Contributions of tropical-extratropical oceans to the prediction skill of ENSO after 2000

SHI Ling nd DING RuiqingSchool of Science,Lnzhou University of Technology,Lnzhou,Chin; Stte Key Lortory of Erth Surfce Processes nd Resource Ecology,Beijing Norml University,Beijing,Chin

ABSTRACT The skill of most ENSO prediction models has declined significantly since 2000. This decline may be due to a weakening of the correlation between tropical predictors and ENSO.Moreover, the effects of extratropical ocean variability on ENSO have increased during this period. To improve ENSO predictability, the authors investigate the influence of the extratropical Atlantic and Pacific oceans on ENSO during the pre-2000 and post-2000 periods, and find that the influence of the northern tropical Atlantic sea surface temperature (NTA SST) on ENSO has significantly increased since 2000. Furthermore, there is a much earlier and stronger correlation between NTA SST and ENSO over the central-eastern Pacific during June-July-August in the post-2000 period compared with the pre-2000 period. The extratropical Pacific SST predictors for ENSO retain an approximate 10-month lead time after 2000. The authors use SST signals in the extratropical Atlantic and Pacific to predict ENSO using a statistical prediction model. This results in a significant improvement in ENSO prediction skill and an obvious decrease in the spring predictability barrier phenomenon of ENSO. These results indicate that extratropical Atlantic and Pacific SSTs can make substantial contributions to ENSO prediction, and can be used to enhance ENSO predictability after 2000.

KEYWORDS ENSO predictability;Northern tropical Atlantic SST;Victoria mode;South Pacific quadrupole

1. Introduction

As the dominant mode of interannual variability in the atmosphere, the El Ni?o-Southern Oscillation (ENSO) has an enormous impact on the global climate and on societies worldwide(e.g.,Bjerknes 1969;Rasmusson and Carpenter 1982;Jin 1997a,1997b;Fang and Mu 2018;Timmermann et al.2018).Thus,accurate ENSO simulations and forecasts are of great importance. Over the past three decades,substantial work has been devoted to ENSO prediction strategies, dramatically improving short-term climate predictions (Aaron and Jin 2010; Tseng et al. 2016). Previous studies have suggested that the influence of some tropical phenomena,such as the warm water volume(WWV)along the equatorial Pacific(2°S-2°N,120°E-80°W),westerly wind events/easterly wind surges,or their combined effects,on intrinsic ENSO dynamics, are effective indicators and key precursors of ENSO(Latif et al.1998;Barnston,Glantz,and He 1999;Tseng et al.2016;Tang et al.2018;Ren,Zuo,and Deng 2019).However,because of a change in the relationship between the WWV and ENSO sea surface temperature(SST)anomalies since 2000,the prediction skill for models that use WWV as a precursor has declined significantly(Horii, Ueki, and Hanawa 2012; Hu et al. 2017). Barnston et al. (2012) and Wang et al. (2017) found a significant decrease in the forecast skill of ENSO prediction models since the start of the 21st century.

Interestingly, observations and model simulations both indicate that the extratropical Pacific, Atlantic,and Indian oceans play an important role in the occurrence and development of ENSO (e.g., Vimont, Battisti,and Hirst 2003; Vimont, Wallace, and Battisti 2003; Yeh et al.2009;Terray 2011;Wang et al.2011;Ham,Kug,and Park 2013; Ham et al. 2013; Ding et al. 2015).Extratropical Pacific signals (e.g., the Victoria mode(VM); Bond et al. 2003; Ding et al. 2015) affect ENSO variation through a seasonal footprinting mechanism(Vimont,Battisti,and Hirst 2003).The South Pacific quadrupole (SPQ; Ding, Li, and Tseng 2014) mode also has substantial effects on ENSO variability through a tradewind charging mechanism (Anderson 2004; Anderson,Perez, and Karspeck 2013). Ham, Kug, and Park (2013)found that SSTAs in the northern tropical Atlantic(NTA;80°W-20°E,0°-15°N)during boreal spring can trigger the warm-pool type of El Ni?o and phase reversal. These results indicate that they may be good precursors for ENSO prediction.

Here,the influence of extratropical Pacific and Atlantic Ocean signals on ENSO variability is evaluated.We modify the prediction model of Tseng et al.(2016)by adding new precursors to predict ENSO SSTAs.The remainder of this paper is arranged as follows:Section 2 presents the datasets and methods.The statistical predictability of the new prediction model is assessed in section 3. Finally,a summary and discussion are given in section 4.

2. Data and methods

2.1. Observational datasets

The ocean temperature data used in this study are from the monthly Global Ocean Data Assimilation System(GODAS), which provides data from 1980 to 2017(Behringer and Xue 2004). The SST data for the same period are from the UK Met Office Hadley Centre Sea Ice and Sea Surface Temperature(HadISST)dataset on a 1°×1° grid (Rayner 2003). Near-surface wind data (0.995 sigma level) from the National Centers for Environmental Prediction-National Center for Atmospheric Research reanalysis project on a 2.5° ×2.5° horizontal grid are used to analyze ocean-atmosphere interactions(Kalnay et al.1996).

2.2. Methods

To calculate the number of effective degrees of freedom between two different time seriesXandY,the effective sample sizeN* is calculated as follows (Pyper and Peterman 1998):

Table 1. Definitions of the Ni?o3.4 index and precursors from the tropical Indian and Atlantic oceans.

whereNis the size of the sample, and ρxand ρyare the autocorrelations of variablesxandyat time lagj(j=1,...,N),respectively.

Correlation and multi-regression analyses are used to examine the relationship between the Ni?o3.4 index and these indices. The statistical significance of correlations is determined using a two-tailed Student’st-test.

2.3. Definitions of indices

Following Bond et al. (2003) and Ding et al. (2015),the second orthogonal function mode(EOF2)of monthly SSTAs over the North Pacific(20°-60°N,120°E-150°W)for the period 1979-2017 is defined as the VM.The EOF2 of monthly SSTAs over the South Pacific poleward of 20°S(Ding, Li, and Tseng 2014) is defined as the SPQ mode.The VMI(SPQI)is the second principal component(PC2)time series associated with the VM (SPQ) pattern (Bond et al.2003;Ding et al.2015).

Other indices used in this study are calculated using monthly SST data from the HadISST dataset on a 1°×1°grid,as listed in Table 1.

3. Results

The Warm Water Volume (WWV) is typically the most important tropical precursor in current ENSO prediction models.Therefore,we first analyze seasonal variations in the standard deviations of the monthly time series of the WWV and Ni?o3.4 indices during the pre-2000 (1980--1999) and post-2000 (2000-2017) periods (Figure 1). As evident in Figure 1(a), there is no distinct difference in Ni?o3.4 index variance between these two periods in winter (November-January, NDJ) because of so-called phase-locking. In contrast, the magnitude of WWV variance throughout the year changes noticeably after 2000(Figure 1(b); the annual variance changes from 1.42 before 2000 to 0.72 after 2000). In addition, the WWV variance pattern in the post-2000 period is similar to that during the pre-2000 period, but the Ni?o3.4 index variance pattern shows a clear change after 2000, with larger differences in variance during March?May. Next,we investigate the lead-lag correlations of the WWV and Ni?o3.4 indices during these two periods (Figure 1(c)).The peak correlation coefficient between these two indices decreases from 0.75 before 2000 to 0.66 after 2000 (significant at the 95% confidence level), and the time by which WWV leads the Ni?o3.4 index also decreases (from seven to four months). These results indicate that the WWV, the dominant signal from the tropical Pacific,is not sufficient to predict ENSO variation after 2000, in agreement with previous studies (Horii,Ueki, and Hanawa 2012; Clarke 2014; Neske and Mcgregor 2018;Zhang et al.2019).

Figure 1.(a)Seasonal variations in the standard deviation of the Ni?o3.4 index during the pre-2000(red bars)and post-2000(green bars)periods.(b)As in(a),but for the time series of WWV.(c)Lead-lag correlation of the WWV with the Ni?o3.4 index during the pre-2000(red line)and post-2000(blue line)periods.The dashed lines represent 95%confidence levels.

Figure 2.(a)Time series of the FMA(0)NTAI(black line)and DJF(1)Ni?o3.4 index(red line)for the period 1980-2017,after removing interdecadal variability. The correlation between these two indices is ?0.36 (significant at the 95% confidence level) and ?0.57(significant at the 99% confidence level) for the pre-2000 and post-2000 periods, respectively. (b) The 15-year sliding-window correlation between the FMA(0) NTAI and NDJ(+1) Ni?o3.4 index for the period 1980-2017. The dashed line indicates the 95%confidence level.(*)and(***)indicate the 90%and 99%confidence levels,respectively.

We further investigate the relationship between major interannual variations from the extratropical oceans.The February-April(FMA)NTAI(Table 1)is compared with the NDJ Ni?o3.4 index for the two study periods (Figure 2). The magnitude of the correlation coefficient for these two indices is larger after 2000(?0.36 before 2000;?0.59 after 2000).Results using a 15-year sliding-window correlation analysis also indicate that the relationship between the FMA NTAI and the NDJ Ni?o3.4 index becomes much stronger after 2000(Figure 2(b)).

To identify major changes in the relationship between NTA SSTs and ENSO SSTs,we compare correlation maps of three-month averaged SST and surfacewind anomalies with the FMA(0) NTAI when the NTA SST is most active (after removing the signal of the former ND(?1)J(0) Ni?o3.4 index) for a range of leadlag times during different periods (Figure 3). Compared with the correlation maps before 2000 (Figure 3(a-d)),the most distinct changes are the range of the negative correlations of the FMA NTAI with the SST over the central-eastern Pacific (black box; 10°S-15°N, 155°E-135°W), which widens in June-August (JJA),September-November (SON), and NDJ (Figure 3(f-h)).In addition, the correlation of the FMA NTAI with SST after 2000 (Figure 3(f-h)) is stronger during JJA, SON,and NDJ, compared with the pre-2000 period.Interestingly, the response of the FMA NTAI to central Pacific SST becomes earlier (JJA; Figure 3(f)) after 2000,prior to which positive correlations exist during SON(Figure 3(c)). Northeastern and southwestern wind anomalies also become more significant in JJA after 2000, increasing SST and strengthening the Hadley circulation. (Bjerknes 1969). Thus, the ENSO predictability can be expected to improve when considering the NTA signal in the model,especially for the period after 2000.

Figure 3.(a-d)Correlation maps of the FMA(0)NTAI with SST anomalies and near-surface winds during MAM,JJA,SON,and NDJ for the pre-2000 period. (e-h) As in (a-d), but for the post-2000 period. The black box indicates the area of significantly negative correlation of NTA SST with ENSO SST(10°S-15°N,155°E-135°W).Correlations that exceed the 95%confidence level are shaded.

The correlation coefficients of the oceanic signals from the North and South Pacific (i.e., the VM and SPQ,respectively)with the Ni?o3.4 index are both significant at the 95% confidence level during both study periods(Figure 4).Although the peak correlation of the VM with the Ni?o3.4 index decreases after 2000,the lead time of the peak correlation remains at ～10 months(Figure 4(a)).We also note that the peak correlation of the SPQI with the Ni?o3.4 index is smaller after 2000,but the lead time(8.5 months) shows no significant change after 2000(Figure 4(b)).

Figure 4.(a)Lead-lag correlation of the VMI with Ni?o3.4 index for 1980-1999(red line)and 2000-2017(blue line).(b)As in(a),but for the relationship between SPQI and Ni?o3.4 index.The dashed lines indicate 95%confidence levels.

Applying these results, a new term is readily established based on the extratropical signals using a simple multi-regression method,and serves as an‘extratropical precursor’of ENSO for real-time forecasts:

where EPIextratropicalis the new term and comprises contributions from three extratropical precursors(NTAI,VMI,and SPQI),andtis the lead time used for the precursors(t= 1, 2, ..., 12 months). The correlations among these three indices are relatively weak, which indicates that they are largely independent of each another. The weighting parameter pairs α,β, and γ are calculated by a least-squares fit to NTAI, VMI, and SPQI, respectively.These parameters may vary with lead time.

To test whether the prediction skill of the new ENSO prediction model is improved by considering signals from the extratropical oceans,we incorporate this‘extratropical term’into the statistical ENSO prediction model of Tseng et al. (2016), which only considers tropical signals:

where EPItropicalis the ENSO prediction index.The two terms on the right-hand side of Equation (2) are applied as in Tseng et al.(2016),and are defined as‘tropical precursors’and comprise the precursor (EPIwwv) in the central Pacific and an ocean-atmosphere coupling feedback term(EPIO-A). A more detailed description of these terms and how they are calculated is provided by Tseng et al.(2016).

To assess the forecasting skill after including the extratropical term, we compare the prediction skill of the model before and after adding the new term(EPIextratropical), using the correlation coefficient (R),the normalized root-mean-square error (RMSE), and the sign consistency (SC) for two different lead times (Figure 5). Figure 5(a,b) show results of hindcast and predicted EPI (blue line) with lead times of 6 months (Figure 5(a)) and 10 months (Figure 5(b)).The correlation coefficient of the hindcast with the Ni?o3.4 index during the pre-2000 period with a lead time of 6 months (10 months) is 0.73 (0.75). The correlation coefficient of the predicted results with the Ni?o3.4 index during the post-2000 period is 0.68 (0.72) with a 6-month (10-month) lead time,which is much higher than that for most prediction models (Barnston et al. 2012; McPhaden 2012).

The correlation coefficient (0.68) and the SC (70%)are also higher than those before adding the new term (R= 0.59; SC = 59%). Furthermore, the RMSE(0.61°C) is significantly lower than that of the original model (RMSE = 0.81°C), in which extratropical variability is not considered (Figure 5(c)). Figure 5(d)shows the prediction skill for a lead time of 10 months. TheR, RMSE, and SC indicate that the prediction skill has improved compared with the original model. In addition, because the peak correlations of the extratropical precursors with the Ni?o3.4 index occur at a lead time of ～10 months, we find that the skill for the pre- and post-2000 periods is much better when the prediction is made with a 10-month lead time compared with a 6-month lead time(Figure 5(b,d)). Therefore, we conclude that the hindcast skill and the prediction skill are significantly improved after the extratropical precursors are taken into account.

Figure 5.(a)Time series of the EPI(blue line)for a lead time of 6 months superimposed on the Ni?o3.4 index(red line).(b)As in(a),but for a lead time of 10 months.(c)The correlation coefficient(R),root-mean-squared error(RMSE),and sign consistency(SC;%)for the new(red)and original(green)ENSO prediction models with lead times of(c)6 and(d)10 months.

4. Summary and discussion

In this study,we investigate for the first time the seasonal variations in the standard deviations of monthly WWV and Ni?o3.4 indices during pre-2000 and post-2000 periods.Results suggest that the maximum magnitude of the Ni?o3.4 index variance during the post-2000 period in winter (NDJ) is similar to that of the pre-2000 period(Figure 1(a)). However, the WWV variance after 2000 is noticeably lower than that before 2000.The lead time at which the peak correlation occurshas also decreased since 2000 (from three seasons to one), suggesting that as a principal oceanic predictor of ENSO, the WWV is less reliable after 2000.

Previous studies have found that the NTA SST plays an important role in the occurrence and development of ENSO (Ham, Kug, and Park 2013; Ham et al. 2013). Our results further indicate that the enhanced influence from NTA SST may force an earlier and stronger response over the central-eastern Pacific in JJA during the post-2000 period than during the pre-2000 period.In addition,although the correlation between the VMI(SPQI)and the Ni?o3.4 index decreases after 2000,the lead times of the peak correlation remain approximately 10 months, indicating that these two extratropical indices are important precursors of ENSO during both periods.

Based on the noticeably stronger relationship between the Ni?o3.4 index and three extratropical oceanic signals (NTAI, VMI, and SPQI) after 2000, we include a new term (EPIextratropical) in the ENSO prediction model of Tseng et al. (2016) with the aim of improving the ENSO prediction skill. Results show that the hindcast and prediction skills for the Ni?o3.4 index are better than when only considering tropical signals, in terms of monthly correlation,RMSE, and SC. Notably, results indicate that the prediction skill (R, RMSE, and SC) with a 10-month lead time is better than that with a 6-month lead time.Results also indicate that our new ENSO prediction model can effectively predict ENSO events during the last decade with a lead time of 10 months.

Disclosure statement

No potential conflict of interest is reported by the authors.

Funding

This research was supported by the National Natural Science Foundation of China [grant number 41975070] and the Identification and mechanism study of global warming‘hiatus’phenomenon of 973 project of China [grant number 2016YFA0601801].