Improving the seasonal forecast for summertime South China rainfall using statistical downscaling

Authors


Abstract

[1] The performance of various seasonal forecast systems in predicting the station-scale summer rainfall in South China (SC) was assessed and was compared with that based on a statistical downscaling scheme. Hindcast experiments from 11 dynamical models covering the period of 1983–2003 were taken from the Asia-Pacific Economic Cooperation Climate Center multimodel ensemble. Based on observations, singular value decomposition analysis (SVDA) showed that SC precipitation is strongly related to the broad-scale sea level pressure (SLP) variation over Southeast Asia, western north Pacific, and part of the Indian Ocean. Analogous covariability was also found between model hindcasts and the observed station precipitation. Based on these results from SVDA, a statistical downscaling scheme for predicting SC station rainfall with model SLP as predictor was constructed. In general, the statistical scheme is superior to the original model prediction in two geographical regions, namely, western SC (near Guangxi) and eastern coastal SC (eastern Guangdong to part of Fujian). Further analysis indicated that dynamical models are able to reproduce the large-scale circulation patterns associated with the recurrent modes of SC rainfall, but not the local circulation features. This probably leads to erroneous rainfall predictions in some locations. On the other hand, the statistical scheme was able to map the broad-scale SLP patterns onto the station-scale rainfall anomalies, thereby correcting some of the model biases. Overall, our results demonstrate how SC summer rainfall predictions can be improved by tapping the source of predictability related to large-scale circulation signals from dynamical models.

1 Introduction

[2] Floods and droughts are a cause of serious social and economic losses in China [Chen, 1991]. This is particularly true for South China (SC), which is densely populated with a fast developing economy. Previous studies indicated that the summertime SC precipitation is affected by a number of climate systems, making its prediction very challenging. Wu and Wang [2000] and Wang et al. [2001] suggested that the western north Pacific summer monsoon (WNPSM) activity affects the seasonal mean precipitation over SC. Mao et al. [2011] showed that the dominant patterns of interannual variation of early summer SC rainfall themselves depend on the phase of the Pacific decadal oscillation (PDO). Ashok et al. [2004], Wang and Guan [2007], and Zhao et al. [2011] investigated the impact of Indian Ocean dipole (IOD) [Saji et al., 1999] on the summer SC precipitation. Yu et al. [2001] also argued that warming in the Indian Ocean affects summer monsoon rainfall over mid-eastern China and southeastern China. Tang and Sun [2005] illustrated how IOD and the subtropical high influence the summer SC precipitation. Other climate modes or climate elements such as El Niño–Southern Oscillation and the location/strength of the western Pacific subtropical ridge also have an impact on the SC circulation [Gong and Wang, 1999; Chang et al., 2000; Lu et al., 2002; Gong and Ho, 2002].

[3] Nowadays, general circulation model (GCM)–based dynamical forecast systems are commonly used for seasonal prediction. Although they have the ability to reproduce large-scale circulation, GCMs show very low skill in predicting the local rainfall. This is because the latter is strongly affected by topography or land-sea contrasts which are usually not well represented in coarse-scale models [e.g., Rodwell, 1998; von Storch et al., 1993]. In order to obtain regional climate information from coarse-scale GCM products, either dynamical or statistical downscaling can be employed [von Storch et al., 1993; Wilby and Wigley, 1997]. For statistical downscaling, the empirical relation between one local climate variable (the predictand) and the other (the predictor) is first sought [Zorita and von Storch, 1999; Wilby et al., 2004]. Forecasts of the local variable are then produced by projecting predictor values from GCMs into the corresponding statistical relationship. Liu et al. [2011] used the Antarctic Oscillation Index [Thompson and Wallace, 2000], 500 hPa geopotential height, 850 hPa humidity, and sea surface temperature as predictors and found increased skill in predicting the summer precipitation in southeastern China. Other predictors such as mean sea level pressure (SLP) [Cavazos, 1999; Wetterhall et al., 2005], geopotential height [Zhu et al., 2008], relative vorticity [Wilby et al., 1998], or even precipitation itself [Widmann et al., 2003] from coarse-resolution models were used for predicting the local-scale rainfall.

[4] Perfect prognosis (PP) and model output statistics (MOS) are the two commonly used statistical downscaling approaches [Wilks, 1995; see also the review by Maraun et al., 2010]. In the former approach, downscaling is constructed based on relationships between observed predictors and observed local predictands [Klein et al., 1959]. Such relationships can be found by, say, regression related methods, while the choice of predictors is often motivated by the associated atmospheric dynamics. By using model-generated variables as predictors in PP, one assumes that these variables are realistically simulated. The advantage of PP is that prediction equations (established from observational records) can be directly used to downscale GCM products without any modification. On the other hand, the performance of PP can be sensitive to model biases. In the MOS approach, statistical relationships between GCM products and observed predictands are established directly. By construction, therefore, statistical downscaling based on MOS can correct GCM errors. In other words, MOS combines bias correlation and downscaling in one step [Maraun et al., 2010]. In this study, an MOS-based statistical downscaling scheme for predicting the SC station-scale precipitation is constructed based on singular value decomposition analysis (SVDA). Later, it will be seen that there are biases in the GCM-simulated large-scale circulation such that the use of MOS, instead of PP, is necessary for this problem. The predictor values are taken from hindcasts of models comprising the Asia-Pacific Economic Cooperation Climate Center (APCC) multimodel ensemble (MME) [Min et al., 2011; Lee et al., 2011; Sohn et al., 2011]. It is well recognized that the MME approach can lead to more accurate forecasts owing to better sampling of model uncertainties [Krishnamurti et al., 1999; Doblas-Reyes et al., 2000; Shukla et al., 2000; Palmer et al., 2000]. The skill of both the statistically downscaled prediction and the spatially interpolated direct model output (DMO), from both individual models as well as their MME average, will be analyzed. Note that our approach is different from that of Liu et al. [2011], which is based on multiple linear regression between rainfall at each station and a predetermined set of climate indices. This is similar to the 1-D maximum covariance analysis approach [Widmann, 2005], which is an alternative to pattern-based methods. However, in this study, the pattern-based SVDA is preferred because of its ability to reveal any linkages between large-scale variables and local precipitation [Wilby and Wigley, 1997]. The rest of this paper is organized as follows. The description of the observational and model hindcast data sets, as well as the methodology, is given in section 2. In section 3, the relationship between SC rainfall and the large-scale circulation is presented, and the skills of seasonal prediction from dynamical models and statistical downscaling are compared. Finally, discussions and summary can be found in section 4.

2 Data and Methodology

2.1 Observations and Model Hindcast Data

[5] The observational data used in this study mainly consist of station precipitation during the season of June, July, and August (JJA) for the period of 1983–2003. Observations from 740 stations in SC within the domain of 18°N–27°N, 105°E–120°E, which includes Guangdong, Hainan, Fujian, Hunan, Jiangxi, and Guizhou provinces, as well as Guangxi, Hong Kong, and Macau, were examined. After screening out stations with missing values in the observational records, 89 stations in the SC region were selected. From Figure 1b, it can be seen that they have a rather uniform spatial distribution with about one to two stations within each 1º × 1º subregion. Besides station precipitation, mean sea level pressure (SLP) and 500 hPa geopotential height (Z500) from the National Centers for Environmental Prediction–Department of Energy (NCEP-DOE) Atmospheric Model Intercomparison Project (AMIP-II) reanalysis [Kanamitsu et al., 2002] with 2.5º × 2.5º resolution were also used.

Figure 1.

(a) Geographical map of China and provinces covering the study area (shading). (b) Location of selected stations (open circle).

[6] The models examined are the 11 climate models participating in the APCC MME seasonal forecast. Table 1 gives a description of the hindcast experiments. The experimental types are those consistent with either the Seasonal Model Intercomparison Project/Historical Forecast Project (SMIP/HFP) or the Coupled Model Intercomparison Project (CMIP). The former type includes forecasts from the Canadian Climate Centre second-generation [McFarlane et al., 1992] and third-generation general circulation models [Scinocca et al., 2008], and also the multilevel spectral primitive equation model [Ritchie, 1991] of the Meteorological Service of Canada (MSC), the Global Data Assimilation, and Prediction System (GDAPS) of the Korea Meteorological Administration (KMA) [Park et al., 2002], the Meteorological Research Institute Atmospheric General Circulation Model [Back et al., 2002] of the National Institute of Meteorological Research (NIMR), the Global Climate Prediction System (GCPS) from the Seoul National University (SNU) [Kang et al., 2004], and the second-generation global forecast system at the Central Weather Bureau (CWB) in Taiwan [Liou et al., 1997]. The latter type of experiments is those from the Predictive Ocean-Atmosphere Model for Australia (POAMA) of the Bureau of Meteorology Research Centre (BMRC), Australia [Zhong et al., 2005], the coupled general circulation model (CGCM) of the Beijing Climate Center (BCC), China [Ding et al., 2000], the Pusan National University (PNU) CGCM, Korea [Sun and Ahn, 2011], and the NCEP Coupled Forecast System (CFS) [Saha et al., 2006]. The common model hindcast period is from 1983 to 2003. For each set of model run, historical predictions for JJA were initialized in May with slightly different initial conditions for each member in the ensemble integration. Finally, meteorological variables including SLP and Z500 from each individual models as well as their MME average (defined as the simple average of outputs from all models) were considered. All model data were interpolated on a 2.5º × 2.5º regular grid.

Table 1. Description of the Model Hindcast Experiments Used in This Study
Member EconomyInstituteModelResolutionEnsemble SizeExperimental TypeReference
AustraliaBureau of Meteorological Research Centre (BMRC)Predictive Ocean-Atmosphere Model for Australia (POAMA)T47 L1710CMIPZhong et al. [2005]
CanadaMeteorological Service of Canada (MSC)MSC-GM2T32 L1010SMIP/HFPMcFarlane et al. [1992]
CanadaMeteorological Service of Canada (MSC)MSC-GM3T63 L3210SMIP/HFPScinocca et al. [2008]
CanadaMeteorological Service of Canada (MSC)MSC Spectral Primitive Equation Model (MSC-SEF)T95 L2710SMIP/HFPRitchie [1991]
ChinaBeijing Climate Center (BCC)BCC CGCMT63 L168CMIPDing et al. [2000]
South KoreaKorean Meteorological Administration (KMA)Global Data Assimilation and Prediction System (GDAPS)T106 L2120SMIP/HFPPark et al. [2002]
South KoreaNational Institute of Meteorological Research (NIMR)Meteorological Research Institute AGCM5º × 4º L1710SMIP/HFPBack et al. [2002]
South KoreaPusan National University (PNU)PNU CGCMT42 L185CMIPSun and Ahn [2011]
South KoreaSeoul National University (SNU)Global Climate Prediction System (GCPS)T63 L2112SMIP/HFPKang et al. [2004]
Chinese TaipeiCentral Weather Bureau (CWB)CWB AGCMT42 L1810SMIP/HFPLiou et al. [1997]
United StatesNational Centers for Environmental Prediction (NCEP)NCEP Climate Forecast System (CFS)T62 L6415CMIPSaha et al. [2006]

2.2 Statistical Downscaling

[7] SVDA [Bretherton et al., 1992; Widmann, 2005; Tippett et al., 2008] was employed in order to unveil any relationship between variability in the station precipitation and that in the large-scale circulation. Before applying SVDA, the linear long-term trends in the JJA mean rainfall as well as the gridded reanalysis and hindcast data were removed to minimize the influence of any decadal changes. (Note that common hindcast period of 1983–2003 is within the same positive phase of PDO, such that the SVDA results should not be affected by any change of the PDO phase.) In this study, SLP was chosen as the large-scale variable (or predictor). This is because GCMs in general can reasonably capture the large-scale SLP variations over the Indo-Pacific region (figures not shown). Moreover, it has strong covariability with the SC regional precipitation (predictand) (see Appendix A). Both the anomalous SLP and station precipitation can be expanded according to SVDA as follows:

display math

[8] Here the anomaly fields SLP(t,x) and precipitation(t,x) are normalized with unit standard deviation. N is the total number of SVD modes. SLPj(x) and Pi(x) represent the singular vectors for the ith and jth SVD mode, while Rj(t) and Qi(t) are the time expansion coefficients corresponding to the SLP and precipitation, respectively. Finally, for downscaling prediction, the following transfer function will be used:

display math

where bij are the coefficients relating precipitation and SLP. In general, [bij] is a matrix that can be determined by multiple linear regression (MLR) [Widmann, 2005; Tippett et al., 2008]. Chu et al. [2008] did not carry out MLR and approximated [bij] by a unit matrix (i.e., bij equals 1 when i = j, otherwise bij = 0). This is equivalent to saying that Qi and Rj are perfectly correlated; hence, precipitation prediction is downscaled by multiplying the precipitation singular vectors with the corresponding SLP expansion coefficients. The expansion coefficients Rj can be computed from the large-scale circulation anomalies, which are supposed to be well captured by GCMs. In this study, statistical downscaling approaches with and without using MLR are evaluated and compared. Finally, N is set to be 18 in the SVDA expansion equation, although our results are practically independent of the number of modes if N > 15.

3 Results

3.1 Relationship Between SC Rainfall and Model Variables

[9] SVDA between observed station precipitation and model SLP was first carried out, and the results for the leading mode are given in Figures 2-4. Figure 2 shows the leading precipitation pattern from SVDA with SLP taken from 11 different models and their MME average. It is worth mentioning that most of the singular vectors for rainfall compare well with observations (see Figure A1). We have also computed the pattern correlation between the observational and model results for the rainfall singular vector. Among the 11 models and MME, 9 of them give a pattern correlation higher than 0.6. From Figure 2, it can be seen that most models give strong rainfall-model SLP covariability with a pattern of surplus rainfall in coastal to eastern SC, and suppressed precipitation over the northwestern part of SC. The only exceptions are the NIMR model and POAMA. For NIMR, the strong positive rainfall signal in eastern SC is missing, while for POAMA, a clear eastern/coastal to western/inland dipole structure of rainfall cannot be found. The MME average also gives positive (negative) anomalous rainfall in eastern/coastal (western) SC, reflecting the leading rainfall pattern from the majority of models. Finally, this leading SVD mode explains about 50% or more of the squared covariance between observed rainfall and model SLP (with a maximum value of 77%) for all models except POAMA (which gives 38% of the fractional squared covariance).

Figure 2.

The leading singular vector for precipitation based on SVD analysis between observed station precipitation and model SLP from (a) BCC, (b) CWB, (c) GCPS, (d) GDAPS, (e) MSC-GM2, (f) MSC-GM3, (g) MSC-SEF, (h) NCEP, (i) NIMR, (j) PNU, (k) POAMA, and (l) the MME average. Upper right of each panel shows the fraction of squared covariance between station precipitation and model SLP explained by this SVD mode.

Figure 3.

Same as Figure 2 except for singular vectors for model SLP.

Figure 4.

Normalized time series of the expansion coefficient for precipitation (solid line) and model SLP (dashed line) from (a) BCC, (b) CWB, (c) GCPS, (d) GDAPS, (e) MSC-GM2, (f) MSC-GM3, (g) MSC-SEF, (h) NCEP, (i) NIMR, (j) PNU, (k) POAMA, and (l) the MME average, corresponding to the leading SVD mode.

[10] The model SLP patterns for this SVD mode are given in Figure 3. In general, they are also consistent with observations, with most models giving a prominent low-pressure anomaly extending from Indochina to South China Sea. However, details of the SLP pattern are different from one model to another. For instance, only BCC and PNU models can capture the positive signal in northwestern SC. Moreover, the high-pressure system located over the Indian Ocean is not found in BCC, MSC-SEF, NIMR, and POAMA hindcasts. The MME average also gives a large-scale low pressure anomaly over Indochina to SC, which is consistent with most models. Figure 4 gives the expansion coefficients for this SVD mode for each individual model and the MME average. The correlation between expansion coefficients of station precipitation and SLP is rather high for all models (ranging from 0.57 to 0.79). This is noteworthy because it indicates that the large-scale circulation in models is linked to the observed station precipitation. We also examined the second and third SVD modes between SC rainfall and model SLP. There is again relatively high correlation between the two sets of expansion coefficient time series (from the value of ~0.5 to 0.8), suggesting that the observed SC rainfall and the SLP field from model hindcasts have strong covariability.

3.2 Predictions Based on Direct Model Output and Statistical Downscaling

[11] Before assessing the performance of models in predicting the local SC rainfall, DMO of precipitation was first interpolated onto each station location. The correlation between the observed and DMO precipitation at 89 stations is shown in Figure 5. The 89-station averaged correlation coefficient is also given at the bottom right of each panel. It is striking that many models show no skill in rainfall prediction in the western part of SC. The exceptions are the hindcasts from BCC and NCEP, which perform poorly in eastern or eastern-to-central SC. Finally, the MME mean gives the highest skill score, as can be seen from its high value of the correlation coefficient averaged over all stations. However, even for the MME average, the skill in some western SC locations remains low; this might be related to the very low skill (and probably variance) of most models in this region. The performance of individual models can strongly affect the skill of MME, and hence, the MME technique cannot increase the skill in this particular subdomain.

Figure 5.

Correlation coefficients between the JJA precipitation at station locations based on observations and the interpolated DMO of precipitation from (a) BCC, (b) CWB, (c) GCPS, (d) GDAPS, (e) MSC-GM2, (f) MSC-GM3, (g) MSC-SEF, (h) NCEP, (i) NIMR, (j) PNU, (k) POAMA, and (l) the MME average. The correlation coefficient averaged over all stations is provided in the bottom right corner.

[12] Following the downscaling scheme outlined in section 2.2, the station-scale rainfall in SC was predicted based on model SLP outputs. Here the downscaled rainfall predictions were produced and validated based on a “leave-one-out” cross-validation framework. It involves making a single-year prediction with the target year excluded from the training period based on which the statistical scheme was constructed. This procedure was repeated each year, and the precipitation prediction was validated based on observations. The cross-validated correlation coefficient was used to evaluate the performance of downscaling SC rainfall prediction based on both methods (with and without MLR). The difference between the correlation coefficients given by DMO and those from non–MLR-based statistical downscaling is shown in Figure 6. For downscaling without MLR, the cross-validated correlation coefficient shows that most models perform better than DMO in many stations. For CWB and BCC, DMO gives the maximum correlation coefficient of just below 0.3 (see Figures 5a and 5b); downscaling can increase its value by ~0.4 or even more. For GCPS and GDAPS, there is an improvement in western SC, with the correlation coefficient increased by ~0.5 (see Figures 6c and 6d). Very similar improvement in the northwestern part of SC is also seen in NIMR, PNU, MSC-GM2, POAMA, and the MME mean hindcasts. For BCC and NCEP, the prediction skill is enhanced in the eastern part of SC. These results are consistent with previous studies showing that prediction skills can be greatly improved at locations where DMO performs poorly [Chen et al., 2012].

Figure 6.

Correlation coefficient difference between predictions based on non–MLR-based statistical downscaling and DMO.

[13] Figure 7 shows the difference between correlation coefficients for DMO and those from MLR-based statistical downscaling. Again, rainfall prediction at some stations has better skill than DMO. For instance, downscaling can increase the correlation coefficient by about 0.4 in MSC-SEF. For GDAPS and POAMA, their 89-station averaged correlations are increased by 0.10 and 0.11, respectively. Similar enhancement of skill in the western part of SC is seen in other models except BCC and NCEP, which show improvement in eastern-coastal SC. Finally, the improvement of skill brought about by MLR-based downscaling is comparable to that due to the non–MLR-based method. For the MME mean, MLR-based downscaling slightly outperforms that without MLR. (Besides temporal correlation, RMS error (RMSE) of DMO and the downscaling predictions were also computed. It was found that both downscaling methods give similar RMSE, but smaller than that for DMO (figures not shown).) Also note that there are a number of stations over which the statistical method underperforms DMO. Thus, depending on the location, statistical downscaling may or may not give better predictions compared to DMO (see also section 4).

Figure 7.

Same as Figure 6 except for the difference between MLR-based statistical downscaling and DMO.

[14] Further inspection reveals that improvement is mainly seen in two separate regions. For one group of models (referred to as type 1 models), improvement is found in northwestern SC; they are CWB, GCPS, GDAPS, MSC-GM2, MSC-GM3, MSC-SEF, NIMR, PNU, and POAMA. For BCC, NCEP, and POAMA, referred to as type 2 models, improvement is found in the eastern part and coastal area of SC. (Note that POAMA belongs to both type 1 and type 2, meaning that its skill is improved substantially in both areas after downscaling.) Based on such classification, ensembles comprising type 1 and type 2 models were formed, and the respective rainfall prediction based on DMO and statistical downscaling were compared. The difference between the correlation coefficients for downscaling and DMO for type 1 and type 2 model ensembles are shown in Figures 8a and 8b, respectively. It is obvious that type 1 (type 2) model ensemble predictions from statistical downscaling perform much better than DMO in northwest (eastern and coastal) SC. Finally, we have also compared DMO and downscaling predictions using the mean square skill score (MSSS) [WMO, 2002].

display math

where MSEfor and MSEclim refer to the mean square error of the forecast and mean square error in model climatology, respectively. It can be seen that MSSS is zero if the forecast error equals to the error in climatology. If MSSS is positive, it indicates that MSEfor is smaller than MSEclim. Conversely, a negative MSSS indicates MSEfor that is higher than MSEclim. Marked improvement in MSSS is observed for both type 1 model (from 0.04 to 0.12) and type 2 model (from 0.05 to 0.17) brought about by MLR-based statistical downscaling for rainfall prediction in western and eastern-coastal SC, respectively.

Figure 8.

Difference between the temporal correlation coefficients for DMO and MLR downscaled precipitation for (a) type 1 and (b) type 2 model ensemble.

[15] To shed light on how statistical downscaling can give better rainfall prediction in certain subregions in SC, we further examined the circulation features associated with the recurrent SC rainfall modes in both observations and model simulations. First, empirical orthogonal function (EOF) analysis of the observed SC rainfall was carried out. Then, both the observed and DMO data were regressed upon the second PC time series of SC rainfall from station observations. Figures 9a and 9b show the regression coefficients for the observed and type 1 model ensemble rainfall, respectively. (Note that the regression map for the observed rainfall has the same pattern as its second EOF.) It is noteworthy that this pattern resembles the leading singular vector of the SC rainfall (see Figure A1; the pattern correlation between the second EOF and the leading singular vector for rainfall is 0.74). In contrast, this rainfall pattern is not fully captured in the type 1 model environment. Corresponding to the same temporal variations of the observed second PC, models fail to predict suppressed rainfall in many of the inland SC stations (Figure 9b). Figures 9c and 9d give the corresponding regression coefficients for SLP. It can be seen that the second SC rainfall mode is associated with negative centers of action in northern Bay of Bengal and western north Pacific, and a positive anomaly in southwest China in observations (Figure 9c). This latter feature is consistent with the suppressed rainfall over the western to inland part of SC (see Figure 9a). On the other hand, such a positive SLP anomaly is not reproduced by the type 1 models (Figure 9d). This is probably the reason why there is no negative rainfall anomaly in the model ensemble. Finally, notice that the broad-scale features of the model SLP map are very similar to those from observations. This supports the notion that models have the ability to capture the large-scale circulation signals associated with the SC rainfall variations. Statistical techniques such as SVDA can map these circulation patterns to changes in the local rainfall, thereby producing skillful forecasts. For the type 2 models, similar analysis was also carried out by comparing the observed and model rainfall and SLP regression, based on the PCs of SC rainfall. Again the downscaling scheme can map the large-scale SLP anomaly and provides a “bias-corrected” rainfall prediction in the eastern and coastal SC locations (figures not shown). Downscaling can therefore enhance the rainfall prediction skill by type 2 models in this region.

Figure 9.

Regression coefficients of the JJA mean (a and b) rainfall (units: mm/d) and (c and d) SLP (contours in interval of 0.05 hPa, with negative values denoted by dashed lines) from (Figures 9a and 9c) observations and (Figures 9b and 9d) type 1 model ensemble average based on the second PC of the observed SC station rainfall.

Figure A1.

The dimensionless leading singular vector for (a) station precipitation and (b) SLP based on SVD analysis for station precipitation and SLP in JJA. Both rainfall and SLP fields are taken from observations. The upper right of Figure A1a shows the fraction of squared covariance between the two fields explained by this leading mode. (c) Normalized time series of the expansion coefficient for precipitation (solid line) and SLP (dashed line), corresponding to the leading SVD mode. Upper right shows the correlation coefficient between the two time series.

4 Discussion and Summary

[16] The relationship between summertime SC rainfall and the large-scale circulation over the Indo-Pacific sector has been studied, and a statistical downscaling scheme based on their covariability has been developed. SVDA was applied in order to examine the covariability between observed precipitation and model outputs. For the leading SVD mode, suppressed (enhanced) precipitation was found over the northwestern (eastern to southeastern) part of SC. This is accompanied by a large-scale SLP pattern with anomalously low pressure over Indochina and a large region in the western north Pacific. The correlation between expansion coefficients of station precipitation and model SLP is rather high (ranging from about 0.6 to 0.8). For the second and third singular vectors, there is still a strong resemblance with their observational counterpart in some models.

[17] The above gave us the confidence to construct a statistical scheme for predicting the SC summertime rainfall based on SVDA. In particular, the performance of both MLR-based and non–MLR-based downscaling prediction was assessed and compared to that from DMO. The latter was found to be skillful over some southern coastal locations, but otherwise, the skill is low in the inland region, especially over the western part of SC. On the other hand, the statistical downscaling schemes can greatly improve the prediction skill in western SC in most models; in coastal eastern SC, statistical downscaling also outperforms DMO for some models. We have also found that, for some models, non–MLR-based downscaling gives better results than that using MLR. This is the case despite MLR being able to capture the maximum amount of variance. It is plausible that MLR-based downscaling may not give the best performance under the cross-validation framework. More work needs to be done to understand the skill of such statistical downscaling scheme for actual predictions.

[18] Based on the area in which downscaling can improve the rainfall prediction, models were classified into two groups: for the type 1 models (including CWB, GCPS, GDAPS, MSC-GM2, MSC-GM3, MSC-SEF, NIMR, PNU, and POAMA), prediction over the western part of SC is improved significantly, while for type 2 models (including BCC, NCEP, and POAMA), the prediction skill is increased in eastern SC by statistical downscaling. Further analysis showed that, while models can reproduce the basin-scale circulation patterns associated with the recurrent SC rainfall modes, they have difficulties in capturing the details of the circulation pattern. For instance, type 1 models have skill in reproducing the regional circulation in the western north Pacific. However, these models have no skill in capturing the anomalous circulation pattern over western SC to Indochina, which can be important for the local rainfall variation. On the other hand, statistical downscaling can map the large-scale SLP patterns on local rainfall variability, thereby tapping the source of predictability from the large-scale circulation signals to enhance the prediction skill at some station locations. One way to utilize our result for actual forecasts, therefore, is to adopt statistical downscaling only in some selected regions. How DMO and statistical prediction products can be combined to improve the overall SC rainfall prediction should be further explored.

[19] We have also made use of Z500 to develop statistical downscaling using the same SVDA technique. It was found that analysis based on the Z500 variable gives a set of rainfall singular vectors similar to those based on SLP. Despite the variable being well simulated by GCMs, the skill of Z500-based downscaling is lower. This is probably because the variable cannot fully capture circulation changes in the low latitudes.

[20] Finally, we found that a number of models have considerable skill in capturing WNPSM activity (figures not shown). In view of its strong linkage with the circulation over SC, it seems likely that WNPSM is one major factor affecting the predictability of SC summertime rainfall. How WNPSM other climate models and PDO-related decadal variability affect the Asian monsoon rainfall and its predictability will be the subject of further studies.

Appendix A

Relationship Between Rainfall and Large-Scale Circulation From Observations

[21] Here the relationship between SC rainfall and the large-scale circulation from observations is examined. Based on correlation analyses between station precipitation and various circulation variables, the area of 10ºS–35ºN and 60ºE–180ºE was adopted for SVDA. Figure A1 shows the precipitation and SLP singular vectors associated with the first SVD mode. This mode explains more than 50% of the squared covariance between station rainfall and SLP. The precipitation pattern indicates strong positive anomalies (solid circles) over the coastal SC and Hainan Island, whereas suppressed rainfall is found (open circles) over the northwestern part of SC (Figure A1a). Associated with this rainfall pattern, anomalous low pressure is found over Hainan, Vietnam, and Indochina, and a large-scale anomaly of the same sign is seen over the western north Pacific. In the north-western SC area, south of Japan, and over the Indian Ocean to Indonesia, high pressure anomalies can be seen (Figure A1b). The aforementioned high-pressure system in northwestern SC is consistent with the suppressed precipitation there. Also, over eastern to coastal SC, above normal precipitation is observed where the anomalous SLP is negative. In other words, the placement of the anomalous highs (lows) and the suppressed (enhanced) rainfall in SC are consistent with each other.

[22] The standardized expansion coefficients for the station rainfall and SLP corresponding to this leading mode are given in Figure A1c. It is noteworthy that the two time series are highly correlated (with a correlation coefficient of 0.77), meaning that the SC precipitation is strongly coupled with large-scale SLP in the Indo-Pacific region. Note that the SLP pattern over the western north Pacific resembles the recurrent circulation associated with anomalous WNPSM activity [Wang et al., 2001]. In fact, the SLP expansion coefficient is highly correlated with the JJA Western North Pacific Monsoon Index (WNPMI, defined as the difference between the 850 hPa zonal wind averaged over 5ºN–15ºN, 100ºE–130ºE, and that over 20ºN–30ºN, 110ºE–140ºE) [Wang et al., 2001], with a correlation coefficient of 0.81. This indicates a strong relationship between SC rainfall and WNPSM activity, consistent with previous studies.

[23] For the second SVD mode, which accounts for about 18% of the squared covariance, a pattern of suppressed (enhanced) rainfall at almost all stations is associated with positive (negative) anomalous SLP covering SC and Indochina (figure not shown). This mode is related to developing IOD events, as evidenced by the correlation of −0.52 between the SLP expansion coefficient and the JJA mean dipole mode index (DMI) [Saji and Yamagata, 2003]. The third SVD mode explains about 12% of the rainfall-SLP covariability. It shows a connection between a northeast-to-southwest dipole SC precipitation pattern and SLP with a center of action covering Taiwan/western north Pacific (figure not shown). The correlations between rainfall and SLP expansion coefficients for the second and third mode are 0.62 and 0.66, respectively. Overall, it can be seen that there is strong covariability between SC rainfall and the large-scale circulation in the Indo-Pacific region.

Acknowledgments

[24] The authors appreciate those institutes participating in the APCC multimodel ensemble operational system for providing the hindcast experiment data. We thank Prof. Joong-Bae Ahn and Drs. Congwen Zhu and Hongwen Kang for discussions. Tony Tung is supported by the City University of Hong Kong (grant 7002512).

Ancillary