Sea level anomalies in the southeastern tropical Indian Ocean as a potential predictor of La Niña beyond one-year lead

Most climate forecast agencies failed to make successful predictions of the strong 2020/2021 La Niña event before May 2020. The western equatorial Pacific warm water volume (WWV) before the 2020 spring failed to predict this La Niña event because of the near neutral state of the equatorial Pacific Ocean in the year before. A strong Indian Ocean Dipole (IOD) event took place in the fall of 2019, which is used as a precursor for the La Niña prediction in this study. We used observational data to construct the precursory relationship between negative sea level anomalies (SLA) in the southeastern tropical Indian Ocean (SETIO) in boreal fall and negative Niño 3.4 sea surface temperature anomalies index one year later. The application of the above relation to the prediction of the 2020/2021 La Niña was a great success. The dynamics behind are the Indo-Pacific “oceanic channel” connection via the Indian Ocean Kelvin wave propagation through the Indonesian seas, with the atmospheric bridge playing a secondary role. The high predictability of La Niña across the spring barrier if a positive IOD should occur in the previous year suggests that the negative SETIO SLA in fall is a much better and longer predictor for this type of La Niña prediction than the WWV. In comparison, positive SETIO SLA lead either El Niño or La Niña by one year, suggesting uncertainty of El Niño predictions.


Introduction
ENSO is one of the most important climate phenomena on interannual time scales, involving variations in sea surface temperature (SST) in the central and eastern tropical Pacific Ocean, influencing temperature and precipitation globally through variations in atmospheric circulation. Substantial efforts have been dedicated to predicting its onset and development because of its extensive repercussions worldwide. The ENSO has been shown to be predictable with one-to two-year lead times in certain numerical hindcast experiments (Kirtman and Schopf, 1998;Fedorov and Philander, 2000;Chen et al., 2004;Luo et al., 2008). However, considerable uncertainties still exist in realtime forecasts (Chen and Cane, 2008;Jin et al., 2008;Duan and Wei, 2012;Tang et al., 2018), and practical prediction skills are currently limited to six months (Latif et al., 1998;Barnston et al., 1999; suggested that SSTA in the Atlantic and Indian Ocean may initiate the onset of these two La Niña events (Luo et al., 2017;Hasan et al., 2022).
Prediction of ENSO using the Indian Ocean Dipole (IOD, Saji et al., 1999;Webster et al., 1999) index as a precursor has the potential to increase the lead times of successful prediction to as long as one year across the spring barrier. Existing studies have focused on atmospheric teleconnection via the Walker circulation (Kug and Kang, 2006;Izumo et al., 2010;Ha et al., 2017). Other studies suggest that the "oceanic channel" of the Indonesian Throughflow (ITF) may provide an important connection between the Indian and Pacific Ocean climate variabilities (Yuan et al., 2011;Yuan et al., 2013;Mayer et al., 2018;Yuan et al., 2018;Wang et al., 2022;Xu et al., 2022;Yuan et al., 2022). The cold tongue SSTA were found to correlate significantly with sea level anomalies (SLA) in the southeastern tropical Indian Ocean (SETIO), the eastern pole of the IOD, one year before. The dynamics are through Kelvin wave propagation from the equatorial Indian Ocean into the western Pacific Ocean via the Indonesian seas. ENSO is characterized by asymmetry in amplitude, duration and transitions (Okumura and Deser, 2010). El Niño tend to decay rapidly followed by La Niña, whereas La Niña sometimes last for two years (Ohba and Ueda 2009;Hu et al., 2014Hu et al., , 2016DiNezio et al., 2017;Okumura et al., 2017;Iwakiri and Watanabe, 2020). So far, the role of the Indo-Pacific "oceanic channel" dynamics in the asymmetric predictability of El Niño and La Niña has not been investigated.
In this study, a simple linear regression model is developed to apply the "oceanic channel" dynamics to ENSO prediction, in which only the SLA in the SETIO are used as a precursor. Compared to the WWV in the western Pacific in fall, which could not transpire the spring barrier in general, the upwelling anomalies (negative SLA) in the SETIO is a more effective and longer-lead predictor of La Niña. The model predicts essentially all of the La Niña events following the outbursts of the positive IOD (pIOD) events successfully at the lead time of one year, including the 2020/2021 cold event. These results underline the importance of the "oceanic channel" dynamics in the predictability of one type of La Niña event preceded by pIOD.
The remainder of this paper is organized as follows. Section 2 introduces the observational dataset and methods used in this study. Section 3 presents the asymmetry in the lead relationship between the SETIO SLA and Niño 3.4 SSTA, which suggests that the negative SETIO SLA in the boreal fall is a better long-lead precursor for one type of La Niña event (those preceded by pIODs) at a lead time of one year. The paper closes with conclusions and discussions in Section 4.
2 Data and method 2.1 Data

Reanalysis and observation data
The sea level data from 1993 to 2021 was obtained from the Aviso satellite altimeter. The SST data included the ERSST v5 (Huang et al., 2017). Wind data were obtained from NCEP data (Kalnay et al., 1996). Other wind data were obtained from the ERA5 reanalysis, which have been regridded to a regular lat-lon grid of 0.25 degrees (Hersbach et al., 2020). The surface current data is provided by the Estimating the Circulation and Climate of the Ocean, Phase II (ECCO2) reanalysis. The horizontal resolution is 1/ 4-deg. ORA-S5 was not used here because data after 2019 were not available. The annual cycle of each variable was removed by subtracting the monthly climatological means. Subsurface temperature anomalies from the global gridded Argo dataset are used, which has a 1°×1°horizontal resolution and 58 vertical layers from the surface to 1975 dbar (Roemmich and Gilson, 2009). The WWV index was derived from the ocean analyses of the Bureau National Operations Centre (BNOC) at the Australian Bureau of Meteorology (Smith, 1995), which are based on Tropical Atmosphere Ocean (TAO) array, Argo floats and XBT measurements. It is well known that the sea level rise in the tropical Indian Ocean in the last two decades has been significantly higher than before (Unnikrishnan et al., 2015). The SLA was detrended as we focused on interannual variability.

Linear regression model of La Niña prediction
The La Niña prediction model was built as follows: SLA_SETIO(t − 1) is the SLA over SETIO during the boreal fall when pIOD occurs. SSTA_Nino3.4 (t) is the SSTA over the Niño 3.4 region during the next fall. b is the regression coefficient and a is a constant term. The data during 1993-2019 were used to calculate and b and to make the 2020/2021 La Niña prediction using the negative SETIO SLA during the fall of 2019.

Cross-validation method
To test the predictive capability of the linear regression model in La Niña prediction, the cross-validation method (Michaelsen, 1987) was used to hindcast the Niño 3.4 index during La Niña events. We chose the leaving-one-out strategy. This strategy systematically deletes one year from 1993 to 2021, builds a forecast model from the remaining years, and tests the prediction on the deleted case. For example, for the 2020/21 case, the 2020 data were deleted, and a forecast model was built from the remaining years to predict the 2020/2021 La Niña event. The assessment of the prediction was evaluated using the correlation between the linear regression model prediction (using the cross-validation method) and observed data from 1993 to 2021.

SETIO SLA calculated from the wind-forced Kelvin wave for the longer period
A linear coastal-trapped Kelvin wave model (Hu et al., 2019b) was used to calculate the Indian Ocean SLA since 1950. This model is based on an analytical Kelvin wave model (Gill, 1982;Sprintall et al., 2000), in which wind-forced Kelvin wave propagation along the Kelvin wave pathway from the western Indian Ocean (0°, 45°E) to the Maluku Channel (2°N, 126.5°E). The model use the momentum equation in the alongshore direction: where A k is the SLA along the Kelvin wave characteristics, g is the acceleration due to gravity, X is the projection of alongshore wind stress onto each baroclinic mode, and c is the wave speed (2.71 m/s for the first baroclinic mode). Equation (2) is integrated along the Kelvin wave characteristics to get the solution of A k (Hu et al., 2019b).

Definition of indices
In this study, we defined the SLA index over the SETIO as the average SLA in the area of 10°S-0°N, 90°E-110°E. The Niño 3.4 index was calculated as the average SSTA in the area of 5°S-5°N, 170°W-120°W. The Indian Ocean dipole mode index (DMI) was constructed by the differences in SSTA in the western tropical Indian Ocean (50°E-70°E, 10°S-10°N) and southeastern tropical Indian Ocean (90°E-110°E, 0°-10°S). The WWV index is defined as the volume of water above the 20°C isotherm in the equatorial Pacific (120°E-80°W, 5°N-5°S). The WWV_w index is defined as the WWV in the western equatorial Pacific (155°E-120°E, 5°N-5°S ).

Results
3.1 A better long-lead precursor of the La Niña: negative SETIO SLA at the lead time of one year Figure 1A shows the lead relation between SLA in the SETIO (10°S-0°, 90°E-110°E) in the boreal fall (September to November) and SSTA in the Niño 3.4 region (5°S-5°N, 170°W-120°W) oneyear later. Only the relationship after 1993 is indicated because of the short time series of AVISO SLA data. Tide gauge sea level data along the Sumatra-Java Island chain are extremely rare and are unable to expand the time coverage of the SLA-Niño 3.4 relationship much longer in history.
A noteworthy feature is an asymmetry in the lead relationship between the SETIO SLA and Niño 3.4 SSTA. For a pIOD event, following the upwelling anomalies (negative SLA) off Java in the fall, a La Niña event will occur in the next fall (in the third quadrant of Figure 1A). Their correlation coefficient is 0.97 (exceeding the 99% confidence level according to Student's t-test), which is much greater than the correlation coefficient of 0.39 (exceed the 95% confidence level) for the whole dataset. The relationship between the negative SETIO SLA and La Niña one year later is one-to-one since no strong warm events in the equatorial Pacific were found following pIOD events (no points far away from the origin in the second quadrant, Figure 1A). In contrast, for the negative phase of the IOD (nIOD), following the downwelling anomalies (positive SLA) off Java in the fall, the Niño 3.4 SSTA may be either positive (in the first quadrant) or negative (in the fourth quadrant) in the following year, suggesting either an El Niño or La Niña event following the nIOD. This result reveals an asymmetric lead relation between the SETIO SLA and Niño 3.4 SSTA, with negative SLA (upwelling anomalies) in the SETIO in the fall being a better predictor of La Niña one year later.
When strong upwelling anomalies (negative SLA) occurred over the SETIO during the boreal fall, strong La Niña events occurred in the following year. Since 1993Since , 1995Since /1996Since , 1998Since /1999Since , 2007Since /2008Since , 2016Since /2017Since , and 2020Since /2021 La Niña events followed the strong pIOD events in 1994,1997,2006,2015, and 2019 ( Figure 1A). Among them, three events are the multi-year lingering La Niña events, 1998Niña events, /99/2000Niña events, , 2016Niña events, /17/18, and 2020. Notably, the relationship between the SETIO SLA in fall and the Niño 3.4 index a year later was different in the first and second years of the 2-year La Niña. La Niña in the first year was preceded by the pIOD (in the third quadrant of Figure 1A), whereas the SETIO SLA is positive in the previous year for the La Niña in the second year for the 1999/2000, 2017/18, and 2021/22 La Niña events (in the fourth quadrant of Figure 1A). This means that one type of La Niña events (those are preceded by pIODs) could be predicted at a lead time of one year based on the negative SETIO SLA during the pIOD events, whereas predictability of another type of La Niña events, the 2-year La Niña in the second year, does not arise from the factor in the Indian Ocean in the previous year. Anyway, the negative SLA in SETIO is a better precursor for most strong La Niña events, including the 2-year La Niña evens in the first year. Here, 2-year La Niña event on 2010/11/12 is a special case. The SETIO SLA are both positive in the previous first and second years.
The above lead relation was found to be not as good between SETIO SSTA and Niño 3.4 SSTA as between SETIO SLA and Niño 3.4 SSTA ( Figure 1B). The correlation coefficient between the negative SETIO SSTA and Niño 3.4 SSTA one year later was 0.83. This is expected because the SSTA are a less direct representation of upwelling anomalies off Sumatra-Java than the SLA. The use of SSTA in SETIO as the precursor generally does not lead to successful La Niña predictions as SLA used as a precursor. Moreover, an asymmetric lead relation with Niño 3.4 SSTA does not exist for the SETIO SSTA in the longer dataset (see Section3.5), although the relationship between the negative SETIO SSTA in fall and La Niña one year later is one-to-one after 1993. Hasan et al. (2022) indicated that the Indian Ocean may contribute to the initiation and predictability of the 2020/21/22 La Niña, but they did not find that the negative SLA in the SETIO was a better longlead precursor for a strong La Niña event.
The ENSO phase transition is asymmetric (Kessler, 2002;Larkin and Harrison, 2002;Okumura and Deser, 2010). La Niña often follows El Niño, but the reverse is not always true. Furthermore, El Niño often forces a response in the Indian Ocean in the form of IOD Xue et al., 2022). Does the good relationship between La Niña and negative SETIO SLA at the lead time of one-year result from the preceding El Niño? As shown in Figure 2A, the asymmetry in the ENSO phase transition was also found in the relationship between the Niño 3.4 SSTA in the previous and following years. However, the correlation between positive Niño 3.4 SSTA in the previous fall and negative Niño 3.4 SSTA in the following year was not significant (-0.23). Thus, the Niño 3.4 SSTA in the fall is not a good precursor for La Niña events one year later.
Furthermore, according to the ENSO recharge oscillator (Jin, 1997), ENSO events tend to be preceded by anomalous WWV in the equatorial Pacific. The release of WWV in the equatorial Pacific can lead to La Niña events in the following year (Meinen and McPhaden, 2000). The negative WWV index in fall, especially the WWV in the western equatorial Pacific (WWV_w), has been consistently considered the best precursor of La Niña one-year before (Planton et al., 2018). As shown in Figures 2B, C, for the type of the La Niña events preceded by pIOD, 1995pIOD, /96, 1998pIOD, /99, 2007pIOD, / 08, and 2016 La Niña events, negative WWV anomalies in the western equatorial Pacific (in the third quadrant in Figure 2C) occurred in the previous fall but with positive WWV anomalies across the entire equatorial Pacific (in the fourth quadrant in Figure 2B). Although the correlation of the negative Niño 3.4 index with the discharged WWV_w one year before (0.43) was higher than that with the positive Niño 3.4 SSTA at the one-year lead time (-0.23), it was still not significant and far less than that with the negative SETIO SLA at a lead time of one year (0.97, with 99% confidence level). This result suggests that, at longer lead times, before the spring predictability barrier, the WWV or Niño 3.4 index is not a good predictor of the type of La Niña events preceded by the pIOD. Especially for the 2020/21 La Niña, WWV_w is not negative, meaning that the equatorial western Pacific is not in a discharge state.
Therefore, although the negative SLA in the SETIO in the fall may not be an independent predictor of La Niña (which is partly correlated with preceding El Niño), it is a better long-lead precursor for the type of La Niña events (preceded by pIOD) than the positive SSTA in Niño 3.4 and discharge WWV in the western Pacific as a predictor.

Prediction of the 2020/21 La Niña event
We noticed this long-lead relationship between SETIO SLA and Niño 3.4 SSTA in 2019 and applied it to the 2020/21 ENSO forecasting as soon as the 2019 pIOD was in a mature phase. We constructed a simple linear regression model based on the negative SETIO SLA in the fall and the Niño 3.4 index one year later during 1993-2019 (see section 2.2.1). Then, using the negative SETIO SLA during the fall of 2019 already observed as a precursor, we predicted the Niño 3.4 SSTA in the coming fall of 2020 to be -1.34°C (orange color dot in Figure 1A) with 95% confidence interval (-1.56°C, -1.12°C). One year later, the observed Niño 3.4 index (-1.11°C, red dot in Figure 1A) was very close to our prediction made one year earlier (with only 0.23°C difference), suggesting the success of the simple linear regression model in predicting the 2020/2021 La Niña beyond the lead time of 1 year. Our prediction is evidently much better than the real-time forecasts of the IRI models using initial conditions in earlymid-2020 (blue dots in Figure 1A), which were made much later than our prediction and were much weaker than the observation (red dot in Figure 1A) in approximately six months. The IRI forecasts made in January through May 2020 had predicted the Niño 3.4 SSTA to be near zero in the coming fall of 2020, while forecasts made in June through August of the majority of the models in the IRI plume continue to favor an ENSO-neutral state (Niño 3.4 index near 0.5°C) in the coming fall. Therefore, the 2020/ 21 La Niña event was not successfully predicted by the IRI models at lead times of one to three seasons.

Dynamics of the 2020/21 La Niña predictability
The 2020/21 La Niña followed a weak El Niño; thus, the discharge process in the tropical Pacific was not strong enough to energize the event (Figure 2). In the summer-fall of 2019, associated with the climax of the 2019 super pIOD, the anomalous easterly winds were strong to the west over the equatorial Indian Ocean, forcing the SSTA and SLA in the southeastern basin to grow quickly into a strong pIOD state (Figure 3). The equatorial Indian Ocean easterly wind anomalies in late 2019 forced strong upwelling Kelvin waves to propagate to the east along the Sumatra-Java island chain, which induces negative SLA in the SETIO and alters the pressure gradient across the Indonesian Seas. ITF transport increased significantly during the summer-fall of 2019 in response to pIOD forcing. Shortly thereafter, the SLA in the western equatorial Pacific Ocean began to propagate eastward along the equator around January 2020 (Figures 3, 4). The upwelling Indian Ocean Kelvin waves forced by easterly wind anomalies during the 2019 pIOD are thus suggested to propagate eastward through the Indonesian seas and reach the western equatorial Pacific, which triggered or enhanced the development of a strong La Niña in 2020. These results underscore the importance of inter-basin dynamic interactions between the Indian and Pacific Oceans through the "oceanic channel" in forcing the 2020/21 La Niña. None of the IRI dynamical coupled climate models can predict the 2020/21 La Niña one year in advance probably because they do not simulate the Indo-Pacific "oceanic channel" dynamics well although they include this oceanic channel.
The propagation of Kelvin waves into the Indonesian seas has been demonstrated by in situ direct mooring observations and ocean reanalysis data. Using moored current observations in the Makassar Strait and the Maluku Channel combined with highresolution ocean modeling validated with satellite data, it has been shown that Indian Ocean Kelvin waves propagated into the Makassar Strait and arrived at the northeastern Indonesian seas during 2015-2016 (Hu et al., 2019b;Pujiana et al., 2019). The propagation of the strong upwelling equatorial Kelvin waves forced by the Indian Ocean easterly anomalies during the 2019 pIOD was simulated successfully by Estimating the Circulation and Climate of the Ocean Phase II (ECCO2) reanalysis ( Figure 5), showing strong southward current anomalies in the Makassar Strait from October 2019 through February 2020. The currents and SLA in the Makassar Strait were evidently not forced by local winds, because the surface wind anomalies were southeasterlies (hollow arrow in Figure 5). The high resolution of ECCO2 simulations has resolved the channels and straits in the Indonesian seas well, showing a negative SLA propagating into the Indonesian seas along the south Java coasts in the winter of 2019 in agreement with the altimeter sea level anomalies. This suggests the propagation of Indian Ocean upwelling Kelvin waves through the Indonesian seas.
The upwelling Kelvin waves arrived in the western Pacific Ocean, as suggested by the strong westward current anomalies in the Sulawesi Sea in December 2019, coincides with the southward current anomalies in the Makassar Strait. The upwelling equatorial Kelvin waves propagated further eastward along the equator (Figure 4). From April to May, upwelling Kelvin waves induced negative subsurface temperature anomalies in the eastern equatorial Pacific, which moved to the surface associated with the shoaling thermocline in the eastern equatorial Pacific. These anomalies are suggested to trigger Bjerknes positive feedback, leading to an outburst of the 2020/21 La Niña event.
It is known that interbasin interaction between the tropical Indian and Pacific also involves an atmospheric bridge. The easterly wind anomalies in the western Pacific persist from spring to winter, triggering the cold SSTA in the eastern Pacific. Previous studies demonstrated that the pIOD can generate easterly wind anomalies in the western Pacific through the Indo-Pacific atmospheric bridge (e.g., Hasan et al., 2022). However, the lead relation between the SETIO SLA in fall and the easterly wind anomalies in the western equatorial Pacific during the following summer is not significant with the lag correlation coefficient of 0.4 (Figure 6), indicating that the atmospheric bridge is not the dominant process contributing to the long-lead good predictor of La Niña events preceded by pIOD. Moreover, Wang et al. (2022) compared the relative roles of the atmospheric bridge and the oceanic channel between the Indian and Pacific in ENSO forecasts through pacemaker experiments. Their model results suggested that the "oceanic channel" dynamics is dominant, whereas the 2020/21 La Niña cannot be generated by the atmospheric bridge alone. They demonstrated that La Niña event is induced by subsurface temperature anomalies in the tropical Indian Ocean propagating to the eastern equatorial Pacific Ocean at the end of 2020.
The westerly anomalies in the western equatorial Pacific also forced upwelling Rossby waves with negative SLA in the far western Pacific Ocean from November 2019 to February 2020 ( Figures 7A,  B). The Rossby waves propagated westward and were reflected into upwelling Kelvin waves at the western boundary. This process may have contributed to the onset of the 2020/21 La Niña event ( Figure 7C). However, the SLA off the east coasts of Mindanao (5°N -8°N, 127°E-135°E), representing the equatorial Rossby waves before reflection at the Pacific western boundary, shows poor correlation with the Niño 3.4 index at the one-year time lag   1995/1996, 1998/1999, 2007/2008, and 2010/2011 La Niña events were not successfully predicted by the NMME models at one-year lead ( Figure 9). A simple linear regression model, based on the lead relation between the negative SETIO SLA in fall and Niño 3.4 one year later, has been tested for its predictive capability to hindcast the historical La Niña events preceded by pIOD since 1993 (eight cases in the third quadrant of Figure 1A). Here, the cross-validation method (see section 2.2.2) was used to hindcast these La Niña events. Finally, for the eight La Niña events preceded by pIOD, the correlation between the predicted and observed Niño 3.4 indices is 0.95, above the 99% significance level. The small root-mean-squareerror (RMSE) of 0.14°C suggests successful predictions of the La Niña events at a lead time of at least 12 months.
The forecasts of our linear regression model are not only good for the mature phase, but also for the evolution of strong La Niña since 1993. Figure   The propagation of the Kelvin waves over the Indo-Pacific Ocean during 2019-2020. Monthly Argo temperature anomalies (°C) over the equatorial Indian and Pacific Ocean (5°S-5°N) during 2019-2020.
SETIO SLA in the fall and the Niño 3.4 indices from June through March of the third year, the initial cold SSTA in the equatorial eastern Pacific in June are predicted successfully, which are strengthened gradually in the summer through winter owing to the Bjerknes positive feedback induced by the upwelling Kelvin waves. After reaching the mature state in winter, the La Niña anomalies begin to decay, presumably due to the arrival of negative feedbacks, including reflected Kelvin waves from the western boundary. The time evolution of La Niña predicted by our linear regression model was in reasonable agreement with the observations. The Indo-Pacific "oceanic channel" dynamics still work to trigger/enhance the development of a strong La Niña preceded by 1994, 1997, and 2006 pIOD events, because 1) it can been seen that, from the SLA evolution (Figure 11), the upwelling Kelvin wave propagate eastward through the Indonesian seas and reach the western equatorial Pacific, and 2) the Makassar Strait southward currents is also identified during the pIOD events of 1994, 1997, and 2006 (Figure 12), suggesting the propagation of Indian Ocean upwelling Kelvin waves through the Indonesian seas.
Although easterly wind anomalies in the western Pacific followed the 1994, 1997, and 2006 pIOD (Figure 6), the lead relationship between the SETIO SLA in fall and the surface zonal wind anomalies in the western Pacific in the following summer is not significant. Thus, the atmospheric bridge is not the dominant process contributing to the long-lead good predictor of La Niña events in 1995Niña events in /96, 1998Niña events in /99, and 2007 Similar to the 2020/21 La Niña event, the 1995/96 and 2007/08 La Niña events followed weak El Niño events in 1994/95 and 2006/ 07, respectively, which were insufficient to produce significant discharge processes (Figures 2A, C). The Indo-Pacific interaction potentially plausibly energized these La Niña events through "ocean channel". However, the 1998/99 La Niña event was preceded by a strong El Niño event in 1997/98, which may be controlled by the discharge process. Meantime, the Indo-Pacific interaction may also play an important role, in which "ocean channel" dynamics potentially contribute to the La Niña event. The combined contributions of the discharge process in the tropical Pacific and remote forcing from the Indian Ocean may enhance the prediction of La Niña events.

Application the statistical model to longer time series in history
The statistical model above using oceanic signals from the Indian Ocean can be used to make La Niña forecasts back in The propagation of the Kelvin waves through the Indonesian seas during 2019-2020. SLA (cm, shading) and surface current anomalies (m s -1 , vectors) from ECCO2 from October 2019 to February 2020. The hollow arrow indicates the direction of the surface wind anomalies over Indonesian Sea. The relationship between the SETIO SLA in fall and the surface zonal wind anomalies in the western equatorial Pacific (130°E -180°E, 5°N-5°S) in the following summer. Correlation coefficient in the third quadrant is shown in the lower left of this Figure. history. To overcome the short time series of AVISO SLA data, wind data from a much longer time series were used to estimate the Kelvin wave SLA (Equation 2), which represent the Indo-Pacific "oceanic channel" connection through the Indonesian seas. Here, wind data from the ERA5 reanalysis of the ECMWF were used. Consistent results were obtained for the NCEP product (figures not shown). The Kelvin wave SLA, represented by integrating the wind stress anomalies along the characteristic line starting from the western Indian Ocean to the Maluku Channel, has shown a relationship with the Niño 3.4 index at a one-year time lag similar to that of AVISA SLA during 1993-2020 ( Figure 13A).
The high predictability based on the SETIO SLA also holds for the La Niña events back in history. For the longer period since 1950, the asymmetric relationship between the SETIO SLA and Niño 3.4 index at a one-year lag basically holds ( Figure 13B). In particular, the linear relation between the negative SETIO SLA and Niño 3.4 SSTA one year later still exists for the longer data, suggesting that strong upwelling anomalies in the SETIO precede strong La Niña in the following year. The negative SLA (upwelling anomalies) clearly favored the occurrence of La Niña one year later with a correlation coefficient of 0.57 exceeding the 99% confidence level. The much larger number of samples in the longer time series (72 years in ERA5 versus 29 years in AVISO data) yielded a considerably smaller correlation coefficient still above the statistical significance. In addition, when downwelling anomalies dominate the SETIO, the Niño 3.4 index one year later may be either positive or negative, which is similar to the uncertainty during 1993-2021. The above analysis suggests that the simple linear regression model of La Niña prediction at a lead time of one year, based on Indo-Pacific "oceanic channel" dynamics, is likely useful notwithstanding the decadal variability of the Indo-Pacific Ocean and climate.
Historical SLA from oceanic data assimilation products extending back to 1950 is also examined for the IOD-ENSO predictability. The correlation coefficient between the negative SETIO SLA and Niño 3.4 SSTA one year later is 0.51 for the ECMWF Ocean Reanalysis System 4 (ORA4), but is only 0.37 for the German contribution to the Estimating the Circulation and Climate of the Ocean project 3 (GECCO3) (figures not shown). We speculate that the differences are caused by the coarse resolution of the ocean model in resolving the narrow channels of the Indonesian Frontiers in Marine Science frontiersin.org seas and topography, which could induce artificial eastern boundary reflection in the eastern Indian Ocean to contaminate the SLA signals. The linear relationship between the negative SETIO SSTA and Niño 3.4 SSTA since 1950 (correlation coefficient is 0.34 in Figure 14) is not as good as between the negative SETIO SLA and the Niño 3.4 index (correlation coefficient is 0.57, Figure 13B). Although the correlation coefficient between the negative SETIO SSTA and Niño 3.4 SSTA exceeded the 99% significance level, it did not capture the one-to-one relationship between the negative SETIO SLA and Niño 3.4 SSTA one year later. When a strong negative SSTA occur over the SETIO during boreal fall, the Niño 3.4 SSTA can be either positive or negative in the following year, suggesting that La Niña prediction based on the SSTA in the Indian Ocean contains large uncertainties.

Conclusion and discussion
ENSO predictors used in most exiting statistical forecasts are variables in the tropical Pacific Ocean, especially WWV indices, which have proven to provide a precursor for the onset of La Niña. Other factors, including inter-basin interactions over the tropical Indo-Pacific Ocean, have been shown to play important roles in ENSO predictability. The previous studies have demonstrated that SSTA in the Indian Ocean may initiate the onset and contribute to the predictability of the 2010/2012 and 2020/2021 La Niña events (Luo et al., 2017;Hasan et al., 2022), which follow weak-tomoderate El Niño without large discharged WWV anomalies. These studies have only focused on atmospheric teleconnection via Walker circulation. The present study emphasizes that the upwelling anomalies in the STEIO provide a better predictor beyond the one-year lead for La Niña events preceded by the pIOD. The dynamics are suggested to occur through the "oceanic channel" of the ITF.
Our analysis suggests a clear asymmetry in the lead correlation between the SETIO SLA in boreal fall and Niño 3.4 SSTA one year later: negative SETIO SLA lead to La Niña, while positive SLA lead to either La Niña or El Niño. The negative SLA (upwelling anomalies) in the SETIO in boreal fall is a better long-lead precursor for one type of La Niña events (which are preceded by pIODs) at the lead time of one year, with a correlation coefficient of  Forecasts of linear regression model at the evolution of the strong La Niña events. Niño3.4 index during 2020Niño3.4 index during /2021Niño3.4 index during , 2007Niño3.4 index during /2008Niño3.4 index during , 1998Niño3.4 index during /1999Niño3.4 index during , 1995Niño3.4 index during / 1996 La Niña events from the observation (solid lines) and prediction (dot lines) using the SLA in the SETIO in fall of 2019,2006,1997,1994. Grey columns indicate the root mean square error (RMSE). 0.97 (exceed the 99% confidence level) since 1993. This relationship is applicable to longer time series in history since 1950, with a correlation coefficient of 0.57 (exceed the 99% confidence level).
It is worth mentioning that some studies, e.g. Jiang et al. (2021), have argued that the IOD-ENSO link is a result of ENSO cycling and, therefore, "adds no additional information to ENSO prediction". However, the auto-correlation of Niño3.4 at one year time lag is only -0.12 during 1950-2020, way below the statistical significance level. Moreover, the auto-correlation at one year lag is not statistical significant in any decade after 1970 (figure not shown). In contrast, the IOD-ENSO lag correlation is strong during the 1990s . Clearly, the ENSO autocorrelation is not the major cause of the IOD-ENSO lead-lag relationship. Furthermore, the correlation of the negative Niño 3.4 index with the discharged WWV in the previous fall is insignificant. These results suggest that, at longer lead times across the SPB, the WWV or Niño3.4 index could not be good predictors of La Niña. The negative SLA in SETIO, which are partly correlated with the preceding El Niño, is mainly produced by the strong oceanatmosphere coupling over the tropical Indian Ocean with intrinsic nonlinearity. The model of Jiang et al. (2021) does not include this process nor the IOD feedback on ENSO. The "oceanic channel" dynamics during the historical La Niña events. SLA (cm) over the equatorial Indian and Pacific Ocean (1°S-1°N) and alongshore the Sumatra-Java island chain during the 1994(A), 1997(B), 2007(C) pIOD. The red dashed arrows indicate the eastward-propagating Kelvin wave. Here, we emphasize that the dynamics behind this long-lead predictability are the Indo-Pacific "oceanic channel" connection ( Figure 15). When a pIOD occurs during fall, anomalous upwelling Kelvin waves propagate from the eastern Indian Ocean to the western Pacific through the Indonesian Seas and further to the eastern equatorial Pacific Ocean, elevating the thermocline and forcing a La Niña event to develop during the following year. However, the atmospheric bridge was not the dominant process contributing to the high long-lead predictability of strong La Niña events in 1995/96, 1998/99, 2007/08, and 2020/21. Although easterly wind anomalies in the western Pacific followed the 1994, 1997, 2006 and 2019 pIOD events, the lead correlation between the The propagation of the Kelvin waves through the Indonesian seas during the historical pIOD events. SLA (color shading, cm) and surface current anomalies (vectors, ms -1 ) from ECCO2 during October and December of the (A) 1994, (B) 1997, (C) 2006 pIOD events.
negative SETIO SLA in fall and the easterly wind anomalies in the western Pacific during the following summer was not significant. Recently, pacemaker experiments  depicted that the Indo-Pacific "oceanic channel" are dominantly important in ENSO forecasts, whereas the atmospheric bridge is secondary to the generation of 2020/21 La Niña. The present analysis emphasizes that the "oceanic channel" is likely to influence the predictive skill of La Niña.
Our results suggest that the lead relationship between SETIO SLA and Niño 3.4 SSTA one year later is asymmetric between El Niño and La Niña. A strong pIOD event triggered the onset of a La Niña event in the following year. A warm event in the eastern equatorial Pacific in the year following a pIOD event does not exist. In contrast, after a nIOD event the tropical Pacific may evolve into either an El Niño or La Niña event or a neutral state. The predictability of this type of La Niña events, including the 2-year La Niña in the second year (1999/2000, 2017/18, 2021/22), did not result from the Indian Ocean in the previous year. The Indo-Pacific interaction, especially the "oceanic channel" dynamics, may partly explain the asymmetry of the warm and cold phases of ENSO predictability.
The asymmetry in predictive skills of negative and positive SETIO SLA may be due to the stronger coupling between the cold tongue SSTA and the eastern equatorial Pacific thermocline depth anomalies during La Niña than during El Niño (Yuan and Rienecker, 2003;Yuan, 2009;Tang et al., 2015). It is known that the surface heat balance is dependent on the depth of the surface mixed layer in the eastern equatorial Pacific. Upwelling currents elevate thermocline and mixed layer in the eastern equatorial Pacific, making the cold tongue SSTA more sensitive to vertical advection. In contrast, downwelling currents depress the thermocline and mixed layer in the eastern equatorial Pacific, resulting in the cold tongue SSTA de-coupled from the thermocline anomalies. Thus, stronger cooling anomalies occur following the pIOD than the warming anomalies produced by the downwelling of the same amplitudes following the nIOD. In addition, some of the La Niña events, such as the 2-year La Niña in the second year, are independent of the Indian Ocean interannual variations. For this type of La Niña events, the cold tongue SSTA are largely controlled by the internal dynamics in the Pacific basin, which explains the events in the fourth quadrant in Figure 1A. The physical mechanisms leading to the asymmetry in the lead correlation between the SETIO SLA and Niño 3.4 SSTA one year later deserve further investigations.
From the previous and current works, it can be observed that ENSO predictability results from a combination of the internal oceanic processes in the Pacific and interbasin interactions through the atmospheric bridge and oceanic channel. For some ENSO events, oceanic channels between the tropical Indian and Pacific Ocean were not the dominant process for the onset of El Niño/La Niña one year later. In any case, the present study emphasizes that an early upwelling anomalies (negative SLA) in the SETIO is an effective and long-lead predictor of La Niña onset timing and amplitude, especially for the 1995/96, 2007/08, and 2020/21 La Niña events following strong pIOD events with weak concurrent El Niño events.
Because of the limited number of La Niña events in the observational record, our statistical results still have some uncertainties. Model result of Wang et al. (2022) has demonstrated the dominate role of "oceanic channel" in the generation of 2020/21 La Niña event. Numerical model experiments are needed in the next step to test the relative roles of the "oceanic channel" and the atmospheric bridge between the tropical Indian and Pacific Ocean for the 1995/96, 1998/99, 2007/08 and 2016/17 La Niña events.

Data availability statement
The 2020 La Niña forecast were obtained from the IRI/CPC [https://iri.columbia.edu/our-expertise/climate/enso/]. The OISST, ERSST, and NCEP data were provided by NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, through their website [https:// www.esrl.noaa.gov/psd/]. Satellite altimeter sea level data were obtained fr om the Aviso project web page [http:// www.aviso.altimetry.fr]. ECCO2 data were downloaded from the  Schematic illustration for the negative SETIO SLA as a predictor of the La Niña events based on the tropical Indo-Pacific "ocean channel" dynamics. The upper (lower) part of this figure shows the SLA (subsurface temperature anomalies). The left (right) part of this figure shows the condition during the peak of pIOD (following spring). The red dashed arrows indicate the eastward-propagating Kelvin wave.

Author contributions
DY developed the scientific hypotheses, directed the research, and attracted fundings to support the study. XZ conducted the analysis. XZ and DY wrote the manuscript. JW worked on the reanalysis data, helped interpret the "oceanic channel" dynamics. All authors contributed to the article and approved the submitted version.

Funding
This work is jointly supported by the National Key R&D Program of China (2019YFA0606702, 2020YFA0608801), CAS (XDB42010000), NSFC (42175027, 41720104008, 91858204), and Shandong Natural Science Foundation Project (ZR2019ZD12). DY is supported by the National Key R&D Program of China