On the Delayed Coupling Between Ocean and Atmosphere in Recent Weak El Niño Episodes

The recent borderline El Niño events of 2014/2015 and 2018/2019 provided operational centers with unique challenges because of the apparent absence of typical coupling between the tropical atmosphere and ocean before onset. The mismatch between atmosphere and ocean raises questions about its causes and predictability. Here we analyze observational data since 1979 to show that a sea surface temperature pattern characterized by an anomalous gradient in the western and central equatorial Pacific played a critical role in inhibiting the expected onset of central tropical Pacific deep convection during these events. This sea surface temperature pattern, which produces an atmospheric response that opposes the response to elevated eastern Pacific sea surface temperatures, has become more prevalent over the past 40 years.


Introduction
The borderline El Niño events of 2014/2015 and 2018/2019 provided a unique challenge to forecasters because of the apparent mismatch between tropical atmospheric and oceanic conditions (McPhaden, 2015;Santoso et al., 2019). This mismatch generated uncertainty for operational centers regarding event classification and forecasting their potential impacts. The National Oceanic and Atmospheric Administration (NOAA) declares an El Niño Advisory (i.e., the occurrence of El Niño conditions) when the sea surface temperature (SST) anomaly (SSTA) in the Niño 3.4 region (5°S-5°N, 120°-170°W; the Niño 3.4 index hereafter) exceeds 0.5°C, the anomaly is expected to exceed that threshold for at least five consecutive overlapping three-month seasons, and the tropical atmosphere exhibits clear evidence of coupling with the ocean in a way that is consistent with the dynamics of the El Niño-Southern Oscillation (ENSO). By boreal autumn in 2014 and in 2018, the Niño 3.4 index had reached the threshold for NOAA to declare an El Niño Advisory (see NOAA's official three-month running mean Niño 3.4 index, or Oceanic Niño Index, at www.cpc.ncep.noaa.gov/products/analysis_monitoring/ensostuff/ONI_v5.php), and both statistical and dynamical models had forecast a high probability of the Niño 3.4 index persisting above the 0.5°C threshold for at least several seasons (see archived forecasts at https://iri.columbia.edu/ our-expertise/climate/forecasts/enso/current/). Despite these positive indicators from the tropical ocean, and the actual persistence of the elevated SSTs in the eastern tropical Pacific, NOAA did not declare the occurrence of El Niño conditions until they became evident in February 2015 and January 2019, respectively, because tropical atmospheric conditions were not clearly consistent with El Niño before then.
One of the primary indicators of coupling considered by forecasters and the focus of this study is the pattern of anomalous tropical convection that accompanies SST anomalies in the Niño 3.4 region. During an El Niño episode, the Walker circulation weakens as the eastern edge of the western equatorial Pacific warm pool and the region of strongest deep convection expand eastward to the International Date Line, and sometimes farther east during stronger events (Johnson & Kosaka, 2016;Takahashi et al., 2011). These changes are critical for the ocean-atmosphere coupling that promotes further El Niño development (Gill & Rasmusson, 1983) and for the excitation of global atmospheric teleconnection patterns (Chiodi & Harrison, 2013;Trenberth et al., 1998). The expected eastward shift of deep convection was conspicuously absent in the boreal fall of 2014 and 2018 (as discussed more thoroughly in section 3.1).
The absence of this expected atmospheric response raises several questions. Was the atypical atmospheric pattern the result of the noise of climate variability primarily internal to the atmosphere, such as that of the Madden-Julian Oscillation? Given the weak amplitude of these recent events, it is plausible that the noise of internal atmospheric variability, with little predictability beyond a few weeks, overwhelmed the weak SST-forced atmospheric signal. Alternatively, did SST patterns independent of those we typically use to monitor ENSO interfere with the canonical tropical atmospheric El Niño pattern? If so, then the hope is elevated for potential skillful prediction of such a pattern on seasonal time scales. This possibility also would open other lines of inquiry, including whether dynamical forecast models can predict this pattern and the response to it. In addition, this viewpoint may indicate the importance of an SST pattern that is separate from the "central Pacific" (CP) and "eastern Pacific" (EP; e.g., Yu & Kao, 2007) ENSO types that dominate the ENSO "flavor" discussion, since both EP and CP El Niño types are expected to produce anomalously strong precipitation in the central equatorial Pacific (e.g., Ashok et al., 2007;Santoso et al., 2019).
Previous studies have focused on coupled ocean-atmosphere processes that regulated the SST evolution during El Niño events, including the sequence of events that limited the 2014 episode to a borderline event (Chiodi & Harrison, 2017;Hu & Fedorov, 2016;Levine & McPhaden, 2016;Menkes et al., 2014). In contrast with these studies, we instead focus on the reasons why the tropical atmosphere failed to manifest a typical atmospheric response despite tropical Pacific SSTAs reaching the threshold for El Niño. Specifically, we focus on the possible role of SST patterns in modulating the tropical atmospheric circulation over the central equatorial Pacific for a given Niño 3.4 index value. Through an analysis of observational data since 1979, we find that a zonal SST gradient pattern in the western and central tropical Pacific plays a significant role in modulating the atmospheric response to ENSO. We focus on the boreal autumn because this is the season when Niño 3.4 index first crossed the nominal threshold for El Niño during these weak events, providing unique challenges for monitoring and forecasting potential impacts, but we show that the basic findings apply to all other seasons as well. We also show that this modulation is related to statistically significant trends in Walker circulation strength and central tropical Pacific convection.

Data
We analyze monthly tropical observational data covering the period of 1979-2018. The SST data are from the Extended Reconstructed SST version 5 (Huang et al., 2017). We focus on Extended Reconstructed SST version 5 because of its prominent role in NOAA's ENSO monitoring, but we also carried out all analyses with Hadley Centre Sea Ice and SST version 1 data set (Rayner et al., 2003). All conclusions are insensitive to SST data set, and we do not discuss the Hadley Centre Sea Ice and SST version 1 results further except in the context of trends in section 3.2. As a proxy for tropical deep convection, we use outgoing longwave radiation (OLR) data available from the University of Maryland OLR Climate Data Record website (Lee et al., 2007). We calculate anomalies by removing the calendar month means. For OLR, we use a base period of 1981-2010. For SST, we follow the procedure used by the NOAA Climate Prediction Center for real-time ENSO monitoring and use centered 30-year base periods updated every five years (e.g., 1981-1985 anomalies use a 1966-1995 base period). The Climate Prediction Center adopted this procedure to counter the influence of long-term trends that do not reflect interannual variability in ENSO monitoring, but as discussed in the section 3.2, multidecadal trends still have a significant influence on the relationship between the Niño 3.4 index and the tropical atmosphere.
We also analyze relationships between ENSO and the Walker circulation strength, as measured by the Equatorial Southern Oscillation Index (EQSOI) derived from reanalysis data. The EQSOI is calculated as the normalized difference between the standardized sea level pressure anomalies averaged between the eastern equatorial Pacific (5°S-5°N, 80°W-130°W) and the equatorial region centered on Indonesia (5°S-5°N, 90°E-140°E). Anomalies are calculated with respect to the 1981-2010 base period. To sample the uncertainty associated with the choice of reanalysis data set, we consider the EQSOI derived from the European Centre for Medium-Range Forecasts ERA-Interim Reanalysis (Dee et al., 2011), National Centers for Environmental Prediction-Department of Energy Reanalysis 2 (Kanamitsu et al., 2002), National Centers for Environmental Prediction/National Center for Atmospheric Research Reanalysis (CDAS; Kalnay et al., 1996), and the Climate Forecast System Reanalysis (Saha et al., 2010). For analyzing the relationship between tropical SST indexes and the spatial pattern of anomalous low-level wind, we use gridded National Centers for Environmental Prediction/National Center for Atmospheric Research reanalysis 850-hPa wind data.

Methods
Our analysis focuses on the use of least squares linear regression to diagnose the boreal autumn (September-November (SON)) relationship between the central tropical Pacific OLR and ENSO, as monitored by the Niño 3.4 index, and between the central tropical Pacific OLR and SSTs independent of the Niño 3.4 index. We define a central Pacific OLR index (OLR DL ) as the area-averaged OLR anomaly for the equatorial region centered on the Date Line (5°S-5°N, 160°E-160°W), a key action center for anomalous deep convection associated with ENSO (e.g., Deser & Wallace, 1990;Johnson & Kosaka, 2016). We also considered the central Pacific OLR (CP-OLR) index of L'Heureux et al. (2015), which is based on a box average farther east (5°S-5°N, 170°E-140°W), but we choose a box centered slightly west of this region because the relationship between the Niño 3.4 index and OLR is more linear near the Date Line. However, the conclusions are insensitive to the focus on OLR DL or CP-OLR indexes.
In this study, we seek to determine if a second predictor independent of the Niño 3.4 index, either a second SST index or the long-term trend, has a significant influence on OLR DL . For this assessment, we compare linear regression models and determine the significance of the second predictor with an F test, for which an effective sample size accounts for autocorrelation following Santer et al. (2000). The null hypothesis is rejected; that is, the addition of a second predictor is significant, if F is greater than the critical value of the F distribution for a false rejection probability of 5% (see Text S1 for more details).

Relationship Between Niño 3.4 Index and OLR DL
We first focus on the relationship between the Niño 3.4 index and OLR DL in SON, with an emphasis on the recent weak El Niño episodes of 2014 and 2018. In Figure 1 Are these OLR DL deviations from the linear Niño 3.4 relationship attributable solely to internal atmospheric variability, or do slowly varying and potentially more predictable SST variations also play a role? To begin to address this question, we follow a procedure similar to partial least squares regression analysis (Black et al., 2017;Smoliak et al., 2010;Wold, 1966)

10.1029/2019GL084021
Geophysical Research Letters Shu & Clarke, 2002). Figure 2 indicates, however, that this Niño 3.4 index influence can be counteracted by an anomalously strong SST gradient between~150°E and 150°W. The physical justification is that elevated SSTs near Papua New Guinea and suppressed SSTs near and east of the Date Line potentially can prevent the eastward extension of the eastern edge of the warm pool that typically accompanies elevated Niño 3.4 index values. Therefore, the warmest Indo-Pacific warm pool waters, which anchor the heaviest tropical Pacific rainfall, may remain near Indonesia and away from the Date Line even when the Niño 3.4 index crosses the threshold necessary for an El Niño event to be declared.
To quantify the influence of the anomalous western/central equatorial Pacific SST gradient, independent of the Niño 3.4 index, on boreal autumn central Pacific convection, we first define a zonal SST gradient index (gradient index hereafter) as the difference between the standardized SSTA averaged over a box near Papua New Guinea (10°S-10°N, 130°E-170°E) and the standardized SSTA averaged over a box in the central Pacific (10°S-10°N, 180°-140°W). These boxes are chosen based on the strongest correlations in Figure 2 and on the physical justification described above.
To evaluate the role of the potentially competing influences of the gradient and Niño 3.4 indexes, we next quantify the impact of both the Niño 3.4 index and the linearly independent contribution from the gradient index on the SON tropical OLR and 850-hPa wind fields. For the latter calculations, we first linearly remove the Niño 3.4 index from both the gradient index (which we call the residual index) and the gridded OLR and wind anomaly fields. To facilitate a direct comparison between the Niño 3.4 and gradient index contributions, we standardize the Niño 3.4 and residual gradient indexes prior to regressions with the OLR and wind anomaly fields (Figures 3a and 3b). Both the unfiltered and residual OLR DL and gradient index time series are illustrated in Figures S2 and S3. The OLR anomaly contribution from the Niño 3.4 index (Figure 3a) provides the familiar east-west dipole, with suppressed convection (positive OLR anomalies) over the maritime continent, and enhanced convection (negative OLR anomalies) centered near the Date Line. The 850-hPa wind regression also provides a familiar pattern (e.g., Deser & Wallace, 1990), with anomalous equatorial westerlies, signifying a weakened Walker circulation.
The gradient index contribution (Figure 3b) is nearly opposite to that of the Niño 3.4 index influence, with suppressed convection over the Date Line, as expected, and anomalous equatorial easterlies in the central Pacific. Note that the contour interval is halved in Figure 3b, which indicates that the Niño 3.4 index influence generally is stronger. This largely reflects the much stronger variance of the Niño 3.4 index (prior to standardization) relative to the residual gradient index. However, in years like 2018 when the Niño 3.4 SSTA is relatively modest (standardized value of 0.67) in relation to the gradient index (standardized residual index value of 1.86), then the two predictors can have an effect of similar strength.
This point is illustrated clearly in the time series of observed and linearly regressed OLR DL anomalies (Figure 3c). The linear regression with both the Niño 3.4 and gradient indexes as predictors captures the OLR DL variability more accurately (R 2 = 0.93; r between the residual OLR DL and residual gradient index is 0.76) than with the Niño 3.4 index alone (R 2 = 0.84); the SST gradient predictor is highly significant (p value <10 −4 ) according to an F test. Table S1 indicates that the SST gradient index is a significant predictor at the 5% level for all other seasons as well. Notably, the inclusion of the gradient index as a predictor Returning to the questions posed in the introduction, we conclude that an SST pattern independent of the Niño 3.4 index played a critical role in the absence of the typical atmospheric response for recent borderline El Niño events, including SON 2018.

Linear Trends of the Tropical Atmosphere
Close inspection of Figure 3c indicates that the linear regression of OLR DL on the Niño 3.4 index alone (orange line) tends to overestimate OLR DL early in the period and underestimate it late in the period. This visual impression is substantiated by a statistically significant downward trend in the difference between the observed and Niño 3.4-only regression lines ( Figure S4). This suggests that the influence of the gradient index captures, at least in part, the influence of a long-term trend. We elaborate on this observation in Figure 4, which shows linear trends of SST, Niño 3.4-adjusted OLR DL , and Niño 3.4-adjusted EQSOI.
The 1979-2018 SON SST trend (Figure 4a) reveals enhanced warming in the western equatorial Pacific, particularly near Papua New Guinea, and reduced warming or cooling in the eastern equatorial Pacific, which is consistent with L' Heureux, Collins, et al. (2013). This SST trend pattern resembles the correlation pattern between Niño 3.4-adjusted SST and OLR DL (Figure 2; uncentered pattern correlation (DelSole & Shukla, 2006) between Figures 2 and 4a is 0.76), although with differences in the location of the negative correlations/cooling east of the Date Line. Despite these differences, both patterns feature an enhanced zonal SST gradient between the western and central equatorial Pacific, which is expected to promote a tendency for the warm pool convection to remain anchored in the western equatorial Pacific. Therefore, the enhanced warming near Papua New Guinea and suppressed warming to the east over the past 40 years, especially in the second half of the period ( Figure S3), have inhibited the eastward migration of the western Pacific warm pool that typically occurs when the Niño 3.4 index rises, supporting a speculation noted in McPhaden (2015). The impact of this SST trend is reflected in the Niño 3.4-adjusted OLR DL trends (Figure 4b), which are positive in all seasons and statistically significant in SON and OND. Therefore, despite efforts to mitigate against the effects of longer-term trends in ENSO monitoring by modifying the Niño 3.4 index base period every five years, the effects of the OLR DL trends cannot be neglected. Consistently, the Niño 3.4-adjusted CDAS EQSOI has undergone positive trends in all seasons (Figure 4c), indicating a strengthening of the Niño 3.4-independent Walker circulation. All trends are similar if Hadley Centre Sea Ice and SST version 1 data are used instead for all calculations ( Figure S5), and the EQSOI trends are consistent in the Department of Energy Reanalysis 2 and ERA-Interim reanalysis but not the Climate Forecast System Reanalysis (see Text S2 and Figure S6). The negative Climate Forecast System Reanalysis EQSOI trends are likely spurious owing to an overly strong trade wind bias that was greatly reduced after 1998 (Xue et al., 2011), and so we conclude that the EQSOI trends derived from the other reanalyses are more reliable.

Conclusions
In this study, we address the absence of the expected boreal autumn tropical atmospheric conditions during recent borderline El Niño episodes, as in 2014 and 2018. We provide observational evidence that tropical SST patterns independent of the Niño 3.4 index, namely, the anomalous western/central Pacific zonal SST gradient, played a critical role in inhibiting the coupling between the atmosphere and the ocean during these events. Anomalous warming in the vicinity of Papua New Guinea inhibited the eastward migration of the eastern edge of the western Pacific warm pool that typically accompanies an elevated Niño 3.4 index, keeping the heaviest tropical rainfall anchored in the western Pacific rather than near the Date Line.
The positive deviations from the expected central Pacific OLR and the associated EQSOI during these recent events relate to a significant multidecadal trend of increasing central Pacific OLR and a strengthening Walker circulation, which is consistent with several recent studies L'Heureux, Lee, et al., 2013;Sandeep et al., 2014;Seager et al., 2019;Sohn & Park, 2010;Solomon & Newman, 2012;Zhao & Allen, 2019) but inconsistent with the expected response to rising global temperatures from global climate models and through basic thermodynamic arguments Sohn et al., 2013;Vecchi et al., 2006). Given that this study is limited to a 40-year observational record, the reasons for this discrepancy are beyond the scope of this study (see, for example, Liu & Xie, 2018;Seager et al., 2019;Xie & Kosaka, 2017). Most relevant here, we find that a significant component of this trend in boreal autumn is independent of the Niño 3.4 index. This result is consistent with studies that have reported a pronounced Walker circulation strengthening after filtering out the influence of ENSO (L'Heureux, Lee, et al., 2013;Solomon & Newman, 2012). The implication is that the atmospheric response to typical El Niño (La Niña) SSTs has weakened (strengthened) over the past 40 years, based on a typical index for monitoring ENSO. This trend opposes the projected strengthened atmospheric response to El Niño with changing background SSTs under increasing greenhouse gases (Cai et al., 2014).
In this study, we focus on the atmospheric response to tropical SSTAs, holding the Niño 3.4 index fixed, but the zonal SST gradient pattern also may influence tropical SST evolution. For example, westerly wind events in the equatorial Pacific, an important forcing of ENSO (e.g., Vecchi & Harrison, 2000), are more likely to occur when the warm pool is extended eastward (Eisenman et al., 2005;Yu et al., 2003). Therefore, the increased zonal SST gradient during recent borderline El Niño events, which inhibited the eastward extension of the warm pool, potentially may have influenced the statistics of these stochastic wind events that regulated the tropical SST evolution, but this possibility requires additional study.
In addition, recent studies (Choi et al., 2011;Xiang et al., 2013) suggest that the change in mean state toward a stronger zonal SST gradient may be responsible for the more frequent occurrence of CP relative to EP El Niño. However, both EP and CP El Niño are expected to generate enhanced convection near the Date Line (e.g., Ashok et al., 2007;Santoso et al., 2019;Xiang et al., 2013), and so the modulating influence from the zonal SST gradient pattern (Figure 2) identified here is distinct from the influences described in those studies. This suggests that the zonal SST gradient pattern deserves attention separate from any of its potential influences on EP or CP ENSO. Given that ENSO's global impacts are driven by its effects on tropical convection, this SST gradient pattern may have important implications for ENSO teleconnection patterns, particularly in boreal fall (Bladé et al., 2008). These overall findings also confirm the utility of indexes that target the relationship between SST and tropical convection (L'Heureux et al., 2015;Williams & Patricola, 2018) to complement the Niño 3.4 index for monitoring ENSO.
To conclude, the delayed onset of the full coupling between atmosphere and ocean for the weak El Niño events of 2014/2015 and 2018/2019 may be rooted in a tropical SST pattern rather than internal atmospheric variability. The oceanic origins provide hope that these tropical atmospheric deviations from the canonical Niño 3.4 index relationships may be predictable on seasonal and possibly longer time scales, especially since the enhanced western Pacific zonal SST gradient is linked with a multidecadal trend. However, the ability of current dynamical or statistical forecast models to predict the Niño 3.4 index-independent component of this SST gradient pattern, and precisely how this gradient pattern may have impacted ENSO development, remain open questions.