal that the simulation skill of the SDMME is relatively better than that of the MME, as well as the downscaling of most individual models.

(4) The projections of the SDMME show that, under the RCP4.5 scenario, the projected domainav eraged relative changes in summer precipitation for the early, middle, and late 21st century are -1.8%, 6.1%, and 9.9%, respectively. For the early part of the century, the increasing trends of precipitation in the western region are relatively weak, while the precipitation in the east shows a decreasing trend. More over, the reliability of the projected changes over the region east of 115°E is higher than that of the west. For the middle of the century, more intense precipitation is projected at most stations in the western part. Decreasing trends at some stations in the east turn to opposite sign. The end of the century is expected to have more intense precipitation, with obvious increasing trends occurring in Hunan, Hubei, and Zhejiang provinces. High reliability mainly appears in the region north of 28.5°N, in both the middle and late periods.

Our study reveals that the uncertainty in projections of summer precipitation over the Yangtze-Huaihe River basin can be reduced to some extent by our method of the SDMME. It gives us a degree con-of confidence in applying the SDMME to other local regions to make higher-resolution projections. Nevertheless, uncertainties do still exist in the precipitation projections due to a complicated range of factors, including not only the large-scale circulations involved but also local factors and the assumption that the relationship between the large-scale fields and local scales will be the same in the future as in the past. It is clear that more research is needed to reduce these uncertainties in future climate projections, such as by increasing the number of coupled models for statistical downscaling or adopting different downscaling ensemble schemes, and so on.

Acknowledgments: We acknowledge the modeling groups listed in Table 1 of this paper for making their simulations available for analysis, the PCMDI for collecting and archiving the CMIP5 model output, and the World Climate Research Programme’s Working Group on Coupled Modelling.
References
DOI:10.1175/BAMS-84-12-1783 Barnston A. G., Mason S. J., Goddard L., et al ,2003: Multimodel ensembling in seasonal climate forecast-ing at IRI. Bull. Amer. Meteor. Soc. , 84 , 1783–1796. DOI:10.1175/BAMS-84-12-1783
DOI:10.1002/(ISSN)1097-0088 Benestad R. E. ,2001: A comparison between two em-pirical downscaling strategies. Int. J. Climatol. , 21 , 1645–1668. DOI:10.1002/(ISSN)1097-0088
DOI:10.1175/1520-0442(2002)015<3008:EDMETA>2.0.CO;2 Benestad R. E. ,2002a: Empirically downscaled multi-model ensemble temperature and precipitation sce-narios for Norway. J. Climate , 15 , 3008–3027. DOI:10.1175/1520-0442(2002)015<3008:EDMETA>2.0.CO;2
DOI:10.3354/cr021105 Benestad R. E. ,2002b: Empirically downscaled temperature scenarios for northern Europe based on a multi-model ensemble. Climate Research , 21 , 105–125. DOI:10.3354/cr021105
DOI:10.1175/1520-0442(2004)017<1816:OTCOTT>2.0.CO;2 Buishand T. A., Shabalova M. V., Brandsma T. ,2004: On the choice of the temporal aggregation level for statistical downscaling of precipitation. J. Climate , 17 , 1816–1827. DOI:10.1175/1520-0442(2004)017<1816:OTCOTT>2.0.CO;2
DOI:10.3354/cr024181 Busuioc A., von Storch H. ,2003: Conditional stochastic model for generating daily precipitation time series. Climate Research , 24 , 181–195. DOI:10.3354/cr024181
DOI:10.1007/s00382-010-0910-3 Chen W. L., Jiang Z. H., Li L., et al ,2011a: Simulation of regional climate change under the IPCC A2 scenario in Southeast China. Climate Dyn. , 36 , 491–507. DOI:10.1007/s00382-010-0910-3
DOI:10.1175/2011JCLI4102.1 Chen W. L., Jiang Z. H., Li L. ,2011b: Probabilistic projections of climate change over China under the SRES A1B scenario using 28 AOGCMs. J. Climate , 24 , 4741–4756. DOI:10.1175/2011JCLI4102.1
DOI:10.1175/WAF-D-11-00079.1 Chen H. P., Sun J. Q., Wang H. J. ,2012: A sta-tistical downscaling model for forecasting summer rainfall in China from DEMETER hindcast datasets. Wea. Forecasting , 27 , 608–628. DOI:10.1175/WAF-D-11-00079.1
DOI:10.1029/2007JD009424 Chu J. L., Kang H. W., Tam C. Y., et al ,2008: Sea-sonal forecast for local precipitation over northern Taiwan using statistical downscaling. J. Geophys. Res. , 113 , D12118. DOI:10.1029/2007JD009424
Cui Yan, Jiang Zhihong, Chen Weilin ,2010: Projec-tion of extreme precipitation events in the 21st cen-tury in the Changjiang-Huaihe River valley based on canonical correlation analysis. Adv. Climate Change Res. , 6 , 405–410.
DOI:10.1007/s00376-015-4152-5 Dong S. Y., Xu Y., Zhou B. T., et al ,2015: Assessment of indices of temperature extremes simulated by multiple CMIP5 models over China. Adv. Atmos. Sci. , 32 , 1077–1091. DOI:10.1007/s00376-015-4152-5
Fan Lijun, Fu Congbin, Chen Deliang ,2005: Re-view on creating future climate change scenarios by statistical downscaling techniques. Adv. Earth Sci. , 20 , 320–329.
DOI:10.1016/j.jhydrol.2004.03.037 Georgakakos K. P., Seo D. J., Gupta H., et al ,2004: Towards the characterization of streamflow simula-tion uncertainty through multimodel ensembles. J. Hydrol. , 298 , 222–241. DOI:10.1016/j.jhydrol.2004.03.037
DOI:10.1007/s00376-013-3016-0 Han L. Q., Li S. L., Liu N. ,2014: An approach for improving short-term prediction of summer rainfall over North China by decomposing interannual and decadal variability. Adv. Atmos. Sci. , 31 , 435–448. DOI:10.1007/s00376-013-3016-0
Huang Hailing, Jiang Zhihong, Wang Zhifu, et al ,2015: The evaluation of the 500hPa geopotential height's main modes in East Asia as done by the CMIP5 models. Acta Meteor. Sinica , 73 , 110–127.
DOI:10.3354/cr013091 Huth R. ,1999: Statistical downscaling in central Europe:Evaluation of methods and potential predictors. Cli-mate Res. , 13 , 91–101. DOI:10.3354/cr013091
IPCC, 2013:Climate Change 2013:The Physical Science Basis. Contribution of Working Group Ⅰ to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. Stocker, T. F., D. Qin, G. K. Plattner, et al., Eds., Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, 649-740.
Jiang Zhihong, Chen Weilin, Song Jie, et al ,2009: Pro-jection and evaluation of the precipitation extremes indices over China based on seven IPCC AR4 cou-pled climate models. Chinese J. Atmos. Sci. , 33 , 109–120.
DOI:10.1007/s10584-011-0090-0 Jiang Z. H., Song J., Li L., et al ,2012: Extreme climate events in China:IPCC-AR4 model evaluation and projection. Climatic Change , 110 , 385–401. DOI:10.1007/s10584-011-0090-0
DOI:10.1175/JCLI-D-15-0099.1 Jiang Z. H., Li W., Xu J. J., et al ,2015: Extreme precipitation indices over China in CMIP5 models. Part I:Model evaluation. J. Climate , 28 , 8603–8619. DOI:10.1175/JCLI-D-15-0099.1
DOI:10.1029/2007GL030730 Kang H. W., An K. H., Park C. K., et al ,2007: Multimodel output statistical downscaling prediction of precipitation in the Philippines and Thai-land. Geophy. Res. Lett. , 34 , L15710. DOI:10.1029/2007GL030730
DOI:10.1023/A:1026054330406 Katz R. W., Parlange M. B., Tebaldi C. ,2003: Stochastic modeling of the effects of large-scale cir-culation on daily weather in the southeastern U.S. Climatic Change , 60 , 189–216. DOI:10.1023/A:1026054330406
Li S. L. ,2008: Projecting the summer climate of main-land China in the middle 21st century:Will the droughts in North China persist?. Atmos. Oceanic Sci. Lett. , 1 , 12–17.
DOI:10.1175/WAF-D-13-00045.1 Liu N., Li S. L. ,2014: Predicting summer rainfall over the Yangtze-Huai region based on time-scale decomposition statistical downscaling. Wea. Fore-casting , 29 , 162–176. DOI:10.1175/WAF-D-13-00045.1
DOI:10.1098/rstb.2005.1750 Palmer T. N., Doblas-Reyes F. J., Hagedorn R., et al ,2005: Probabilistic prediction of climate using multi-model ensembles:From basics to applications. Philos. Trans. Roy. Soc. Lond. B Biol. Sci. , 360 , 1991–1998. DOI:10.1098/rstb.2005.1750
DOI:10.1175/1520-0442(2002)015<3529:GRATSA>2.0.CO;2 Palutikof J. P., Goodess C. M., Watkins S. J., et al ,2002: Generating rainfall and temperature scenarios at multiple sites:Examples from the Mediterranean. J. Climate , 15 , 3529–3548. DOI:10.1175/1520-0442(2002)015<3529:GRATSA>2.0.CO;2
Raftery A. E., Gneiting T., Balabdaoui F., et al ,2003: Using Bayesian model averaging to calibrate fore-cast ensembles. Mon. Wea. Rev. , 133 , 1155–1174.
DOI:10.3354/cr00836 Semenov M. A., Stratonovitch P. ,2010: Use of multi-model ensembles from global climate models for assessment of climate change impacts. Climate Res. , 41 , 1–14. DOI:10.3354/cr00836
DOI:10.1007/s00703-012-0195-7 Sun J. Q., Chen H. P. ,2012: A statistical down-scaling scheme to improve global precipitation fore-casting. Meteor. Atmos. Phys. , 117 , 87–102. DOI:10.1007/s00703-012-0195-7
DOI:10.1029/2000JD900719 Taylor K. E. ,2001: Summarizing multiple aspects of model performance in a single diagram. J. Geophys. Res. , 106 , 7183–7192. DOI:10.1029/2000JD900719
DOI:10.1038/nature04503 Thomson M. C., Doblas-Reyes F. J., Mason S. J., et al ,2006: Malaria early warnings based on seasonal cli-mate forecasts from multi-model ensembles. Nature , 439 , 576–579. DOI:10.1038/nature04503
DOI:10.1007/s00704-006-0275-z Tomozeiu R., Cacciamani C., Pavan V., et al ,2007: Climate change scenarios for surface temperature in Emilia-Romagna (Italy) obtained using statistical downscaling models. Theor. Appl. Climatol. , 90 , 25–47. DOI:10.1007/s00704-006-0275-z
DOI:10.1007/s11069-013-0552-y Tomozeiu R., Agrillo G., Cacciamani C., et al ,2014: Statistically downscaled climate change projections of surface temperature over northern Italy for the periods 2021-2050 and 2070-2099. Natural Hazards , 72 , 143–168. DOI:10.1007/s11069-013-0552-y
DOI:10.1175/1520-0442(1999)012<3505:OTUOII>2.0.CO;2 von Storch H., Agrillo G., Cacciamani C., et al ,1999: On the use of "inflation" in statis-tical downscaling. J. Climate , 12 , 3505–3506. DOI:10.1175/1520-0442(1999)012<3505:OTUOII>2.0.CO;2
Wilby R. L., Wigley T. M. L. ,1997: Downscaling general circulation model output:A review of meth-ods and limitations. Progress in Physical Geography , 21 , 530–548.
DOI:10.3354/cr014129 Xoplaki E., Luterbacher J., Burkard R., et al ,2000: Connection between the large-scale 500 hPa geopo-tential height fields and precipitation over Greece during wintertime. Climate Res. , 14 , 129–146. DOI:10.3354/cr014129
Xu Y., Xu C. H. ,2012: Preliminary assessment of simulations of climate changes over China by CMIP5 multi-models. Atmos. Oceanic Sci. Lett. , 5 , 489–494.
Zhao Zongci ,2006: Latest advances in global climate projections. Adv. Climate Change Res. , 2 , 68–70.
DOI:10.1175/JCLI-D-13-00761.1 Zhou B. T., Wen Q. H., Xu Y., et al ,2014: Projected changes in temperature and precipitation extremes in China by the CMIP5 multimodel ensembles. J. Climate , 27 , 6591–6611. DOI:10.1175/JCLI-D-13-00761.1
DOI:10.1007/s00376-008-0867-x Zhu C. W., Park C. K., Lee W. S., et al ,2008: Statisti-cal downscaling for multi-model ensemble prediction of summer monsoon rainfall in the Asia-Pacific re-gion using geopotential height field. Adv. Atmos. Sci. , 25 , 867–884. DOI:10.1007/s00376-008-0867-x