Detection of external influence on sea level pressure with a multi-model ensemble



[1] Over the past fifty years, December–February mean sea level pressure has decreased markedly over both poles, corresponding to a trend toward strengthened westerlies in both hemispheres. In this study we compare observed sea level pressure trends with those simulated in response to natural and anthropogenic influence in a suite of eight up-to-date coupled general circulation models. A global analysis indicates that sea level pressure trends may be attributed to external influence. However, while simulated Southern Hemisphere sea level pressure trends are consistent with those observed, simulated Northern Hemisphere sea level pressure trends are not: Observations show a large negative trend in the Arctic and a positive trend over the subtropical North Atlantic and Mediterranean which is not reproduced in the simulations.

1. Introduction

[2] Over recent decades, trends in boreal winter sea level pressure have been observed, with decreases in sea level pressure (SLP) over the polar regions, and increases in the subtropics [e.g., Thompson et al., 2000; Gillett et al., 2003b]. These trends correspond to changes toward the positive phase of the Northern and Southern Annular Modes [e.g., Thompson et al., 2000]. Since these trends were first identified, many attempts have been made to link them to changes in external forcings. Many researchers have identified a possible role for greenhouse gases in inducing the positive phase of the Northern Annular Mode [e.g., Fyfe et al., 1999; Shindell et al., 1999], although ozone depletion and natural forcings have also been cited as possible causes [e.g., Gillett et al., 2003a]. The North Atlantic Oscillation index (NAO), which correlates highly with the Northern Annular Mode index, exhibited a maximum during the early nineties, and has since decreased somewhat [Osborn, 2004]. Trends in the Southern Annular Mode have been robustly linked to stratospheric ozone depletion [Sexton, 2001; Thompson and Solomon, 2002; Gillett and Thompson, 2003], although greenhouse gas changes [Marshall et al., 2004; Shindell and Schmidt, 2004], and even changes in natural forcing [Marshall et al., 2004] have also been cited as contributing factors.

[3] Looking beyond indices of circulation, Gillett et al. [2003b] analyzed time-space patterns of SLP change by comparing a suite of observational and reanalysis SLP data sets with simulated SLP from four climate models with greenhouse gas and sulfate aerosol forcing only. Consistent with the observations they found that the climate models simulated decreases in SLP over both poles and a small increase over the subtropical North Atlantic, although the magnitude of the changes was much smaller than that observed. They concluded that observed SLP trends are inconsistent with simulated internal variability and also inconsistent in magnitude with the simulated response to greenhouse gas and sulfate aerosol.

[4] Overall, greenhouse gases, stratospheric ozone depletion, volcanic aerosol and solar forcing have all been cited as having an impact on SLP [Shindell et al., 2001; Gillett et al., 2003a; Marshall et al., 2004]. Therefore, in order to obtain the best comparison with the observations it is important to compare with model simulations which contain all these forcings. In this study, we therefore made use of integrations with all the major anthropogenic and natural forcings from a large suite of state-of-the-art coupled models collected in preparation for the IPCC Fourth Assessment Report. We examined trend patterns and went on to apply a detection and attribution analysis similar to that applied by Gillett et al. [2003b].

2. Data

[5] We based our analysis on decadal mean December–February (DJF) gridded SLP (1955–2005) from the recently-completed HadSLP2.0 data set (R. J. Allan and T. J. Ansell, A new globally complete monthly historical gridded mean sea level pressure data set (HadSLP2): 1850–2003, submitted to Journal of Climate, 2005), the NCEP/NCAR reanalysis [Kalnay et al., 1996], and the ERA-40 reanalysis [Uppala et al., 2005]. Compared to the period 1948–1998 analysed by Gillett et al. [2003b], our use of 1955–2005 data ameliorates some of the problems associated with biases in the NCEP reanalysis over Antarctica during the early part of the record [Marshall et al., 2004], and also addresses the issue of a recent reversal in SLP trends in the Northern Hemisphere (NH) [Osborn, 2004]. HadSLP2.0 is a selective release of a new monthly mean gridded SLP data set based on station and ship measurements with no infilling of missing data. It includes more observations than were used in the previous HadSLP1.0 version of the data set used by Gillett et al. [2003b], and is updated to 2003 using new terrestrial and marine observations. We further updated it to 2005 using the NCEP reanalysis (allowing for differences in the long term means of the two data sets). When calculating decadal means at each grid point we required at least half the monthly anomalies to be present, otherwise the decadal mean was marked as missing. Note that the ERA-40 reanalysis only extends from 1957 to 2002, and therefore when calculating decadal means from 1955 to 2005 we used only eight years when calculating the first decadal mean and seven years when calculating the last. We chose to focus on the boreal winter, since this is when simulated SLP changes are largest in both hemispheres, and it is also the season examined in previous studies [e.g., Gillett et al., 2003b].

[6] We compared observed SLP with that simulated by a suite of eight up-to-date coupled general circulation models in response to changes in greenhouse gases, the direct sulfate aerosol effect, stratospheric ozone, volcanic aerosol and solar irradiance. The models used were UKMO-HadCM3, CCSM3, PCM, GFDL-CM2.0, GFDL-CM2.1, MIROC3.2(medres), GISS-EH, and GISS-ER. We chose all the models included in the IPCC Data Archive which had an initial condition ensemble of at least three 20th century simulations, and which included all the forcings listed above (output from UKMO-HadCM3 was obtained directly from the Hadley Centre). Some of the models also included the indirect sulfate aerosol effect, black carbon aerosol and land use changes. In total, 39 individual 20th century simulations were available. Since the 20th century simulations generally finished in 1999 or 2000, and we wished to compare with observations up to 2005, it was necessary to use output from SRES scenario integrations for the last part of the simulated period: Simulations with either SRESA2, SRESA1B or SRESB1 emission scenarios were used as available, since these scenarios differ little in the first five years. As these scenario integrations include all the major anthropogenic forcings, and there have been no large volcanic eruptions in the past five years, the forcing in the period 2000–2005 in these scenario integrations is close to the true forcing. Control simulations from each model were used to assess statistical significance. R. L. Miller et al. (Forced variations of annular modes in the 20th century IPCC AR4 simulations, submitted to Journal of Geophysical Research, 2005) examined 1950–1999 sea level pressure variability in these models, and found the structure and amplitude of Southern Hemisphere (SH) variability to be generally realistic, while that in the NH was less realistic, although the discrepancy may be partly due to differences in trends.

3. Results

[7] Figure 1 shows December–February SLP trends over the period 1955–2005 based on decadal means of (a) the NCEP reanalysis, (b) the ERA-40 reanalysis, and (c) HadSLP2.0. The trends are characterized by decreases over the two polar regions, and increases in parts of the subtropics. The 1955–2005 NCEP reanalysis trends shown in Figure 1a are similar to those shown over the period 1948–1998 by Gillett et al. [2003b], and are broadly consistent with the trends in the ERA-40 reanalysis and HadSLP2.0. The three data sets are least consistent in the SH, where they are constrained by fewer observations. Figure 1d shows the multi-model mean of the simulated response to anthropogenic and natural forcings. The simulated response is dominated by a decrease in SLP southward of 55°S, and an accompanying increase at around 45°S: A comparable SH response was simulated by all the models individually. By contrast, in the NH, although the mean simulated high latitude response is predominantly negative, it is generally very close to zero: The PCM and GISS models simulated the largest decreases over the Arctic, but even these were much smaller than the observed changes. Although the GISS models have a high upper boundary at 0.1 hPa, the PCM does not, and no systematic relationship between upper boundary height and NH SLP response was found across the ensemble.

Figure 1.

December–February sea level pressure trends based on decadal means over the period 1955–2005 are shown for (a) the NCEP reanalysis, (b) the ERA-40 reanalysis, (c) the HadSLP2.0 data set, and (d) for the mean simulated response to greenhouse gas, sulfate aerosol, stratospheric ozone, volcanic aerosol, and solar irradiance changes in eight coupled climate models.

[8] In order to test the consistency of simulated and observed SLP trends globally, we applied a detection and attribution analysis identical to that applied by Gillett et al. [2003b], using the multi-model technique of Gillett et al. [2002]. After masking with the observational coverage we regridded the 1955–2005 decadal mean DJF mean SLP onto a course grid (60° × 25°), truncated onto the first ten common EOFs of the controls, and applied a total least squares regression [Allen and Stott, 2003], using a mean over all available ensemble members for the model response. Concatenated control from all the models was used to assess the uncertainty ranges on the regression coefficients. Regression coefficients based on the NCEP, ERA-40 and HadSLP2.0 data sets are shown with their associated uncertainty intervals in Figure 2. In all three cases the uncertainty range on the regression coefficients does not include zero, indicating that the observed SLP changes are inconsistent with simulated internal variability and the response to external forcing is detected, as Gillett et al. [2003b] also concluded. However, Gillett et al. [2003b] (using the NCEP reanalysis and HadSLP1.0) and Hegerl et al. [2005] (using ERA-40) found that the observed SLP changes were not consistent with the simulated response to greenhouse gas and sulfate aerosol changes only (the calculated regression coefficient was >3). In our results, based on a regression onto the response to all the major forcings, we find a regression coefficient much closer to one, and consistent with one in the case of the NCEP reanalysis and HadSLP2.0. A response to external forcing was also detected in all three data sets using six of the eight models individually. Since 20th century integrations with individual forcings were not available from the IPCC Data Archive, it was not possible to attribute the changes to individual forcings over the whole ensemble. Examination of the responses to individual forcings in UKMO-HadCM3 and PCM, however, indicated that ozone was the dominant contributor to the SH SLP trends, consistent with previous findings [Gillett and Thompson, 2003], with greenhouse gas and natural forcings giving smaller contributions.

Figure 2.

Regression coefficients, β, of observed sea level pressure changes against changes simulated in response to greenhouse gas, sulfate aerosol, stratospheric ozone, volcanic aerosol and solar irradiance changes. Regression coefficients were derived using a multi-model mean of simulated 1955–2005 sea level pressure changes, and each of three observational data sets. The black bars represent 5–95% uncertainty ranges derived from control variability, and the horizontal lines represent best estimates of the regression coefficients.

[9] Although the residual consistency test [Allen and Tett, 1999] did not indicate any inconsistency between simulated and observed SLP changes, it has a liberal bias [Allen and Stott, 2003], and a comparison of the observed and simulated trend patterns (Figure 1) indicates that the two are very different in the NH. A detection analysis applied separately to each hemisphere using the NCEP reanalysis indicated a detectable influence of external forcing in the SH only, with a best estimate of the regression coefficient of 1.2 (90% confidence interval 0.5–1.8). The global result is dominated by the SH response because the simulated response is close to zero in the NH, and the apparent inconsistency in the NH is not identified by the residual test because of its liberal bias. Gillett [2005] examined this issue in more detail by comparing observed NH DJF zonal index trends over the past fifty years with those simulated in the same ensemble of anthropogenic and naturally forced integrations as was used here (one additional simulation of the MIROC3.2(hires) model was also included). The observed trend was found to be inconsistent both with simulated internal variability and with the simulated response to all the major forcings. By contrast an analysis of zonal index trends in the SH indicated that simulated and observed zonal index trends are consistent (not shown). Thus our demonstration of consistency between simulated and observed SLP changes at the global level should be interpreted with caution given the inconsistency between simulated and observed zonal index trends in the NH.

4. Conclusions

[10] Over the past fifty years, boreal winter SLP has decreased over both poles, and increased over the subtropics. We find that these observed SLP changes are inconsistent with simulated internal variability, and we detect a response to external forcing, consistent with the results of Gillett et al. [2003b]. When we compare observed trends with those simulated in response to greenhouse gas, sulfate aerosol, stratospheric ozone, volcanic aerosol and solar irradiance changes, we find that SH trends are consistent: The simulations also show a pronounced decrease in SLP over the Antarctic, and an increase in southern midlatitudes. In contrast to observed trends, however, in the NH simulated SLP trends are close to zero. These trends appear, if anything, smaller than those simulated in response to greenhouse gas and sulfate aerosol changes only [Gillett et al., 2003b], although the absence of individual forcing simulations for the IPCC ensemble prevents us from pursuing this issue further.

[11] A detection analysis applied to global decadal mean SLP indicated a detectable response to external forcing, with regression coefficients much closer to one than in a previous study in which only the response to greenhouse gas and sulfate aerosol changes was considered [Gillett et al., 2003b]. Indeed using the NCEP reanalysis and HadSLP2.0 data set regression coefficients were consistent with one, indicating that observed global SLP trends are attributable to external influence. This is likely due to the inclusion of stratospheric ozone depletion in the simulations described here. However, while observed SLP has exhibited a large decrease over the Arctic over the past five decades, simulated SLP changes over the Arctic are much smaller [Gillett, 2005]. Thus our results should be interpreted with caution given the inconsistency between simulated and observed NH SLP trends. This inconsistency indicates that either the simulated NH SLP response to external forcing is underestimated in the models, or the simulated internal variability is too small. One recent study found that when a trend in stratospheric circulation similar to that observed was imposed in a model, a trend in the NAO consistent with that observed was simulated [Scaife et al., 2005]. This suggests that the observed trend may relate to coincident stratospheric changes which are poorly simulated by the models. Two of the models included in this study, however, extend to 0.1 hPa (GISS-ER and GISS-EH), yet are unable to reproduce the observed NH SLP changes, thus any such shortcoming in the models' simulation of a response to stratospheric changes is unlikely to relate purely to their limited vertical extent [Shindell et al., 1999].


[12] We thank Peter Stott (Hadley Centre, UK Met Office) for providing HadCM3 data and for advice; Daithi Stone (University of Oxford) for comments and advice; and Myles Allen (University of Oxford) for providing his optimal detection code. We acknowledge the international modeling groups for providing their data for analysis, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) for collecting and archiving the model data, the JSC/CLIVAR Working Group on Coupled Modelling (WGCM) and their Coupled Model Intercomparison Project (CMIP) and Climate Simulation Panel for organizing the model data analysis activity, and the IPCC WG1 TSU for technical support. The IPCC Data Archive at Lawrence Livermore National Laboratory is supported by the Office of Science, U.S. Department of Energy. ECMWF ERA-40 data used in this study were obtained from the ECMWF Data Server. This work was partly funded through the Climate Change Detection and Attribution Project by NOAA's Office of Global Programs and the U.S. Department of Energy's Office of Science. Rob Allan and Tara Ansell were supported by the UK Government Meteorological Research (GMR) contract.