El Niño , La Niña , and the global sea level budget

Previous studies show that nonseasonal variations in global-mean sea level (GMSL) are significantly correlated with El Niño–Southern Oscillation (ENSO). However, it has remained unclear to what extent these ENSO-related GMSL fluctuations correspond to steric (i.e., density) or barystatic (mass) effects. Here we diagnose the GMSL budget for ENSO events observationally using data from profiling floats, satellite gravimetry, and radar altimetry during 2005–2015. Steric and barystatic effects make comparable contributions to the GMSL budget during ENSO, in contrast to previous interpretations based largely on hydrological models, which emphasize the barystatic component. The steric contributions reflect changes in global ocean heat content, centered on the Pacific. Distributions of ocean heat storage in the Pacific arise from a mix of diabatic and adiabatic effects. Results have implications for understanding the surface warming slowdown and demonstrate the usefulness of the Global Ocean Observing System for constraining Earth’s hydrological cycle and radiation imbalance.


Introduction
Sea level is an informative index of climate and serious concern for coastal communities.Hence, understanding the modern altimetry record is important from scientific and societal vantage points.The most apparent signals in the altimetric global-mean sea level (GMSL) data are the annual cycle and linear trend (e.g., Fig. 4 in Masters et al., 2012).In principle, these changes in the global ocean's water volume relate to the ocean's mass and its density, referred to as "barystatic" and "steric" sea level changes, respectively (e.g., Gregory et al., 2013;Leuliette, 2015).Past studies have successfully used in situ hydrography and satellite gravity data to assess ocean mass and density changes and to evalu-ate barystatic and steric effects on the annual cycle and the linear trend in GMSL (e.g., Lombard et al., 2007;Willis et al., 2008;Cazenave et al., 2009;Leuliette and Miller, 2009;Leuliette and Willis, 2011;Leuliette, 2014Leuliette, , 2015)).
Although the annual cycle and linear trend are the most prominent signals in the record, altimeter data also evidence more subtle GMSL variations superimposed on those signals.In particular, it has long been reported that nonseasonal GMSL anomalies are significantly correlated with El Niño-Southern Oscillation (ENSO), such that the GMSL is anomalously positive during warm El Niño phases and anomalously negative during cool La Niña phases (Nerem et al., 1999(Nerem et al., , 2010;;Chambers et al., 2002;Ngo-Duc et al., 2005;Landerer et al., 2008;Merrifield et al., 2009;Llovel et al., 2010Llovel et al., , 2011;;Boening et al., 2012;Cazenave et al., 2012Cazenave et al., , 2014;;Meyssignac and Cazenave, 2012;Stammer et al., 2013;Fasullo et al., 2013;Haddad et al., 2013;Meyssignac et al., 2013;Calafat et al., 2014;Dieng et al., 2014Dieng et al., , 2015;;Pugh and Woodworth, 2014).Recent papers argue that ENSO-related GMSL changes are essentially of barystatic origin, related to changes in the hydrological cycle, and patterns of precipitation and evaporation (Llovel et al., 2011;Boening et al., 2012;Cazenave et al., 2012Cazenave et al., , 2014;;Fasullo et al., 2013).However, these papers are based on either observations during an isolated event or correlation analysis of model output, and the extent to which barystatic or steric effects are responsible for ENSO-related GMSL fluctuations more generally has not been firmly established based on observations.In fact, conflicting accounts of the GMSL budget during ENSO events are given in the literature.For example, based on altimetry, sea-surface temperature data, and ocean model output, Nerem et al. (1999) reason that the anomalous GMSL rise during the 1997-1998 El Niño was due to thermal expansion of the upper ocean.In contrast, using altimetry and global hydrological models, Ngo-Duc et al. (2005), Llovel et al. (2011), andCazenave et al. (2012) argue that this anomalous rise in GMSL was owing to an increase in global ocean mass.On the one hand, based on satellite data and in situ observations, Boening et al. (2012) and Fasullo et al. (2013) conclude that the anomalous fall in GMSL during the 2010-2011 La Niña was related to a decrease in global ocean mass.On the other hand, and based on very similar datasets, Dieng et al. (2014) conclude differently, finding that this anomalous GMSL fall was owing in approximately equal parts to barystatic and steric contributions.
The literature thus paints a confusing picture.Clarifying the nature of ENSO-related GMSL variations is important for understanding the ocean's role in Earth's hydrological cycle and energy imbalance (e.g., Fasullo et al., 2013;Leuliette, 2015).Here we exploit the growing record length of the Global Ocean Observing System, analyzing satellite gravity, radar altimetry, and in situ hydrographic observations using linear estimation (regression) to elucidate observationally the nature of the altimetric GMSL budget for ENSO events.

Satellite altimetry
We study GMSL records from four groups: AVISO (Ablain et al., 2009), Colorado (Nerem et al., 2010), NOAA (Leuliette and Scharroo, 2010), and CSIRO (Church and White, 2011).Time series derive from the reference altimetry missions (TOPEX/Poseidon, Jason-1, -2).The standard corrections (postglacial rebound, wet troposphere, inverted barometer) are made and a 60-day filter is used to remove a spurious 59-day signal (Masters et al., 2012).Time series are interpolated onto regular monthly intervals over 1993-2015 and we use the ensemble average across the interpolated records.A standard error (Table 1) is estimated based on variances in differences between time series (cf.Ponte and Dorandeu, 2003).

Profiling floats
Monthly Argo in situ temperature and salinity grids produced by Scripps Institution of Oceanography (SIO) and International Pacific Research Center (IPRC) are also employed.The grids are generated using objective analysis applied to quality controlled float profiles (Roemmich and Gilson, 2009).Fields span from 65 • S to 65 • N latitudinally, and down to ∼ 2000 m, but do not cover marginal shelf seas.We use the data for the period 2005-2015, since float coverage was not sufficient before then (Leuliette, 2015, and references therein).We use these gridded fields to evaluate steric sea level following Gill and Niiler (1973).And as with altimetry data, we use the average of the SIO and IPRC time series, deriving a standard error using the difference between these products (Ponte and Dorandeu, 2003).

Gravimetric retrievals
Monthly estimates of the barystatic sea level term based on retrievals from the Gravity Recovery and Climate Experiment (GRACE) (e.g., Tapley et al., 2004) are also considered.Values are from Release-05 data processed by the three main science data system centers at the University of Texas at Austin Center for Space Research (CSR; Bettadpur, 2012), the Jet Propulsion Laboratory (JPL; Watkins and Yuan, 2012), and the GeoForschungsZentrum Potsdam (GFZ; Dahle, 2013).These data are then postprocessed by Don P. Chambers at the University of South Florida following the methods detailed in Chambers and Bonin (2012) and Johnson and Chambers (2013).We consider the ensemble mean across the estimates, deriving an estimate of the standard error according to variances in the differences between series (Ponte and Dorandeu, 2003).To be overlapping with Argo, we consider the GRACE ocean mass data over 2005-2015.

Results and discussion
Figure 1a shows nonseasonal anomalies of GMSL (i.e., annual cycle and trend removed) alongside the Multivariate ENSO Index (MEI) (Wolter and Timlin, 1998) over 2005-2015.As in earlier papers cited above, there is a tight relation between GMSL and MEI curves, such that the GMSL is higher during El Niño periods and lower during La Niña periods.The Pearson product-moment correlation coefficient (hereafter simply referred to as the correlation) between these two records (0.73) is significant at the 95 % confidence level and suggests that approximately half of the nonseasonal anomalous GMSL variance over this period corresponds to ENSO.More generally, we observe that correlation between the nonseasonal GMSL and MEI anomalies is significant for all other 11-year periods during the altimeter record, as well as for the entire 23-year altimetric record itself (not shown).
Nonseasonal GMSL anomalies from satellite altimetry data are consistent with the sum of barystatic and steric components from GRACE and Argo (Fig. 1b).The correlations between GMSL from GRACE and Argo and from altimetry (0.89), and between MEI and the sum of GRACE and Argo (0.67), are both significant.Correlation values between GRACE and the MEI (0.54; Fig. 1c) and Argo and the MEI (0.65; Fig. 1d) are also significant.In fact, all pairs of time series displayed in Fig. 1 are significantly correlated (not shown).These results suggest that GMSL fluctuations tied to ENSO and seen by satellite altimetry are independently corroborated by the other ocean observing platforms and that barystatic and steric terms both contribute to the significant relationship between GMSL and ENSO.
To consider the GMSL budget related to ENSO more formally, we use linear estimation, namely ordinary least squares (OLS).We model the data as linear combinations Table 1.Results of OLS applied to altimetric GMSL (η), GRACE barystatic sea level (p b ), Argo steric sea level (η ρ ), and linear combinations thereof.Values are given as 90 % confidence intervals as described in Appendix B. Note that, while the predictors of the OLS fit include an annual sine and cosine, we present results here transformed into the amplitude and phase of a sine term using standard trigonometric transformations.Note also that n * is the effective number of data points (evaluated following Eq.A3 in the Appendix), whereas δ Y is the standard error evaluated for the different data as outlined in Sect. 2.  of decadal trend, annual cycle, and MEI regressors, simultaneously solving for the regression coefficients for all predictors by minimizing the residual.This particular form of linear regression is motivated by previous studies referenced in the Introduction.(Indeed, the regression explains 90 % of the variance in the GMSL, barystatic, and steric curves over 2005-2015, and the coefficients of the regressors are all statistically significant, as revealed in Table 1 and discussed in more detail below, suggesting that this form of regression model is justified.)While OLS assumes the residuals behave as white noise, in practice we find that residuals are serially correlated (not shown).Thus, we inflate the standard errors according to the lag-1 autocorrelation and the effective degrees of freedom as detailed in Chambers et al. (2012) and Calafat and Chambers (2013).More technical details of our methods are found in Appendix A.
Table 1 shows results of this OLS procedure applied to altimetry, GRACE, and Argo.All quoted values are 90 % confidence intervals as described in Appendix B. (Since they are not our focus here, we defer discussion of results for the annual cycle and linear trend to Appendix C.) Per unit MEI change, altimetric GMSL changes by 2.76 ± 1.87 mm, C. G. Piecuch and K. J. Quinn: ENSO and global sea level which is close to the value of 2.97 ± 1.47 mm given by the sum of Argo steric and GRACE barystatic terms.Indeed, the residual value is not statistically distinguishable from zero (−0.20 ± 0.64 mm), showing that the GMSL budget related to ENSO can be closed using observational data.Closure of the budget implies that steric contributions from regions not sampled by Argo (shelf seas, Arctic Ocean, below 2000 m) cannot be detected over the study period.Llovel et al. (2014) reach a similar conclusion regarding deep ocean steric contributions to the GMSL trend budget over 2005-2013.Significant regression coefficients are also determined for Argo steric (1.42 ± 0.53 mm) and GRACE barystatic (1.54 ± 1.50 mm) components.The error bars on the barystatic term are comparatively wider than on the steric term, agreeing with the relatively stronger correlation between Argo and MEI than between GRACE and MEI seen above (Fig. 1).
The OLS regression coefficients demonstrate that steric and barystatic effects generally make comparable contributions to the ENSO-related GMSL changes over the study period.Judging from Monte Carlo simulations performed using values in Table 1 (see Appendix D), it is as likely as not (33-66 % likelihood) that barystatic effects are responsible for 45-58 % of the sum of barystatic and steric contributions to GMSL variations linked to ENSO, and very unlikely (< 10 % likelihood) that the barystatic term amounts to > 68 % (Fig. 2).This is at odds with the emphasis placed on the barystatic contribution by recent studies (e.g., Llovel et al., 2011;Cazenave et al., 2012Cazenave et al., , 2014)), revealing that, at least over this time period, the steric component is equally as important.
Regional distributions of ENSO-related terrestrial water storage, which are ultimately coupled to the barystatic contributions to GMSL fluctuations through mass conservation, are explored in past papers (Llovel et al., 2011;Boening et al., 2012;Phillips et al., 2012;Fasullo et al., 2013;de Linage et al., 2013;Eicker et al., 2016); they are not revisited here.However, ENSO-related GMSL behavior owing to steric effects is not as well understood.The steric contributions to the GMSL fluctuations related to ENSO arise from changes in ocean heat content.Arguments based on mass conservation (Munk, 2003) suggest that any global steric contributions resulting from salinity changes would be exceedingly small.To elucidate ocean heat content changes potentially contributing to GMSL changes related to ENSO, we apply the OLS method to Argo vertical potential temperature profiles, averaging horizontally over the global ocean as well as individual ocean basins (Fig. 3).
There is significant warming of the global ocean's surface waters (0-100 m) and cooling within its main thermocline (130-320 m) during El Niño periods.Marginally significant warming also occurs at some intermediate depths (600-650 m).On the whole, the global upper ocean (0-2000 m) gains 5.5 ± 5.2 ZJ (ZJ ≡ 10 21 J) of heat per unit of MEI increase (equivalent to a uniform global ocean temperature The thick black curve is the likelihood that the barystatic contribution to ENSO-related GMSL changes will exceed a certain fraction of the sum of barystatic and steric terms based on Monte Carlo runs, where the steric term is evaluated based on the average of the SIO and IPRC gridded data products.The thin dark gray (light gray) curve is that same likelihood but with the steric term assessed using only the SIO (IPRC) product.
variation on the order of 0.001 • C).While there are some significant thermal changes related to ENSO observed in other basins at some depths (< 60 m in the Indian; > 1350 m in the Atlantic; see Fig. 3 caption for basin definitions), the vertical structure of the global ocean's ENSO-related thermal variations derives from the Pacific, where there is similar warming near the surface (0-110 m), cooling in the thermocline (130-320 m), and warming of intermediate waters (500-1150 m).Indeed, only the Pacific shows significant net thermal changes during ENSO, which is hardly surprising as ENSO originates from coupled air-sea interactions in the Pacific (e.g., Clarke, 2008, and references therein).
Given only the Argo data, one cannot unambiguously assess heat budgets for the various layers over the different basins.One possible interpretation is that net Pacific Ocean heat storage is owing to local surface heat exchanges with the atmosphere.This interpretation assumes no contributions from the deep ocean (> 2000 m) and no fluxes between basins, and demands heat fluxes from the thermocline layer to the surface and intermediate layers (Fig. 4).Our interpretation is supported by Mayer et al. (2014), who argue that ocean heat storage over the tropical Pacific (30 • S-30 • N) during ENSO is balanced by surface heat exchanges.Other interpretations are possible given the data, but would imply that surface heat fluxes over every other basin are balanced and compensated by ocean heat transports out of or into that basin.Any more definitive diagnosis of the heat budgets would require a more advanced approach.For example, future studies could use an ocean state estimate covering the altimetric era (e.g., Forget et al., 2015), not only to investi-Ocean Sci., 12,[1165][1166][1167][1168][1169][1170][1171][1172][1173][1174][1175][1176][1177]2016 www.ocean-sci.net/12/1165/2016/gate a longer time period and corroborate or refute the purely observational results presented here, but also to better understand the physical processes contributing to the global and regional steric changes (cf.Piecuch andPonte, 2011, 2014).
Previous studies suggest that both the global ocean and climate system lose heat during El Niño events (e.g., Roemmich and Gilson, 2011;Loeb et al., 2012;Trenberth et al., 2014).This would appear to conflict with our finding that the ocean is warmer during El Niños.However, the discrepancy is only apparent, since we consider ocean heat content and those past studies focus on the ocean heat content tendency (i.e., its rate of change).Moreover, scrutinizing visual examination of the earlier results (e.g., Fig. 8 in Trenberth et al., 2014) suggests that there is a phase lag between ENSO and the heat content tendency, such that warming precedes El Niño peaks and cooling follows peaks.This would be fully consistent with our findings, and those of von Schuckmann et al. (2014), who show a negative global ocean heat content anomaly during the 2010-2011 La Niña.Future studies should investigate in closer detail the coherence between variations in ocean heat content and ENSO.
The vertical structure of ocean temperature changes during ENSO events found here (Fig. 3) has implications for understanding which ocean regions and depth levels contributed to the recent "surface warming slowdown", which some partly relate to the dominant La Niña phase of the 2000s relative to the 1990s (Kosaka and Xie, 2013;Cazenave et al., 2014;England et al., 2014;Risbey et al, 2014).Nieves et al. (2015) determine that the slowdown was caused by a decadal shift 300 m warmed from the 1990s to the 2000s, but that the rate of global ocean heat storage above 1500 m did not change during that time.Our results (Fig. 3) suggest that cooling of the surface Pacific between the 2 decades is consistent with phasing of ENSO, but subsurface Indian warming and lack of net ocean warming or cooling are not, hinting that processes unrelated to ENSO also contributed to the surface warming slowdown, consonant with papers showing an important role for the Interdecadal Pacific Oscillation (Meehl et al., 2013;Trenberth and Fasullo, 2013;Steinman et al., 2015;Fyfe et al., 2016).In this study, SIO and IPRC Argo datasets were considered.While reflected in the standard errors, differences between these two products are apparent.For example, while both curves evidence an overall increase from the beginning of 2011 to the middle of 2015, the SIO and IPRC global steric height series diverge thereafter, with IPRC turning down and decreasing, and SIO continuing to rise through the latter half of 2015 (Fig. 1d).These global differences stem from regional discrepancies (Fig. 5).Nonseasonal steric height patterns over the global ocean from SIO and IPRC from July to December 2015 are generally similar, but mani-fest clear discrepancies in the North Pacific, such that SIO shows more negative values than IPRC near the Equator towards the west, and more positive values over the tropics more broadly (Fig. 5c).Differences between the datasets could be due to different data sources, vertical resolution, or processing strategies, and more detailed future studies should more definitively attribute such discrepancies.Results shown in Llovel et al. (2014) attest to similar differences between SIO and IPRC datasets with regard to the global steric height trend over 2005-2013.Our qualitative conclusions are robust to such quantitative differences between the Argo datasets; for example, employing either SIO or IPRC only, the GMSL budget related to ENSO closes (not shown), and it is unlikely (< 33 % likelihood) that the barystatic term contributes > 68 % to the sum of barystatic and thermosteric contributions to the GMSL changes linked to ENSO (Fig. 2).
Finally, nonseasonal anomalous GMSL was considerably higher during the 2014-2015 El Niño than during the 1997-1998 El Niño (Fig. 6), which is noteworthy because these two El Niño events were comparable in amplitude.(In addition to the distinct axis limits, Figs.1a and 6 differ in that the removed linear trend and annual cycle are estimated for 2005-2015 in the former and 1993-2015 in the latter.)This could suggest that the relationship between GMSL and ENSO is a complicated function of time period and frequency band, in which case the results presented here apply strictly to the study period.However, it could also suggest that other climate modes (e.g., Pacific Decadal Oscillation, e.g., Hamlington et al., 2016) exert an influence on GMSL that has yet to be discussed.

Conclusions
It has long been known that nonseasonal variations in global-mean sea level (GMSL) are correlated with El Niño-Southern Oscillation (ENSO), but the nature of such GMSL fluctuations tied to ENSO, whether steric or barystatic, has remained unclear.We used linear estimation to consider a decade's worth of altimetry, GRACE, and Argo data processed by different research centers, thus clarifying the nature of the GMSL balance related to ENSO.Fluctuations in ENSO, GMSL, and barystatic and steric terms are significantly correlated (Fig. 1).Barystatic and steric components render comparable contributions to GMSL changes during ENSO events (Table 1).The steric contributions reflect ocean heat storage across various depths in the Pacific Ocean (Fig. 3).We offered a heuristic interpretation of the Pacific heat budget during ENSO periods in terms of diabatic exchanges at the sea surface and adiabatic redistributions within the ocean interior (Fig. 4), but more work is needed in the future to diagnose more definitively the relative contributions of surface fluxes, interbasin exchanges, vertical transports, and the deep ocean to the heat budgets.More work is also needed to understand differences between grid-Ocean Sci., 12,[1165][1166][1167][1168][1169][1170][1171][1172][1173][1174][1175][1176][1177]2016 www ded Argo datasets (Fig. 5) and to determine why the anomalous GMSL response to ENSO was apparently much stronger during the 2014-2015 El Niño than during the 1997-1998 El Niño (Fig. 6).Our results corroborate previous suggestions made based on models (Landerer et al., 2008) or observations during an isolated event (Dieng et al., 2014(Dieng et al., , 2015) ) that steric contributions to ENSO-related GMSL fluctuations are not negligible relative to barystatic contributions.These findings also have implications more generally for understanding the ocean's role in the planet's radiation imbalance and hydrological cycle.

Data availability
Data used in this study are available from the sources detailed in Appendix Table E1.Matlab code for processing these data and creating the figures here are available from the first author upon request.

Figure 1 .
Figure 1.Monthly time series over 2005-2015 of (a) altimetric GMSL (black) and the MEI (shading), (b) GMSL from altimetry (black) and from GRACE and Argo (blue), (c) GRACE barystatic sea level (green) and the MEI (shading), and (d) Argo steric sea level (orange) and the MEI (shading).The thin dark gray (light gray) curve in (d) is Argo steric sea level based on the SIO (IPRC) gridded dataset.Linear trends and annual cycles have been removed from all time series.The MEI record has been scaled to have variance equal to that of the respective sea level time series.
Figure2.The thick black curve is the likelihood that the barystatic contribution to ENSO-related GMSL changes will exceed a certain fraction of the sum of barystatic and steric terms based on Monte Carlo runs, where the steric term is evaluated based on the average of the SIO and IPRC gridded data products.The thin dark gray (light gray) curve is that same likelihood but with the steric term assessed using only the SIO (IPRC) product.

Figure 3 .
Figure 3. Coefficients of regressions of Argo potential temperature on the MEI ( • C per MEI) over 2005-2015 over (a) the global ocean and the (b) Pacific, (c) Indian, (d) Atlantic, and (e) Southern (south of 30 • S) basins.Solid lines are the regression coefficients and dashed lines mark the 90 % confidence interval.Bold lines mark significance at the 95 % confidence level (i.e., one-tailed test).Note the different horizontal axis limits between the top and bottom panels.The colored values between the top and bottom panels represent the total ocean heat storage (units of ZJ per MEI; 1 ZJ ≡ 10 21 J) integrated over 0-2000 m in the different basins given as 90 % confidence intervals.For this figure, the Indian Ocean was defined as between 31.5 and 122.5 • E and north of 30.5 • S, the Atlantic Ocean between 76.5 • W and 14.5 • E and north of 30.5 • S, the Pacific Ocean between 118.5 • E and 69.5 • W and north of 30.5 • S, and the Southern Ocean south of 30.5 • S.

Figure 4 .
Figure 4. Hypothesized Pacific heat budget during El Niño events.The blue blocks are the ocean surface (0-110 m), main thermocline (120-380 m), and intermediate water (400-2000 m) layers.The red arrows are heat exchanges between the ocean layers or with the overlying atmosphere.Black values are either the total ocean heat storage within the layers as given by Argo data or the required heat exchanged between them under the stated assumptions of no transports between ocean basins and no contributions from the deep (> 2000 m) ocean.Units are ZJ per unit MEI.(Note that all arrows and signs, shown here for El Niño, would be reversed for La Niña.)