Observed Antarctic oscillation trends

Several DJF AAO indices derived from observed data and reanalysis slp data are displayed in Fig. 1a. Generally speaking, these AAO indices from different sources show consistent year-to-year variability over the entire period. Their similarities are evident as suggested by the co-variations in the overlapping periods. For example, a low point that occurred in 1977 appears in most of the AAO indices and a high event that occurred in 2000 is manifest in these time series. Generally, there are more positive AAO events after the 1980s.

Fig. 1 a. Observed December–January–February Antarctic oscillation (DJF AAO) indices since 1958. b. Sea level pressure (slp)-based DJF AAO indices since 1850s. c. Principal component reconstruction method (PCR-recon) DJF AAO indices since 1500. d. Composite plus scale reconstruction method (CPS-recon) DJF AAO indices since 1500. All the series are normalized with respect to the period 1961–90. The smoothed lines denote low-frequency (> 10 years) variations from a Butterworth filter. Shading indicates the range of±2 x standard error derived from the unresolved variance in the calibration period in c. and d. DJF in 1958 denotes December 1957, January 1958 and February 1958.

The smoothed lines in Fig. 1a denote the interdecadal variations (> 10 years) of each AAO index during the last 50 years. There is an obvious positive phase in the early 1960s which is consistent with the series in Jones & Widmann (Reference Jones and Widmann2004), and another significant positive phase occurred in the mid-1990s. From the mid-1960s to 2000s, the station-based Marshall-AAO shows a positive trend. The persistent increasing trends are also evident for the other three AAO model-based records (Fig. 1a). The linear trends of each DJF AAO index over different periods are estimated and shown in Table I. It is clear that the summer AAO has undergone a strengthening since the 1960s, and the most intense enhancement occurred from the mid-1980s to the end of the 1990s. Although the linear trend varies from a minimum 0.34 σ 10 years^-1 to a maximum of 0.58 σ 10 years^-1 during the period 1980–2011, the trends for six different AAO series are all significant at the 0.05 level. Similar conditions can also be found for the period 1970–2000. The positive increasing trends described here are basically consistent with several previous studies.

Table I Linear trends of December–January–February Antarctic oscillation (DJF AAO) indices in different periods (unit is σ 10 years^-1, all the series are normalized with respect to the period 1961–90).

**Significant at 0.01 level, *significant at 0.05 level, based on a two-tailed student's t-test. ± standard error. ^#The HadSLP-AAO, Fogt-AAO and Visbeck-AAO end in 2004, 2004 and 2005, respectively.

Trends and interdecadal variations of the December–January–February Antarctic oscillation in reconstructions

As mentioned in the previous section, the DJF AAO during the last 50 years is characterized by significant positive trends, especially in the 1980–2000 period. We wish to assess whether the trend over the last several decades is unusual from a long-term perspective. We show five reconstructed DJF AAO indices in Fig. 1b–d, including Visbeck-AAO, Fogt-AAO, HadSLP-AAO, and two optimal reconstructions of AAO conducted by Zhang (Reference Zhang2010) and Zhang et al. (Reference Zhang, Gong, He, Lei and Feng2010). These are labelled as PCR-recon (by using the principal component reconstruction method) and CPS-recon (by using the composite plus scale reconstruction method). The latter two optimal reconstructions were based on multiple proxies distributed widely in the Southern Hemisphere (including tree rings, corals and ice cores) and extend back to 1500 AD. Both the PCR-recon and CPS-recon AAO time series focused only on the interannual to interdecadal components (< 50 years). In historical periods, there are apparent differences among these reconstructed AAO indices on both interannual and interdecadal time scales (see Appendix A, http://dx.doi.org/10.1017/S0954102013000734). The AAO time series inevitably contains uncertainty, whether the AAO is based on multiple proxies or slp (Jones et al. Reference Jones, Fogt, Widmann, Marshall, Jones and Visbeck2009, Visbeck Reference Visbeck2009, Zhang Reference Zhang2010). However, the statistics for the multiple-proxy optimal reconstruction indicate that more skilful proxies used in reconstruction would be helpful to reduce the uncertainty or error. For example, with the available proxy increases from 41 in the period of 1500–1518 to 263 in the period of 1894–1956, the standard error (s.e.) of the CPS-recon (shading in Fig. 1c) decreases from 1.14 to 0.97, respectively. For the PCR-recon (shading in Fig. 1d) the s.e. decreases from 0.79 to 0.50. Generally, the multiple-proxy based DJF AAO indices since 1500 AD are skilful and reliable according to the s.e., reduction of error and other statistics (see Appendix A, http://dx.doi.org/10.1017/S0954102013000734).

Figure 2 presents the 30-year running linear trends of the five reconstructed DJF AAO indices. The Visbeck-AAO, Fogt-AAO and HadSLP-AAO indices have experienced persistent positive trends since the 1960s, and exhibited negative trends in the 1950s, though the evolutions of the three curves are not consistent before the 1920s. Moreover, the positive trends of the Visbeck-AAO index (Fig. 2a) in the 1890s are close to the trends found since the 1960s. The positive trends of the HadSLP-AAO index (Fig. 2c) in the 1870s and from 1930 to the 1940s are close to or higher than the trends in the period since the 1970s. Taking the PCR-recon (Fig. 2d) and CPS-recon (Fig. 2e) DJF AAO series over last 500 years into account, it is clear that, under the half-century timescale (< 50 years), all of the positive trends of PCR-recon in the 1650s, 1870s, 1890s and 1940s, and of CPS-recon in the 1580s, 1650s, 1750s, 1790s, 1830s and 1940s are greater than the trends during 1970–2000, which are near or exceed 0.4 σ 10 years^-1. Moreover, comparing the trends derived from the raw Marshall-AAO series (not filtered, the red crosses in Fig. 2d & e) with PCR-recon and CPS-recon, we find that, even though the trends of the raw AAO series are slightly larger than filtered values, the highest value is still not beyond the maxima of the optimal reconstruction trend. Thus, it can be concluded that the intense positive trends of the DJF AAO on the 30-year timescale have occurred not only in the past 150 years but also occurred in the past 500 years. This implies that the persistent positive trend of the DJF AAO since the 1970s is unlikely to be unprecedented from a historical perspective.

Fig. 2 The 30-year running linear trends for reconstructed December–January–February Antarctic oscillation (DJF AAO) series. a. Visbeck-AAO. b. Fogt-AAO. c. HadSLP-AAO. d. Principal component reconstruction method (PCR-recon). e. Composite plus scale reconstruction method (CPS-recon). The red lines in d. and e. denote the 30-year running trends derived from the raw (not filtered) AAO series. Trends significant at the 0.05 level are marked by the cross for all trend lines. The frequency of 30-year trends for 1000 Monte Carlo reconstructions. f. For PCR from randomly sampling proxies. g. For CPS from randomly sampling proxies. h. For PCR from randomly sampling time coefficients of the principal component analysis (PCs). The black lines in f.–h. denote the corresponding trends from the optimal reconstructions.

The uncertainty of the reconstructions may exaggerate or bias the trend estimations. We investigated this issue and performed multiple AAO reconstructions through Monte Carlo simulation. Here we take into account two kinds of data sampling. One is proxy data sampling and the other is PC (time coefficients of the principal component analysis, PCA) sampling. The proxy sampling test was performed for both PCR and CPS. In these multi-reconstructions, we firstly re-sample the original proxy data by 1000 times. To harness as much information as possible, we confine the number of proxy candidates to being the maximum (i.e. varying between 236 and 262). Based on each set of re-sampling data, the PCR reconstruction and CPS reconstruction are conducted respectively. Then the 30-year running linear trends of each AAO series were examined and plotted as the frequency distribution in Fig. 2f & g. As expected, the AAO series from Monte Carlo simulations show larger variability than the optimal PCR-recon and CPS-recon. The maximum increasing trends of DJF AAO index from Monte Carlo simulation are generally greater than the optimal PCR-recon. There are only 23 simulations showing smaller maximum linear trends than the optimal reconstruction. For the Monte Carlo CPS simulations, all maximum linear trends are larger than the maximum trend of the corresponding optimal AAO. Additionally, we repeated PCR but re-sampling the PCs. According to the proxy data availability, we divided the whole reconstruction period into 150 segments. Given a segment, all available proxies were utilized for PCA analysis. Then the Monte Carlo simulation was performed by randomly selecting PCs. As we did in the optimal reconstruction, only the significant PCs are considered. Note that in the early period the number of proxy is lower, to successfully re-sample 1000 times we intently selected more PC candidates which meet a slightly lower significance level than the later periods. The running trends are shown in Fig. 2h. The probability distribution is similar to that derived from proxy re-sampling simulations. There are much broader tails on both the positive and negative sides of the distribution than for the optimal reconstructions. This implies that the possibility of the maximum trends in our optimal AAO reconstructions exceeding the maximum trends of the re-sampling reconstructions is very low.

Trends and interdecadal variations of the December–January–February Antarctic oscillation in modelling

To investigate the possible contributions of anthropogenic forcings to the positive trend of the DJF AAO over the past several decades, three group multi-model ensemble (MME1, MME2, MME3) mslp fields obtained from the 20C3M and PICTL simulations were developed. Only the first runs of each model were used, and an ensemble denotes the average of a group of models. MME1 is an ensemble of 14 models from the 20C3M scenario, in which the experiments were run with the time-varying greenhouse gases (GHGs) and ozone (O₃) concentration as observed during the twentieth century (denoted as O₃+GHGs run). MME2 is an ensemble of ten model runs forced by only time-varying greenhouse gases (GHGs run). MME3 is an ensemble of eight models from the PICTL runs forced with constant pre-industrial levels of greenhouse gases and ozone (control run).

The differences of mslp climatology in the late nineteenth century (1969–98) and early twentieth century (1901–30) for summer are shown in Fig. 3. The slp difference in each case can be regarded as the results of the greenhouse gas or ozone and greenhouse gas increment. In Fig. 3a & b, the increment of greenhouse gases for MME1 and MME2 is the same between the two analysis periods; the only forcing difference between Fig. 3a & b is the ozone increment. Similarly, the difference between Fig. 3b & c should be related to the greenhouse gases increment. Therefore, their slp differences are suggestive of ozone and greenhouse gas forcings. In the GHGs run, the positive and negative mslp anomalies dominate in the mid-latitudes and high latitudes, respectively. This pattern of pressure indicates that GHGs play a role in strengthening AAO over the twentieth century. In the O₃+GHGs run, the spatial feature is similar but with much larger values, suggesting that ozone and greenhouse gases are both effective at enhancing the positive AAO trend. The control run (Fig. 3c) does not show an obvious AAO feature suggesting that AAO trends cannot be simulated without varying greenhouse gases or/and ozone concentration.

Fig. 3 The differences of sea level pressure (slp) (1969–98 minus 1901–30). a. Multi-model ensemble 1 (MME1). b. MME2. c. MME3. Areas significant at the 0.05 level are shaded. The corresponding time series of AAO. d. MME1. e. MME2. f. MME3. Time series are normalized with respect to the base period of 1961–90, the smoothed lines denote low-frequency (> 10 years) variations from a Butterworth filter, and the red lines indicate linear trends during 1969–98. g. The means of the linear trends plus the error bars (one standard error) for 1969–98.

The normalized DJF AAO indices derived from the three group simulation ensembles are shown in Fig. 3d–f. There are distinct interdecadal variations in all three MME AAO series. The linear trends are evident for the GHGs and O₃+GHGs runs, especially during the last several decades. During 1969–98 the linear trends of MME1 and MME2 are 0.29 σ 10 years^-1 and 0.14 σ 10 years^-1, respectively. In contrast, MME3 shows no evident trend (only a value near zero). We note that the ozone and greenhouse gas forced trend in the IPCC AR4 simulations is approximately 63% and 85% of the trend of station-based Marshall Index for 1970–2000 and for 1980–2011, respectively. This suggests that the anthropogenic forcing may dominate AAO trends in recent decades, although natural forcings, such as solar activity, may also play a role (Kuroda & Shibata Reference Kuroda and Shibata2006).