NAO implicated as a predictor of Northern Hemisphere mean temperature multidecadal variability

Authors

  • Jianping Li,

    Corresponding author
    1. State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China
    • Corresponding authors: J. Li and C. Sun, State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, 100029, China. (ljp@lasg.iap.ac.cn; scheng@lasg.iap.ac.cn)

    Search for more papers by this author
  • Cheng Sun,

    Corresponding author
    1. State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China
    • Corresponding authors: J. Li and C. Sun, State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, 100029, China. (ljp@lasg.iap.ac.cn; scheng@lasg.iap.ac.cn)

    Search for more papers by this author
  • Fei-Fei Jin

    1. Department of Meteorology, University of Hawaii at Manoa, Honolulu, Hawaii, USA
    Search for more papers by this author

Abstract

[1] The twentieth century Northern Hemisphere mean surface temperature (NHT) is characterized by a multidecadal warming-cooling-warming pattern followed by a flat trend since about 2000 (recent warming hiatus). Here we demonstrate that the North Atlantic Oscillation (NAO) is implicated as a useful predictor of NHT multidecadal variability. Observational analysis shows that the NAO leads both the detrended NHT and oceanic Atlantic Multidecadal Oscillation (AMO) by 15–20 years. Theoretical analysis illuminates that the NAO precedes NHT multidecadal variability through its delayed effect on the AMO due to the large thermal inertia associated with slow oceanic processes. An NAO-based linear model is therefore established to predict the NHT, which gives an excellent hindcast for NHT in 1971–2011 with the recent flat trend well predicted. NHT in 2012–2027 is predicted to fall slightly over the next decades, due to the recent NAO decadal weakening that temporarily offsets the anthropogenically induced warming.

1 Introduction

[2] It is well-known that the Earth's climate is warming, which has major global implications for human well-being. Observations show that in contrast with the monotonic increase in global annual mean concentration of atmospheric carbon dioxide (CO2) (Figure 1a, black line), the oscillatory multidecadal variability in the twentieth century Northern Hemisphere (NH) mean surface temperature (NHT) (Figure 1a, red line) is obvious [Schlesinger and Ramankutty, 1994; Swanson et al., 2009], superimposed upon a long-term warming trend. The NHT exhibits a warming-cooling-warming (WCW) pattern in the twentieth century [Thompson et al., 2010] and a noticeable flat trend over the past decade [Keenlyside et al., 2008; Easterling and Wehner, 2009; Swanson and Tsonis, 2009; Solomon et al., 2010]. This NHT multidecadal variability is difficult to understand, model, and predict and its source remains elusive, although several explanations have been reported. For instance, the change in sulphate aerosols in the troposphere may have contributed to the WCW pattern in NHT [Hegerl et al., 2007; Booth et al., 2012; Zhang et al., 2013]. Alternatively, oceanic multidecadal variability may also be implicated in the nonmonotic rise in NHT [Delworth and Mann, 2000; Knight et al., 2005; Zhang et al., 2007]. However, the NHT response to sulphate aerosols is model dependent due to large uncertainties in the representation of aerosol-related physics [Hegerl et al., 2007; Zhang et al., 2013], and the in-phase relationship between the NHT and multidecadal variability in sea surface temperature (SST) [Knight et al., 2005; Oldenborgh et al., 2012] cannot be used in the advance prediction of the NHT multidecadal variations. Meanwhile, the role of atmospheric low-frequency variability in contributing to the NHT multidecadal variability is not fully understood.

Figure 1.

NHT, NAO, and AMO indices from 1900 to 2011. (a) The 11 year running mean (red line) NHT (averaged poleward of the equator) anomalies from the HadCRUT4 data set, relative to the base period 1961−1990. The black line shows the global atmospheric CO2 concentrations. (b) The 11 year running means of DNHT (red), NAO (blue), and AMO (green) indices. The pink shaded areas in Figures 1a and 1b show the 2-sigma uncertainty ranges of the NHT series for the HadCRUT4 data set estimated using the 100 realizations [Kennedy et al., 2011b]. The vertical green line in Figure 1b shows the 2-sigma uncertainty ranges of the AMO series for the HadSST3 data set.

[3] The North Atlantic Oscillation (NAO) is the dominant mode of atmospheric variability over the North Atlantic region [Hurrell, 1995; Li and Wang, 2003] and displays fluctuations on multiple timescales ranging from interannual to multidecadal [Luterbacher et al., 1999; Polyakov and Johnson, 2000; Wanner et al., 2001]. It is significantly associated with many meteorological factors in the North Atlantic sector [Hurrell et al., 2003 and references therein]. Previous studies have also suggested that the NAO variability significantly contributes to the oceanic variability in the North Atlantic, such as the SST tripole pattern and the strength of the ocean circulation [Visbeck et al., 2003 and references therein]. In the present paper, through both observational and theoretical analyses, we demonstrate that the NAO, through its delayed effect on the ocean, leads NHT multidecadal variability and serves as a useful predictor for decadal NHT.

2 Data and Methodology

[4] We obtained the HadSST3 SST data set [Kennedy et al., 2011a, 2011b] and the HadCRUT4 combined land and SST data set [Morice et al., 2012] from http://www.cru.uea.ac.uk. Data for 100 ensemble members are available, which sample the estimated observational uncertainty. The median of the ensemble has been used here. Global CO2 concentrations were obtained from the suite of atmospheric concentration observations and emissions estimates for greenhouse gases [Meinshausen et al., 2011]. NCAR sea level pressure (SLP) data set used is available during 1899–2011 [Trenberth and Paolino, 1980]. The NAO index (NAOI) used here is defined as the difference in the normalized SLP zonally averaged over the North Atlantic sector from 80°W to 30°E between 35°N and 65°N with the base period 1961−1990, derived from the NCAR SLP data set [Li and Wang, 2003]. In addition, the principal component (PC)-based NAOI [Hurrell, 1995] and the NAOI from 1880 derived from the Hadley SLP data set [Allan and Ansell, 2006] were employed to verify the reliability of the results. The Atlantic Multidecadal Oscillation (AMO) index during 1900–2011 was calculated by averaging detrended annual mean SST anomalies over the extratropical North Atlantic region (30°N–65°N, 75°W–7.5°W) with the base period 1961−1990 [Trenberth et al., 2007].

[5] The significance of the correlation between two autocorrelated time-series was accessed using the effective number of degrees of freedom, Neff, which can be given by the following approximation [Pyper and Peterman, 1998; Li et al., 2012]

display math(1)

where N is the sample size and ρXX(j) and ρYY(j) are the autocorrelations of two sampled time-series X and Y at time lag j, respectively. The confidence intervals for values predicted by the multiple linear regression are determined by von Storch and Zwiers [1999] (see Methods in supporting information).

3 Results

[6] Figure 1b shows the 11 year running mean NAOI and detrended NHT (hereafter, DNHT) time-series based on annual averages during 1900–2011. It can be seen from Figure 1b that both the annual mean NAO and DNHT show pronounced multidecadal variability and a phase lag of about 10–20 years (DNHT lagging NAO). As shown in Figure 2a, the maximum correlation coefficients (0.45 for unfiltered data, 0.95 for 11 year running means, both significant at the 98% confidence level) occur at a lag of approximately 16 years (NAO leading DNHT); the simultaneous correlations are near zero, and no significant correlations are found when DNHT leads the NAO. Similar results are found for other choices of running means (supporting information Figure 1). Data from longer (1880–2011) and shorter (1950–2011) periods of sampling also confirm the significant correlation found when the NAO leads the DNHT by around 16 years (supporting information Figure 2). Furthermore, detrending the NHT time-series using the best guess radiative forcing due to CO2 and other well-mixed greenhouse gases gives qualitatively similar results (supporting information Figure 3), implying this significant lead-lag correlation does not depend on the method of detrending. These results suggest that the annual mean NAO precedes the DNHT, and it can be considered as a precursor to the multidecadal variability of DNHT.

Figure 2.

(a) Lead-lagged correlation between the annual mean NAO and DNHT indices (1900–2011). The red (blue) line is for the annual mean (11 year running means) time-series. Negative (positive) lags mean that the NAO leads (lags) DNHT, and the red (blue) dashed lines denote the 98% confidence levels for unsmoothed (smoothed) time-series using the effective numbers of degrees of freedom. (b) As in Figure 2a, but for the annual mean NAO and AMO indices. (c) Lead correlation map between annual mean NAOI during 1900–1995 and detrended SST anomalies over the extratropical North Atlantic during 1916–2011 based on 11 year running mean data. The dotted area denotes significant correlations at the 95% confidence level.

[7] Both observational and numerical simulation studies have shown that the AMO affects surface air temperatures over much of the NH, and therefore contributes to multidecadal variability in NHT [Sutton and Hodson, 2005; Knight et al., 2005; Zhang et al., 2007]. The fluctuations in the annual AMO appear to be largely in phase with those in DNHT on interdecadal timescales (Figure 1b). High correlation coefficients (0.7 for unfiltered data, 0.9 for 11 year running means, both significant at the 98% confidence level) are observed between the annual AMO and DNHT indices at zero lag based on both unfiltered and 11 year running mean data (1900–2011). Meanwhile, after subtracting the AMO signal from the NHT, the residual shows a nearly monotonic increase, which is very much like the linear trend in the NHT (supporting information Figure 4), and the multidecadal variability disappears. Previous studies also reported that the response of NHT to the warming/cooling in the extratropical North Atlantic is simultaneous [Knight et al., 2005; Zhang et al., 2007; Wyatt et al., 2012]. On the other hand, as seen in Figure 2b (see also supporting information Figure 2), the most significant positive correlations are found when the annual mean NAO leads the AMO by around 16 years for both unsmoothed and smoothed indices, similar to the lagged correlation between DNHT and the NAO (Figure 2a). The 16 year lead of the NAO compared to the AMO is verified by using different AMO indices [e.g., Enfield et al., 2001; Sutton and Hodson, 2005] (see supporting information Figures 3 and 5). Furthermore, the correlations between SSTs over the extratropical North Atlantic and the NAOI 16 years earlier are significantly positive with a basin wide homogeneous pattern resembling the AMO (Figure 2c). These results imply that the AMO might play a “bridge” role in the 16 year lead of the NAO relative to the DNHT. In the following, we will focus on the associated physical interpretations.

[8] The NAO influence on the AMO has been discussed extensively in numerical model studies [Delworth et al., 1993; Visbeck et al., 1998; Delworth and Greatbatch, 2000; Eden and Jung, 2001; Eden and Greatbatch, 2003]. There is substantial modeling evidence that NAO-related surface turbulent heat flux anomalies over the North Atlantic Ocean, particularly over the Labrador Sea, have an important impact on multidecadal fluctuations of the Atlantic meridional overturning circulation (AMOC), which in turn produce the SST signatures of the AMO. Thus, the AMO may be a delayed response to NAO variability at time lag of decadal timescales [Visbeck et al., 1998; Delworth and Greatbatch, 2000; Eden and Jung, 2001]. The NAO impact on the AMOC was also demonstrated in a recent observational study, and the long-term change in the AMOC (as indicated by the Atlantic dipole) is thought to be related to the low-frequency variations of NAO through changes in Labrador Sea convection [Latif et al., 2006]. At the same time, once the North Atlantic Ocean is warming, this warming will in turn warm the entire NH. Here we defined the NAO signal in NHT as the damped response of the NH climate system to the NAO variability, and simulated the NHT response to variability in NAO using a simple dynamical model.

[9] Analogous to the approach proposed by Hasselmann [1976] to examine the ocean response to atmospheric forcing, and used by Thompson et al. [2009] to examine the global mean surface temperature response to ENSO and volcanic eruptions, we model the NHT response to NAO variability (TNAO) following the Hasselmann climate model (see Methods in supporting information). The simulated response of NHT to year-to-year NAO variability is shown at the top of Figure 3. The simple Hasselmann climate model acts to integrate the impact of the NAO and lag it by decades compared to the NAOI shown in Figure 1b. Thus, the large thermal inertia associated with slow oceanic processes mainly accounts for the decadal lag, and this is basically consistent with previous modeling conclusions [Eden and Jung, 2001]. As is evident in Figure 3, TNAO has similar phase and amplitude to the smoothed DNHT in Figure 1b, particularly for the period after the 1910s. The discrepancy of the simulation before the 1910s may be due to the relatively large uncertainty in NHT in the early twentieth century [Morice et al., 2012] and initialization as well (Methods in supporting information). The correlation between TNAO and the smoothed DNHT is high (r = 0.80) and significant at the 95% level in 1905–2011. It is also noted that one third of the variance in the smoothed DNHT is left unexplained by TNAO, indicating that the NAO is probably not the only factor contributing to the NHT multidecadal variability. Thus, although the NAO is shown to have a significant impact on the NHT, this does not rule out other possible factors (e.g., solar forcing, aerosol loadings, etc.) that influence on the NHT variations (Hegerl et al. [2007] and references therein). A recent study also suggested that the rapid freshening of the subpolar North Atlantic around 1970 also has an impact on the long-term change in the NHT [Thompson et al., 2010]. The AMO signal in the NHT was obtained by linear regression of the annual mean NHT onto the AMO index, and is referred to as AMO-related NHT (the top of Figure 3). The correlation between TNAO and smoothed AMO-related NHT is higher (r = 0.84) for 1905–2011, providing further evidence that the AMO plays an important role in connecting the NAO and DNHT. The agreement of TNAO with the observed DNHT and AMO-related NHT and the phase lag between NAO variability and NHT response demonstrates that the NAO precedes multidecadal variability in NHT through its delayed effect on the AMO.

Figure 3.

TNAO and residual NHT. (a) The contribution of NAO variability to NHT obtained by the Hasselmann climate model (TNAO, blue, see Methods in supporting information) and the smoothed AMO-related NHT obtained by the linear regression of the annual mean NHT onto the AMO index (green). The grey area shows the 2-sigma uncertainty ranges of the AMO-related NHT for the HadSST3 data set estimated using the 100 realizations. (b) The residual NHT obtained by subtracting TNAO from the smoothed NHT. The pink area shows the 2-sigma uncertainty ranges of the residual NHT for the HadCRUT4 data set.

[10] The response of NHT to NAO variability was removed by subtracting TNAO from the smoothed NHT, and the resulting residual NHT is shown at the bottom of Figure 3. Evidently, the residual NHT is dominated by a nearly monotonic rise pattern, particularly the rise during the first half of the century and the persistent rise since 1950, further implying the multidecadal variability in NHT may be explained mainly by TNAO.

[11] The results obtained from the Hasselmann climate model provide an explanation for the phase lead-lag between the NAO and DNHT. This lead-lag relationship offers a simple but useful way to predict NHT around a decade and half in advance. Namely, the NAOI being shifted by 16 years can serve to predict NHT. An NAO-based linear model for predicting decadal NHT is therefore established as follows:

display math(2)

where t is time in years and the coefficients a, b, and c are determined empirically by linear regression based on the data over the historical period, so that the regression error of equation ((2)) is minimized. To test the performance of the model, we first performed twenty-six 16 year hindcast experiments for the smoothed NHT during 1971–2011. The empirical model used for each hindcast was constructed only with knowledge of the training period before the hindcast period, and the model coefficients are shown in supporting information Figure 7. High correlations and relatively small root mean square errors between the hindcast and observed NHT are found in each set of hindcast experiment (supporting information Figure 7). Figure 4a shows two 16 year hindcasts of NHT, starting at 1976 and 1996, respectively, by this model. As shown in Figure 4a, the persistent and strong NHT warming trend starting from the middle of the 1970s and the recent flat trend are very well predicted, suggesting that the model does have hindcast skill in predicting the decadal variations of NHT. Fitted NHT using the linear model (Figure 4b) follows the observed NHT in 1916–2011 and captures the twentieth century WCW pattern and the recent flat trend in the NHT very well. The persistent decline in the NAO during the 1920s to 1960s led to the NHT cooling during the 1940s to 1970s. Likewise, the recent decadal weakening of the NAO starting in the 1990s may result in a hiatus of the NHT warming over the next decade. Furthermore, the fitted NHT and the hindcasts using the longer NAOI since 1880 (supporting information Figures 2 and 8) are also in good agreement with observations, showing the robustness of the model.

Figure 4.

The observed, modeled, hindcast, and predicted NHT. (a) The 11 year running mean of the observed NHT (red) from 1916 to 2011, and two sets of 16 year hindcast NHT starting at 1976 and 1996, respectively, using the NAO-based linear model. The solid circle represents the starting year of the hindcast and the fitted NHT before each hindcast is shown as a dotted line in the same color as the corresponding hindcast. The model coefficients for each hindcast experiment are listed in supporting information Table 1. Vertical bars denote the 95% confidence intervals for the linear model. (b) The 11 year running mean of the observed NHT (red) from 1916 to 2011, the model fit of NHT (blue), and the predicted NHT in 2012−2027. The model coefficient a is identical to that used in Figure 4a, and coefficients b and c for the training period 1900−2011 are 8.47 × 10−3 and −16.65, respectively. The pink shaded areas in Figures 4a and 4b show the 2-sigma uncertainty ranges of the NHT series for the HadCRUT4 data set.

[12] With evidence that the NAO precedes multidecadal variability in NHT via a delayed AMO “bridge,” we carry out a prediction for NHT in 2012–2027. The model prediction suggests that NHT will fall slightly over the next decades (Figure 4b). The recent decadal weakening of the NAO, a signal that has already occurred, will thus act to temporarily offset the anthropogenically induced warming trend and lead to a steady or even slight cooling of NHT.

4 Summary and Discussion

[13] This paper demonstrates that the NAO is a useful predictor of NHT up to around two decades in advance. The NAO signal leads the observed DNHT and AMO by about 16 years, and the multidecadal variability in NHT can be explained and modeled very well by using the NAO signal. An NAO-based linear model is then established to predict the NHT, and the hindcasts of NHT in 1971–2011 compare well with the observations. The NHT in 2012–2027 is predicted to fall slightly over the next decades, due to the recent NAO decadal weakening that temporarily offsets the anthropogenically induced warming. In addition, similar results could be obtained by using the PC-based NAOI (supporting information Figures 9–12).

[14] Although the origin of multidecadal variations in the NAO remains unclear [Cohen and Barlow, 2005], the results presented here provide compelling evidence that multidecadal variability in the NAO contributes significantly to the NHT multidecadal variability, and they also illuminate that understanding and better modeling of this NAO multidecadal variability are instrumental for interpretation and future projections of the NH climate changes on multidecadal timescale.

Acknowledgments

[15] The authors wish to acknowledge Dr. Clara Deser from NCAR for her helpful discussion and suggestion, and thank two anonymous reviewers for their constructive comments. J.L. and C.S. are supported by 973 Program (2010CB950400), NSFC grant (41030961), and CAS Project (XDA5090403). F.J. is supported by NSF grant (ATM 1034798), NOAA grant (NA10OAR4310200), and DOE grant (DESC0005110).

[16] The Editor thanks two anonymous reviewers for assistance evaluating this manuscript.

Ancillary