Climate Dynamics

, Volume 38, Issue 9, pp 1757–1773

Twentieth century Walker Circulation change: data analysis and model experiments

Authors

  • Qingjia Meng
    • Leibniz-Institut für Meereswissenschaften
    • River and Coastal Environment Research CenterChinese Research Academy of Environmental Sciences
    • Key Laboratory of Ocean Circulation and Waves, Institute of OceanologyChinese Academy of Sciences
    • Leibniz-Institut für Meereswissenschaften
  • Wonsun Park
    • Leibniz-Institut für Meereswissenschaften
  • Noel S. Keenlyside
    • Leibniz-Institut für Meereswissenschaften
  • Vladimir A. Semenov
    • Leibniz-Institut für Meereswissenschaften
    • A. M. Obukhov Institute of Atmospheric Physics, Russian Academy of Sciences
  • Thomas Martin
    • Leibniz-Institut für Meereswissenschaften
Article

DOI: 10.1007/s00382-011-1047-8

Cite this article as:
Meng, Q., Latif, M., Park, W. et al. Clim Dyn (2012) 38: 1757. doi:10.1007/s00382-011-1047-8

Abstract

Recent studies indicate a weakening of the Walker Circulation during the twentieth century. Here, we present evidence from an atmospheric general circulation model (AGCM) forced by the history of observed sea surface temperature (SST) that the Walker Circulation may have intensified rather than weakened. Observed Equatorial Indo-Pacific Sector SST since 1870 exhibited a zonally asymmetric evolution: While the eastern part of the Equatorial Pacific showed only a weak warming, or even cooling in one SST dataset, the western part and the Equatorial Indian Ocean exhibited a rather strong warming. This has resulted in an increase of the SST gradient between the Maritime Continent and the eastern part of the Equatorial Pacific, one driving force of the Walker Circulation. The ensemble experiments with the AGCM, with and without time-varying external forcing, suggest that the enhancement of the SST gradient drove an anomalous atmospheric circulation, with an enhancement of both Walker and Hadley Circulation. Anomalously strong precipitation is simulated over the Indian Ocean and anomalously weak precipitation over the western Pacific, with corresponding changes in the surface wind pattern. Some sensitivity to the forcing SST, however, is noticed. The analysis of twentieth century integrations with global climate models driven with observed radiative forcing obtained from the Coupled Model Intercomparison Project (CMIP) database support the link between the SST gradient and Walker Circulation strength. Furthermore, control integrations with the CMIP models indicate the existence of strong internal variability on centennial timescales. The results suggest that a radiatively forced signal in the Walker Circulation during the twentieth century may have been too weak to be detectable.

Keywords

Tropical atmospheric circulationTwentieth century Walker Circulation

1 Introduction

Global Warming is well underway and at least half of the surface warming observed during the twentieth century can be attributed to anthropogenic forcing (IPCC 2007). The global-scale surface air temperature (SAT) exhibits a strong land-sea contrast, an inter-hemispheric asymmetry with stronger warming of the Northern Hemisphere, and a poleward amplification of the warming in the latter, features global climate models successfully simulate. On regional scales, however, climate models do not agree with regard to the SAT response (Meehl et al. 2008). In particular, Latif and Keenlyside (2009) by reviewing the refereed literature describe a large uncertainty in the Tropical Pacific response to Global Warming. They investigated not only the changes in the interannual variability by studying the El Niño/Southern Oscillation (ENSO), but also the changes in the mean state. Concerning the latter, climate models simulate both an El Niño- and a La Niña-like response, and there seems not to be a clear preference for either of them in the ensemble of models. Such behaviour is expected given the large spread of representing mean state in the models, since there is a subtle balance of the processes governing the sea surface temperature (SST) along the Equatorial Pacific. Both the shallow mixed layer depth in the east and the temperature dependence of the nonlinear cloud-albedo feedback, for instance, support an eastward amplification of the Global Warming response. On the other hand, the negative feedback by the mean equatorial upwelling is strongest in the east, which may favour a westward amplification. One can expect the nature of the SST response to have a strong impact on the thermodynamically driven Walker Circulation, one of the most important circulation systems in the Tropics and the focus of this study.

Observational studies also differ strongly among each other even concerning the sign of the change in the SST gradient along the Pacific Equator. Meehl and Washington (1996), for instance, find evidence for an El Niño-like, while Cane et al. (1997) and very recently Karnauskas et al. (2009) for a La Niña-like SST trend pattern during the twentieth century. Furthermore, Xue et al. (2003) report strong interdecadal SST variability in the tropical Indian and Pacific Oceans during the last 150 years and differences between datasets, which hampers the estimation of the twentieth century inter-basin SST gradient trend. Finally, Deser et al. (2010) report a large uncertainty in the SST trends of the equatorial East Pacific during the twentieth century. As different signs in the change of the SST gradient along the Equator may drive completely different atmospheric responses in the Walker and Hadley Circulation, it is of primary importance to obtain further insight into the observed changes.

A number of studies investigated decadal and longer trends in the tropical circulation (e.g., Clarke and Lebedev 1996; Zhang and Song 2006; Power and Smith 2007; Bunge and Clarke 2009; Deser et al. 2010) and most of those have argued that the Walker Circulation has weakened over the twentieth century. In particular, Vecchi et al. (2006) investigated sea level pressure (SLP) observations and results from one climate model driven with observed twentieth century external forcing. They inferred a weakening of the Walker Circulation from both observations and their simulations. This is consistent with the study of Held and Soden (2006) who investigated the effect of radiative forcing on the hydrological cycle from a number of climate change experiments generated for the Fourth Assessment of the Intergovernmental Panel on Climate Change (IPCC). They conclude that by the fundamental measure provided by the average vertical exchange of mass between the boundary layer and the free troposphere, the atmospheric circulation must slow down. On the contrary, Sohn and Park (2010) analyzing water vapour transports described strengthened tropical circulations in the past three decades. Likewise the trend pattern of altimetry sea level during the two recent decades featuring a rather strong increase in the western Equatorial Pacific and even a slight drop in the east implies a strengthened Walker Circulation, as the sea level gradient across the Equator is strongly controlled by the zonal surface winds. Individual tide gauge stations support the steepening of the Equatorial Pacific sea level gradient with high statistical significance from the 1940s onward (http://www.tidesandcurrents.noaa.gov/sltrends/sltrends.shtml). Many processes, however, may they be internal or external, atmospheric or oceanic, control the strength of the Walker Circulation, including changes in the tropical hydrological cycle in response to radiative forcing (see the discussion by DiNezio et al. 2010 and references therein), which partly explains the competing results.

Here we re-investigate the behaviour of the Walker Circulation during the twentieth century using different observational datasets and ensemble integrations with an atmospheric general circulation model (AGCM) forced by the observed history of SST. We conducted also a number of sensitivity experiments in order to test the experimental setup and to investigate the sensitivity to the choice of the SST dataset used to drive the AGCM and to radiative forcing. Finally, we investigate pre-industrial control and twentieth century runs with global climate models in order to obtain further insight into the dynamics of the tropical atmospheric circulation, specifically the link between the Indo-Pacific SST gradient and Walker Circulation strength. Our main focus is the determination of the spatial structure of the long-term trend in tropical SST and the atmospheric response to it during the twentieth century.

In Sect. 2, we describe the data, the statistical methods used in this study, and the AGCM. The results of the statistical analyses of the observed data are presented in Sect. 3. The atmosphere model results are shown in Sect. 4. The outcome of the different sensitivity experiments with the AGCM is described in Sect. 5. The results obtained concerning the link between the SST and SLP gradient in the Indo-Pacific Region in the different AGCM experiments and in the twentieth century integrations with the Coupled Model Intercomparison Project (CMIP) models are summarized in Sect. 6. A comparison of the SST-forced AGCM results with observations is given in Sect. 7. The main conclusions are drawn in Sect. 8.

2 Data, atmosphere model, and statistical method

2.1 Data

We used two observational SST datasets in this study. The first is the ERSST.v3b dataset (Smith et al. 2008; http://www.ncdc.noaa.gov/oa/climate/research/sst/ersstv3.php, in the following ERSST), which was available for the period 1854–2007, providing monthly data on a 2° × 2° grid. These data were used in the statistical analyses described below. The second SST dataset used in this study is HadISST (Rayner et al. 2006; http://www.hadobs.metoffice.com/hadisst/data/download.html) providing monthly data on 1° × 1° grid.

The atmospheric data are described next. We used observed near-surface zonal wind velocity, precipitation and sea level pressure (SLP) data that are provided as timeseries at stations or on regular grids, constructed via simple interpolation or more complex reconstruction techniques (Table 1). Near-surface zonal wind velocities are obtained from International Comprehensive Ocean-Atmosphere Data Set (ICOADS) (http://www.cdc.noaa.gov/data/gridded/data.coads.2deg.html). Observations of precipitation and SLP at stations are provided by NCDC/NOAA GHCN (ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/v2/), in which no reconstruction methods are applied. We used gridded precipitation data of University of East Anglia (Mitchell and Jones 2005; http://www.cru.uea.ac.uk/cru/data/hrg/cru_ts_2.10), NASA/GISS-DAI (Dai et al. 1997; http://www.data.giss.nasa.gov/precip_dai/), and NCEP/NCAR Reanalysis (Kalnay et al. 1996; ftp://ftp.cdc.noaa.gov/Datasets/ncep.reanalysis.derived/surface_gauss). We note, however, that precipitation from re-analysis products is not well constrained by observations. We also used the SLP datasets from ICOADS, Hadley Centre Sea Level Pressure dataset (HadSLP2) (Allan and Ansell 2006; http://www.hadobs.metoffice.com/hadslp2), and NCEP/NCAR Reanalysis (ftp://ftp.cdc.noaa.gov/Datasets/ncep.reanalysis.derived/surface). Gridded data are reconstructed from the station data and ship observations and the methods are described in Table 1. All data are provided as monthly mean values. These were averaged to annual means for calculation of the linear trend. For the station SLP data, we considered the stations with more than 70 years to calculate the linear trend of yearly mean SLP.
Table 1

List of atmospheric observations used in this study

Variables

Name

Period

Space and time intervals

Reconstruction method

Near-surface zonal wind velocity

ICOADS

1800–2007

2° × 2°

Monthly

Box averages

Precipitation

NOAA station

1900–2007

Station

Monthly

No

CRU

1901–2002

0.5° × 0.5°

Monthly

Simple interpolation

GISS-DAI

1850–1995

2.5° × 2.5°

Monthly

Distance-weighted interpolation

NCEP/NCAR

1948-present

T62 Gaussian grid

Monthly

Model reanalysis output

Sea level pressure (SLP)

NOAA station

1900–2002

Stations

Monthly

No

ICOADS

1800–2007

2° × 2°

Monthly

Box averages

HadSLP2

1850–2008

2° × 2°

Monthly

Reduced-space optimal interpolation

NCEP/NCAR

1948-present

2.5° × 2.5°

Monthly

Model reanalysis output

Note that period shown in the table represents the whole period that is provided. Analyzed periods are indicated in figures and text

2.2 Atmosphere model experiments

The atmospheric data, however, are rather sparse both in time and space. In order to derive a consistent atmospheric response to the observed SST variability, we analyzed the output of an AGCM forced by the history of the observed global monthly (HadISST) SSTs and sea ice distributions for the period 1870–2007. Time varying external forcing is not considered in this particular ensemble experiment. Greenhouse gas concentrations are fixed at “present-day” values, with a carbon dioxide concentration of 348 ppm. As described below, several additional integrations were performed in order to test the experimental strategy, the sensitivity to the choice of the SST dataset and to radiative forcing.

The model, ECHAM5 (Roeckner et al. 2003), is a state-of-the-art AGCM and was used with a horizontal resolution of T106, corresponding to a 1.125° × 1.125° grid resolution with 31 vertical levels. Five integrations were performed with different initial, but identical boundary conditions. This allows to distinguish the boundary (SST) forced signal, estimated by the ensemble mean, from the internal (chaotic) variability of the atmosphere. The same model runs were used by Latif et al. (2007) to study the competing effects of warming in the different tropical oceans on Atlantic hurricane activity. In particular, the model successfully simulated the vertical wind shear over the tropical North Atlantic, an important region for tropical storm development. The AGCM also successfully reproduces the decadal fluctuations in Sahel rainfall during the twentieth century (not shown), which provides some additional confidence in the ability of the model to realistically simulate the main tropical circulation systems and their variability.

A relatively large number of sensitivity integrations were performed. Most of them were conducted with a coarse-resolution version of the AGCM (ECHAM5-T31) that employs a horizontal resolution of T31 (3.75° × 3.75°) with 19 vertical levels. First, in order to test the applicability of the uncoupled experimental setup three positive and three negative centennial SST trends were computed from a pre-industrial control simulation with the Max-Planck-Institut für Meteorologie (MPI) climate model obtained from the Coupled Model Intercomparison Project (CMIP3) database, and ECHAM5-T31 was driven with these. Second, in order to investigate the sensitivity of the results to the choice of the SST dataset the long-term SST trend patterns for each calendar month were derived from the two SST datasets (ERSST, HadISST) and added to the SST climatology to drive the coarse-resolution model in equilibrium-response simulations. Third, in order to investigate the influence of the SST trends in the Indian Ocean on the model results, these two runs were repeated with (1) SST trends specified only in the Pacific and Atlantic, and (2) with trends prescribed only in the Indian Ocean. All equilibrium response integrations are each 25 years long, and the last 20 years were used in the subsequent analyses. Furthermore, we repeated the ensemble of transient ECHAM-T106 integrations (with HadISST) using ERSST. Finally, fourth, in order to investigate the role of time-varying external forcing the ECHAM5-T106 ensemble experiments (with HadISST) were repeated with ECHAM5-T31 with and without time-varying external forcing. The latter includes variable greenhouse gas, aerosol, and ozone forcing. The two ensembles consist of seven realizations each.

2.3 Statistical method

Our statistical investigation of the tropical SSTs is based on the Principal Oscillation Pattern (POP) method, which is designed to extract the dominant modes of variability from a multidimensional dataset (Hasselmann 1988; Storch et al. 1988). In contrast to Empirical Orthogonal Function (EOF) analysis that considers only the spatial co-variability in a dataset, the POP method considers the full space–time structure. Furthermore, the POP method is a dynamical approach, in contrast to the purely statistical EOF method. Mathematically, the POPs are the eigenvectors of the system matrix obtained by fitting the data to a multivariate first-order Markov process. The latter can be regarded as a null-hypothesis for internal climate variability (e.g., Hasselmann 1976). POPs are in general complex with real part P1 and imaginary part P2. The corresponding complex coefficient timeseries satisfy the standard damped harmonic oscillator equation, so that the evolution of the system in the two-dimensional POP space spanned by the real and imaginary part can be interpreted as a cyclic sequence of spatial patterns:
$$ \cdots \to P_{1} \to - P_{2} \to - P_{1} \to P_{2} \to P_{1} \to \cdots $$
Complex POPs describe oscillatory modes, propagating or stationary. POPs can be also real. In this case they describe exponentially decaying modes. The characteristic period to complete a full cycle is referred to as rotation period and the e-folding time for exponential decay as damping time. The two timescales are estimated as part of the POP analysis.

One advantage of the POP method relevant here is that no priori information has to be provided about the space–time structure of the long-term SST trend. Prior to the POP analysis, the monthly values were averaged to annual means, and the long-term mean computed from the complete record was removed. No other time filtering was applied to the SST observations. As will be shown below and as expected, the long-term trend is represented by a real POP mode. We note that we applied POP analysis to both SST datasets (ERSST and HadISST) to investigate the robustness of the results. The findings, however, were rather similar and we concentrate here on the results from the POP analysis of HadISST. The time evolutions of the “trend-POPs” obtained from the two different SST datasets, for instance, are rather similar, as shown below.

3 Results of the statistical analyses

3.1 Trend analysis

Linear trends during the twentieth century of observed SST (HadISST and ERSST), precipitation, SLP, and near-surface zonal wind velocity are shown in Fig. 1. The SST trend pattern from HadISST (Fig. 1a) displays a clear strengthening of the Equatorial Pacific SST gradient, with cooling in the eastern and central Equatorial Pacific and warming to west, while the tropical Indian and Atlantic Oceans also exhibit warming. The SST trend pattern derived from ERSST (Fig. 1b) displays some important differences to that derived from HadISST and is dominated by warming everywhere. However, the SST trend along the Equatorial Pacific also exhibits a zonal gradient, with less warming in the eastern and more warming in the western part. The strengthening of the SST gradient, however, is less pronounced than in HadISST.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig1_HTML.gif
Fig. 1

Linear trend patterns of a SST from HadISST, b SST from ERSST, c precipitation of climate research unit (CRU) station data, d SLP from NOAA station data, and e near-surface zonal wind velocity from ICOADS

A cooling trend in the eastern Equatorial Pacific has been previously described by Cane et al. (1997) and Karnauskas et al. (2009) who used different SST datasets. Cane et al. (1997) originally proposed such behaviour on the basis of the so called dynamical thermostat (Clement et al. 1996), in which the mean equatorial upwelling acts as a negative feedback on the SST and thus on Global Warming. We find a similar cooling trend only in HadISST (see Fig. 1a). This indicates a strong sensitivity of the equatorial SST trend pattern during the twentieth century to the choice of the dataset (Vecchi et al. 2008). The existence of an enhanced zonal SST gradient across the Equatorial Pacific, however, appears to be a robust result, and this is important to the following discussion.

The trend pattern in land precipitation (Fig. 1c) is rather patchy. We would like to note the reduction in rainfall over western North Africa, the reduction over western South America and a consistent increase around the Indian Ocean. The SLP trend (Fig. 1d) can be estimated only at rather few stations. Most obvious is the reduction in SLP over India and eastern Asia. The near-surface zonal wind trend pattern (Fig. 1e) is also rather noisy and shows easterly anomalies over most of the Equatorial Pacific and westerly anomalies over the Equatorial Indian Ocean. We note, however, that this trend is computed over a much shorter time period. As will be shown below, the change in SST gradient evolved rather gradually during the twentieth century, which partly justifies this. The Equatorial Pacific SST trend pattern derived from the different SST products suggests that the Walker Circulation may have strengthened, since such a change in the SST gradient would drive a stronger Walker Circulation considering the type of large-scale ocean–atmosphere interactions acting on interannual timescales. On the other hand, as has been shown by Held and Soden (2006), the radiative forcing during the twentieth century may have a competing effect. The trend analyses of the land precipitation, station SLP, and ship-based near-surface zonal wind, however, are inconclusive, but not entirely inconsistent with the picture of a strengthened Walker Circulation.

3.2 POP analysis of observed SST

Next we investigate the observed SST in more detail. The POP analysis of the annual mean SST anomalies from HadISST in the region 20°N–20°S yields two modes that are of importance here. We note that the POP analysis of ERSST yields very similar results. The leading POP mode (Fig. 2) from HadISST accounting for about 30% of the total variance is ENSO. It is complex and describes an almost stationary evolution in the Equatorial Pacific with some indication of westward phase propagation (Fig. 2a, b). The well-known teleconnections to the Tropical Indian Ocean and Tropical Atlantic are also seen in the POP patterns. The rotation period amounts to about 4 years, the decay time to 1 year, which agrees well with other observational estimates of the characteristic ENSO timescales. The two POP coefficient timeseries (Fig. 2c) are highly correlated with standard ENSO indices such as the Niño 3.4 SST anomaly index. The imaginary part of the POP coefficient timeseries describing the “mature” phase, for instance, correlates with the Niño 3.4 index at −0.8. The use of annual means, however, is not optimal to describe the space–time structure of the interannual ENSO phenomenon. This is probably the reason for the finding that the correlations between the two POP timeseries and the Niño 3.4 index are both most significant at zero lag, which was revealed from lag-correlation analyses (not shown). More importantly, the two ENSO POP timeseries do not show any sustained long-term trend or any strong multidecadal variability. Thus the leading POP mode mostly represents the interannual variability in the Tropics associated with ENSO and its decadal modulation. The latter is reflected by the increase of the level of interannual variability from the first to the second half of the twentieth century.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig2_HTML.gif
Fig. 2

Leading (oscillatory) POP mode from HadISST (1872–2007): a Real and b imaginary part, c the POP coefficients. This mode explains 30% of the annual SST anomalies in the tropics (20°S–20°N). The POP analysis was done with annual mean SST anomalies on a 2° × 2° grid

The second most energetic POP mode of HadISST (Fig. 3) is real and accounts for almost 7% of the variance in the whole domain under consideration. Like the linear trend pattern (Fig. 1a), the POP pattern (Fig. 3a) is characterized by a strong zonal gradient in the Pacific, with cooling in the eastern and central part of the Equatorial Pacific and warming in the western part, the tropical Indian and Atlantic Oceans. The corresponding coefficient timeseries (Fig. 3d) is dominated by a downward trend until around 1900 and a gradual upward trend starting thereafter. Thus the POP analysis justifies the use of the linear trend patterns in the following investigation of the long-term climate evolution during the twentieth century. The quality of the SST data, however, can be questioned, as described above. The trend-POP mode obtained from ERSST, for instance, displays no net cooling in the eastern and central Equatorial Pacific (Fig. 3c). What is important to our discussion, however, is the fact that large parts of the eastern half of the Equatorial Pacific warmed less than the western half. Results below indicate that the change in the SST gradient affects the Walker Circulation (i.e., a net cooling of the eastern and central Equatorial Pacific is not necessarily required to strengthen the Walker Circulation).
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig3_HTML.gif
Fig. 3

a Second most energetic (real) POP mode explaining 7% of the variance of the annual SST anomalies in the Tropics (20°S–20°N) of HadISST, b the second EOF mode of HadISST, c the second POP of ERSST, and d their timeseries (HadISST POP, black; EOF, red; ERSST POP, blue)

The question arises whether the trend structure obtained by the POP method is more reliable than that derived using other statistical methods when applied to the same SST dataset. Standard Empirical Orthogonal Function (EOF) analysis yields quite different results and compounds the interannual variability and the long-term trend. The leading EOF (48%) of HadISST is mostly associated with ENSO (not shown), with strong loadings in the Equatorial East Pacific. However, the corresponding principal component displays a long-term trend, on which strong interannual variability is superimposed. The second EOF mode of HadISST accounts for 22% of the variance. It is also ENSO-like in the Pacific. In fact, in the Pacific, the first two EOF modes are mirror images, with the same pattern but with opposite sign. Its pattern (Fig. 3b) shares some similarities to that of the second POP mode (Fig. 3a). However, the meridional extent of the cooling in the Pacific is much larger. Furthermore, the warming in the Indian Ocean is considerably weaker than indicated by the linear SST trend pattern (Fig. 1a). The principal component of the second EOF mode (Fig. 3d) depicts (like that of the leading EOF mode) a long-term trend, but it also displays considerable interannual variability. In contrast, the interannual variability in the timeseries of the second POP mode is much suppressed. All this indicates that the POP method appears to be superior to the EOF method in distinguishing between the interannual (ENSO) variability and the long-term trend in the observed SST. The results of the POP analysis are, however, largely consistent with those of Guan and Nigam (2008) who investigated SST from HadISST by means of rotated extended EOF analysis. This method follows a similar philosophy as POPs, because it also considers the full space–time structure of the variability.

4 Atmosphere model simulations with observed SSTs

We now turn to the changes in atmospheric circulation which go along with the long-term changes in SST, as obtained from the AGCM forced by the history of the observed SSTs. We have chosen HadISST to drive the model, as it exhibits the stronger change in the Equatorial Pacific SST gradient. The sensitivity of the model results to the choice of the SST dataset is discussed further below.

The uncoupled approach assumes that the ocean and atmosphere are strongly coupled in the Equatorial Region, and that the model and driving SST are realistic. It is well accepted that the ocean and atmosphere are tightly coupled in the Equatorial Pacific and Atlantic, at least at interannual timescales. In the case of the Equatorial Indian Ocean, this has been questioned at decadal timescales by Copsey et al. (2006) reporting a clear discrepancy between the observed and AGCM-simulated trends in SLP in the Indian Ocean Region during the second half of the twentieth century. The results presented next, however, indicate that even on centennial scales the tight relationship between the SST and SLP is supported by a rather large number of global climate models obtained from the CMIP database.

The transient ECHAM5-T106 ensemble integrations forced by HadISST using constant radiative forcing yield a detailed picture of the atmospheric anomalies for the period 1870–2007, provided the made assumptions hold. As time-varying radiative forcing is not considered in this particular ensemble experiment, we can obtain only the SST-forced component of the complete response during the twentieth century. The time evolution of the second most energetic POP mode of HadISST is almost linear during the twentieth century (Fig. 3d), which justifies the calculation of linear trends (1900–2007) to derive the centennial-scale changes in atmospheric circulation. As a consistency check, the map of linear trend coefficients (Fig. 1a) for the SST itself which drives the model can be compared with the pattern of the second most energetic POP mode (Fig. 3c), and they are indeed very similar. The explained variances amount to up to 60% in the western Equatorial Indian Ocean and the Tropical Southeast Atlantic (not shown). Not much variance is explained by the SST trend in the Equatorial East Pacific, which is not surprising given the strong interannual (ENSO) variability in this region, which considerably reduces the signal-to-noise ratio.

The model SLP trend pattern (Fig. 4a) is dominated by a wave number one response with mostly enhanced pressure over the Pacific and mostly reduced pressure over the Indian Ocean and the Atlantic, consistent with a strengthening of the Walker Circulation. The explained variances amount to about 40% over the western Equatorial Indian Ocean. Consequently, anomalous easterly near-surface winds are simulated over the western Pacific and eastern Indian Ocean and anomalous near-surface westerly winds over the western Indian Ocean (Fig. 4b). The model rainfall trend pattern (Fig. 4c) displays enhanced precipitation in the region where the two anomalous flow regimes converge in the central Indian Ocean. The explained variances here amount to up to 60%. The SLP and precipitation model patterns over the Indian Ocean are somewhat consistent with the trend patterns obtained from the land observations, which yielded evidence for reduced pressure over India and enhanced rainfall around the Indian Ocean (Fig. 1).
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig4_HTML.gif
Fig. 4

Linear trend pattern 1900–2007 of a SLP, b near-surface wind velocity, and c precipitation simulated by the high-resolution (T106) atmosphere model forced by observed SST (HadISST) using fixed “present-day” radiative forcing. Contours denote explained variances

Similar analyses were performed for the divergence at 200 hPa (Fig. 5a) and the zonally averaged mass streamfunction (Fig. 5b). The upper level divergence trend pattern (Fig. 5a) basically reflects the precipitation pattern (Fig. 4c) with enhanced divergence over the Indian Ocean and reduced divergence over the Maritime Continent. All this indicates that the high-resolution AGCM forced by HadISST simulates a reproducible intensification of the Walker Circulation during the twentieth century. The trend pattern of the zonal-mean mass streamfunction (Fig. 5b) indicates also an intensification of the Hadley Circulation as it exhibits similarities to climatology. These above described changes are rather insensitive to the choice of model horizontal resolution and to radiative forcing, as described next. There is, however, sensitivity to the choice of the forcing SST. The sign of the change, however, remains unchanged.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig5_HTML.gif
Fig. 5

Linear trend pattern 1900–2007 of a divergence deviation from zonal-mean at 200 hPa and b zonal-mean mass stream function from the high-resolution (T106) atmosphere model forced by observed SST (HadISST) using fixed “present-day” radiative forcing. Contours denote explained variances

5 Sensitivity experiments

Several assumptions are made when deriving the atmospheric circulation changes during the twentieth century from an AGCM which was run in forced mode with an estimate of the observed boundary conditions. We therefore performed a number of sensitivity experiments and investigate in the following the sensitivity of the results to the experimental setup, forcing SST dataset, and radiative forcing. Most of the sensitivity experiments were conducted with the coarse-resolution model ECHAM5-T31. All experiments described in this study are presented in terms of the changes in Equatorial Pacific SST and SLP gradient in a scatter diagram (Fig. 6a) to ease the interpretation of the large-scale changes. Furthermore, the results of different types of experiments are shown separately in additional panels (Fig. 6b–d). The boxes chosen to compute the gradients are adopted from Vecchi et al. (2006) who used them to obtain a surrogate for the Walker Circulation strength: The data were averaged over the two boxes 80°–160°E, 5°N–5°S and 160°–80°W, 5°N–5°S to define the SST and SLP differences. We note that the SLP gradient has been inverted (east minus west), so that a value in the upper two quadrants indicates a strengthened Walker Circulation. Land points were not removed in the computation of the temperature in the western box. The results, however, are virtually unchanged when the land points are excluded (not shown). The mean of the HadISST forced ECHAM5-T106 ensemble experiment described above amounts to 0.8 hPa/century and is denoted by the black pentagram in Fig. 6a and b. The spread is relatively small: The changes (in hPa/century) obtained from the individual five realizations amount to: 0.776, 0.811, 0.798, 0.782 and 0.876.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig6_HTML.gif
Fig. 6

Linear trend of the surface temperature (Tsurf) and SLP contrast along the Equator in a all experiments (coupled and uncoupled), b forced AGCM runs, c climate of the twentieth century experiments (20c3m), and d the pre-industrial control experiments (picntrl). Small symbols denote picntrl, and big symbols 20c3m in a. The black line in a is the linear regression of Tsurf trend and SLP trend based on picntrl. See text for details

5.1 Sensitivity to experimental setup

Copsey et al. (2006) have suggested that the response of some atmospheric AGCMs to the Indian Ocean warming may not provide a reliable guide to the behavior of the real world. A considerable part of Indian Ocean SST variability may be driven by the atmosphere and not by the ocean, so that the experimental setup of driving an AGCM with observed SSTs can be questioned. Schneider and Fan (2007) demonstrated that the SST forced AGCM experiments are completely consistent, as long as it is recognized that they are giving the atmospheric feedback to the SST, and are missing the “weather noise” part of the actual evolution. We tested the experimental setup by conducting equilibrium-response integrations with the coarse-resolution (T31) version of ECHAM5. To this end we chose three positive and three negative centennial trends in SST contrast (one of the positive is shown in Fig. 7a) simulated in a pre-industrial control run by one particular global climate model (labeled mpi echam5 in Fig. 6a), and forced the ECHAM5-T31 model by the corresponding trend patterns in stand-alone mode. Only the response patterns from the experiment with the SST anomaly pattern (labeled mpi uncoupled 3 and displayed as bold red star in Fig. 6a) are displayed here (Fig. 7). The SST gradient change in this experiment is closest to the observed change. The results of the other five sensitivity experiments are shown in terms of the two gradient indices as bold stars in Fig. 6a (labeled mpi uncoupled 1 and mpi uncoupled 2), together with the indices obtained from the original coupled model integration (labeled mpi coupled 1, mpi coupled 2, and mpi coupled 3). We note that we used only three numbers in Fig. 6a to distinguish the six experiments, because positive and negative SST gradient changes are well separated in the scatter diagram.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig7_HTML.gif
Fig. 7

Sensitivity experiment with the coarse-resolution (T31) atmosphere model (ECHAM5-T31) driven with SST forcing from the MPI coupled model: a 100-year SST trend pattern from the coupled model used as forcing, b SLP and near-surface wind velocity responses of ECHAM5-T31, and c SLP and near-surface wind velocity responses simulated by the original coupled model

The uncoupled results are entirely consistent with those from the coupled model. This can be inferred from the symbols forming pairs (uncoupled/coupled) being reasonably close to each other in the scatter diagram (Fig. 6), which provides some confidence in our uncoupled experimental setup. Figure 6a provide all cases that are used in the present study, and Fig. 6c–d for individual cases that will be discussed later. The SLP and near-surface wind velocity responses indicate clear changes in the Walker Circulation (Fig. 7b), similar to those in the original coupled model simulation (Fig. 7c). We would like to point out, however, two caveats. First, there are noticeable differences between the patterns simulated in forced mode and in the coupled model. As suggested by Fig. 6a, however, the large-scale structures are rather similar. Also, the SST patterns exhibiting a positive change in the SST gradient are not that similar to the observed twentieth century trend pattern, with relatively strong changes in the model found in the Equatorial Pacific (Fig. 7a), where we know ocean and atmosphere are tightly coupled. Thus the wind and SLP patterns are quite different to those simulated using observed SSTs (Fig. 4), but the results basically agree in terms of the two gradient indices (Fig. 6a). Second, the mechanism in the MPI coupled model that produces the SST anomalies in the Indian Ocean could be different to that in reality, which would enable a reproduction of the coupled model results in uncoupled mode with prescribed SST. Further investigation of this issue will be the topic of future research and is in our view beyond the scope of this paper.

5.2 Sensitivity to forcing SST

We next investigated the sensitivity of the ECHAM5 model to two estimates of the observed SST variability during the twentieth century. This was performed in transient and equilibrium response experiments. Monthly linear trend patterns were computed first from both SST datasets (HadISST and ERSST) and used to force the coarse-resolution (T31) version of the AGCM. The linear trend patterns are shown in Fig. 8 (ERSST, Fig. 8a; HadISST, Fig. 8b). The annual mean SLP and near surface wind response (Fig. 8c, d) is similar in the two integrations, especially concerning the inter-basin SLP gradient between the Pacific and the Indian Ocean. Differences are seen, however, over the Pacific that are also reflected in the near-surface zonal wind response. Anomalous westerlies over the western Equatorial Pacific are simulated when the ERSST SST trend drove the model and anomalous easterlies when using the SST trend from HadISST. Both forced simulations, however, yield an intensification of the Walker Circulation using the definition applied in Fig. 6 and in previous studies (Vecchi et al. 2006), which can be readily seen in Fig. 6b (displaying only the SST-forced ECHAM5 results) as blue (HadISST) and red pentagrams (ERSST). The intensification in the transient and equilibrium response experiments with HadISST, however, is considerably stronger than those in the runs with ERSST. Obviously, there is some sensitivity of the model results to the choice of the SST forcing. Yet the sign of the Walker Circulation change is robust and independent of the forcing SST used.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig8_HTML.gif
Fig. 8

Sensitivity experiment with the coarse-resolution (T31) atmosphere model: a SST forcing from ERSST, b SST forcing from HadISST, c SLP and near-surface wind velocity response simulated with ERSST forcing, and d SLP and near-surface wind velocity responses simulated with HadISST forcing

What is the role of the Indian Ocean SST trend in our model? In order to isolate the effect of the Indian Ocean on the Walker Circulation, we repeated the two ensemble integrations shown in Fig. 8 with (1) the SST trend restricted only to the Pacific and Atlantic (Fig. 9) and (2) only to the Indian Ocean (Fig. 10). The results (given by the four crosses in Fig. 6b) highlight the prominent role that the Indian Ocean SST plays in our model. The Walker Circulation weakens, when the SST forcing is restricted only to the Pacific and Atlantic (Figs. 6b, 9c, d) independent of the SST forcing dataset, in contrast to the two full SST forcing cases where the Walker Circulation strengthens (Figs. 6b, 8). However, the absolute change of the Walker Circulation is rather small when we use HadISST (red cross in Fig. 6b), whereas the relative change compared to the full forcing case (red pentagram) is similar to that in the ERSST run (corresponding blue symbols).
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig9_HTML.gif
Fig. 9

As Fig. 8, but with no SST forcing in the tropical Indian Ocean

https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig10_HTML.gif
Fig. 10

As Fig. 8, but with SST forcing only in the Indian Ocean

When the SST forcing is restricted only to the Indian Ocean (Fig. 10), the model simulates a strengthened Walker Circulation in both experiments (black and light blue crosses in Fig. 6b). We find a strong sensitivity of the SLP over the Indian Ocean: While the model simulates a high pressure anomaly in the case with no Indian Ocean SST forcing (Fig. 9c, d), it simulates a low pressure anomaly when the forcing is restricted only to the Indian Ocean (Fig. 10c, d). These results show that the Indian Ocean SST basically determines the sign of the Walker Circulation response in our model.

5.3 Sensitivity to radiative forcing

The last set of sensitivity experiments concerns the role of time-varying radiative forcing. First, we repeated the transient ECHAM5-T106 integrations without time-varying external forcing (with HadISST SST forcing) with the ECHAM5-T31 version. In a second ensemble experiment with the coarse-resolution model, we included time-varying external forcing as observed. All known radiative forcing was used to force the model according to the CMIP3 protocol. The results (Figs. 6a, b, 11a, b) indicate the inclusion of radiative forcing does indeed have an effect: There is a weakening of the negative anomalies over the Indian Ocean and far western Pacific (Fig. 11b) compared to the run without time-varying external forcing (Fig. 11a). However, there is a compensating effect over the central Equatorial Pacific, where the positive SLP anomalies strengthen. The reduction in the (inverted) SLP gradient is therefore modest, and the two entries in Fig. 6b overlap each other. There is, however, some difference between the transient and equilibrium response experiments of the order of about 10%, with less SLP gradient change in transient mode. Finally, the two SLP patterns are consistent with the trend pattern obtained from the transient ECHAM5-T106 integration (Fig. 4a). In summary, the changes in the SST gradient clearly dominate the atmospheric response in our set of uncoupled numerical experiments, while the effect of the radiative forcing is relatively small.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig11_HTML.gif
Fig. 11

Transient experiments with the coarse-resolution (T31) atmosphere model: a SLP trend pattern 1900–2007 with fixed “present-day” radiative forcing and b SLP trend pattern 1900–2007 with time-varying (observed) radiative forcing. Both experiments were forced by SST from HadISST

6 Relationship between SST and SLP changes

We additionally investigate now the behaviour of a number of atmosphere–ocean coupled general circulation models, to assess the performance of AGCM ECHAM5 in forced mode and to study the link between the gradients in SST and SLP in the Equatorial Pacific Sector. We computed 100-year trends of SST differences and corresponding SLP differences [using the definition of Vecchi et al. (2006) from various twentieth century integrations (20C3M) with global climate models forced by all known external forcing and from pre-industrial control runs (PICNTRL)]. The data were obtained from the CMIP3 database. We computed trends for the entire twentieth century from the 20C3M-ensemble, and from the pre-industrial control runs we chose the strongest positive and negative 100-year SST trend in each simulation and the corresponding SLP trends.

All analyzed climate model simulations are shown together with all forced AGCM experiments in a scatter diagram (Fig. 6a) with SST difference and (inverted) SLP difference as axes. The trend line in this summarizing scatter plot is adopted from Fig. 6d which shows the results from the unforced (pre-industrial) control runs with the coupled models. The forced AGCM integrations alone are presented in Fig. 6b. All twentieth century integrations with the coupled models are shown separately in Fig. 6c and all pre-industrial control runs with the coupled models separately in Fig. 6d. The SLP gradient change from the transient HadISST-forced ECHAM-T106 experiment amounts to about 0.8 hPa/century (Fig. 6a, b). This constitutes the strongest change in terms of the SLP gradient, which is not surprising given the rather strong SST gradient change. The major result, however, is the existence of a consistent linear relationship: An anomalously strong SST gradient goes along with an anomalously strong (inverted) SLP gradient, and this is independent of the type of model experiment that is considered. This finding indicates that the SST gradient is a good measure of the strength of the Walker Circulation during the twentieth century in virtually all climate models, independent of whether or not time-varying external forcing is included. Even the model (gfdl cm 2.1) that could explain the observed twentieth century SLP gradient change (Vecchi et al. 2006) supports the SST/SLP relationship. The SST gradient change in this model is inconsistent with observations (blue square in Fig. 6c), suggesting a stronger role of SST and that external forcing may have been less important than stated by Vecchi et al. (2006). In fact only one (hadgem1, light blue diamond in Fig. 6a) of the 24 analyzed twentieth century integrations is somewhat consistent with the changes inferred from the gridded datasets shown as black and red hexagons in Fig. 6a. This indicates that there is no externally driven signal in the Walker Circulation during the twentieth century which can be robustly simulated with global climate models when prescribing observed radiative forcing.

We note that the SST gradient change obtained from HadISST is considerably larger than that from ERSST, and also exceeds those simulated in the 20C3M integrations with the CMIP3 models. Yet the magnitude of the change can be simulated internally, as shown in Fig. 6d displaying the results from the pre-industrial control runs. We further note a sensitivity of the SST trends to the horizontal model resolution. This is due to the box averages used in the calculation of the SST gradients that are not exactly the same at the two used horizontal resolutions.

The trends derived from the gridded SLP datasets (HadSLP2 and ICOADS) are inconsistent with those simulated by the vast majority of the global climate models in twentieth century simulations (Fig. 6c), as noted above. On the other hand, the NCEP estimate (blue hexagon in Fig. 6a) is consistent with the models, which may have been expected as NCEP is also a model-based estimate. Moreover, the results from the pre-industrial control integrations (Fig. 6d) display a relatively large internal variability on centennial timescales in the two indices. All this may suggest that that a radiatively forced signal in the Walker Circulation strength remains too weak to be detected. The trend line in Fig. 6c, however, is shifted downward relative to that obtained from the (unforced) control runs (Fig. 6d), which suggests some influence of the radiative forcing along the lines suggested by Held and Soden (2006).

7 Comparison with data

Our SST-forced AGCM results are inconsistent with different observational datasets, as described above in terms of the simple gradient indices. We now compare the simulated trend patterns. However, we consider in the following only the trends during the second half of the twentieth century, because the trend in SST seems to be relatively steady throughout the twentieth century and we expect the data quality to increase with time. There are, however, some differences to the model patterns obtained from the full twentieth century, which can seen by comparing the trends to those shown in Fig. 4. The drop in SLP over the Indian Ocean simulated by the AGCM ECHAM5-106 (Fig. 12a) and in all other forced AGCM integrations (with different forcing SST, in equilibrium response and transient modes, at different horizontal resolutions, and with and without time-varying external forcing) is seen at some stations over India (Fig. 12b) but not as a large-scale feature in three gridded datasets (ICOADS, HadSLP2, and NCEP). These indicate anomalously high pressure over the Indian Ocean (Fig. 12c–e). In fact the SLP gradient obtained from ICOADS and HadSLP2 is positive (Fig. 6a) as opposed to negative as simulated in our forced AGCM experiments (we note again the different definition east minus west in Fig. 6). At least the direction of the zonal pressure gradient at the Equator between the Indian Ocean and the Pacific seems to be consistent with that simulated by the AGCM and in NCEP. We note, however, the large differences between the different datasets. Even the sign of the trends often do not agree, which is most pronounced over the Pacific where opposite trends are found over large regions.
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig12_HTML.gif
Fig. 12

Linear trend pattern of annual mean SLP from a ECHAM5-T106, b NOAA station observations, c ICOADS observations, d Hadley Centre data, and e NCEP

A similar situation exists for the different observational precipitation datasets (NOAA station precipitation, GISS-DAI, CRU station precipitation, and NCEP), the trend patterns of which are shown in Fig. 13. There is a large uncertainty concerning the precipitation trends, which inhibits definitive statements from data about a long-term change in the Walker Circulation. The station data (Fig. 13b–d), however, tend to agree better with each other than with NCEP. The trend in NCEP precipitation (Fig. 13e) during the second half of the twentieth century agrees reasonably well with that simulated by the AGCM (Fig. 13a) in the sense that it implies a strengthening of the Walker Circulation, with enhanced rainfall over large parts of the Equatorial Indian Ocean and parts of the Maritime Continent, and reduced precipitation over the Equatorial Western and Central Pacific. However, NCEP precipitation is not well constrained by observations and mostly reflects the model’s response to observed SST. Large uncertainties are also found in different near-surface wind products (not shown).
https://static-content.springer.com/image/art%3A10.1007%2Fs00382-011-1047-8/MediaObjects/382_2011_1047_Fig13_HTML.gif
Fig. 13

Linear trend pattern of annual mean precipitation from a ECHAM5-T106, b NOAA station data, c GISS-DAI observations, d climate research unit station data, and e NCEP

8 Conclusions

We have shown that the evolution of SST in the Equatorial Region during the twentieth century is zonally asymmetric with a relatively steady increase of the zonal SST contrast between the Maritme Continent and the eastern and central Equatorial Pacific. Although there are large uncertainties in the observational datasets concerning SST trends during the twentieth century, the strengthening of the large-scale SST gradient seems to be robust. The enhanced SST gradient may have strengthened the Walker Circulation during the twentieth century, which was demonstrated by SST-forced experiments with the ECHAM5 AGCM.

We note a sensitivity of the AGCM to the forcing SST dataset: The HadISST forcing is characterized by a stronger SST gradient change and yields a stronger intensification of the Walker Circulation than the ERSST forcing. The changes in the SST gradient force near-surface wind, precipitation and upper air circulation anomalies over the Equatorial Pacific and the Equatorial Indian Ocean in both experiments that are entirely consistent with the picture of a strengthened Walker Circulation during the twentieth century. Furthermore, the model integrations also suggest an intensification of the Hadley Circulation.

We find that the model results are strongly influenced by the warming of the Indian Ocean. Atmosphere models may suffer from serious problems in simulating the response to SST variations in warm pool regions, and we cannot rule out the possibility that our model results are flawed. The uncoupled experimental setup, however, was carefully tested and proven reliable within a coupled model framework. Moreover, the forced AGCM used in this study exhibits similar response characteristics to those simulated by a number of global climate models in twentieth century simulations with observed radiative forcing or in pre-industrial control integrations. In particular, the role of time-varying external forcing in driving changes of the Walker Circulation is found to be weak in both our uncoupled AGCM simulations and twentieth century integrations with the coupled models. Aside from that the internal variability simulated by the coupled models is relatively large, even on multidecadal and longer timescales. This makes detection of an externally driven signal in the Walker Circulation a challenge.

With the caveat in mind that climate models may suffer from large biases and given the extremely large uncertainties in the various observational datasets, any conclusion about the behaviour of the Walker Circulation during the twentieth century is subject to large uncertainties and therefore more or less hypothetical. Many twentieth century integrations cluster around zero change and internal variability in control integrations is rather large. Only one of the 24 analyzed twentieth century integrations is consistent with the changes inferred from the gridded datasets, although its sea level pressure response is weaker and the magnitude of the change could be internally simulated.

Acknowledgments

This work was supported by the DFG-supported project SFB 754 (www.sfb754.de), the TROIA Project of the Deutsche Forschungsgemeinschaft, and by the European Project ENSEMBLES. Qingjia Meng was supported by the China Scholarship Council. The model runs were performed at the North-German Supercomputing Alliance (HLRN). This is a contribution to the Excellence Cluster “The Future Ocean” (www.ozean-der-zukunft.de).

Copyright information

© Springer-Verlag 2011