1Introduction

The East Asian monsoon affects approximately one-third of the global population, influencing the climate over large parts of Indochina, the Philippines, China, Korea, and Japan (Zhu et al., 1986; Ding, 1992). The South China Sea (SCS) summer monsoon (SCSSM) is an important subsystem of the East Asian summer monsoon. Since the SCS monsoon experiment in 1998 (Lau et al., 2000; Ding et al., 2004), researchers have paid much attention to the onset and variability of the SCSSM (Li and Zhang, 1999; Xu et al., 2001; He et al., 2003; Wen et al., 2006a). Along with the onset of the SCSSM, the integrated humidity and vertical velocity increase at 500 hPa over the SCS (Li and Qu, 2000), and monsoon flow transports moist air from the Indian Ocean and Pacific to East Asia (Wang et al., 2009a). Usually, a strong (weak) monsoon over the SCS leads to less (more) precipitation over the middle and lower reaches of the Yangtze River basin and more (less) precipitation in North China (Li et al., 2001; Ding et al., 2004; Li and Pan, 2007). Besides, the SCSSM variability can affect the weather and climate in North America via atmospheric teleconnection (Li and Zhang, 1999). Therefore, accurately predicting the SCSSM variability is crucial for improving the prediction of the weather and climate in East Asia and North America.

Because of the unique geographical location of the SCS (a boundary region between the Asian continent and the western Pacific), the SCSSM is under the influence of some key ocean regions, including the tropical Pacific (Chang et al., 2000; Wang et al., 2000; Gao et al., 2002; Zhou and Chan, 2007), North Pacific (Chan and Zhou, 2005; Wang et al., 2008c; Mao et al., 2011), and Indian Ocean (Wen et al., 2006b; Yang et al., 2007; Ding et al., 2010). After the late 1970s, both the intensity and the variability of the SCSSM undergo significant changes, with a number of researchers pointing out that the intensity of the SCSSM is weakening while its variability is increasing (Liang et al., 1999, 2007; Li et al., 2007). The increasing air-sea interactions are responsible for the change. After the late 1970s, El Niño-Southern Oscillation (ENSO) has a closer relationship with the SCSSM because the stronger air-sea interaction induces a longer lasting anomalous western North Pacific (WNP) anticyclone (WNPAC) in the decaying summer of El Niño events (Wang et al., 2008b, 2009a). The anomalous SST over the tropical Indian Ocean in the following summer of El Niño events also contributes to the persistence of the anomalous WNPAC via exciting the tropospheric Kelvin wave (Yang et al., 2007; Xie et al., 2009; Huang et al., 2010; Jiang et al., 2013). Moreover, after the mid-1970s, the strengthening intensity of spring North Pacific synoptic-scale eddy activity induces a stronger eddy feedback to the low-frequency mean flow. This leads to a stronger spring Arctic Oscillation (AO)-related cyclonic circulation over the tropical WNP. Thus, the relationship between the spring AO and the following SCSSM becomes closer after the late 1970s than before (Chen et al., 2015). Noticeably, the convection intensity over the Maritime Continent (MC) also strengthens after the late 1970s, which may weaken the cross-equatorial flow over the MC area and ultimately lead to a weak SCSSM (Gao and Xue, 2006; Zhang et al., 2015).

The prediction of the atmospheric seasonal climate over the Asian monsoon region mainly arises from atmospheric teleconnection associated with ENSO forcing and monsoon-ocean interactions in the Indian Ocean and WNP (Wang et al., 2008a, 2009b; Yang et al., 2008; Lee et al., 2011). Because the relationship between the SCSSM and key ocean areas is complicated and unstable, it is a great challenge to understand the variability and prediction of the SCSSM.

To provide accurate prediction, scientists concentrate their efforts in developing new climate models and proposing effective statistical forecasting methods. With the development of climate models, coupled ocean-atmosphere general circulation models (CGCMs) have been recognized as essential tools for climate prediction and diagnosis. In recent years, as scientists begin to realize that multi-model ensemble (MME) predictions can decrease the error caused by the uncertainty of a single model forecast (Palmer et al., 2004; Park et al., 2008; Wang et al., 2009b), a series of projects comparing state-of-the-art climate models have been conceived. Among these projects, the Ensemble-based Predictions of Climate Changes and Their Impacts (ENSEMBLES) and Development of a European Multimodel Ensemble System for Seasonal to Interannual Prediction (DEMETER) projects are two European Union funded integrated projects that aim to develop ensemble prediction systems for climate change based on the main state-of-the-art, high-resolution global models developed in Europe (Ingleby and Huddleston, 2007). Compared with DEMETER, the models used in ENSEMBLES are characterized by their higher resolution; longer hindcasts; more reasonable representation of sub-grid-scale physical processes; more accurate land, sea-ice, and greenhouse gas boundary forcing; and the use of assimilation for ocean initialization (Weisheimer et al., 2009; Li et al., 2012). Therefore, the ENSEMBLES prediction skill is more sophisticated over the tropical ocean (Alessandri et al., 2011). Although CGCMs exhibit positive skill for prediction in the tropics, they show poor performance in extratropical climate predictions. This is because the extratropics is situated far from the impacts of ENSO, and the positive prediction skill in CGCMs depends mainly on ENSO-monsoon teleconnection (Wang et al., 2008a, 2009b; Yang et al., 2008; Lee et al., 2011). Additionally, a number of studies have shown that the prediction skill of CGCMs for monsoon variability depends on the prediction of the climate mean state (Turner et al., 2005), the local heat exchange influenced by ENSO (Wang et al., 2003; Wu and Kirtman, 2005; Annamalai et al., 2007; Kim et al., 2012), and ENSO intensity (Wang et al., 2008a; Yang et al., 2008; Chowdary et al., 2010; Rajeevan et al., 2012). To improve the prediction skill for the East Asian summer monsoon and rainfall in China, a number of effective schemes have been proposed based on the year-to-year increment prediction (Fan et al., 2008), and effective hybrids of dynamical and statistical schemes have been developed (Lang and Wang, 2010; Fan et al., 2012; Sun and Chen, 2012; Liu and Fan, 2014). Among these studies, the yearto-year increment approach has been proven to significantly improve the prediction of summer rainfall in eastern China (Fan et al., 2008), the activity of WNP typhoons (Fan and Wang, 2009), East Asian summer monsoon (Fan et al., 2012), and Asian-Pacific Oscillation (Huang et al., 2014).

Based on DEMETER and ENSEMBLES, the present study seeks to address whether the prediction of the SCSSM has improved since the late 1970s. And if so, what factors determine the prediction of the SCSSM in CGCMs, and how can predictions of the SCSSM be further improved? The models and datasets used in this study are described in Section 2. A comprehensive assessment of the prediction of the SCSSM is presented in Section 3. The reasons for the improvement in the prediction of the SCSSM are presented in Sections 4. The year-to-year increment approach is employed to improve SCSSM prediction in Section 5. Discussion and summary is provided in Section 6.

2Models and datasets

The ENSEMBLES project is supported by the European Commission Seventh Framework Program and comprises five global CGCMs from the United Kingdom Met Office (UKMO), Météo France (MF), the ECMWF, the Leibniz Institute of Marine Sciences at Kiel University (IFM-GEOMAR), and the EuroMediterranean Center for Climate Change (CMCCINGV) in Bologna (van der Linden and Mitchell, 2009). The common hindcast period of all five models, with 2.5° × 2.5° horizontal resolution, is from 1960 to 2005. For each year, the seasonal forecasts are initialized on the first day of February, May, August, and November, and then run for seven months with nine members for each of the five individual models. In this study, we select the hindcast results starting on the first day of February, May, and November. The MME results are calculated through simple composite analysis, by applying equal weights to each of the five models in ENSEMBLES. Further details can be found in Weisheimer et al. (2009). Three CGCMs in the DEMETER project, with hindcast data initiated on the first day of May 1960-2001, are also analyzed in this study. The models are those developed by the ECMWF, the Centre National de Recherches Météorologiques (CNRM) of France, and the UKMO (Palmer et al., 2004). The MME of DEMETER is calculated as the average of the three models.

In order to evaluate the prediction abilities of the five models, several reanalysis datasets are also employed in the present study. The monthly mean reanalysis atmospheric datasets for the period 1960-2005, with 2.5° × 2.5° horizontal resolution, are derived from the NCEP-NCAR (Kalnay et al., 1996). The SST data from 1960 to 2005, with 1° × 1° horizontal resolution, are from the Centennial Observation-Based Estimates of SST dataset, version 2 (Hirahara et al., 2014). The monthly precipitation data, covering the period from 1960 to 2005 on a 2.5° × 2.5° grid, are from the combination of the ECMWF 40-yr Reanalysis (ERA-40; Uppala et al., 2005) and the interim ECMWF Reanalysis (ERA-Interim; Dee et al., 2011) of precipitation.

In this study, we use the SCS summer monsoon index (SCSSMI) defined by Wang et al. (2009a): SCSSMI=U850(5°-15°N, 110°-120°E)-U850(20°-25°N, 110°-120°E). The first and second terms on the right-hand side represent 850-hPa zonal wind (U850) averaged over 5°-15°N, 110°-120°E and 20°-25°N, 110°-120°E, respectively. A positive (negative) SCSSMI represents an active (a break) phase of the SCSSM. The Niño3.4 index is calculated through the seasonal mean of the SST anomalies in the Niño3.4 region (5°S-5°N, 170°-120°W). The WNPAC index used in this paper is that defined by Wang et al. (2001); namely, the meridional difference of the 850-hPa zonal wind anomalies between the southern (5°-15°N, 100°-130°E) and northern regions (20°-30°N, 110°-140°E), but with the sign reversed. The Indian Ocean Basin Mode (IOBM) index is defined as the SST anomaly averaged within 20°S-20°N, 40°-110°E (Jiang et al., 2013). The AO index is obtained from the website of the Climate Prediction Center, NOAA (http://www.cpc.ncep.noaa.gov/products/precip/CWlink/daily_ao_index/ao.shtml), which is defined on the basis of the standardized leading principal component time series of monthly mean Northern Hemisphere sea level pressure (SLP) for all months of the year. The MC index is defined as the precipitation averaged within 15°S-10°N, 95°-145°E (Zhang et al., 2015).

In this study, the year-to-year increment prediction approach proposed by Fan et al. (2008) is applied to the CGCMs, improving the prediction of SCSSM rainfall. This approach utilizes the year-to-year increment (or DY of a variable, i.e., a difference in a variable between the current year and the previous year) as the predictand. First, the DY of the variable is predicted; and then, the original variable is obtained by adding the predicted DY of the variable to the observed value from the previous year. The year-to-year increment prediction approach can amplify prediction signals and capture the interannnual and interdecadal variability of the variable well. Following Fan et al. (2012), the predicted SCSSM is obtained by adding the predicted DY of the SCSSM from the CGCM to the observed one.

In this paper, winter means the monthly average of December, January, and February; spring means the monthly average of March, April, and May; and summer refers to the monthly average of June, July, and August (JJA).

3Interdecadal change of the predictability of the SCSSM around the late 1970s 3.1SCSSMI

The SCSSMIs from 1960 to 2005 derived from the reanalysis data and the hindcast of the five individual models in ENSEMBLES and their MME are shown in Fig. 1. It is found that the predictability of the SCSSMI undergoes an interdecadal change around the late 1970s. As shown in Table 1, the correlation coefficients between the hindcast and reanalysis data for the periods 1960-1978 (the first period) and 1979-2001 (the second period) are calculated separately. Hindcasts from both DEMETER and ENSEMBLES show that the predictability of the SCSSMI is better after the late 1970s than before. For the first period, the correlation coefficient between the ENSEMBLES MME hindcast and the reanalysis SCSSMI is 0.28, while the coefficient derived from the DEMETER MME is 0.38. The results from the individual models range from -0.07 (MF in ENSEMBLES) to 0.42 (UKMO in DEMETER). For the second period, the correlation coefficient between the ENSEMBLES MME hindcast and the reanalysis SCSSMI is 0.72, and that produced by the DEMETER MME is 0.76. The correlation coefficients derived from the single models range from 0.56 (MF in ENSEMBLES) to 0.78 (ECMWF in DEMETER). The models and their MMEs can also capture the larger interannual variability of the SCSSM after the late 1970s than before. As shown in Table 1, the standard deviation (STD) of the SCSSMI in the first period is 0.71, while the value in the second period is 1.17. For the EMSEMBLES MME, the STDs before and after the late 1970s are 0.79 and 1.13, and they are 0.81 and 1.14 for the DEMETER MME. The five models also exhibit reasonable skill in predicting the larger interannual variability of the SCSSM after the late 1970s.

As indicated by the predictions, the predictability of the SCSSMI becomes better after the late 1970s than before. It is found that the models and their MME in DEMETER behave better than those in ENSEMBLES, despite the models in ENSEMBLES being superior to those in DEMETER, indicating that the skills of climate dynamic models have large room for improvement (Table 1). Therefore, we attempt to improve the prediction of the SCSSM by using the year-to-year increment prediction approach.

3.2Hindcasts of the interannual variation of circulation and precipitation before and after the late 1970s

First, we explore the ENSEMBLES MME’s performance in predicting the interannual variation of the atmospheric circulation, SLP, and precipitation related to the SCSSM. Because the SCSSM undergoes an interdecadal variation in the late 1970s, we divide the data period into two sub-periods: 1960-1978 and 1979-2005. The geographic distributions of time correlation coefficients (TCCs) between the reanalysis and the MME hindcast in terms of U850, zonal wind at 200 hPa (U200), SLP, and precipitation are shown in Fig. 2. After the late 1970s, TCCs are apparently higher than before, and the areas where TCCs exceed the 95% confidence level are larger than before, showing better skill in predicting the circulation from the surface to the upper troposphere, as well as the precipitation. For the lower-and upper-tropospheric circulation fields, compared with the period before the late 1970s, the MME better reproduces the SCSSM climate systems, such as the interannual variation of the SCSSM trough (Figs. 2a and 2b) and the Southern Hemisphere subtropical ridge (Figs. 2c and 2d), after the late 1970s. The prediction skill with respect to the interannual variation of SLP over the south MC and WNP is better after the late 1970s (Figs. 2e and 2f) than before. For the precipitation, before the late 1970s, the MME is barely able to predict the interannual variation of precipitation over Indonesia (Fig. 2g); while after the late 1970s, the MME can reasonably predict the interannual variation of precipitation over Indonesia and WNP (Fig. 2h). Overall, the MME apparently performs better for the period after the late 1970s than before.

3.3Hindcasts of anomalous circulation and precipitation in strong and weak SCSSM years before and after the late 1970s

We further explore the prediction of anomalous summer circulation and precipitation in strong and weak SCSSM years during 1960-1978 and 1979-2005, separately. Because the anomalous circulation and precipitation in strong and weak SCSSM years show opposite polarities, the composite difference in wind vectors at 200 hPa (UV200), wind vectors at 850 hPa (UV850), and precipitation, between the strong and weak SCSSM years, are presented. The strong (weak) SCSSM years are defined by the normalized values of the SCSSMI derived from the reanalysis data being greater (less) than 1.0 (-1.0) of the STD for 1960-1978 and 1979-2005, respectively. According to this criterion, for the period 1960-1978, three strong SCSSM years (1967, 1972, and 1978) and three weak SCSSM years (1964, 1966, and 1969) are identified. For the period 1979-2005, three strong SCSSM years (1981, 1990, and 2001), and four weak SCSSM years (1988, 1995, 1996, and 1998) are identified.

For the period 1960-1978, Fig. 3 shows the composite differences of atmospheric circulation and precipitation derived from the ENSEMBLES MME hindcast and the reanalysis results. In strong SCSSM years, the 200-hPa circulation difference field is characterized by southward cross-equatorial flow over the MC area, an anomalous cyclone over the west of Australia, and an anomalous anticyclone over the east of Australia (Fig. 3a). The composite differences of UV200 in the hindcast are slightly different from the characteristics mentioned above: the position of the major differences is to the east of those shown in the reanalysis (Fig. 3b). At 850 hPa, the hindcast also exhibits different characteristics compared with the reanalysis results: the anomalous cyclone at 850 hPa over the tropical WNP cannot be reproduced (Figs. 3c and 3d). As for difference in precipitation between the strong and weak SCSSM years, significant negative anomalies appear over Southwest China, the Bay of Bengal, India, and the South Indian Ocean, while significant positive anomalies are mainly distributed over the SCS, tropical WNP, and the east of Australia (Fig. 3e). However, the MME prediction cannot reasonably reproduce the precipitation anomalies (Fig. 3f).

For the period 1979-2005, the composite difference of the circulation and precipitation between the strong and weak SCSSM are predicted reasonably (Fig. 4). At 200 hPa (Figs. 4a and 4b), remarkable anomalous easterlies appear over the MC area in the strong SCSSM years. In addition, the southward cross-equatorial flow over East Asia and the subtropical high in the Southern Hemisphere are enhanced. At 850 hPa (Figs. 4c and 4d), an increasing westerly anomaly prevails over the Bay of Bengal, the SCS, and east of the Philippines. Meanwhile, an easterly anomaly occurs over the north of 20°N, forming an anomalous cyclone over the WNP. For the difference in the precipitation field, significant positive anomalies appear over the east of the Philippines, the South Indian Ocean, and the southeast of Australia, while negative anomalies occur over the region from North India to the islands of Indonesia (Fig. 4e). The MME can reasonably predict the characteristics and location of the rainfall anomalies, although the intensity is overestimated (Fig. 4f).

As shown in Figs. 3 and 4, the prediction skill in terms of the composite difference of the atmosphere and precipitation between the strong and weak SCSSM years is much better after the late 1970s than before. Further, we calculate the pattern correlation coefficients (PCCs) and root-mean-square errors (RMSEs) between the hindcast and the reanalysis fields. Excluding the PCC of meridional wind at 850 hPa (V850), all the PCCs of the anomalous circulation fields cannot pass the 95% confidence level before the late 1970s, whereas the PCCs all pass the 99% confidence level after the late 1970s. Furthermore, the PCCs of the precipitation anomalies are 0.11 for the period 1960-1978, and 0.39 for the period 1979-2005. As for the RMSEs of circulation and precipitation, there are no statistically significant differences between the two sub-periods.

3.4Hindcasts of anomalous circulation and precipitation in strong ENSO years before and after the late 1970s

We also evaluate the ENSEMBLES MME prediction skill regarding the anomalous circulation in summer after strong ENSO events in the two subperiods. As the CGCMs exhibit good skill in predicting the ENSO cycle, with the correlation coefficient of predicted winter Niño3.4 index and the reanalysis one reaching 0.97, statistically significant at the 95% confidence level (figure omitted), during the period 1960-2005, years with the previous winter’s Niño3.4 index being greater than 1.0 are marked as El Niño years, and those with Niño3.4 index lower than -1.0 are marked as La Nina years. According to this criterion, for the period 1960-1978, there are two El Niño years (1966 and 1973) and three La Nina years (1971, 1974, and 1976). For the period 1979-2005, there are four El Niño years (1983, 1987, 1992, and 1998) and three La Nina years (1989, 1999, and 2000).

Accordingly, we evaluate the forecasting capacity for the anomalous circulation (UV850 and UV200) and precipitation after El Niño and La Nina events for both of the sub-periods. The specific values of the PCCs and RMSEs between the hindcast and the reanalysis anomalies are shown in Table 2. In the summer after an El Niño event, for the period 1960-1978, there are no statistically significant differences in both the upper-and lower-tropospheric circulation (UV200 and UV850) over the area (40°S-30°N, 60°-180°E). The MME cannot predict the main feature of anomalous UV200 (Figs. 5a and 5b), but can reproduce the anomalous anticyclone over the WNP in the 850-hPa wind field (Figs. 5c and 5d). For the period 1979-2005, the influence of El Niño events on the summer circulation is stronger than before, and the predictions of anomalous circulation are better. For UV200, the MME can reasonably reproduce the significant northward cross-equatorial flow over Indonesia, as well as a pair of cyclones over the WNP and east of Australia (Figs. 6a and 6b). For UV850, the MME can reasonably predict both the strength and the location of the significant anomalous anticyclone over the WNP (Figs. 6c and 6d). The PCCs and RMSEs shown in Table 2 also demonstrate that the prediction of the anomalous wind fields during 1979-2005 is better than that during 1960-1978. For U200, the PCCs for 1960-1978 and 1979-2005 are 0.08 and 0.39, respectively. For U850, the PCCs for 1960-1978 and 1979-2005 are -0.29 and 0.16, respectively. As for precipitation during 1960-1978, significant positive anomalies occur over the tropical area (60°-170°E), while significant negative anomalies mainly occupy a small area in the middle of the Indonesian islands (Fig. 5e). However, the hindcast shows a large bias with respect to this pattern (Fig. 5f). As a result, the PCC between the reanalysis and the hindcast precipitation is negative (Table 2). For the period 1979-2005, significant positive anomalies of precipitation occur over the east of the Ryukyu islands, the tropical Indian Ocean, the Indonesian islands, and the east of Australia, while significant negative anomalies occur over 150°-180°E, east of the Philippines (Fig. 6e). The MME overestimates the values but it reproduces the location of the precipitation anomalies reasonably well (Fig. 6f). As shown in Table 2, the PCC is 0.53 and the RMSE is 0.84 mm day^-1.

In the summer after La Niña events, the prediction of anomalous circulation and precipitation for the period 1979-2005 is also better than that for the period 1960-1978. The PCCs and RMSEs of UV200, UV850, and precipitation between the reanalysis and the hindcast in two sub-periods are shown in Table 2. For the period 1960-1978, the PCC of U200 passes the 95% confidence level, and the PCCs of meridional wind at 200 hPa (V200), U850, V850, and precipitation are negative. For the period 1979-2005, except for U850, the PCCs for all the fields mentioned above pass the 95% confidence level. The RMSEs for the two sub-periods show no statistically significant difference.

4Possible reasons for the improved prediction of the SCSSM after the late 1970s

Previous studies have revealed that the WNPAC acts as a bridge between the tropical Pacific SST anomalies and the SCSSM, which can persist during the El Niño mature winter and the following spring through a positive wind-evaporation-SST feedback (Wang et al., 2000). As the anomalous WNPAC weakens the westerly over the SCS, the WNP subtropical ridge will extend farther to the west than usual. As a result of this change, a weaker SCSSM will appear (Chang et al., 2000; Wang et al., 2000, 2008b). The remote forcing of the warming Indian Ocean also contributes to the maintenance of the WNPAC in the summer of ENSO decaying years via exciting the equatorial Kelvin wave (Yang et al., 2007; Xie et al., 2009; Jiang et al., 2013). After the late 1970s, with global warming, as the air-sea interaction over the warm pool of the WNP is stronger, the relationship between ENSO and SCSSM has strengthened via the longer-lasting WNPAC. Besides, the linkage between the spring AO and the SCSSM also enhances after the late 1970s. This is because the stronger spring North Pacific synoptic-scale eddy activity induces a stronger eddy feedback to the low-frequency mean flow after the mid 1970s, and the spring AO-related cyclonic circulation anomaly over the tropical WNP is subsequently stronger and located more southward than before (Chen et al., 2015). But how well can the relationship between the aforementioned key oceans (tropical Pacific, North Pacific, and Indian Ocean) and the SCSSM be predicted in the ENSEMBLES models? As shown in Fig. 7, in both the reanalysis and the hindcast results, the correlation coefficients between anomalous SSTs and the SCSSMI for the period 1979-2005 are larger than those for the period 1960-1978. Before the late 1970s, the SST anomalies over the Pacific Ocean and the Indian Ocean have no statistically significant influence on the SCSSM. In contrast, after the late 1970s, the SSTs over the tropical Pacific, North Pacific, and tropical Indian Ocean have a statistically significant correlation with the SCSSM (all exceeding the 95% confidence level). Therefore, the intensified linkage between the SCSSM and key SSTs contributes to the better prediction of the SCSSM after the late 1970s.

Sun and Wang (2013) pointed out that larger variability will lead to better predictability. After the late 1970s, the SST variability over the North Pacific, tropical Pacific, and Indian Ocean becomes larger than before (Fig. 8a). The difference in SST variability derived from the ENSEMBLES MME has the same feature compared with the reanalysis (Fig. 8b). As a result, the ENSEBMLES models show better prediction of the SST over these key ocean areas after the late 1970s than before (figure omitted).

It is noted that the anomalous wind in the upper and lower troposphere in the following summer of El Niño is better predicted for the period 1979-2005 (Figs. 5 and 6). As shown in Figs. 9a and 9b, the regression of Niño3.4 index against the circulation at 850 hPa over the SCS shows a more significant anticyclone anomaly after the late 1970s, representing the longer lasting of the WNPAC in the decaying summer of El Niño events and the closer relationship between ENSO and the SCSSM. This change can be reasonably predicted by the ENSEMBLES MME (Figs. 10a and 10b). The regression of IOBM index against the circulation at 850 hPa features a stronger WNPAC in both the reanalysis (Figs. 9c and 9d) and hindcast (Figs. 10c and 10d), wherein the position of the warming Indian Ocean-induced WNPAC is more southward and westward after the late 1970s than before. As for the regression of spring AO index against the circulation at 850 hPa after the late 1970s, the position of the AOrelated cyclonic circulation anomaly over the tropical WNP is more southward and westward (Figs. 9e and 9f), showing the larger influence of the spring AO on the SCSSM. The hindcast can reasonably predict the southward and westward change of the cyclone over the WNP (Figs. 10e and 10f).

Notably, the intensity of MC convection can also change the SCSSM rainfall via influencing the intensity of the cross-equatorial flow over the MC and the related water vapor transport (WVT) (Gao and Xue, 2006; Zhang et al., 2015). As shown in Figs. 9g and 9h, the relationship between the MC rainfall and the SCSSM also changes after the late 1970s, which is characterized by a stronger cyclonic WVT anomaly induced by anomalous MC rainfall at 850 hPa over the SCS. This changing relationship can also be reasonably predicted (Figs. 10g and 10h). In conclusion, relative to the forecasting results for 1960-1978, the key SSTs, sea-atmosphere processes, as well as the WVT, associated with the SCSSM variability, can be predicted better after the late 1970s, which contribute to the higher prediction skill of the SCSSM.

5Improving the prediction of SCSSM rainfall

As discussed above in Section 3.1, although the models in ENSEMBLES are superior to those in DEMETER, the MME prediction skill for the SCSSM is lower than the DEMETER MME (Table 1), suggesting a large room for improvement in the skill, especially for SCSSM rainfall prediction. Therefore, in this section, we attempt to apply the year-to-year increment approach to improve the prediction skill for the SCSSM rainfall of DEMETER and ENSEMBLES. According to Fan et al. (2012), we develop an interannual increment SCSSM rainfall scheme, in which the predicted precipitation is obtained by adding the yearto-year increment of precipitation derived directly from the DEMETER and ENSEMBLES hindcasts to the observed one.

The TCCs between the original MME hindcast and the reanalysis of precipitation during 1960-1978 and 1979-2005 are calculated to evaluate the scheme skill score. The spatial distribution of the difference between the year-to-year increment approach improved TCCs and the original hindcast TCCs is shown in Fig. 11. The results show that the year-to-year increment approach indicates a higher accuracy level than the raw hindcasts of ENSEMBELS and DEMETER before and after the late 1970s, for most areas within the SCSSM domain. It is found that the prediction of SCSSM rainfall over most areas of the SCS is improved with the year-to-year increment approach. However, the precipitation over the southeast of China (Fig. 11a), tropical WNP (Fig. 11b), Southeast SCS, Southeast China (Fig. 11c), South Philippines, and South China (Fig. 11d), needs to be further improved. This is because the level of improvement of prediction is partly dependent on the skill of the CGCM in predicting the year-to-year increment of the climate variable (Fan et al., 2012).

6Summary and discussion

Our results show that the prediction of the SCSSM is much better after the late 1970s than before, and several probable reasons have been discussed. Because the skills of the ENSEMBLES and DEMETER models still have a large room for improvement, we develop a year-to-year increment scheme to improve the SCSSM rainfall.

It is found that the relationship between the key SSTs over most parts of the Pacific and Indian oceans are more significantly correlated with the SCSSM after the late 1970s than before (see Fig. 7). Moreover, the larger interannual variability of these key SSTs associated with the SCSSM after the late 1970s may contribute to the better prediction (see Figs. 9 and 10).

The WNPAC is a key bridge that connects the ocean anomalies over the tropical Pacific with the SCSSM, and modulates the SCSSM via influencing the position of the WNP subtropical ridge and the westerly over the SCS (Chang et al., 2000; Wang et al., 2000, 2008b). After the late 1970s, the tropical Pacific, North Pacific, and tropical Indian Ocean have strongly influenced the SCSSM, which is due to the stronger and more southward WNPAC than before. It is found that the position and intensity (see Figs. 5, 6, 9, and 10) of the anomalous WNPAC can be better predicted after the late 1970s. This may be the key reason for the improvement in the prediction of the SCSSM.

Besides, the convection intensity over the MC has strengthened after the late 1970s, resulting in a cyclonic WVT anomaly over the SCS (see Figs. 9h and 10h). The more reasonable prediction of the MC convection-induced WVT anomaly is also favorable for the better prediction of the SCSSM after the late 1970s.

Since the late 1970s, data quality has improved with the advent of satellite data being assimilated into models. Furthermore, the models in ENSEMBLES feature new approaches to mitigate the uncertainties in observations, and encompass better assimilation systems. The improvement in data quality and models may also be a reason for the better prediction of the SCSSM.

In addition, the interannual variability of the SCSSM for the period 1979-2005 is larger than that for the period 1960-1978. The reason for the larger variability may be due to the increasing amplitude of ENSO events since the late 1970s (Dai et al., 2000; Li et al., 2007; Wang et al., 2009a). The MME captures the larger interannual variability of both the key SSTs (see Fig. 8) and the SCSSM (Fig. 1 and Table 1) for the period 1979-2005, as larger variability of a climate variable leads to better prediction (Fan et al., 2008, 2012; Sun and Wang, 2013). The increasing of the interannual variability of the SCSSM may also contribute to the better prediction of the SCSSM for the period 1979-2005.

This study evaluates the prediction skill of the SCSSMI using two multi-model ensemble projects: DEMETER and ENSEMBLES. It is found that, although ENSEMBLES has made many improvements in aspects including higher resolution, more reasonable representation of sub-grid-scale physical processes, and better assimilation for initialization, its prediction skill of the SCSSMI is poorer than DEMETER’s. This indicates that there are more challenges in the prediction of the SCSSM. Therefore, we develop an interannual increment scheme applied to the CGCMS in DEMETER and ENSEMBLES to further improve the prediction skill for the SCSSM rainfall (Fig. 11). However, as indicated by Fan et al. (2012), the improvement of prediction is partly dependent on the ability of the CGCM to predict the year-to-year increment of the climate variable. Hence, there still exists poor prediction in the tropical WNP, Southeast SCS, and South Philippines. Thus, in subsequent work, we intend to develop a more effective scheme-for instance, a hybrid dynamical and statistical scheme-to improve the prediction of the SCSSM, focusing especially on the poor rainfall prediction over the aforementioned regions. In addition, we will investigate how well the variation in land-air interaction, sea ice, and mid-high latitude systems can influence the prediction of the SCSSM.