The NCEP Climate Forecast System (CFS) is an important source of information for seasonal climate prediction in many Asian countries affected by monsoon climate. The authors provide a comprehensive analysis of the prediction of the Asian summer monsoon (ASM) by the new CFS version 2 (CFSv2) using the hindcast for 1983–2010, focusing on seasonal-to-interannual time scales. Many ASM features are well predicted by the CFSv2, including heavy monsoon rainfall centers, large-scale monsoon circulation patterns, and monsoon onset and retreat features. Several commonly used dynamical monsoon indices and their associated precipitation and circulation patterns can be predicted several months in advance. The CFSv2 has better skill in predicting the Southeast Asian monsoon than predicting the South Asian monsoon. Compared to CFS version 1 (CFSv1), the CFSv2 has increased skill in predicting large-scale monsoon circulation and precipitation features but decreased skill for the South Asian monsoon, although some biases in the CFSv1 still exist in the CFSv2, especially the weaker-than-observed western Pacific subtropical high and the exaggerated strong link of the ASM to ENSO.
Comparison of CFSv2 hindcast with output from Atmospheric Model Intercomparison Project (AMIP) and Coupled Model Intercomparison Project (CMIP) simulations indicates that exclusion of ocean–atmosphere coupling leads to a weaker ASM. Compared to AMIP, both hindcast and CMIP show a more realistic annual cycle of precipitation, and the interannual variability of the ASM is better in hindcast. However, CMIP does not show any advantage in depicting the processes associated with the interannual variability of major dynamical monsoon indices compared to AMIP.
The Asian summer monsoon (ASM), which is characterized by the heaviest seasonal precipitation and the largest heat source for the atmospheric circulation over the globe, exhibits variability on a wide range of time scales (e.g., Webster et al. 1998; Ding and Sikka 2006; Goswami 2006; Waliser 2006; Yang and Lau 2006; Li et al. 2010; Jiang and Li 2011). Owing to its strong economic and societal impacts, prediction of the ASM is of central importance (e.g., Yang et al. 2008; Wang et al. 2008).
Dynamical predictions of the ASM have advanced significantly in recent decades. Charney and Shukla (1981) hypothesized monsoon predictability based on the influence of boundary forcing at the earth’s surface. Bengtsson et al. (1993) proposed a two-tier approach in which sea surface temperature (SST) was first predicted by coupled models and then atmospheric anomalies were predicted by atmospheric general circulation models (AGCMs) forced by the predicted SST. However, assessment of the ASM predicted by simulations of the Atmospheric Model Intercomparison Project (AMIP)-type integrations indicated that uncoupled models may have apparent deficiencies (Kang et al. 2002; Wang et al. 2004; Zhou et al. 2009; Li and Zhang 2009). In recent years, one-tier systems using coupled ocean–atmosphere models have been used in dynamical monsoon prediction. Coupled seasonal prediction systems have shown reasonable prediction of the ASM, although the coupled ocean–atmosphere models still have large biases (e.g., Wang et al. 2008; Yang et al. 2008). At present, both one-tier and two-tier systems are used for operational climate prediction. Ongoing assessment of AGCM predictions forced by prescribed SST and of ocean–atmosphere coupled model predictions by operational prediction models provides further understanding about the advantages and disadvantages of the two types of prediction systems.
The National Centers for Environmental Prediction (NCEP) Climate Forecast System (CFS), which is a fully coupled forecast system, provides operational prediction of the Asian monsoon climate (Saha et al. 2006). Yang et al. (2008) provided a comprehensive assessment of simulation and prediction of the ASM by the first version of the CFS (CFSv1), which became operational in August 2004. The CFSv1 demonstrated reasonable skills in predicting the ASM, although deficiencies were also noted (Yang et al. 2008; Drbohlav and Krishnamurthy 2010; Gao et al. 2011). In March 2011, a new version of the CFS, CFS version 2 (CFSv2), replaced CFSv1. Compared to CFSv1, the CFSv2 incorporates a number of new physical packages for cloud–aerosol–radiation, land surface, ocean, and sea ice processes and a new atmosphere–ocean–land data assimilation system (Saha et al. 2010). Previous studies (e.g., Yuan et al. 2011) have shown increased skill in predicting global land precipitation and surface air temperature from CFSv1 to CFSv2. However, many features about the simulation and prediction of ASM by the CFSv2 remain undocumented: they are the focus of this study. We focus on the climatological features, annual cycle, and interannual variability of the monsoon and compare the major monsoon features between CFSv1 and CFSv2.
Previous efforts have been devoted to identifying the source of predictability of the ASM. Skill of the model’s prediction of the ASM generally increases with a decrease in lead time, increase in model resolution, inclusion of atmosphere–ocean–land coupling, and improvement in model physics (Yang et al. 2011; Wen et al. 2012). Ocean–atmosphere coupling, in particular, is considered as a very important factor of monsoon simulation and prediction (Wang et al. 2005; Kumar et al. 2005). While coupled models simulate air–sea interaction over the ASM region more reasonably compared to AMIP experiments (Wang et al. 2005), to what extent ocean–atmosphere coupling affects the ASM in the prediction model CFSv2 remains unclear. In this paper, we will show the differences in ASM simulations by CFSv2 hindcasts, AMIP-type simulations, and Coupled Model Intercomparison Project (CMIP)-type free runs to understand how ocean–atmosphere coupling affects the ASM in the dynamical prediction model.
Because of the complexity of the monsoon, measuring monsoon variability appropriately is not an easy task. Given the superiority of monsoon indices in measuring major monsoon features (e.g., Miyakoda et al. 2003), previous studies have applied commonly used monsoon indices to evaluate a model’s ability in predicting monsoon variability (e.g., Yang et al. 2008; Drbohlav and Krishnamurthy 2010). However, a good prediction of monsoon indices does not necessarily guarantee a good prediction of their related monsoon processes (e.g., Wang et al. 2008; Zhou et al. 2009). The ability of models in monsoon prediction is often based on how accurately they simulate the physical processes responsible for monsoon variability. Thus, it is necessary to understand how well the atmospheric features that are related to monsoon index variability are also predicted.
The rest of this paper is organized as follows. Descriptions of the CFSv2, three different types of CFSv2 integrations, and observations used to verify CFSv2 output are given in section 2. Predictions of monsoon precipitation, monsoon circulation patterns, and monsoon indices in the CFSv2, as well as a comparison between CFSv1 and CFSv2, are discussed in section 3. In section 4, we discuss monsoon predictions with time leads. In section 5, we investigate the role of ocean–atmosphere coupling in monsoon prediction by comparing hindcast, AMIP-type simulation, and CMIP-type simulation. A summary and further discussion of the results obtained are provided in section 6.
2. Model, experiments, and observational data
The NCEP CFSv2 is a fully coupled dynamical prediction system (Saha et al. 2012, manuscript submitted to J. Climate). It consists of the NCEP Global Forecast System at T126 resolution, the Geophysical Fluid Dynamics Laboratory Modular Ocean Model version 4.0 at 0.25°–0.5° grid spacing coupled with a two-layer sea ice model, and the four-layer Noah land surface model.
Outputs from three types of simulations with the CFSv2 are used: hindcast, AMIP-type simulation, and CMIP-type simulation, which are referred to as hindcast, AMIP, and CMIP for brevity. Outputs from a CFSv2 9-month hindcast are analyzed over a 28-yr period from 1983 to 2010. Beginning on 1 January, 9-month hindcast runs were initiated from every fifth day and run from all four cycles of that day. The initial days vary from one month to another. A detail about the initial time can be found at http://cfs.ncep.noaa.gov/cfsv2.info/ (see file “Retrospective CFSv2 Forecast Data Information”). An ensemble mean of the monthly mean values of 24 members is used, with initial dates after the 7th of the particular month used as the ensemble member for the next month. For a June–August (JJA) 0-month lead forecast, the ensemble mean of the runs initialized from 5 June and 11, 16, 21, 26, and 31 May is used as the forecast. The longest 7-month lead forecast for JJA is an ensemble mean of the runs initialized from 2 and 7 November and 8, 13, 18, 23, and 28 October. The AMIP simulations are an ensemble mean of 11 integrations by the atmospheric component of CFSv2, which were all initialized from January 1950 with different atmospheric initial conditions. Monthly SST and sea ice from the Hadley Centre Sea Ice and Sea Surface Temperature dataset (HadISST) (Rayner et al. 2003) and optimally interpolated (OI) SST (OISST) (Reynolds et al. 2007) are used as the boundary conditions for 1950–2008 and 2009–10. These runs are forced with observed monthly CO2 concentration as well. The CMIP integration, initialized on 1 January 1988, is run for 48 years, and the data from the last 28 years are analyzed.
The observations used for model verification include the Climate Prediction Center Merged Analysis of Precipitation (CMAP) (Xie and Arkin 1997), the winds and temperature from the NCEP Climate Forecast System Reanalysis (CFSR) (Saha et al. 2010), and the SST from the NOAA OISST analysis (Reynolds et al. 2007). With the first guess from a coupled atmosphere–ocean–sea ice–land forecast system, the CFSR has improved the climatological precipitation distribution over various regions and the interannual precipitation correlation with observations over the Indian Ocean (IO), the Maritime Continent, and the western Pacific compared to several previous reanalyses (Wang et al. 2011).
3. Prediction of monsoon precipitation, circulation, and dynamical indices
In this section, we discuss the climatological features and interannual variations of the ASM. We assess the performance of CFSv2 by comparing the 0-month lead predictions with observations.
a. Climatological features of precipitation and atmospheric circulation
Figure 1 shows the JJA mean precipitation and winds at both 850 and 200 hPa, in hindcast and observations, along with their differences. Several precipitation centers are observed over western India, the northern Bay of Bengal (BoB), Bangladesh/Burma, the Indo-China Peninsula, the South China Sea (SCS), and east of the Philippines (Fig. 1a). There is also heavy precipitation over the tropical central IO. The mei-yu rainband over eastern China, the baiu over Japan, and the changma over Korea can be seen as well. Comparison between Figs. 1a and 1b indicates that the hindcast captures the general features observed, including the locations and magnitude of precipitation centers. However, the hindcast overestimates the precipitation over the west of India, the eastern BoB, west of the Philippines, the equatorial central and eastern IO, the Himalayas, and Japan and underestimates the precipitation over the northwestern BoB, northern India, and the SCS (Fig. 1c).
The ASM is characterized by lower-tropospheric southwesterly monsoon flow from the Arabian Sea to the western North Pacific, the cross-equatorial flow over the western IO (the Somali jet) and the Maritime Continent, the monsoon troughs over the BoB and the Philippines, the western Pacific subtropical high (WPSH) (see Fig. 1a), and the upper-tropospheric easterly monsoon flow associated with the South Asian high (Fig. 1d). The hindcast captures the general features of monsoon flows in both the lower and upper troposphere (Figs. 1b,e). However, it predicts weaker-than-observed WPSH and monsoon trough over the Indian subcontinent, as well as a stronger-than-observed Somali jet and westerlies over the equatorial central and eastern IO in the lower troposphere (Fig. 1c). Compared with observation, the hindcast has biases in the southerly flow over South Asia in the lower troposphere and the anticyclonic circulation over the Asian continent and the western North Pacific in the upper troposphere (Fig. 1f). Comparison between Figs. 1c and 1f shows that the bias of hindcast in predicting monsoon flow is dynamically consistent with the deficiency of the model in predicting monsoon precipitation.
Yang et al. (2008) reported that the CFSv1 overestimated the precipitation over the eastern Arabian Sea and simulated a weaker-than-observed upper-tropospheric easterly monsoon flow associated with a weaker South Asian high. Compared with CFSv1 (Figs. 1 and 2 in Yang et al. 2008), the CFSv2 demonstrates an apparent improvement in simulating monsoon flows and precipitation. The CFSv2 clearly improves the deficiency in which the CFSv1 overestimated the precipitation over the eastern Arabian Sea and the associated lower-tropospheric cyclonic circulation and underestimated the intensity of the South Asian high, with weaker-than-observed 200-hPa easterlies over tropical Asia and westerlies over the midlatitudes. The CFSv2, however, shows some noticeable new biases—for example, overestimation of precipitation over the equatorial central and eastern IO, which is dynamically coupled with convergence at the lower troposphere and divergence at the upper troposphere (Figs. 1c,f). This bias is accompanied by a weaker-than-observed South Asian monsoon, with the weaker-than-observed southerlies in the lower troposphere and northerlies in the upper troposphere. More discussion of the weak South Asian monsoon in CFSv2 will be provided later.
b. Annual cycle of monsoon precipitation
The Asian monsoon exhibits a strong annual cycle, and the onset and retreat of the monsoon vary from one place to another (Jiang and Li 2011; Yang et al. 2011). Here, the annual cycles of monsoon precipitation along three different longitudinal sections are examined. Over 70°–90°E (Figs. 2a,d), the onset of observed monsoon precipitation features heavy rainfall that advances northward from the equator in spring to subtropical India in July and then retreats southward to the equator in winter, maximizing around 15°N during summer. The hindcast captures the general features of monsoon onset and retreat (Fig. 2d). However, it fails to capture the shift of the maximum precipitation center from the equator to around 15°N, owing to stronger-than-observed rainfall over the equatorial central and eastern IO (Figs. 1c, 2d). A deficiency in simulating the rainfall over the Himalayas is also seen in the annual cycle of precipitation (Fig. 2d).
Over 90°–110°E (Fig. 2b), the annual cycle of observed monsoon precipitation is characterized by an abrupt onset and a gradual retreat, with a rapid northward jump of the maximum rainfall center from the equator to north of 10°N in May. This jump of the rainfall center is caused by a quick northward shift of the meridional maximum SST (Jiang and Li 2011). The hindcast predicts the abrupt onset and the gradual retreat very well including the timing and the magnitude of precipitation, except for the stronger-than-observed precipitation to the north of the equator during October and November (Fig. 2e).
Over 110°–130°E (Figs. 2c,f), the structure of the annual cycle of observed monsoon rainfall is more complicated than that of the previous two longitude bands. The maximum precipitation center also shows an abrupt northward jump from the equator to around 15°N in May, corresponding to the monsoon onset over the SCS (Lau and Yang 1997; Qian and Yang 2000). After the monsoon onset, one maximum precipitation center advances northward, corresponding to commencement of the rainfall season in East Asia. However, the other remains around 15°N, corresponding to the eastward onset of the western Pacific monsoon (Wu and Wang 2001). The former rainfall center shows an abrupt southward retreat in September, while the latter retreats gradually after August. There is a premonsoon rainfall belt around 25°N, resulting mainly from frontal systems (Wan and Wu 2007). Although the annual cycle of monsoon precipitation is complicated, the hindcast captures the major features reasonably well, except for the southward shift of the maximum rainfall center around 15°N.
Compared with CFSv1 (Fig. 3 in Yang et al. 2008), the CFSv2 predicts a more realistic intensity of precipitation in July over 70°–90°E, with an amount of 10 mm. This advancement is attributed to the CFSv2’s improvement in simulating the precipitation over the eastern Arabian Sea. The CFSv2 also captures a more realistic retreat of the precipitation, while it depicts a more southward precipitation maximum. Over 110°–130°E, the CFSv2 predicts a more realistic monsoon advance, compared to CFSv1, with maximum precipitation in August. However, the amount of maximum precipitation is underestimated by CFSv2. In both CFSv2 and CFSv1, the precipitation maximum over 110°–130°E is located too far southward.
c. Dynamical monsoon indices
We further analyze the interannual variability of the ASM measured by several popular dynamical monsoon indices, which include the Webster–Yang (WY) (Webster and Yang 1992), the South Asian (SA) (Goswami et al. 1999), and the Southeast Asian (SEA) (Wang and Fan 1999) monsoon indices. The WY index is defined as the vertical shear of zonal winds between 850 and 200 hPa averaged over 0°–20°N, 40°–110°E, and the SA index as the vertical shear of meridional winds between 850 and 200 hPa averaged over 10°–30°N, 70°–110°E. The SEA is defined as the horizontal zonal wind shear at 850 hPa between 5°–15°N, 90°–130°E and 22.5°–32.5°N, 110°–140°E.
As shown in Fig. 3, the hindcast captures the WY index very well, including both magnitude and interannual variability (R = 0.76) (Fig. 3a). The interannual variability of the SEA index is also well predicted in the hindcast (R = 0.77), while the magnitude of the regional monsoon is stronger than observation, caused mainly by the weaker-than-observed WPSH (Figs. 1c, 3c). However, the hindcast shows weaker-than-observed SA index, with a moderate correlation with the observed SA index (R = 0.40) (Fig. 3b). As mentioned previously, the weaker-than-observed SA index, that is, weaker-than-observed monsoon meridional circulation, is associated with the stronger-than-observed convection over the equatorial central and eastern IO, which weakens the SA index by strengthening the local Hadley circulation. (Note that the monsoon meridional circulation is opposite to the Hadley cell.)
Yang et al. (2008) reported that the CFSv1 captured the interannual variations of the three indices very well but produced weaker-than-observed monsoon circulation measured by WY, associated with a systematic cold bias over the Asian continent. Compared with the dynamical monsoon indices predicted by CFSv1 (Fig. 4 in Yang et al. 2008), the CFSv2 demonstrated better (worse) skill in predicting the WY (SA) index. The deterioration in predicting the SA index from CFSv1 to CFSv2 may lead to difficulty in studying the Indian summer monsoon using the CFSv2. To understand more about the improvement of WY prediction by CFSv2, we further examine Fig. 4, which shows the surface air temperature in CFSv2. Overall, the systematic cold bias over the Asian continent in CFSv1 (Fig. 5 in Yang et al. 2008) is no longer a problem in CFSv2, which even produces a slight warm bias in the northwest of the Tibetan Plateau. The warmer-than-observed temperatures over the Arabian Peninsula, Africa, and the extratropical North Pacific agree with the features in CFSv1. The overall improvement in surface air temperature may be attributed to the replacement of the two-layer Oregon State University land surface model in CFSv1 by the four-layer Noah model in CFSv2 (see also Yang et al. 2011).
4. Monsoon predictions of different time leads
In this section, we discuss prediction of the ASM as a function of lead time, focusing on the forecast and predictability errors of various monsoon components. A forecast error is defined as the difference between hindcast prediction and observation, and a predictability error is defined as the difference between two model forecasts, which target the same month but with different initial conditions (Drbohlav and Krishnamurthy 2010). We choose the difference between two consecutive leads as the predictability error for providing the smallest error in the initial conditions for ensemble mean. For example, the predictability error of 1-month lead is the difference between 1-month lead and 0-month lead predictions.
a. Forecast errors and predictability errors
As shown in Figs. 5a–d, the forecast errors of precipitation share some major features in forecasts from 1-month lead to 7-month lead, which include the overestimation to the west of India, to the west of the Philippines, over the eastern BoB, the equatorial central and eastern IO, the Himalayas, and Japan. The features also include the underestimation over the northwestern BoB, northern and western India, the SCS, central eastern China, and the Korean Peninsula. These features are consistent with the differences between 0-lead prediction and observation (Fig. 1c). The prediction errors of 850-hPa winds exhibit cyclonic circulation over the western North Pacific, anticyclonic circulation over the Indian subcontinent, and westerlies from the equatorial central IO to the equatorial western Pacific. The excessively predicted precipitation over the eastern IO and the tropical western North Pacific is associated with a convergent wind pattern over these regions. The anticyclonic circulation over the Indian subcontinent is accompanied by deficient precipitation over the most of the Indian subcontinent and the western BoB. Prediction errors of both precipitation and 850-hPa winds over the Indian subcontinent and the western BoB grow as lead time increases. However, the prediction errors over the equatorial central and eastern IO and the western Pacific decrease as the lead time increases, especially the bias of the subtropical easterlies.
As shown in Figs. 5e–h, large predictability errors occur only in the first three lead months. However, the patterns of errors vary in different leads. The predictability errors of 1-month lead are concentrated over tropical Asia and the western Pacific, with anticyclonic circulation over the Arabian Sea, cyclonic circulation over the BoB and SCS, and westerlies from the equatorial eastern IO to the equatorial western Pacific. Deficient precipitation is seen over the eastern Arabian Sea and northern India and excessive precipitation over the eastern SCS and the tropical western North Pacific (Fig. 5e). From 1-month lead to 2-month lead, precipitation is deficient over the eastern Arabian Sea, the Indian subcontinent, and the western North Pacific, accompanied by excessive precipitation over the eastern BoB and the southeastern SCS (figures not shown). Northeasterlies appear over the Arabian Sea, easterlies over the equatorial central and eastern IO, and westerlies over the subtropical and equatorial western Pacific (figures not shown). From 2-month lead to 3-month lead, a large precipitation deficit emerges over the tropical western North Pacific, accompanied by anticyclonic circulation over the western North Pacific and easterlies over the northern IO (Fig. 5f). After a 4-month lead, changes in the model’s climatology occur only over the equatorial western Pacific and decrease as lead time increases (Figs. 5g,h).
The occurrence of large predictability errors in the first three lead months suggests that the bias of ASM prediction is sensitive to the initial conditions of the first few months. Comparison of the magnitude between forecast errors and predictability errors shows that forecast errors originate mainly from model imperfection, for example, unrealistic parameterization of cloud. A decrease in predictability errors with lead time has also been reported for CFSv1 (Drbohlav and Krishnamurthy 2010) and other GCM simulations (e.g., Shukla 1998). Shukla reported that the tropical wind and rainfall patterns in certain tropical regions were strongly determined by SST and did not depend on atmospheric initial conditions. That is, for long lead time predictions, the predicted state is mostly determined by SST. The difference in predicted climatological SST between two consecutive leads decreases as lead time increases (figures not shown). Therefore, the differences in predicted climatological precipitation and winds between two consecutive leads also decrease as lead time increases. Furthermore, the predicted state by an ocean–atmosphere coupled model will converge in model climatology after long integrations. Thus, the difference in predicted climatology between two long-lead predictions is smaller than that of two short-lead predictions.
b. Prediction of dynamical monsoon indices as a function of lead time
As shown in the hindcast of 0-month lead, the CFSv2 can well capture the interannual variations of several popular dynamical monsoon indices (Fig. 4). How do the prediction skills vary as lead time increases? Figure 6 shows the correlation of monsoon indices between observation and hindcast predictions for different lead months. The model predicts the interannual variability of large-scale monsoon circulation very well, including the WY monsoon index and the SEA monsoon index, with correlation coefficients above 0.37 (corresponding to the 95% confidence level for a Student’s t test) when lead time is less than 5 months. However, it shows much lower skill in predicting the interannual variability of the meridional circulation over South Asia, with a correlation coefficient of about 0.2 for just the 2-month lead prediction. The skills of WY and SEA predictions decrease quickly from 0-month lead to 3-month lead. This characteristic is consistent with the above analysis of predictability errors, which shows a large predictability error just in the first three lead months. The hindcast shows overall low skill in the 3- and 4-month leads for WY and SEA indices, corresponding to the predictions initiated in January and February. This low forecast skill should be related to the so-called “spring predictability barrier” phenomenon (Webster and Yang 1992). Compared with CFSv1 (Fig. 10 in Yang et al. 2008), the coefficients of correlation between the observed indices and the predicted indices by CFSv2 decrease more rapidly than those by CFSv1 from 0-month lead to 4-month lead. Thus, the CFSv2 features a more apparent spring predictability barrier.
The hindcast shows high skill in predicting the large-scale circulation of the ASM, but not the precipitation over the domains dominated by the monsoon (figures not shown). Thus, it is also interesting to understand to what extent the variations of precipitation and circulation patterns related to these dynamical indices can be predicted by the model. Figure 7 shows the patterns of correlation between the WY index and precipitation and regression of 850-hPa winds against WY. The variation of observed WY index is associated with strong monsoon flow over South Asia, cyclonic circulation over Southeast Asia, anticyclonic circulation over northeast Asia, easterlies over the equatorial Pacific, and anticyclonic circulation over the tropical central North Pacific (Fig. 7a). The atmospheric circulation pattern over the Pacific shows La Niña–related features. Indeed, the JJA correlation between the WY index and Niño-3.4 SST is −0.37. The WY has significant negative correlation with the precipitation over the Arabian Sea, the equatorial western IO, central eastern China, and central Japan and positive correlation over northeastern India, the eastern SCS, and the Philippine Sea (Fig. 7a). The hindcast captures the major observed features related to the WY index described above (Fig. 7b). The deficiency of the hindcast in predicting the circulation features related to the WY index includes stronger-than-observed westerlies over the equatorial eastern IO, easterlies over the equatorial central and western Pacific, anticyclonic circulation over the western North Pacific, weaker-than-observed cyclonic circulation over Southeast Asia, and lack of anticyclonic circulation over northeast Asia. Correspondingly, the correlation of the WY index with precipitation is higher than observations over the western IO, the central and eastern Pacific, and the Maritime Continent. All features over the tropics show that the hindcast depicts a stronger-than-observed relationship between the WY index and ENSO (R = −0.47 for 0-month lead prediction), which has also been reported for CFSv1 (Yang et al. 2008). The observed significant negative correlation for WY and precipitation over northeast Asia cannot be predicted by the hindcast. The negative correlation of the WY index with precipitation in eastern China is too far south, consistent with the weak cyclonic circulation over Southeast Asia. The observed positive correlation between the WY index and the precipitation in the eastern SCS shifts to the western SCS in the hindcast. From 1-month to 7-month lead, the predicted correlation pattern of the WY index with precipitation and wind pattern regressed on WY are similar to those of 0-month lead.
Figure 8 shows the patterns of correlation between the SA index and precipitation and regression of 850-hPa winds against SA. The SA index is related to southwesterlies over the Arabian Sea and India, southerlies over the western BoB, and easterlies from the equatorial central IO to the equatorial central Pacific (Fig. 8a). There is a wave train pattern over the western North Pacific. A high SA index corresponds to an increase in precipitation over southern and northern India and a decrease over the Philippine Sea. The winds and precipitation related with SA over the western Pacific also show some ENSO-related features, but the correlation between the SA index and Niño-3.4 SST is only −0.27. In the hindcast, there is no wave train pattern in the western North Pacific (Fig. 8b). The observed SA is directly associated with southwesterlies from the Arabian Sea to India; however, the CFSv2 SA is directly associated with southeasterlies from the western Maritime Continent to India. The correlation between the SA index and precipitation is overestimated over the eastern Arabian Sea and northern India. The hindcast also overestimates the link between the SA index and ENSO, with correlation between SA and Niño-3.4 SST of −0.35 for 0-month lead prediction. Wind and precipitation patterns vary from 1-month to 7-month lead (Figs. 8c–f). The hindcast shows some skill in predicting the relationship of the SA index with winds and precipitation when the lead time is less than 3 months. In long-lead predictions, even the significant lower-tropospheric southerlies over India associated with the SA cannot be predicted (Figs. 8d–f). The variation of the SA is contributed mainly from the variation of upper-tropospheric circulation and is closely linked to ENSO in the long-lead predictions (figure not shown).
Figure 9 shows the patterns of correlation between the SEA index and precipitation and regression of 850-hPa winds against SEA. The observed variation of the SEA index is related to strong convection in the vicinity of the Philippines and northwestern BoB, but weak convection over large portions from the Maritime Continent northwestward to the Arabian Sea and over East Asia (Fig. 9a). The regression of winds against the SEA index shows westerlies over South Asia, cyclonic circulation over Southeast Asia, and anticyclonic circulation over northeast Asia. The wind and precipitation patterns related to the SEA index over the western North Pacific are similar to those related to the WY index, being known as the Pacific–Japan pattern (Figs. 7a, 9a) (Nitta 1987). Also, the SEA index is positively correlated with Niño-3.4 SST simultaneously (R = 0.36). The hindcast simulates the tropical component related to SEA very well (Fig. 9b). The anticyclonic circulation related to SEA over northeast Asia is not captured by the hindcast (Fig. 9b). This deficiency leads to the southward shift of the observed rain belt related to the SEA index over East Asia (Figs. 9a,b). In addition, the correlation between the SEA index and Niño-3.4 SST is also overestimated (R = 0.46 for 0-month lead prediction). The general characteristics of winds and precipitation related to SEA in hindcast do not change significantly as lead time increases (Figs. 9b–f). The SEA index, however, shows a closer correlation with winds and precipitation over the tropical Pacific as lead time increases.
To conclude, the hindcast captures the general features of convection and atmospheric circulation related to the WY and SEA indices over the tropics, in spite of an overestimated relationship between the indices and the tropical Pacific climate with increase in lead time. Stronger-than-observed correlations of ENSO with WY and SEA are found in the hindcast. The hindcast shows an inability in reproducing the response of the subtropical East Asian climate to the tropical heating near the Philippines. It also shows apparent deficiency in predicting the convection and atmospheric circulation related to SA, although some of the features are predicted by lead time less than 2 months.
5. Role of ocean–atmosphere coupling in monsoon prediction
In this section, we compare the predictions of the ASM among hindcast and simulations based on AMIP and CMIP. Since the difference between the hindcast and the AMIP is not only in ocean–atmosphere coupling but also in initial conditions, their differences can only partially illustrate the role of ocean–atmosphere coupling in ASM prediction. To understand the possible impact of initial conditions and highlight the importance of ocean–atmosphere coupling, we also analyze the Asian summer monsoon in the CMIP. Although the interannual variability of the ASM in CMIP cannot be compared to observations, an analysis of the features of monsoon process may yield useful information.
a. Climatological features
Figure 10 shows the climatological precipitation and winds in the AMIP and their differences from observation and the hindcast. Heavy precipitation centers are located over the central Arabian Sea, the northern BoB, and the Himalayas in AMIP (Fig. 10a). Differences in precipitation between AMIP and observation indicate that the AMIP seriously underestimates the precipitation over most of the oceans and overestimates the precipitation over the central Arabian Sea (Fig. 10b). Correspondingly, there are weaker-than-observed WPSH and lower-tropospheric westerlies and upper-tropospheric easterlies over South Asia. The patterns of difference in winds and precipitation between AMIP and hindcast are consistent with those between AMIP and observation over oceans, despite their difference in magnitude (Figs. 10b,c). It is of interest to note that AMIP and hindcast show similar biases in the vicinity of India, over the equatorial central and eastern IO, and over the western Maritime Continent (Figs. 1b and 10b,c), further indicating the imperfectness of the model in simulating the atmospheric circulation and precipitation in these regions.
It can be seen from Fig. 11, which shows the climatological precipitation and winds in CMIP and their differences from observation and hindcast, that CMIP depicts precipitation and lower-tropospheric wind patterns that are similar to those in the hindcast (Figs. 1a,b and 11a,b). However, the magnitude is different over the western Arabian Sea, the Indian subcontinent, the central Indian Ocean, and the subtropical western North Pacific. The deficiency of CMIP in simulating tropical upper-tropospheric winds is similar to that of hindcast, but the bias of CMIP is larger (Figs. 1e and 11e). Also, the upper-level trough over East Asia in CMIP is stronger than that in both observation and hindcast (Figs. 11e,f). The difference between CMIP and hindcast indicates that initial conditions are not very important for predicting the climatological features of precipitation and low-tropospheric circulation, although they may be important for a specific year. Comparison of Figs. 10c and 11c illustrates that, for climatology, the difference in precipitation and lower-tropospheric circulation between AMIP and hindcast is mainly from their difference in ocean–atmosphere coupling. Thus, it is reasonable to conclude that the monsoon would be weaker over the BoB, the SCS, and the western Pacific but stronger over the Arabian Sea when ocean–atmosphere coupling is excluded in the CFSv2.
Figure 12 illustrates that the annual cycle of rainfall in the CMIP is overall more reasonable than that in the AMIP, while the Indian monsoon is an exception. Over 70°–90°E, the AMIP shows similar characteristics of the annual cycle of rainfall to those in the observation, capturing the general onset and retreat features of the Indian monsoon. The CMIP, however, simulates a smaller increase in rainfall after the monsoon onset. The maximum rainfall is located too far south (near the equator) all year round. Over both 90°–110°E and 110°–130°E, the CMIP shows similar characteristics of the annual cycle of rainfall to those in the hindcast, despite the overall stronger rainfall in the CMIP (Figs. 2 and 12). The AMIP, however, shows inability in simulating the onset and retreat of the tropical monsoon, while it simulates a northward commence of rainfall season in East Asia (Figs. 12b,c).
b. Interannual variability
We further analyze the interannual variability of the ASM measured by the three dynamical monsoon indices: the WY, the SA, and the SEA. As seen from the climatological features, the AMIP simulates apparent weaker-than-observed WY and SEA, especially the WY (Figs. 13a,c), associated with weaker-than-observed tropical westerlies and easterlies in the lower and upper troposphere, respectively (Figs. 10b,e). Despite the deficiency in simulating their magnitude, interannual variations of the WY and the SEA in AMIP are significantly correlated with observations, with a correlation coefficient of 0.54 for both indices. The SA in AMIP shows an insignificant correlation with observation, while it has nearly the same magnitude as the observed.
Table 1 lists the means and standard deviations of the WY, the SA, and the SEA indices for different simulations, as well as the correlations between the indices from hindcast of 0-month lead and AMIP and those from observations. The hindcast is better than the AMIP in simulating the magnitude of the WY, the standard deviation of the SEA, and the interannual variations of all indices, especially the SA. However, it also shows deficiency in depicting the magnitude and standard deviation of the SA compared to AMIP, presumably attributed to the unrealistic SST over the equatorial and central IO. For the magnitude of the SEA, both hindcast and AMIP show apparent deficiency. We have seen from Fig. 6 that even though the contribution from initial conditions to monsoon prediction has been considered, the correlation between observation and hindcast for this monsoon index decreases significantly in the first two lead months. Then, why does the hindcast involving ocean–atmosphere coupling not show apparent higher skill in predicting the interannual variability of the ASM compared to the AMIP, although the hindcast performs better in simulating the climatological features of ASM? To answer this question, we further analyze the atmospheric circulation and precipitation patterns related to the monsoon indices in AMIP and CMIP.
Figure 14 shows the patterns of correlation between monsoon indices and precipitation and regression of winds against the indices in both AMIP and CMIP. Comparison between Figs. 7a and 14a illustrates that the AMIP simulates some major features related to the WY including westerlies from the Arabian Sea to the Philippine Sea, increase in precipitation near the Philippines, and decrease in precipitation over the Arabian Sea in spite of a westward shift of the position of heavy precipitation. The relationship between the WY and ENSO is insignificant in AMIP, with a correlation coefficient of −0.04. The CMIP shows deficiency in simulating many features related to the WY in tropical Asia and the western Pacific, especially the precipitation in the vicinity of the Philippines (Figs. 7a and 14d). The WY–ENSO relationship in CMIP is stronger than observed, with a correlation coefficient of −0.33. The AMIP simulates some features of the circulation and precipitation related to the SA index in tropical Asia and the western North Pacific, as observed (Fig. 14b). It shows deficiencies in predicting the larger correlations between the SA and precipitation over northern India and the SCS. The AMIP also fails to predict the wave train pattern over the western North Pacific, just as the hindcast. The CMIP, however, does not capture many features of circulation and precipitation associated with the SA (Fig. 14e). The correlation coefficients between the SA and Niño-3.4 SST do not exhibit apparent differences among observation, AMIP, and CMIP. For the SEA monsoon index, the AMIP simulates many related features, including more precipitation over the South China Sea and the Philippine Sea, less precipitation over the Maritime Continent, and easterlies over the tropical eastern North Pacific. The AMIP predicts stronger-than-observed cyclonic circulation over Southeast Asia, while it fails to simulate the anticyclonic circulation over northeast Asia (Fig. 14c). The CMIP shows overall weaker-than-observed features of precipitation and circulation related to the SEA and it fails to capture the SEA-related anticyclonic circulation over northeast Asia and westerlies over the Arabian Sea and the western Indian peninsula (Fig. 14f). Stronger-than-observed correlation between the SEA and Niño-3.4 SST appears in both AMIP and CMIP, with coefficients of 0.45 and 0.40, respectively.
The above analyses indicate that both AMIP and CMIP capture some features related to the dynamical monsoon indices examined. Compared with the CMIP, the AMIP depicts more reasonable processes associated with the variation of the SA, owing likely to the realistic SST. In coupled model runs, the contribution from ocean–atmosphere interaction to simulating the WY and the SEA indices seems to be reduced by the deficiency in SST simulation. The model’s inability in simulating the Pacific–Japan pattern, an extratropical response over East Asia to tropical heating in the vicinity of the Philippines, suggests limited prospects for the seasonal prediction of the East Asian monsoon by directly using the output from the hindcast.
6. Summary and discussion
The NCEP CFSv2, which became operational in March 2011, provides an important source of information about the seasonal climate prediction for many Asian countries affected by monsoon. In this study, we have provided a comprehensive assessment of the prediction of the Asian summer monsoon by the CFSv2, focusing on seasonal-to-interannual time scales. We have investigated the importance of ocean–atmosphere coupling for predicting the ASM and various physical processes of the monsoon by the CFSv2. We have also compared the differences between CFSv2 and CFSv1 to understand the improvement of ASM prediction from CFSv1 to CFSv2.
Many major features of the ASM are well predicted by CFSv2, which include the heavy rainfall centers, large-scale monsoon circulation patterns, monsoon onset and retreat, and the interannual variability of several dynamical monsoon indices. The dynamical indices and their associated patterns of precipitation and circulation can be predicted several months in advance. Large predictability errors occur only in the first three lead months. Large-scale monsoon (measured by the Webster–Yang monsoon index) and the Southeast Asian monsoon are more realistically predicted than the South Asian monsoon. In the CFSv2, the relationships of the dynamical monsoon indices with winds and precipitation over the tropical Pacific strengthen with increase in lead time.
The CFSv2 shows improvements in predicting the large-scale circulation and precipitation patterns associated with the ASM compared to CFSv1. The improvement of large-scale circulation is likely due to the reduction of the cold bias over the Asian continent found in CFSv1. The South Asian monsoon, however, becomes worse from CFSv1 to CFSv2. Some major biases in CFSv1 still exist in CFSv2, especially the weaker-than-observed WPSH and the exaggerated strong link of the ASM to ENSO.
Comparisons among hindcast, AMIP, and CMIP indicate that ocean–atmosphere coupling is important for ASM prediction by CFSv2. AMIP reproduces overall weaker-than-observed ASM, especially from the BoB to the western North Pacific, while the difference between hindcast and CMIP is smaller in most regions. Both hindcast and CMIP show a more realistic annual cycle of precipitation compared to AMIP. The interannual variability of ASM is better in hindcast, compared to AMIP. Comparison between AMIP and CMIP indicates that CMIP does not show any advantage in depicting the atmospheric processes associated with the interannual variability of several major dynamical monsoon indices. The CFSv2 fails to reproduce the extratropical response over East Asia to tropical heating near the Philippines in all three types of simulations.
The CFSv2 has demonstrated skills in predicting dynamical monsoon indices. However, is it also skillful in predicting the regional precipitation indices? Here, we discuss the skill of CFSv2 prediction for five regional precipitation indices: 1) the South Asian rainfall (SAR) index, defined as the precipitation averaged over 10°–30°N, 70°–110°E (Goswami et al. 1999); 2) the Arabian Sea rainfall (ASR) index over 5°–20°N, 50°–70°E; 3) the Bay of Bengal rainfall (BoBR) index over 5°–20°N, 80°–100°E; 4) the Southeast Asian rainfall (SEAR) index over 10°–20°N, 115°–140°E (Wang and Fan 1999); and 5) the East Asian mei-yu index over 27.5°–40°N, 110°–140°E (Gao et al. 2011). Figure 15 shows the prediction skills of CFSv2 for the five rainfall indices. The CFSv2 predicts the rainfall over the Arabian Sea and Southeast Asia very well, but fails to predict the precipitation over South Asia and mei-yu. Since the variation of the Southeast Asian monsoon is significantly dominated by convection over the Philippine Sea (Wang and Fan 1999), the prediction skills for SEAR with lead time are similar to those for the SEA monsoon index. However, the CFSv2 has higher skill for the dynamical SEA monsoon index (Figs. 6 and 15). The CFSv2 inability in predicting the SA monsoon index is consistent with the poor South Asian rainfall prediction. Although the CFSv2 has high skill in predicting the precipitation over East Asia that is related to SEA monsoon index (Fig. 9), it fails to predict the mei-yu over East Asia because the Southeast Asian monsoon is just one of the factors affecting the interannual variation of mei-yu (Gao et al. 2011).
The western Pacific subtropical high is a very important atmospheric system that affects the prediction of monsoons over East Asia and the western North Pacific. Why does the Climate Forecast System predict a weaker-than-observed WPSH? A recent diagnostic study of the biases of Asian monsoon prediction using the CFSv2 daily hindcast data from 1999 to 2010 suggested that the weaker-than-observed WPSH was attributed to the bias of internal variability of the atmosphere since the development of this bias was often found in the lead time at less than two weeks (Liu et al. 2013). Moreover, ENSO-related teleconnection patterns are the basis for seasonal climate prediction. Why does the CFS predict an exaggerated link of the ASM to ENSO? The CFSv2 simulates an unrealistic zonal SST gradient in the tropical Indian Ocean during the ENSO developing phase, with a large warm SST bias over the tropical western IO, accompanied by anomalous easterly flow over the equatorial IO (figures not shown). The easterly flow weakens the WY index, which may contribute to a stronger-than-observed link of the WY to ENSO since, in observation, the WY is weak during the ENSO developing phase.
During the ENSO decaying phase, the CFSv2 also has a cold SST bias over the tropical North Pacific in long-lead prediction. Deficient precipitation occurs with the cold SST bias, accompanied by stronger-than-observed anticyclonic circulation to the northwest (figures not shown). On one hand, the anomalous northeasterly flow of the anticyclonic circulation can sustain the cold SST bias through enhancing surface evaporation. On the other hand, the cold bias leads to deficient precipitation, which strengthens the anticyclonic circulation as a Rossby wave response. This feature means that there may be a positive feedback between the cold bias and the anticyclonic circulation. Because of the close link of the ASM to ENSO, the biases during the ENSO decaying phase may partly explain why the relationships of the dynamical monsoon indices with winds and precipitation over the tropical Pacific strengthen with increase in lead time.
The above analyses indicate that errors of ASM prediction are linked to SST errors over both the Pacific and IO in the CFSv2. Indeed, the observed ASM is affected not only by remote SST in the central and eastern Pacific, but also by the IO SST (Yoo et al. 2006), whose anomalies are partly a response to ENSO, especially in the decay phase of ENSO (Xie et al. 2009). Thus, further studies focusing on Indian Ocean and Pacific SSTs and their impacts on ASM during different phases of ENSO in the CFSv2 are important for shedding more light onto the errors of ASM prediction.
Ocean–atmosphere coupling apparently contributes to monsoon simulation and prediction, especially for the tropical regions from the BoB to the western Pacific. Wu et al. (2008) reported that on a subseasonal time scale the response of SST to the atmosphere was slower in the CFSv1 compared to observation. At present, relevant features simulated and predicted by the CFSv2 are unknown. Thus, further studies of the ocean–atmosphere coupling and monsoon in the CFSv2 are warranted.
The authors thank the two anonymous reviewers for their constructive comments, which improved the overall quality of the paper. This study was jointly supported by the National Natural Science Foundation of China (Grant 41105061), the National Basic Research Program of China (Grant 2012CB417202), the Basic Research and Operation Program of the Institute of Plateau Meteorology, CMA (Grant BROP201215), and the Open Research Fund Program of Plateau Atmosphere and Environment Key Laboratory of Sichuan Province (Grant PAEKL-2011-C2). Xingwen Jiang, who was partially supported by U.S. National Oceanic and Atmospheric Administration and China Meteorological Administration Bilateral Program, thanks NOAA’s Climate Prediction Center for hosting his visit while this study was conducted.