Impact of Indian Ocean Dipole on Atlantic Niño predictive skill

Figure 1 depicts the TCC skills of the ATL3 and DMI indices as a function of lead times for all months combined in the individual models and MME. It is shown that the prediction skills of the ATL3 index decay slightly faster than those of the DMI index in both the individual models and MME when the lead time increases (figures 1(a) and (b)). When 0.5 is chosen as the cut-off value for the TCC skill, following the previous study (Stockdale et al 2006, Wang et al 2021), less than half of the models can predict the ATL3 index skillfully 2 months in advance, while most of the models can predict the DMI index skillfully at 2 month lead. The result regarding the DMI index is consistent with that of Liu et al (2017). For a cut-off value of 0.6, only the MME and ECMWF5 have a predictive skill for ATL3 as long as 2 months. For both indices, the MME shows higher TCC skills than almost all the individual models (figures 1(a) and (b)). Especially for lead times ranging from 2 to 4 months, the MME can predict the ATL3 index much better than the individual models, which may imply that there is still much room for improvement in predicting the ATL3.

Zoom In Zoom Out Reset image size

The seasonal dependence of the TCC skills in the ATL3 and DMI indices is further examined. In general, the TCC skills of the ATL3 index are lower for target months in summer and higher in winter (figure 1(c)). For the MME average, the effective forecast (TCC ⩾ 0.6) of the ATL3 index can reach 5 months in winter but is limited to only 2 months in summer. Although the TCC skills of the DMI index are relatively low for target months in spring and summer, the skills become higher for target months in autumn and are even higher than 0.7 at 5 month lead in late autumn (figure 1(d)). Moreover, a comparison of figures 1(a) and (b) indicates that the individual models generally perform better in predicting the DMI index than the ATL3 index, though the MME has a similar performance when all months are combined. The higher TCC skills of DMI in autumn remind us of that the IOD peaking in autumn may provide a source of predictability for the Atlantic Niño peaking in the following winter.

To examine the impact of the IOD on the prediction skills of Atlantic Niño in winter, figures 2(a)–(d) show the inter-model relationship between the TCC skill of the ATL3 index in DJF and the strength of the IOD–Atlantic Niño connection with a lead of the IOD over the Atlantic Niño in DJF by 0–3 months. Note that the strength of the IOD–Atlantic Niño connection is represented as the correlation coefficient between the ATL3 and DMI indices, where the ATL3 index is fit on the target season of DJF, and the DMI index varies with the lead time. It is indicated that the TCC skill of the ATL3 in DJF has a significant and in-phase relationship with the strength of the IOD–Atlantic Niño connection among the models when the IOD leads the Atlantic Niño in DJF by 1–3 months. Moreover, the majority of the models capture a positive correlation between the DJF ATL3 and the previous autumn DMI index, as seen in the observations, suggesting that these models can reasonably reproduce the observed IOD–Atlantic Niño connection. Considering the difference in the prediction skills of the eastern and western lobes of the IOD (Luo et al 2005), the impacts of the IODW and IODE on the TCC skills of the ATL3 in DJF are further examined, and similar results are obtained, except that the IODW-related inter-model relationship is slightly stronger than that for the IODE (figures 2(e)–(l)). Therefore, these results confirm that the IOD that peaks in autumn provides an important source of seasonal predictability for the Atlantic Niño in the following winter.

Zoom In Zoom Out Reset image size

To further verify the impact of the IOD–Atlantic Niño connection on the prediction skills of DJF ATL3, we compare the prediction skills of the ATL3 index in six models that have the strongest IOD–Atlantic Niño connection (i.e. CMCC-SPS3, GloSea5, CanCM4i, MF6, ECMWF5 and GloSea6, referred to as the strong-connection models hereafter) with those in six models that have the weakest IOD–Atlantic Niño connection (i.e. CFSv2, FIO-ESM, JMA_CPS2, CCSM4, NUIST_CFS1.0, and GEM_NEMO, referred to as the weak-connection models hereafter). Note that the strong- and weak-connection models are chosen according to the correlation coefficient between the ATL3 index in DJF and the DMI index in the previous November. For the strong-connection models, the correlation coefficients between the DJF ATL3 and November DMI indices are all significant at the 95% confidence level (p< 0.05), while for the weak-connection models, the correlation coefficients are below significance at the 90% confidence level (p> 0.1). The correlation coefficients between the ATL3 in DJF and DMI indices in October and September are also verified, and similar results are obtained.

Figure 3 displays the TCC and RMSE skills of the DJF ATL3 index as a function of lead times for the MME of the strong- and weak-connection models, respectively. It is evident that the TCC skills of DJF ATL3 for the MME average of the weak-connection models are lower and decay faster than those for the MME average of the strong-connection models, with a skill value of less than 0.50 for the former but higher than 0.70 for the latter at 3 month lead (figure 3(a)). Consistently, the RMSE of DJF ATL3 for the MME average of the strong-connection models is smaller than that for the MME average of the weak-connection models (figure 3(b)). Moreover, models with a stronger IOD–Atlantic Niño connection generally have higher prediction skills for the winter ATL3 index than those with a weaker IOD–Atlantic Niño connection. These results support the significant impact of the IOD in autumn on the seasonal predictability of Atlantic Niño in the following winter.

Zoom In Zoom Out Reset image size

In both the observations and MME of the strong-connection models, two of the three typical positive IOD events (1994, 1997, and 2006) during 1994–2016 are corresponding to an Atlantic Niño event in the following winter (supplementary figure S2), supporting the significant impact of the IOD on the late-onset Atlantic Niño event. For the three typical negative IOD events (1996, 2010, and 2016), however, only the one in 1996 is followed by an Atlantic Niña event, while the other two correspond to neutral SST anomalies in the eastern equatorial Atlantic in the following winter. Such an asymmetry in the response of the Atlantic Niño/ Niña to the IOD may be attributed to the asymmetry in the amplitude of the IOD. The positive IOD events tend to have a larger amplitude and therefore stronger climate impact than the negative IOD events (supplementary figure S2, Cai et al 2013, Qiu et al 2014, Behera and Ratnam 2018).

We also examined the inter-model relationship between the strength of the IOD–Atlantic Niño connection and the prediction skill of the ATL3 index in other seasons, but no significant relationship was found (figure not shown). Liao and Wang (2021) suggested that the western lobe of the IOD in DJF can trigger an Atlantic Niño event that peaks in the following summer, indicating that the former leads the latter by approximately two seasons. However, the time range for hindcast and forecast is only 6 calendar months for most of the operational seasonal forecast systems used in this study, which may be too short to detect the impact of the WIO SST anomalies on the prediction skill of the Atlantic Niño that peaks in summer.

To investigate the model performance in simulating relationship between the winter Atlantic Niño and the previous autumn IOD, regression analysis is applied to the SST and 850 hPa wind anomalies from SON to DJF and the DMI index in SON for the observations and the MME averages of the strong- and weak-connection models, respectively. Note that all the hindcasts used here are initialized from September, and CCSM4 and NUIST_CFS1.0 are excluded in the analysis due to the lack of wind data. As suggested by Zhang and Han (2021), the positive rainfall anomalies over the western lobe of the IOD in autumn can trigger an anomalous lower-level cyclonic circulation over the southern tropical Atlantic in the following months, accompanied by westerly wind anomalies in the equatorial eastern Atlantic and northern wind anomalies over western Africa (supplementary figures S3(a)–(d)). Consequently, warm SST anomalies appear along the coast of Angola 1 month after the peak of IOD in SON and then extend to the eastern equatorial Atlantic and peaks in DJF. It is found that the MME average of the strong-connection models can reasonably capture the IOD-related lower-level cyclonic circulation anomalies over the southern tropical Atlantic and thus the warm SST anomalies in the eastern equatorial Atlantic in NDJ and DJF, although the amplitude of the anomalies is weaker than the observed one (supplementary figures S3(g) and (h)). In contrast, the MME average of the weak-connection models fails to reproduce the lower-level circulation and SST anomalies in the southern tropical Atlantic, though the spatial pattern of the circulation and SST anomalies in the tropical Indian ocean resembles that in the observation as well as in the MME average of the strong-connection models (supplementary figure S3).

To better understand the factors that affect model performance in simulating the IOD–Atlantic Niño connection, figure 4 displays the differences in the SST and 850 hPa wind anomalies from SON to DJF regressed onto the DMI index in SON between the strong- and weak-connection models. Compared to those in the weak-connection models, significant warm SST anomalies appear in the equatorial Atlantic in the MME average of the strong-connection models (figures 4(a)–(d)), which are associated with the lower-level cyclonic circulation anomalies over the southern tropical Atlantic developing in OND (figures 4(c) and (d)). These wind anomalies are likely originated from the stronger positive response of precipitation to the SST warming in the western lobe of the IOD in the strong-connection models when compared to the weak-connection models (figures 4(f)–(h); supplementary figures S4(f)–(h) and (j)–(l)). The westerly wind anomalies in the equatorial Atlantic associated with the lower-level cyclones induce eastward oceanic currents and ultimately lead to warm SST anomalies in the eastern Atlantic basin through the Bjerknes positive feedback (Zhang and Han 2021). These results suggest that the stronger precipitation response to the SST forcing in the western lobe of the IOD contributes to a stronger IOD–Atlantic Niño connection in the models.

Zoom In Zoom Out Reset image size

A question arises as to why some seasonal forecast models show a stronger precipitation response to the SST anomalies in the western lobe of the IOD and thus a stronger IOD–Atlantic Niño connection than the other models. It can be expected that the mean state bias in the tropical Indian Ocean SST may play a role, as a higher background SST generally benefits more active deep convection over the tropics (Gadgil et al 1984, Graham et al 1987, Waliser and Graham 1993).

As shown in figure 5, when compared to those in the weak-connection models, the MME average of the strong-connection models shows a higher climatological mean of SST in the western Indian Ocean in autumn (figure 5(a)), accompanied by heavier climatological mean of precipitation there (figure 5(c)). Similar results are observed in winter (figures 5(b) and (d)), except that the significant differences in the climatological mean of SST and precipitation mainly occur over the southern tropical Indian Ocean. Comparison of figures 5 and 4 indicates that the stronger precipitation response to the IOD-related SST anomalies is consistent with the warmer mean SST in the western Indian Ocean in the strong-connection models when compared to those in the weak-connection models. Additionally, when compared to the climatological mean of SST in observations, the MME average of the strong-connection models shows a warmer SST bias in the western tropical Indian Ocean both in autumn and winter (supplementary figure S5), which is consistent with the stronger IOD–Atlantic Niño connection in those models than in the observations (figure 2(b)). In contrast, the MME average of the weak-connection models suffers a colder SST bias over the tropical Indian Ocean in winter (supplementary figure S5(d)).

Zoom In Zoom Out Reset image size

In short, the warmer climatological mean SST in the western tropical Indian Ocean benefits a stronger precipitation response to the underlying SST forcing in the strong-connection models. The stronger precipitation response to the western lobe of the IOD then triggers a stronger Gill-type Rossby wave response over the tropical southern Atlantic, which ultimately enhances the westerly wind anomalies over the equatorial Atlantic and thus warm SST anomalies in the eastern Atlantic basin. This is consistent with the result regarding the mechanism responsible for the linkage between the IOD and Atlantic Niño, as proposed by Zhang and Han (2021). Inter-model relationship of the mean SST and precipitation biases in the tropical Indian Ocean with the TCC skill of the DJF ATL3 index as well as the SON DMI–DJF ALT3 correlation is examined (supplementary figure S6), and the results support our conclusion regarding the effect of tropical Indian Ocean mean state bias on the IOD–Atlantic Niño connection and thus prediction skill of the Atlantic Niño in boreal winter. Previous studies have demonstrated that the warmer SST bias in the tropical Indian Ocean in autumn can be traced back to errors in the South Asian summer monsoon and can interact with the equatorial easterly bias over Indian Ocean through the Bjerknes feedback (Li et al 2015, Lu et al 2018b, Long et al 2020), but this is beyond the scope of the present study.