Predictive skills of the Somali cross-equatorial flow (CEF) and the Maritime Continent (MC) CEF during boreal summer are assessed using three ensemble seasonal forecasting systems, including the coarse-resolution Predictive Ocean Atmospheric Model for Australia (POAMA, version 2), the intermediate-resolution Scale Interaction Experiment–Frontier Research Center for Global Change (SINTEX-F), and the high-resolution seasonal prediction version of the Australian Community Climate and Earth System Simulator (ACCESS-S1) model. Retrospective prediction results suggest that prediction of the Somali CEF is more challenging than that of the MC CEF. While both the individual models and the multimodel ensemble (MME) mean show useful skill (with the anomaly correlation coefficient being above 0.5) in predicting the MC CEF up to 5-month lead, only ACCESS-S1 and the MME can skillfully predict the Somali CEF up to 2-month lead. Encouragingly, the CEF seesaw index (defined as the difference of the two CEFs as a measure of the negative phase relation between them) can be skillfully predicted up to 4–5 months ahead by SINTEX-F, ACCESS-S1, and the MME. Among the three models, the high-resolution ACCESS-S1 model generally shows the highest skill in predicting the individual CEFs, the CEF seesaw, as well as the CEF seesaw index–related precipitation anomaly pattern in Asia and northern Australia. Consistent with the strong influence of ENSO on the CEFs, the skill in predicting the CEFs depends on the model’s ability in predicting not only the eastern Pacific SST anomaly but also the anomalous Walker circulation that brings ENSO’s influence to bear on the CEFs.
Significant amounts of moisture and energy are transported from the Southern Hemisphere to the Asian monsoonal area via low-level cross-equatorial flows (CEFs) during the boreal summer (June–August) (e.g., Tao et al. 1962; Findlater 1966; Saha and Bavadekar 1973; Ramesh Kumar et al. 1999), with opposite north-to-south CEFs prevailing in the upper troposphere via the local Hadley cell (e.g., Schneider et al. 2014). The strongest of these flows occurs over the western Indian Ocean, which is known as the Somali CEF (e.g., Findlater 1969). The interannual variation of the Somali CEF is found to be related to broad climate variabilities, including the Mascarene high (e.g., Xue et al. 2003), the Indian summer monsoon (ISM) precipitation (e.g., Halpern and Woiceshyn 2001), and the Pacific–Japan teleconnection pattern (e.g., Wang and Xue 2003). Besides, another substantial CEF has been documented over the Maritime Continent (MC) region (e.g., Wang and Li 1982), which connects the Australian high and the East Asian summer monsoon (EASM) circulation (e.g., Liu et al. 2009). The MC CEF is driven by the interhemispheric thermal contrast between the cold Australian land surface in austral winter and the warm Asian continent in boreal summer, contributing to the moisture convergence in the tropical EASM region (e.g., Wang et al. 2003).
Previous studies have suggested that the Somali CEF and the MC CEF are not independent of each other but rather are significantly negatively correlated (e.g., Cong et al. 2007; Wang and Yang 2008; Zhu 2012). Consequently, the distributions of the precipitation anomaly patterns over Asia associated with the two CEFs are similar but with opposite sign (e.g., Li and Li 2014). Li and Li (2014) proposed a seesaw index (i.e., the difference of the two CEFs) to describe their negative correlation, and found that the CEF seesaw index shows a better correlation to the EASM than the individual CEFs do (e.g., Li and Li 2016). Thus, the CEF seesaw index may provide a useful approach for better understanding of the associated monsoonal precipitation. For instance, the CEF seesaw is related to the out-of-phase variation in precipitation between the ISM region and the tropical EASM area that was documented in previous studies (e.g., Huang et al. 1998; Zhang 2001). In addition, Li and Li (2014) have found that the CEF seesaw shows a high correlation with the convection seesaw over the Indo-Pacific region, which was referred to as the Maritime Continent–Pacific convection oscillation (MPCO) by Li et al. (2013).
The seesaw relation between the Somali and the MC CEFs likely results from their distinct relations with El Niño–Southern Oscillation (ENSO) (e.g., Wang et al. 2001; Zhu and Chen 2002). Previous studies have shown that the Somali CEF tends to be weakened while the MC CEF is strengthened during a warm ENSO episode, and vice versa during a cold ENSO episode. Li et al. (2017) have further investigated the influences of different types of ENSO events upon the two CEFs. It has been shown that a high positive correlation exists between the MC CEF and all the types of ENSO, including the eastern Pacific (EP) El Niño, central Pacific (CP) El Niño, EP La Niña, and CP La Niña events. This is consistent with Zhou and Kim (2015), who suggested a modulation of ENSO on the MC CEF through the ENSO-related local meridional sea surface temperature (SST) gradient around the MC area. In contrast, the negative correlation between the Somali CEF and ENSO is weaker than that between the MC CEF and ENSO, and is subject to the different types of ENSO. The EP El Niño exerts a significant influence on the Somali CEF via the atmospheric teleconnection between the Pacific and the Indian Ocean. During the boreal summer, warm SST anomaly in the eastern Pacific favors an easterly anomaly in the northern Indian Ocean and a weakened Somali CEF in the western Indian Ocean through weaker-than-normal Walker cell. But the impact of the other types of ENSO on the Indian Ocean is much weaker. Thus, the EP El Niño plays a dominant role in generating the CEF seesaw (e.g., Li et al. 2017). Previous studies also investigated the correlations between the CEFs and the Indian Ocean dipole (IOD) based on both observations and climate model experiments (e.g., Li and Li 2014; Li et al. 2017; Li 2017, 16–21). Results suggested that a purely IOD-related SST anomaly has a weak influence on the CEFs and the CEF seesaw.
Seasonal prediction skill of ENSO has been significantly improved over the past 2–3 decades through the development of coupled ocean–atmosphere general circulation models (OAGCMs) (e.g., Cane et al. 1986; Barnston et al. 1999; Jin et al. 2008; Luo et al. 2008, 2015; Graham et al. 2011; Cottrill et al. 2013; MacLachlan et al. 2015). While the forecast skill of ENSO varies with target seasons, ENSO phases, and ENSO strength (e.g., Jin et al. 2008), ENSO can be generally predicted with a useful skill (i.e., anomaly correlation coefficient of above 0.5) out to two or three seasons by many current state-of-the-art OAGCMs (e.g., Palmer et al. 2004; Jin et al. 2008; Xue et al. 2011; Barnston et al. 2012; Luo et al. 2015). Note that correct predictions of ENSO onset across the boreal spring, associated with the spring predictability barrier, remains a long-standing challenge. Owing to the significant influence of ENSO on the MC and Somali CEFs, it is expected that the high ENSO prediction skill may provide a good potential for a skillful prediction of the two CEFs. However, the predictive skill of the CEFs is yet to be assessed. Assessing predictability of the CEF seesaw may improve our understanding of the predictability of the related precipitation anomaly in the Asian and Australian monsoon systems.
The purpose of the present study is to examine the predictive skill of the Somali and MC CEFs using three ensemble seasonal forecast models: the coarse-resolution Bureau of Meteorology Predictive Ocean Atmosphere Model for Australia (POAMA) model (e.g., Alves et al. 2003), the intermediate-resolution Scale Interaction Experiment–Frontier Research Center for Global Change (SINTEX-F) model (e.g., Luo et al. 2005a,b), and the high-resolution seasonal prediction version of the Australian Community Climate and Earth System Simulator (ACCESS-S1) model (e.g., Alves and Hudson 2015; Lim et al. 2016). Details about the three forecast models and the verification datasets are described in section 2. The multimodel prediction skill of the Somali CEF, MC CEF, CEF seesaw index, and the related monsoonal precipitation pattern are assessed in section 3. In section 4, we will explore the possible reasons for the distinct forecast performance of the Somali CEF in the different models. A summary is given in section 5.
2. Multimodel and verification data
1) The POAMA-2 model
POAMA (version 2) is a fully coupled ocean–atmosphere climate model that was developed by the Australian Bureau of Meteorology (BoM) to provide intraseasonal to seasonal predictions (e.g., Hudson et al. 2013). The atmospheric model component of POAMA-2 is the BoM’s Atmospheric Model (BAM3.0), which has a T47 horizontal resolution (approximately a 2.5° × 2.5° grid) with 17 vertical levels (e.g., Colman et al. 2005; Wang et al. 2005; Zhong et al. 2006). The land surface component uses a simple bucket model for soil moisture after Manabe and Holloway (1975) with three soil temperature levels (e.g., Hudson et al. 2011). The Australian Community Ocean Model version 2 (ACOM2) (e.g., Schiller et al. 1997; Oke et al. 2005) forms the ocean component of POAMA-2, based upon the Geophysical Fluid Dynamics Laboratory (GFDL) Modular Ocean Model (MOM version 2) (Schiller et al. 2002). Grid spacing in ACOM2 starts at 2° in the zonal direction and 0.5° in the meridional direction at the equator and gradually increases to 1.5° at the poles. A total of 25 vertical levels are used for the ocean model, with the first 12 levels in the upper 185 m and a maximum depth of 5 km. Atmosphere and ocean models are coupled using the Ocean Atmosphere Sea Ice Soil (OASIS) coupling software (Valcke et al. 2000). The atmospheric model of POAMA-2 is initialized with the atmosphere–land initialization scheme (ALI) that captures the observed atmosphere state. ALI generates realistic atmosphere–land surface initial conditions by nudging the BAM3.0 model toward the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40; Uppala et al. 2005) for the period 1982 to August 2002, and to the global analyses from BoM’s numerical weather prediction system thereafter (e.g., Hudson et al. 2011). Initial conditions for the ocean are obtained from the POAMA Ensemble Ocean Data Assimilation System (PEODAS), which assimilates available ocean observations using an ensemble Kalman filter (e.g., Yin et al. 2011).
2) The SINTEX-F model
The SINTEX-F coupled model (e.g., Luo et al. 2003, 2005a,b; Masson et al. 2005) was developed at Frontier Research Center for Global Change under a European Union–Japan collaboration (e.g., Gualdi et al. 2003; Guilyardi et al. 2003). The atmospheric component consists of the European Center Hamburg version 4.6 (ECHAM4.6) model at T106 (about 1.1° × 1.1°) with 19 vertical levels (Roeckner et al. 1996). The global ocean component is the reference version 8.2 of the Ocean Parallélisé (OPA8.2) (Madec et al. 1998), which has a 2° Mercator horizontal mesh with a meridional resolution that increases to 0.5° at the equator and 31 vertical layers. The atmospheric and oceanic components are coupled every two hours using OASIS (Valcke et al. 2000) without any flux correlation except that sea ice cover is relaxed toward observed monthly climatologies in the ocean circulation model. The coupled model is spun up for 10 years by assimilating observed monthly climatological SST into the coupled model (e.g., Luo et al. 2005a,b). Observed SSTs from 1982 onward are then assimilated into the coupled model to provide initial conditions for the SINTEX-F model forecasts (e.g., Luo et al. 2005b, 2008). That is, model SSTs are strongly nudged toward the National Centers for Environmental Prediction (NCEP) SST dataset. This coupled SST–nudging initialization approach provides a simple but effective way to generate well-balanced realistic initial conditions and provides useful ENSO prediction skill out to two years ahead (e.g., Luo et al. 2008, 2015, 2017). The SINTEX-F model shows good performance in simulating the ENSO variability, including its spatial SST distribution, magnitude, and period (e.g., Yamagata et al. 2004; Luo et al. 2005a,b; Behera et al. 2006; Jin et al. 2008).
3) The ACCESS-S1 model
ACCESS-S1 (the seasonal prediction version of ACCESS) represents the BoM’s next seasonal forecasting system, replacing POAMA-2 in 2017 (e.g., Lim et al. 2016). The coupled model used in ACCESS-S1 is the Unified Model-General Configuration 2 (UM-GC2) (Williams et al. 2015), which is the same model used for the Met Office’s global coupled model seasonal forecast system GloSea5-GC2 (MacLachlan et al. 2015). Components of this coupled model consist of United Model Global Atmosphere 6.0 (GA6.0), Global Land 6.0 (GL6.0) [GA6.0 and GL6.0 are fully documented by Walters et al. (2017)], Global Ocean 5.0 (GO5.0) (Megann et al. 2014), and Global Sea Ice 6.0 (GSI6.0) (Rae et al. 2015). ACCESS-S1 is a high-resolution system with an N216 atmospheric horizontal resolution (approximately 60 km in the midlatitudes) and is coupled to an ocean model with a resolution of 0.25°. The vertical resolution is set by the component definitions, including 85 levels in the atmosphere (with a top at 85 km), four soil levels, 75 levels in in the ocean (with 1-m thickness in the top level), and five sea ice thickness categories. Further information regarding the coupled configuration can be found in Williams et al. (2015). Reanalyses from the ECMWF ERA-Interim project are used to initialize the atmosphere component model and land temperatures (MacLachlan et al. 2015). Initial conditions for the ocean and sea ice are derived from the GloSea Ocean Sea Ice Reanalysis that is based on the Met Office Forecast Ocean Assimilation Model (FOAM) data assimilation system (Blockley et al. 2014). Further details of the initialization are described in MacLachlan et al. (2015).
To address forecast uncertainties, ensemble hindcasts have been generated for each model using different coupling physics and/or initial conditions. The main features of the multimodel retrospective forecasts are listed in Table 1. Model resolution increases from POAMA-2 to SINTEX-F to ACCESS-S1. Note that the ACCESS-S1 has a shorter hindcast period (1990–2012) and fewer lead time forecasts available (2-month and 5-month lead only), in contrast to those of the POAMA-2 and SINTEX-F models, which have a longer hindcast period of 1982–2012 and have forecasts of at up to 7 months lead (Table 1). For this study, hindcasts are verified for the seasonal mean of boreal summer [i.e., the target season is June–August (JJA)]. We define 1-month lead prediction of the JJA mean as the forecast starting on the first day of June, and 2-month lead forecast refers to that initialized on 1 May. Similarly, 7-month lead forecast refers to that initialized on 1 December of previous year. The ensemble mean of individual model forecasts is calculated by the average of all members of individual models. The multimodel ensemble (MME) mean is calculated by the average of the ensemble means of individual models.
b. Verification data and methods
To assess the CEF prediction skill, JJA-mean horizontal winds at the 850-hPa level from the ERA-Interim reanalysis datasets during 1982–2012 (Dee et al. 2011) are compared with the multimodel seasonal predictions. The intensity of the Somali (averaged at 37.5°–62.5°E) and MC (averaged at 102.5°–152.5°E) CEFs are computed using the local meridional wind velocity in the equatorial region (2.5°S–2.5°N). The definition of the MC CEF in this study has two minor differences compared to that in previous studies (e.g., Li and Li 2014; Li et al. 2017). First, the velocity at the 850-hPa level rather than the 925-hPa level is used to measure the MC CEF due to the limit of model data in SINTEX-F. Second, the coarse resolution of the seasonal prediction model (i.e., POAMA2) is unlikely sufficient to resolve the three narrow CEF channels over the MC region, and therefore the mean meridional wind across the entire region (102.5°–152.5°E) is used to calculate the MC CEF. Note that this newly defined MC CEF is highly correlated with the originally defined MC CEF (with a correlation coefficient of 0.95) based on the ERA-Interim data. The CEF seesaw index is defined by the difference between the standardized Somali CEF and the standardized MC CEF (e.g., Li and Li 2014; Li et al. 2017). SST forecasts from the multimodel outputs are verified against the National Oceanic and Atmospheric Administration Optimum Interpolation (NOAA OI, version 2) SST dataset (Reynolds et al. 2002). The JJA mean Niño-3 index (SST anomaly averaged over 5°S–5°N, 90°–150°W) is used to represent EP ENSO conditions. For precipitation, verification is performed against the Global Precipitation Climatology Project (GPCP, version 2; Adler et al. 2003). Two monsoon indices are adopted to represent the Indian summer monsoon and East Asian summer monsoon, respectively. One is the all-Indian summer rainfall index (AISRI), which is measured by the areal mean precipitation anomaly over the Indian peninsula (0°–20°N, 75°–90°E). The other is the East Asian summer monsoon index (EASMI) proposed by Wang et al. (2008), which is defined by the U850 in (5°–15°N, 90°–130°E) minus U850 in (22.5°–32.5°N, 110°–140°E). Forecast anomalies of the multimodels are calculated relative to each model’s hindcast climatology (1990–2012) at each lead time. Two statistical techniques are applied to verify deterministic prediction skill, the simple anomaly correlation coefficient (ACC) and the root-mean-square error (RMSE) score. Projection method and composite analyses are also conducted in this study. All the calculations are performed based on the time series of the observed and forecast anomalies without detrend.
a. Prediction of the CEFs
1) Mean state and variability
A comparison between the observed and predicted climatological CEFs is shown in Fig. 1. First, all the three models demonstrate a good performance in predicting the structure and strength of the Somali CEF, although the maximum speed core of the Somali CEF in ACCESS-S1 is slightly overestimated compared to the observed velocity (Fig. 1c). The sharp increase of the Somali CEF along the western Indian Ocean coast is well reproduced in both the ACCESS-S1 and SINTEX-F, but it is less well captured in the POAMA-2, probably due to its coarse resolution. Similarly, for the MC CEF, POAMA-2 fails to reproduce its fine local structure (Fig. 1a). The MC CEF channels (i.e., narrow maximum meridional wind centers at 102.5° to 110°E, 122.5° to 130°E, and 147.5° to 152.5°E, respectively) are related to the sharp land–sea contrasts over the MC region (e.g., Li and Li 2014). In contrast, this fine local structure of the MC CEF is well reproduced in the higher-resolution models, SINTEX-F and ACCESS-S1. The three local maxima of the MC CEF appear to be overestimated in the highest-resolution ACCESS-S1 model. Figures 1d and 1e show the predicted magnitudes of the climatology of the Somali and MC CEFs at all lead times. The Somali CEF appears to be underestimated in SINTEX-F but overestimated in ACCESS-S1 at all the lead times. In POAMA-2, the intensity of the Somali CEF is overestimated at short lead times, but it is well reproduced at lead times of 4–6 months. The intensity of the MC CEF is best predicted in SINTEX-F model at all the lead times (Fig. 1e). In contrast, both the POAMA-2 and ACCESS-S1 predicted a stronger-than-observed MC CEF at all lead times.
Figure 2 displays biases in the predicted climatological JJA precipitation. The overestimated CEFs in the models appear to be related to their precipitation biases in the Indo-Pacific Ocean. For instance, the too strong MC CEFs in both the ACCESS-S1 and POAMA-2 models are consistent with their strong wet bias in the South China Sea and western North Pacific. Also, the realistic MC CEF in the SINTEX-F is consistent with the small precipitation bias in the South China Sea and western North Pacific (cf. Figs. 1e and 2). The overestimated CEFs across the central-eastern Indian Ocean in the three models appear to be related to a south dry–north wet bias across the equatorial Indian Ocean. However, a clear relation between the precipitation bias and the Somali CEF bias is not visible in the three models (cf. Figs. 1d and 2). Further studies are warranted for a better understanding of the bias in the climatological intensity of the Somali CEF.
Figure 3 shows the standard deviation of interannual variability of the CEFs. The standard deviation in each model is first calculated for every ensemble member, and then the average of all ensemble members gives the ensemble mean for each model. The results show that all the three models overestimate the interannual variance of the Somali CEF at all lead times (Fig. 3d). Among the three models, the coarse-resolution POAMA-2 model severely overestimates the variance of the Somali CEF, particularly at lead times longer than 3 months, even though its mean intensity is predicted well (cf. Figs. 3d and 1d). The advantage of the high-resolution ACCESS-S1 model over the coarse-resolution POAMA2 model is seen clearly in their predictions of the MC CEF. The ACCESS-S1 model shows the best predictions of the variance of the MC CEF at both 2 and 5 months lead (Fig. 3e), except small discrepancies in predicting the fine structure over the MC area (Fig. 3c). The SINTEX-F model also shows a good skill in predicting the MC CEF variance and its local fine structure (Figs. 3b,e). In contrast, the POAMA-2 model shows a poor skill in predicting the variance of the MC CEF and its fine structure; the variance is much overestimated at all lead times (Figs. 3a,e). In general, biases of the predicted CEFs’ variance decrease with the increasing model resolution, particularly over the MC area where complex land–sea distribution exists. The local spikes (i.e., maxima) of the wind variance appear to occur over narrow straits in the MC area, which is best captured by the highest-resolution model (Figs. 3a–c). Thus, it appears that the high resolution could help improve the prediction of the local fine structure of the CEF variance over the MC area. However, it is worth noting that the models’ bias in predicting the climatology and mean structure of the CEFs does not lead to a similar bias in the models’ prediction of the interannual variations of the CEFs.
Similarly, the biases in the predicted CEFs variance tend to be related to the biases in the predicted precipitation variance. Among the three models, the strong variance of the Somali CEF and the MC CEF in the POAMA-2 appears to be related to its strong precipitation variance in the regions from the Indian continent to western-central Pacific (Figs. 4a,d). In the ACCESS-S1 model, the precipitation variance over Indonesia is small compared to the other two models (Figs. 4c,f), which may help avoid an overestimation of the MC CEF variance (recall Fig. 3e). It is worth noting that the variances of the CEFs in the central equatorial Indian Ocean (60°–90°E) are overestimated by all the three prediction systems (Figs. 3a–c). One possible reason is that all the three models overestimate the precipitation variance in the equatorial and northern Indian Ocean and parts of southern Asia. The strong precipitation variance may favor a strong variability of the CEFs in 60°–90°E. However, exact reasons for the overestimated variability of the CEF in 60°–90°E appear to be complicated. For instance, the ACCESS-S1 model, which shows the strongest variability bias of the CEF in 60°–90°E, does not appear to have the strongest bias of precipitation variance in the Indian Ocean and southern Asia. Further studies are needed to explore the exact reasons for the systematically overestimated CEF variance bias in the central Indian Ocean.
In the following skill assessments, the mean-state bias of each model’s forecast is removed in a posterior manner by calculating the model forecast anomaly relative to the climatology of the model at each lead time.
2) Deterministic skill
Interannual variations of the observed and predicted JJA mean CEFs are shown in Fig. 5. To investigate whether the prediction skill of CEFs is related to the model’s performance in predicting the EP ENSO events, Fig. 5 also shows the interannual variations of the Niño-3 index. The results indicate that the interannual variations of the MC CEF appear to be better predicted at both 2 and 5 months lead in all the three models, compared to those of the Somali CEF (cf. the first and second rows of Fig. 5). Besides, the ACCESS-S1 appears to have a higher skill in predicting the interannual variations of the two CEFs, compared to the POAMA-2 and SINTEX-F. The higher skill of the ACCESS-S1 in predicting the CEF seesaw index can also be seen in Fig. 5 (third row). All the three models predict the negative Somali CEF, positive MC CEF, and negative CEF seesaw in 1997 at 2 and 5 months lead, corresponding to their good prediction of the 1997 EP El Niño event (Fig. 5, bottom row). Another two extreme negative CEF seesaws in 1982 and 1987, corresponding to two EP El Niño events in 1982 and 1987, are also realistically predicted in the POAMA-2 and SINTEX-F models. However, the positive Somali CEF in the early 2000s is not predicted in any of the three models. Based on the above analyses, a higher prediction skill is expected for the MC CEF, particularly in the ACCESS-S1 model. In contrast, the Somali CEF appears to have a lower predictability.
It is worth noting that the observed Somali CEF experienced a positive trend from 1982 to 2012 (Fig. 5, top row), consistent with a La Niña–like cooling trend in the Pacific (e.g., McPhaden et al. 2011; Luo et al. 2012). Although POAMA-2 and SINTEX-F can successfully reproduce the cooling trend in the eastern Pacific during 1982–2012 (figures not shown), the two models fail to capture the upward trend of the Somali CEF. The underlying reasons for why the two prediction systems fail to capture the positive trend of the Somali CEF are not yet clear and warrant further studies.
To examine the deterministic prediction skill in the individual models and the MME mean, Fig. 6 shows the anomaly correlation coefficient (ACC) as a function of lead times, including both the ensemble mean prediction skill (red solid line or red dot) and the mean skill of the ensemble members (blue dashed line or blue cross). Note that the ACC scores are calculated based on nondetrended time series and results based on detrended time series are almost identical. In general, the skill of the ensemble mean is higher than the mean skill of the ensemble members in all the three models. Also, the ensemble mean skill in each model is also higher than the skill of most members. The persistence skill (black dashed lines) is also displayed in Fig. 6. To calculate the persistence skill, a 3-month running mean is applied to the observed monthly time series. The persistence skill at 1-month lead refers to the correlations between the target season JJA and the start month May (i.e., April–June mean), and so forth. In general, the persistence skill of the CEFs is lower than that of the ACC prediction skill, except the prediction of the Somali CEF at 1–2-month lead in the POAMA-2 and SINTEX-F models. The low persistence of the CEFs might be related to the fact that the south-to-north CEFs basically occur during boreal summer (May to September).
According to the model ensemble mean results, both POAMA-2 and SINTEX-F do not produce a useful skill (with an ACC of lower than 0.5) in predicting the Somali CEF at all lead times. ACCESS-S1 can predict the Somali CEF skillfully at 2-month lead (with an ACC of greater than 0.5), but the skill drops below 0.3 at 5-month lead. In contrast, the ACC skill of the MC CEF is much higher than that of the Somali CEF (cf. the first and second rows of Fig. 6). This higher skill may be partly due to the fact that the correlation between the MC CEF and ENSO is higher than that between the Somali CEF and ENSO (e.g., Li et al. 2017). The ensemble mean skill of individual models in predicting the MC CEF shows a useful score out to 4–5 months lead. ACCESS-S1 shows the highest skill among the three models with an ACC score of above 0.7 even at 5 months lead.
For the prediction of the CEF seesaw index, the skill in the POAMA-2 is above 0.6 at 1-month lead but drops quickly to below 0.5 at 2-month lead. Surprisingly, the ACC skill is then gradually improved up to 5-month lead, consistent with the same skill behavior of POAMA-2 in predicting the Somali CEF (cf. the first and third panels of the first column of Fig. 6). However, its prediction skill of the Niño-3 index shows a gradual decrease with the increasing lead time (the fourth panel of the first column of Fig. 6). This suggests that other factors may play a role in influencing the predictive skill of the Somali CEF. The SINTEX-F model produces a useful skill for the CEF seesaw index up to 4-month lead, but the skill decreases sharply at 5–6 months lead. Encouragingly, skillful predictions of the CEF seesaw index are produced in the ACCESS-S1 at both 2-month lead (with a high ACC score of 0.8) and 5-month lead (with the ACC score of 0.6).
Comparing the correlation skills among the three models for the common period of 1990–2012, it is shown that the ACCESS-S1 model produces the highest skill in predicting the two CEFs, the seesaw index and the Niño-3 index, except that the skill of the Somali CEF prediction at 5-month lead is lower than that of the POAMA-2 (i.e., rightmost column of Fig. 6). The MME outperforms one or two models but not always all the three models in predicting both the CEF indices and the Niño-3 index. It is interesting to note that the highest prediction skill of the Somali CEF at 5-month lead is found in the POAMA-2, despite the fact that its prediction skill of the Niño-3 index at 5-month lead is the lowest among the three models. In contrast, although SINTEX-F can successfully predict the Niño-3 index with correlation of above 0.6 at 5-month lead, it fails to produce a skillful prediction of the interannual variation of the Somali CEF at 5-month lead. This suggests that the skillful prediction of the EP ENSO events may not guarantee a useful skill for the Somali CEF prediction. A further discussion about this issue is presented in section 4.
Figure 7 displays the root-mean-square error (RMSE) skill of the four indices. The RMSE is calculated based on the observed and forecasted anomalies at each lead time with the model climate drift being removed. Note that a skillful prediction generally requires that the RMSE is smaller than one standard deviation of the observation. The hindcast results show that the RMSE skill of the ensemble mean is overall better than the mean skill of the ensemble members in all the three models. The models show large RMSE values for the Somali CEF and the MC CEF predictions (red solid lines in Fig. 7). The RMSEs are close to or higher than their observed standard deviations, except for the ACCESS-S1 model, which shows useful skill in predicting the MC CEF at both 2-month and 5-month lead times. For the seesaw index, the RMSE values at all lead times are generally smaller than one standard deviation of the observations except those of the POAMA-2 model (third row of Fig. 7). It appears that the RMSE prediction skill of the seesaw index in the three models is generally better than that of the individual CEF. For the JJA mean Niño-3 index, all the models produce useful skill out to 5–7 months lead with the RMSE being smaller than one standard deviation of the observation (bottom row of Fig. 7). The MME always outperforms one or two models but not all the three models at both 2 and 5 months lead (rightmost column of Fig. 7). In general, the RMSE skills at all lead times are consistent with those of ACC score, with lower RMSE values corresponding to higher ACC scores.
To assess the reliability of the model ensemble forecasts, we also examine the spread of model ensembles (red dashed lines and open dots in Fig. 7), which is represented by one standard deviation of the differences between the individual member and the ensemble mean for each model’s predictions. With sufficiently large ensemble sizes in a perfect model, a reliable forecast system requires that the RMSE of the model ensemble mean should be equal to the spread among the ensembles (e.g., Weisheimer et al. 2009; Fortin et al. 2014). The high-resolution ACCESS-S1 model produces reliable predictions for the Somali and MC CEFs as well as the CEF seesaw with the RMSE of the ensemble mean being very close to the ensemble spread (third column of Fig. 7). The POAMA-2 and SINTEX-F models produce reliable predictions for the MC CEF, but they are overconfident in predicting the Somali CEF and the CEF seesaw index, with the ensemble spread being substantially smaller than the RMSE of the ensemble mean (first and second columns of Fig. 7). Predictions of the Niño-3 index in all the three models are overconfident except for the short-lead forecasts in the SINTEX-F model (bottom row of Fig. 7). Interestingly, including all the 53 members in the MME forecasts does not improve the overconfidence problem in the predictions of the Niño-3 and CEF seesaw indices (rightmost column of Fig. 7). For the predictions of the Somali and MC CEFS, however, the MME becomes too much dispersive. The underlying reasons for this interesting behavior are unclear and warrant future studies.
b. Prediction of the CEF seesaw–related monsoonal precipitation pattern
The interannual variation of the CEF seesaw shows a statistically significant positive correlation with the AISRI (with a correlation coefficient of 0.59) and a significant negative correlation with the EASMI (−0.31). Thus, the observations show a robust relation between the CEFs and the two Asian monsoons. In this subsection, the prediction skills of the AISRI, EASMI, and the CEF seesaw–related precipitation anomaly pattern over the entire Asia–Australia monsoon region are assessed.
Figure 8 shows the ACC skills of the AISRI and EASMI prediction in the three models. The results indicate that the prediction skill of the AISRI is low in POAMA-2 and SINTEX-F, in consistent with their low ACC skill of the Somali CEF. The ACC skill of the EASMI is higher than that of the AISRI in all the three models. This may be partly because the prediction skill of the MC CEF is higher than that of the Somali CEF, and partly because precipitation is generally less predictable compared to the monsoonal circulation. Again, the high-resolution ACCESS-S1 model shows a better performance in predicting the two monsoon indices. Encouragingly, all the models’ predictions outperform the persistence forecasts; also, the MME mean tends to (but does not always) produce the best forecasts compared to the three individual models.
Correlated with the CEF seesaw index, a “horseshoe” precipitation anomaly pattern in the Asian–Australian monsoon region during JJA is seen in the observations (Fig. 9a). Related to a positive CEF seesaw index, significant positive anomalies appear in northern Australia, the MC area, the Indian peninsula, and northern China, and negative anomalies prevail from the northern part of Bay of Bengal to southeastern China and the western Pacific. This precipitation anomaly pattern is similar to the first leading EOF (empirical orthogonal function) pattern over the entire Asian–Australian summer monsoon area (with a pattern correlation of 0.70). To assess how well the observed CEF seesaw–related precipitation anomaly pattern can be predicted by the models, we project the predicted precipitation anomaly onto the observed horseshoe pattern (i.e., Fig. 9a). This type of projection has been widely applied in previous studies [e.g., the Madden–Julian oscillation (MJO) predictions in which model forecasts are projected onto the observed MJO EOFs; see http://www.cpc.ncep.noaa.gov/products/precip/CWlink/MJO/CLIVAR/clivar_wh.shtml]. An alternative way is to assess the skill of the modeled CEF seesaw–related precipitation pattern. However, since the models have biases in capturing the observed CEF seesaw–related precipitation pattern, this type of assessment does not tell us the true skill in predicting the observed precipitation anomaly pattern.
The projection coefficients of the precipitation anomaly on the observed horseshoe pattern are then compared between the observed and multimodel prediction results (Fig. 9b,c). Interannual variations of this horseshoe precipitation anomaly pattern are highly predicted in the ACCESS-S1 model, and correlation skills between the observed and predicted projection coefficients reach about 0.85 at 2-month lead and about 0.7 at 5-month lead in ACCESS-S1 (green asterisks in Fig. 10). Comparatively, the other two models with coarse and intermediate resolutions show lower skill in predicting the CEF seesaw–related monsoonal precipitation pattern (Fig. 10). Nevertheless, the horseshoe monsoonal precipitation pattern can still be skillfully predicted out to 4–5 months lead in the two models. In particular, the positive (negative) phase of the horseshoe monsoonal precipitation pattern in 1988, 1995, 1998, and 2010 (1982, 1987, 1991, 1994, 1997, and 2002) is realistically predicted at 2-month and 5-month lead (Fig. 9).
In summary, the above results suggest the MC CEF is generally more predictable compared to the Somali CEF. On average, the correlation skill for the MC CEF reaches above 0.5 out to 5-month lead, with RMSEs being smaller than one standard deviation of the observations. In contrast, the models show a low prediction skill for the Somali CEF, with a low correlation skill and a large RMSE even at 2-month lead. Interestingly, the high-resolution ACCESS-S1 model shows the best skill in predicting the two individual CEFs, the CEF seesaw index, and the CEF seesaw–related monsoonal precipitation anomaly pattern for the common period of 1990–2012. While the superiority of the ACCESS-S1 over the other two models may be related to its high resolution, different initialization and parameterizations implemented in these model forecast systems could also play a role in producing the different prediction skills.
Previous results of Li et al. (2017) have suggested that the interannual variations of the Somali CEF are predominately influenced by the ENSO through the Walker circulation, rather than the SSTA in the western Indian Ocean or other tropical areas. In addition, it is the eastern Pacific SST anomalies that have the strongest impacts on the Somali CEF, and the central Pacific ENSO has little influence on the Somali CEF. Consequently, we may normally expect that a higher predictive skill of the EP ENSO should lead to a better skill in predicting the Somali CEF. However, as shown in the previous section, while SINTEX-F exhibits a higher ACC skill of the JJA Niño-3 index out to 5–6 months lead than POAMA-2 does, SINTEX-F fails to produce a better skill of the Somali CEF. In contrast, POAMA-2 produces an increasing prediction skill of the Somali CEF from 2-month lead to 5-month lead (recall Fig. 6). These results suggest that a higher predictive skill of Niño-3 SST anomaly may not guarantee a better skill in predicting the Somali CEF.
To investigate a possible reason for the different Somali CEF prediction skills across the models, Figs. 11 and 12 show the ACC skills in predicting global JJA mean SST and precipitation anomaly, respectively. Retrospective prediction results show a high SST predictive skill in the tropical Pacific at 2-month lead in the three prediction systems. Useful skills are produced across almost entire tropical Pacific and large parts of the Indian Ocean and the Atlantic (upper row of Fig. 11). The correlation skills in the SINTEX-F and ACCESS-S1 models decrease quickly from 2-month lead to 5-month lead (Figs. 11d,f). In contrast, POAMA-2 shows a smaller skill drop and produces a higher SST skill in the central Pacific at 5-month lead (Fig. 11b). The higher predictive skill of the POAMA-2 model at 5-month lead can also be seen clearly in the prediction of the precipitation anomaly in the western-central Pacific (Fig. 12). In POAMA-2, ACC skill of the precipitation anomaly in the western-central Pacific at 5-month lead decreases a little compared to that at 2-month lead. In contrast, the SINTEX-F model exhibits a large decrease of the precipitation prediction skill in the western-central Pacific, particularly over the MC area, from 2-month lead to 5-month lead (cf. Figs. 12c and 12d). This suggests that SINTEX-F has a poor prediction skill of the Walker circulation over the western Pacific, which may possibly explain its inability to predict the teleconnection that leads to ENSO’s influence on the Somali CEF.
A similar result is obtained from the composite analysis of the three strong EP El Niño years (1982, 1987, and 1997) in the POAMA-2 and SINTEX-F models (Figs. 13 and 14). It is interesting to note that, in POAMA-2, the predicted SST anomaly pattern is closer to the observed pattern at 5-month lead (with a pattern correlation coefficient of 0.78) than that at 2-month lead (with a pattern correlation coefficient of 0.40) (cf. Figs. 13b and 13d), although the intensity of the predicted SST anomaly at 5-month lead is still weaker than the observed. Furthermore, compared to those at 2-month lead, the POAMA-2 predictions at 5-month lead shows a better forecast of the precipitation anomaly in both the central Pacific Ocean and the India area, and a better forecast of the basinwide easterly and northerly anomalies in the equatorial and northern Indian Ocean (cf. Figs. 14b and 14d); this favors a better prediction of the Somali CEF at 5-month lead (Fig. 14d). This may explain the gradually increasing prediction skill of the Somali CEF from 2-month to 5-month lead in POAMA-2. In contrast, SINTEX-F shows better predictions of the EP El Niño SST anomaly at both 2- and 5-months lead (Figs. 13c and 13e) than POAMA-2 does, but SINTEX-F fails to produce a better prediction of the precipitation anomaly over the MC region, the India area, and the central Pacific at 5-month lead (cf. Figs. 14c and 14e). Correspondingly, SINTEX-F fails to predict the easterly and northerly anomalies in the tropical Indian Ocean at 5-month lead, which may lead to a poor prediction of the Somali CEF at 5-month lead. In summary, good prediction skill of both the Niño-3 SST and the Walker circulation anomaly over the MC area are required to produce useful skill of the Somali CEF.
It is worth noting that, while the CEF seesaw index and the associated precipitation pattern over Asia can be skillfully predicted up to 4–5 months ahead (recall Figs. 6 and 10), the precipitation anomalies over individual grid meshes in Asia have low prediction skill (Fig. 12). The discrepancy between the relatively high CEF seesaw skill and the low gridpoint precipitation skill over Asia might be partly due to two factors, namely that the gridpoint precipitation is generally difficult to predict and that the models have biases in reproducing the observed CEF seesaw–related monsoon precipitation pattern.
In this study, predictive skill of the Somali and MC CEFs during boreal summer (JJA) has been investigated based on three seasonal prediction models, the coarse-resolution POAMA-2, intermediate-resolution SINTEX-F, and high-resolution ACCESS-S1. The climatological structure of the Somali CEF is well captured in the three models. In contrast, the local structure of the MC CEF is well reproduced in the SINTEX-F and ACCESS-S1 models but not in the coarse-resolution POAMA-2 model. This deficiency in predicting the fine structure of the MC CEF in POAMA-2 may be attributed to its inadequate horizontal resolution, since the MC region has a complicated and sharp land–sea distribution. ACC and RMSE skill analyses suggest that the prediction of the Somali CEF is more challenging than the prediction of the MC CEF. This may be partly because that the MC CEF has a higher correlation with ENSO than the Somali CEF does. For the MC CEF, the multiple models show useful ACC skill (>0.5) up to about 5-month lead (i.e., the prediction starting from the first day of February). On the contrary, skillful prediction of the Somali CEF is produced only up to 2-month lead (i.e., starting from the first day of May) in the ACCESS-S1 model and MME mean. The comparison of the prediction skill among the multiple models suggests that the ACCESS-S1 model and the MME mean show the highest predictability of the two CEFs, the CEF seesaw index, and the related precipitation anomaly pattern in the Asian and Australian monsoonal region. In addition, the present results suggest that the prediction skill of the Somali CEF depends on not only the model’s ability in predicting the ENSO-related SST anomaly in the eastern Pacific (since EP ENSO has a large influence on the CEF) but also the model’s ability in predicting the Walker circulation anomaly that brings ENSO’s influence onto the Somali CEF.
It is well known that the interannual variation of the CEFs shows a close correlation with the Asian summer monsoon and Australian winter climate. However, few existing studies have examined the predictive skill of the CEFs. And although many previous studies have addressed the predictability of the Asian summer monsoon, correct prediction of the monsoon precipitation is still a big challenge in current state-of-the-art climate prediction systems. The present study provides a preliminary assessment of the predictive skill of the Somali CEF, MC CEF, and monsoonal precipitation anomaly pattern related to the CEF seesaw, based on three seasonal prediction systems. Our results may help improve the understanding of the predictability of interannual variations of the two CEFs and the Asian summer monsoon precipitation.
This study is jointly supported by the Strategic Project of the Chinese Academy of Sciences (Grant XDA11010401), the National Key Research and Development Program of China (2015CB453202 and 2016YFA0601802), and the Natural Science Foundation of China (41421004 and 41528502). The authors thank Drs. Matthew Wheeler, Aurel Moise, Eun-Pa Lim, and Joshua Soderholm for their helpful comments and suggestions.