1. Introduction
The East Asian monsoon (EAM) influences a wide range of regions including eastern China, Korea, Japan, and the adjacent marginal seas. Its anomalous behaviors may result in disastrous winter snows and summer floods (Lau and Li 1984; Ding 2004). The regional economy and society are thus critically influenced by the evolution and variability of EAM. Better understanding and prediction of EAM variability will greatly benefit these populated regions.
In the past two decades, a large amount of efforts have been devoted to investigating EAM variability and its prediction (e.g., Ding 1994; Lau and Weng 2001; Ding 2004; Chang 2004). El Niño–Southern Oscillation (ENSO) has been regarded as a major factor modulating the interannual variability of EAM (e.g., Huang and Wu 1989; Chang et al. 2000; Lau and Weng 2001). A sandwich precipitation pattern with the Yangtze River and Huaihe River valleys being wet and both north and south China experiencing drought was found in ENSO-developing summer (Huang and Wu 1989). The most significant precipitation variability along the Yangtze River valley appeared in the mei-yu season after the ENSO peak phase for the period of 1951–99 (Chang et al. 2000). ENSO shows a delayed impact on EAM through Pacific–East Asia teleconnection (Wang et al. 2000).
The monsoon–warm pool interaction is also an essential process that determines the EAM system. The convectively coupled Rossby wave and sea surface temperature (SST) feedback can maintain both the western North Pacific anticyclone and SST anomalies, which provide a prolonged impact of ENSO on the East Asian summer monsoon (EASM) even when the SST anomalies in the eastern Pacific disappear (Wang et al. 2000). ENSO-induced monsoon anomalies change surrounding warm pool SST through surface heat fluxes or through Ekman transport of ocean heat (Wang 2008). The local atmospheric response to these SST anomalies opposes the remote response of monsoon flow to the SST anomalies in the eastern Pacific, thus providing an important negative feedback in the ocean–atmosphere coupled system.
Besides the influence from the ocean, the atmospheric circulation contributes significantly to EAM variations. The monsoon basic flow not only directly regulates the atmosphere–ocean interaction, but also modifies the monsoon response to remote ENSO forcing (Wang et al. 2003). Among all atmospheric circulations, the western Pacific subtropical high associated with EASM is mostly studied. The western Pacific subtropical high may be modulated by anomalous convection associated with the eastern Pacific SST anomaly through Hadley and Walker circulations (Chang et al. 2000). The western North Pacific wind anomalies develop rapidly in late fall of the years when a strong warm or cold ENSO event matures (Wang et al. 2000). It was found that the anomalous anticyclone increases precipitation over central China in summer by strengthening the western Pacific subtropical high and shifting it westward (Chang et al. 2000; Wu and Zhou 2008; Zhao et al. 2007).
The soil moisture and snow cover have also long been considered as sources of Asian monsoon variability, especially for the rainfall over the continental monsoon regions. Yasunari (1991) and Dirmeyer et al. (1999) pointed out that the land surface conditions in spring have an impact on the following summer monsoon. The spring snow cover has been shown to affect summer monsoon (Bamzai and Shukla 1999; Zhang et al. 2004).
Because of the complexity of EAM, an accurate dynamic prediction of it requires realistic modeling of ENSO, teleconnection associated with ENSO, warm pool ocean–monsoon interaction, and atmosphere–land surface interaction (Wang 2008). These requirements make the dynamic prediction of monsoons extremely challenging. Among all challenges, the forecast of monsoon precipitation remains the most difficult. In the past, substantial efforts have been devoted to the study of monsoon predictability by conducting Atmospheric Model Intercomparison Project (AMIP)-type experiments in which atmospheric general circulation models are constrained by realistic SST and sea ice (Gates et al. 1999). AMIP-type simulations reproduce some major twentieth-century climate events (Scaife et al. 2009), but EASM circulation remains poorly modeled (Zhou et al. 2009b). For example, the AMIP phase I models show a large bias in simulating the annual cycle of the eastern China monsoon rainfall (Liang et al. 2001). Atmospheric general circulation models (AGCMs) participating in phase II of the AMIP (AMIP II) show barely any skill in simulating the interannual variability of summer precipitation over the extratropical western North Pacific and the South China Sea (Zhou et al. 2009a). Since the unsatisfactory rainfall simulation is partly due to neglected local air–sea feedbacks in the AMIP-type simulation (Wang et al. 2005), ocean–atmosphere coupled models should do a better job on monsoon rainfall prediction. However, an evaluation of the performance of a seven-coupled-model ensemble, retrospective seasonal prediction (1980–2004), found nearly zero predictive skill in rainfall over East Asian land area in summer (Wang et al. 2008a). The coupled National Centers for Environmental Prediction (NCEP) climate forecast system (CFS) also performs poorly in capturing the interannual variability of the EASM circulation index (Yang et al. 2008). The predictive skill of coupled models may be compromised by model bias in the mean state (Cherchi and Navarra 2007).
Although the physical processes involved in the EAM system are complicated, the EAM interannual variability is essentially caused by internal feedback processes within the coupled EAM climate system (Liu et al. 2008). Using a multivariate empirical orthogonal function (MEOF) analysis on a set of six meteorological fields for boreal summer, Wang et al. (2008b) suggested that the leading mode of the interannual variation of EASM is primarily associated with the decaying phase of major El Niño, while the second mode is associated with the developing phase of El Niño/La Niña. The first two modes capture a large portion (50%) of the total variance of the precipitation and three-dimensional circulation. The dominance of the first two modes in the EASM interannual variability suggests a possible application of linear Markov model to EAM prediction. In this study, we use a technique combining MEOF analysis and linear Markov prediction. Our model results indicate that the dominant modes of the EAM variability are indeed predictable a few months in advance and that our simple statistical model can serve as a useful tool for EAM precipitation forecast.
2. Markov model construction
The model was constructed in the MEOF space. The base functions of the model's spatial dependence are the MEOFs of the physical variables chosen to define the state of the EAM climate. The temporal evolution of the model is a Markov process with its transition functions calculated from the corresponding principal components (PCs). By retaining only a few leading modes of the MEOFs, we can largely reduce the model space and filter out incoherent small-scale features that are generally unpredictable. This kind of statistical model has been used to predict ENSO (e.g., Xue et al. 1994; Johnson et al. 2000; Xue et al. 2000; Wu and Chen 2010) as well as Antarctic sea ice (Chen and Yuan 2004). While this study adopts a linear approach to making predictions, nonlinear data-driven approaches, such as neural networks, were also used for monsoon prediction (e.g., Navone and Ceccatto 1994; Sahai et al. 2000).
The model in this study was constructed following the same steps used by Chen and Yuan (2004) for their sea ice prediction model. Four atmospheric variables [precipitation, sea level pressure (SLP), zonal and meridional winds at 850 mb] along with SST and soil moisture were chosen to define the coupled EAM climate system. Other variables had also been included in test experiments, but their effects were found to be insignificant. The precipitation data are monthly averaged Global Precipitation Climatology Project (GPCP; Adler et al. 2003), version 2, data on 2.5° × 2.5° grid. Monthly-mean SLP and wind data at 850 mb are from the NCEP reanalysis (Kanamitsu et al. 2002), also on a 2.5° × 2.5° grid. The SST data are from the Extended Reconstructed Sea Surface Temperature (ERSST; Smith et al. 2008), version 3b, data at 2° × 2° grid. The monthly soil moisture data are from the Climate Prediction Center (CPC) with 0.5° × 0.5° resolution (Fan and van den Dool 2004), which were interpolated to the same grid of the precipitation data. The model domain for EAM covers a rectangular region (0°–50°N, 100°–140°E) that contains the cohesive variability among tropics, subtropics, and midlatitudes (Liu et al. 2008). Thirty years (January 1979–December 2008) of observational and reanalysis data were used for our analysis and model building.
The Markov model is considered season dependent in this study. A seasonal Markov model is more useful than a nonseasonal one in EAM precipitation forecast because EAM has distinctive seasonal evolution characteristics relative to other components of the Asian–Australian monsoon system (Wang et al. 2003). Details on Markov transition matrix construction can be found in Chen and Yuan (2004). The procedure to predict EAM consists of the following five steps. First, the climatologies based on the study period were subtracted to obtain anomalies for the six variables, and these anomalies were normalized and combined into a single matrix with all seasons. Second, spatial patterns (MEOFs) and their corresponding time series (PCs) were calculated by decomposing the multivariable matrix. Third, the monthly PCs were grouped into 12 subsets, one for each calendar month, and Markov transition matrices for each of the 12 calendar months were calculated. Fourth, the predictions of the PCs were made at increasing lead times by successively applying the Markov transition matrices. Finally, the predicted PCs were combined with the respective MEOFs to give forecasts of the selected variables.
3. Major modes of interannual variability of the EAM
Figure 1 shows the first three MEOF modes for all model variables, which account for 11.61%, 9.33%, and 6.71% of the total variance, respectively. The spatial patterns of the first mode precipitation are primarily characterized by a dipole structure (Fig. 1a), with an elongated band of positive anomalies extending from the middle and lower reaches of the Yangtze River and Huaihe River valleys (typical Chinese mei-yu region) to the southern Korean Peninsula and southern Japan, and a large area of negative anomalies over the South China Sea (SCS) and the tropical western North Pacific (WNP). The 850-mb southwesterly is considerably strengthened over southeast China and across the East China Sea to the south of Japan. The strength of the WNP anticyclone also increases. The precipitation intensifies accordingly along the East Asia subtropical front that stretches from the Yangtze River and Huai River valleys to South Korea and south Japan.
(left)–(right) First three MEOF modes showing (a)–(c) precipitation (color; mm day−1), SLP (contours; mb), and winds at 850 mb (vectors; m s−1) and (d)–(f) soil moisture (color; mm) and SST (contours; °C).
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
The precipitation of the second mode also shows a dipole pattern, with negative precipitation (high pressure) anomalies over the East Asian continent and positive precipitation (low pressure) anomalies over WNP (Fig. 1b). The pressure anomaly pattern implies an overall weakening of EASM. The 850-mb northeasterly is strengthened over northeastern China and the adjacent East China Sea. The WNP anticyclone weakens and the monsoon trough retreats southward. Thus, reduced precipitation is seen along the East Asia subtropical front. Note that Liu et al. (2008) found a sandwich precipitation pattern in the second MEOF mode of EASM, which is probably associated with the developing phase of ENSO (Wang et al. 2008b; Huang and Wu 1989). However, such a pattern is not present in our analysis of the full EAM. Unlike many previous studies, our focus here is not on EASM alone, but on EAM for all seasons.
In the third MEOF mode (Fig. 1c), precipitation is intensified at latitudes south of 10°N and in the middle reach of the Yangtze River, while it is reduced in southeast China and the adjacent seas. Generally speaking, all three modes indicate that precipitation has the largest variances along the East Asian subtropical front and mei-yu region (20°–33°N, 110°–130°E), and in the tropical WNP monsoon trough (0°–16°N, 110°–140°E), with the anomalies in these two regions tending to be out of phase. Therefore, our hindcast and forecast experiments discussed later are mainly based on these two regions.
As shown in Figs. 1d–f, the MEOF patterns of soil moisture are less well defined. For the second and third modes, there appears to be a good correspondence between soil moisture and precipitation anomalies over land, but such a relationship is not obvious for the first mode. Also shown in Figs. 1d–f are the corresponding MEOF patterns of SST, with cooling in the first and second modes over all oceanic area and weak warming in the third mode over a large portion of the model domain except in regions near shore and north of 33°N.
To build a Markov model with MEOF bases, one practical problem we should consider is the number of leading modes retained for forecast. Using too few modes may miss predictable signals, and too many may contaminate the model with noise. The right number is achieved by trial and error, and seven modes are retained in our forecast experiments shown later. Figure 2 compares the contributions from different numbers of MEOF modes to the precipitation anomalies averaged in the 110°–140°E zonal band. When only the first mode is present, precipitation variability is limited almost entirely to the latitudes of the SCS and tropical WNP. When the first three modes are retained, some anomalies appear at latitudes between 20° and 30°N. The first three modes appear to capture most of the WNP variance. When the first 7 modes are included, all major anomalous precipitation signals in EAM are captured, and the differences between the seven-mode case and the total precipitation field are basically random noise. Large precipitation variations are found during major ENSO events, such as in 1982/83 and 1997/98, showing evidence of teleconnection between the Pacific and East Asia during the extreme phases of the ENSO cycle (Wang et al. 2000). Although the first 7 MEOF modes only account for about 47% of the total variance, they seem to contain most of the predictable signals in the precipitation field.
Averaged precipitation anomalies (mm day–1) in the 110°–140°E zonal band during 1979–2008 period. The leftmost panel is raw data, the four middle panels are MEOFs with different numbers of modes included, and the rightmost panel is the residual of raw data minus the first seven MEOF modes.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
4. Model results
a. Hindcast
Hindcast experiments were performed for the period from January 1979 to December 2008. For each set of experiments a series of Markov models were constructed with different model variables included and with a different number of MEOFs retained. Figure 3 shows the correlation skills of these Markov models in predicting the average precipitation anomaly in the WNP region. It is clear that a multivariate model is always better than the precipitation-only model in the hindcast experiments, and the best model is the one with all variables retained. Including SST information seems to have the largest impact, probably because the ocean contains a large portion of the system's memory. Figure 4 shows the corresponding model skills for the mei-yu region. The model does not do as well over land in the mei-yu region as in the WNP region, and such a regional contrast has also been found in previous numerical model predictions (Wang et al. 2008a). Again, multivariate models are generally better than the model with precipitation alone. It is interesting to note that in the mei-yu region, including soil moisture has the largest impact as compared to other variables, which is consistent with the suggestion that soil moisture is an important initial condition for monsoon forecast (Yasunari 1991). For both WNP and mei-yu regions, the model skill seems to increase with increasing number of retained MEOF modes, but this is not generally true when the skill is cross validated, in which case seven retained modes appear to be the optimal choice (to be shown later).
Model-observation precipitation correlation in WNP region (0°–16°N, 110°–140°E) as functions of lead months for different variables included in the model: (a) precipitation only, (b) precipitation and SST, (c) precipitation and soil moisture, (d) precipitation and SLP, (e) precipitation and wind, and (f) precipitation and all other variables. The different curve colors are for different numbers of retained MEOFs.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
As in Fig. 3, but in the mei-yu region (20°–33°N, 110°–130°E).
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
There may be some artificial skill in the hindcast results shown in Figs. 3 and 4 since the model was trained with the same datasets as those used to initialize and verify the model. A more convincing way to evaluate model hindcast skill is to use a cross-validation scheme (Barnston and Ropelewski 1992), in which the data used to verify the model hindcasts are not used for model training. In a cross-validation scenario, one year of data is removed and a Markov model is trained upon the remaining years and verified at the removed year. The 1-yr window is moved forward month by month until the end of the time series is reached. In our study, there are a total of 360 (360 = 30 × 12) multiple analyses with different years removed. Figure 5 shows the cross-validated correlations between model-predicted and observed precipitation anomalies at different lead times. Here the Markov model has all four atmospheric variables plus SST and soil moisture, with seven MEOF modes retained. The hindcast skill is considerable in the WNP region even at a 5-month lead time, while it is relatively poor over land. Figure 6 shows the cross-validated correlation skills at different lead times and for different seasons. At 0-month lead, the correlation skill is generally high over most of the model domain, indicating the projection of the observed variables into reduced model space was successful. The hindcast skill is quite remarkable in the WNP region, with significant correlations at lead times up to 4–6 months in boreal spring [March–May (MAM)] and winter [December–February (DJF)], and up to 1–3 months in boreal fall [September–November (SON)]. In contrast, the model has very little skill in boreal summer. The relatively poor overall skills in Fig. 5 are mainly caused by the poor skills in boreal summer. This indicates that model predictions can hardly cross boreal spring. In other words, there is a “spring barrier” in monsoon precipitation forecast just as in ENSO SST forecast. This is not surprising since the interannual variability of the tropical precipitation is primarily modulated by ENSO.
Cross-validated correlation between model hindcast and observed precipitation anomalies for different lead times of 1, 3, 5, and 7 months as indicated at the top of each panel. The correlation within a 95% confidence t-test interval is contoured.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
Cross-validated correlation between model hindcast and observed precipitation anomalies for (top)–(bottom) DJF, MAM, JJA, and SON and (left)–(right) 0, 1–3, and 4–6 months lead time. The lead time is defined as the time between the start month and the first month of target forecast season.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
However, the model skill may not be so poor for all summer conditions when individual years were examined. For example, Fig. 7 displays the model hindcasts at different lead times for the summer of 1994, when an ENSO event was developing, and for the summer of 2000, when an ENSO event was decaying. The first column (0-month lead) is simply the observations represented by the first seven MEOF modes and thus can be considered as the target to predict. In the ENSO developing summer, the model shows no prediction skill after one month, while in the ENSO decaying summer there is considerable skill at lead times of 4–6 months. The western North Pacific wind anomalies develop in the late fall of strong ENSO years (Wang et al. 2000), which may lead to good monsoon prediction in the ENSO decaying phase next summer.
Cross-validated model hindcasts at (left to right) 0, 1–3, and 4–6 months lead time for anomalous precipitation in (top) June–August (JJA) 1994 and (bottom) JJA 2000. The 0-month lead can be regarded as the target season to predict.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
b. Sensitivity test
The model sensitivities to the number of retained MEOF modes and to the variables to include have been discussed earlier (Figs. 3, 4) in terms of non-cross-validated hindcast skill. Here we further test these sensitivities using cross-validated hindcasts. Figure 8 shows the cross-validated correlation skill and root-mean-square (rms) error in predicting the average precipitation anomaly in the WNP region. The Markov models used for this test keep the full range of variables, including precipitation, wind, SLP, SST, and soil moisture, but differ in the number of MEOF modes retained. It is obvious that the model hindcasts beat the persistence prediction by a large amount in terms of both correlation skill and rms error. It is also clear that the model performance is not very sensitive to the number of MEOF modes retained, underscoring the fact that the EAM system is dominated by a few distinctive modes, which makes our model approach quite robust. We chose the model with seven modes as our standard model for hindcast and forecast experiments simply because it has a slightly better skill at lead times of a few months. A different choice would not change the outcome of this study in any significant way.
Cross-validated (top) correlation and (bottom) rms error (mm day−1) between hindcast and observed precipitation anomalies averaged in WNP region (0°–16°N, 110°–140°E). Compared are six model hindcast experiments with different numbers of MEOF mode retained. The skill of persistence prediction is also shown for reference.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
The contributions of different model components are evaluated in Fig. 9, where cross-validated model skills in the WNP region are shown for six cases with different variables included in the model. Consistent with Fig. 3, a multivariate model is generally better than the precipitation-only model. Including SST, SLP, wind, and even soil moisture over land can all help to improve the model skill, but SST is the single most effective variable in the WNP region. For comparison, the skill of a Markov model for SST alone is also shown in Fig. 9, which turns out to be comparable to the skill of the precipitation-alone model, indicating the necessity of multivariable approach. The best model performance is achieved when all variables are included, but the contributions of these variables are not accumulative because they are not independent from one another and must have redundant information in them. The same sensitivity test was carried out for the mei-yu region, and generally similar results were obtained (not shown). The main differences are that the model skill is poorer in the mei-yu region and that soil moisture has a stronger impact there (as in Fig. 4).
As in Fig. 8, but with different model variables included. Seven MEOF modes are retained for these experiments.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
c. Forecast
Finally, we present the result of an experiment that mimics a “real-time” forecast in which the target period to predict is beyond the training period of the model. Figure 10 shows the 1–3-month and 4–6-month lead forecasts of precipitation initialized from October–December (OND) of 2008, along with the observations averaged in January–March (JFM) and April–June (AMJ) of 2009. The main features of the observed precipitation anomaly fields are well predicted, but some mismatches remain. The model is able to capture the observed wetter-than-normal condition in the SCS and tropical WNP as well as the drier-than-normal condition in southeast China and along the East Asian trough. The longer lead forecast underestimates the magnitude of the precipitation anomalies, a common problem with all statistical models. The model fails to pick up the small-scale anomalies over land, such as the increased precipitation in northeast China and the decreased precipitation in central China in AMJ. This is probably because of fact that the model only contains a few dominant modes of variability in the EAM region.
Comparison of (top) observations with (bottom) precipitation forecasts (mm day−1). The model was initialized using the observations in each month of OND 2008, and forecasts were made for each month of (left) JFM 2009 (1–3-month lead) and (right) AMJ 2009 (4–6-month lead). The 3-month mean precipitation is averaged after the forecast.
Citation: Journal of Climate 26, 14; 10.1175/JCLI-D-12-00408.1
5. Summary and discussion
In this study, we developed a linear Markov model to predict EAM, with emphasis on precipitation variability. Four atmospheric variables along with SST and soil moisture were chosen to define the state of the EAM climate, and the multivariate empirical orthogonal functions of these variables were used to build the Markov model. The forecast skill of the model was evaluated in a cross-validated fashion, and a series of sensitivity experiments were conducted to further validate the model. In both hindcast and forecast experiments, the model showed considerable skill in predicting the precipitation anomaly a few months in advance, especially in boreal winter and spring. The forecast skill is comparable to present climate forecast systems (e.g., prediction from Met Office in the United Kingdom). Our statistical approach is simpler and less expensive to develop and thus provides a good alternative choice.
The predictability of the EAM interannual variability demonstrated in this study can be attributed to the domination of the regional atmosphere–ocean–land coupled system by a few distinctive, slowly changing modes. The physical processes involved in the EAM system are complicated, but a well-constructed Markov model can capture the statistical characteristics of the system. After picking up the major signals from the observed initial conditions, the model can go on predicting the growth and decay of these initial signals in a statistically meaningful way. Such a statistical model can serve as a useful forecasting tool as long as it contains the dominant climate modes associated with the predictand, as demonstrated by the previous work on ENSO and Antarctic sea ice as well as the present study on EAM.
The model can predict EAM precipitation variability at lead times up to 4–6 months in boreal winter and spring, and up to 1–3 months in fall, but its performance in boreal summer is rather poor. It seems that the model can hardly forecast across spring—a phenomenon similar to the well-known “spring barrier” in ENSO prediction. Since the interannual variability of tropical precipitation is primarily modulated by ENSO, it is not surprising to find a spring barrier in EAM precipitation forecasts. However, it is interesting to note that the model has better skill in ENSO-decaying summers than in ENSO-developing summers. It seems that an ENSO event that matures in winter has a strong and lasting effect on the western North Pacific that not only makes monsoon precipitation more predictable during winter and spring but also in the following summer.
The self-evolving Markov model emphasizes the local atmosphere–ocean–land interaction in the EAM region, but the success of this local model does not compromise the importance of large-scale teleconnections. For instance, the EAM system is influenced by ENSO in the first place, and such influence is contained in the initial conditions of the model. Thus, the subsequent evolution of the initial signals in the model implicitly includes the ENSO-EAM teleconnection. Similarly, the remote effect of spring snow cover on EAM could be accounted for. The snowmelt of the Tibetan Plateau results in a surface cooling over the plateau and the neighboring regions, leading to high-pressure anomalies that cause a more northwestward extension of the western Pacific subtropical high in the subsequent summer (Zhang et al. 2004). Thus, by including these pressure and wind anomalies, the Markov model implicitly takes into account the variability of snow cover over the Tibetan Plateau.
Acknowledgments
Funding for this study is provided by the National Science Foundation of China (40976018, 41276030), National Basic Research Program (2013CB430302), and Ocean Public Welfare Scientific Research Project (201105018). NOAA and NASA are acknowledged for providing data for this study. Helpful comments by three anonymous reviewers are greatly appreciated.
REFERENCES
Adler, R. F., and Coauthors, 2003: The Version-2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 1147–1167.
Bamzai, A. S., and J. Shukla, 1999: Relation between Eurasian snow cover, snow depth, and the Indian summer monsoon: An observational study. J. Climate, 12, 3117–3132.
Barnston, A. G., and C. F. Ropelewski, 1992: Prediction of ENSO episodes using canonical correlation analysis. J. Climate, 5, 1316–1345.
Chang, C. P., 2004: The East Asian Monsoon. World Scientific, 560 pp.
Chang, C. P., Y. S. Zhang, and T. Li, 2000: Interannual and interdecadal variations of the East Asian summer monsoon and tropical Pacific SSTs. Part I: Roles of the subtropical ridge. J. Climate, 13, 4310–4325.
Chen, D., and X. Yuan, 2004: A Markov model for seasonal forecast of Antarctic sea ice. J. Climate, 17, 3156–3168.
Cherchi, A., and A. Navarra, 2007: Sensitivity of the Asian summer monsoon to the horizontal resolution: Differences between AMIP-type and coupled model experiments. Climate Dyn., 28, 273–290.
Ding, Y., 1994: Monsoons over China. Kluwer Academic, 419 pp.
Ding, Y., 2004: Seasonal March of the East Asian Summer Monsoon. World Scientific, 560 pp.
Dirmeyer, P. A., A. J. Dolman, and N. Sato, 1999: The global soil wetness project: A pilot project for global land surface modeling and validation. Bull. Amer. Meteor. Soc., 80, 851–878.
Fan, Y., and H. van den Dool, 2004: Climate Prediction Center global monthly soil moisture data set at 0.5° resolution for 1948 to present. J. Geophys. Res., 109, D10102, doi:10.1029/2003JD004345.
Gates, W. L., and Coauthors, 1999: An overview of the results of the Atmospheric Model Intercomparison Project (AMIP I). Bull. Amer. Meteor. Soc., 80, 1962–1970.
Huang, R., and Y. Wu, 1989: The influence of ENSO on the summer climate change in China and its mechanisms. Adv. Atmos. Sci., 6, 21–32.
Johnson, S. D., D. S. Battisti, and E. S. Sarachik, 2000: Empirically derived Markov models and prediction of tropical Pacific sea surface temperature anomalies. J. Climate, 13, 13–17.
Kanamitsu, M., W. Ebisuzaki, J. Woollen, S.-K. Yang, J. J. Hnilo, M. Fiorino, and G. L. Potter, 2002: NCEP–DOE AMIP-II Reanalysis (R-2). Bull. Amer. Meteor. Soc., 83, 1631–1643.
Lau, K.-M., and M.-T. Li, 1984: The monsoon of East Asia and its global associations—A survey. Bull. Amer. Meteor. Soc., 65, 114–125.
Lau, K.-M., and H. Weng, 2001: Coherent modes of global SST and summer rainfall over China: An assessment of the regional impacts of the 1997–98 El Niño. J. Climate, 14, 1294–1308.
Liang, X. Z., W. C. Wang, and A. N. Samel, 2001: Biases in AMIP simulations of the east China monsoon system. Climate Dyn., 17, 291–304.
Liu, J., B. Wang, and J. Yang, 2008: Forced and internal modes of variability of the East Asian summer monsoon. Climate Past, 4, 225–233.
Navone, H. D., and H. A. Ceccatto, 1994: Predicting Indian monsoon rainfall: A neural network approach. Climate Dyn., 10, 305–312.
Sahai, A. K., M. K. Soman, and V. Satyan, 2000: All India summer monsoon rainfall prediction using an artificial neural network. Climate Dyn., 16, 291–302.
Scaife, A. A., and Coauthors, 2009: The CLIVAR C20C project: Selected twentieth century climate events. Climate Dyn., 33, 603–614.
Smith, T. M., R. W. Reynolds, and J. Lawrimore, 2008: Improvements to NOAA's historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 2283–2293.
Wang, B., 2008: Thrusts and prospects on understanding and predicting Asian monsoon climate. Acta Meteor. Sin., 66, 653–669.
Wang, B., R. Wu, and X. Fu, 2000: Pacific–East Asian teleconnection: How does ENSO affect East Asian climate? J. Climate, 13, 1517–1536.
Wang, B., R. Wu, and T. Li, 2003: Atmosphere–warm ocean interaction and its impact on Asian–Australian monsoon variation. J. Climate, 16, 1195–1211.
Wang, B., Q. Ding, X. Fu, I.-S. Kang, K. Jin, J. Shukla, and F. Doblas-Reyes, 2005: Fundamental challenge in simulation and prediction of summer monsoon rainfall. Geophys. Res. Lett., 32, L15711, doi:10.1029/2005GL022734.
Wang, B., and Coauthors, 2008a: Advance and prospectus of seasonal prediction: Assessment of APCC/CliPAS 14-model ensemble retrospective seasonal prediction (1980–2004). Climate Dyn., 33, 93–117.
Wang, B., Z. Wu, J. Li, G. Wu, J. Liu, C.-P. Chang, and Y. Ding, 2008b: How to measure the strength of the East Asia summer monsoon. J. Climate, 21, 4449–4463.
Wu, B., and T. Zhou, 2008: Oceanic origin of the interannual and interdecadal variability of the summertime western Pacific subtropical high. Geophys. Res. Lett., 35, L13701, doi:10.1029/2008GL034584.
Wu, Q., and D. Chen, 2010: Ensemble forecast of Indo-Pacific SST based on IPCC twentieth-century climate simulations. Geophys. Res. Lett., 37, L16702, doi:10.1029/2010GL044330.
Xue, Y., M. A. Cane, S. E. Zebiak, and M. B. Blumenthal, 1994: On the prediction of ENSO: A study with a low-order Markov model. Tellus, 46, 512–528.
Xue, Y., A. Leetmaa, and M. Ji, 2000: ENSO prediction with Markov models: The impact of sea level. J. Climate, 13, 849–871.
Yang, S., Z. Zhang, V. E. Kousky, R. W. Higgins, S. H. Yoo, J. Liang, and Y. Fan, 2008: Simulations and seasonal prediction of the Asian summer monsoon in the NCEP Climate Forecast System. J. Climate, 21, 3755–3775.
Yasunari, T., 1991: The monsoon year: A new concept of the climate year in the tropics. Bull. Amer. Meteor. Soc., 72, 1331–1338.
Zhang, Y., T. Li, and B. Wang, 2004: Decadal change of the spring snow depth over the Tibetan Plateau: The associated circulation and influence on the East Asian summer monsoon. J. Climate, 17, 2780–2793.
Zhao, P., R. Zhang, J. Liu, X. Zhou, and J. He, 2007: Onset of southwesterly wind over eastern China and associated atmospheric circulation and rainfall. Climate Dyn., 28, 797–811.
Zhou, T., B. Wu, and B. Wang, 2009a: How well do atmospheric general circulation models capture the leading modes of the interannual variability of the Asian–Australian monsoon? J. Climate, 22, 1159–1173.
Zhou, T., and Coauthors, 2009b: The CLIVAR C20C Project: Which components of the Asian–Australian monsoon circulation variations are forced and reproducible? Climate Dyn., 33, 1051–1068.
