Volume 30, Issue 10
Free Access

Tropical Atlantic seasonal predictability: The roles of El Niño remote influence and thermodynamic air-sea feedback

Ping Chang

Ping Chang

Department of Oceanography, Texas A&M University, College Station, Texas, USA

Search for more papers by this author
R. Saravanan

R. Saravanan

National Center for Atmospheric Research, Boulder, Colorado, USA

Search for more papers by this author
Link Ji

Link Ji

Department of Oceanography, Texas A&M University, College Station, Texas, USA

Search for more papers by this author
First published: 20 May 2003
Citations: 36

Abstract

[1] Recent studies suggest that the thermodynamic feedback and the remote influence of El Niño Southern Oscillation (ENSO) are two dominant factors affecting climate variability in the tropical Atlantic sector. Given that both of these processes are included in an atmospheric general circulation model (AGCM) coupled to a mixed layer ocean (ML), such a simple coupled system is expected to possess useful skills in predicting regional climate variability on seasonal time scales. This letter reports some preliminary results from a set of prediction experiments using a coupled AGCM-ML model. These results show promise for prospects of seasonal climate prediction in the tropical Atlantic sector using the simple dynamical seasonal prediction system.

1. Introduction

[2] A number of well known climatic phenomena in the tropical Atlantic sector, such as Nordeste rainfall and subsaharan drought, have long been regarded as potentially predictable climate signals on seasonal or longer time scales (see Hastenrath [1995] for an historic account). Various statistical approaches have been developed for seasonal forecasts of these phenomena. Useful forecast skill has been reported, particularly in the seasonal prediction of Nordeste rainfall [Hastenrath and Greishar, 1994]. With the recent advancement in climate models, it has become increasingly evident that atmospheric general circulation models possess high skill for seasonal climate forecasts in the tropical Atlantic sector provided that an accurate prediction of sea surface temperature (SST) is available [Goddard and Mason, 2002]. However, predicting SST anomalies in the tropical Atlantic on seasonal time scales remains a challenge. At present, routine SST forecasts in the tropical Atlantic are largely based on statistical approaches [Penland and Matrosova, 1998]. Fully coupled climate models have not yet demonstrated useful skills in predicting tropical Atlantic SST.

[3] In this study, we take an alternative approach to fully coupled climate models for examining predictability in the tropical Atlantic sector. The coupled model we use has full atmospheric physics, but the simplest ocean physics—a slab ocean mixed layer. We refer to this type of coupled system as an AGCM-ML model.

[4] Following a brief description of the coupled model and prediction experiments in section 2, major results of the prediction experiments are presented in section 3. In section 4 we discuss the predictable dynamics in the coupled model and its potential as a dynamical seasonal prediction system for Tropical Atlantic Variability (TAV).

2. Coupled Model and Prediction Experiments

[5] The AGCM-ML model used in this study consists of CCM3, version 3.6.6 of the Community Climate Model developed at the National Center for Atmospheric Research (NCAR) [Kiehl et al., 1998] and a simple ocean mixed layer with a spatially varying depth based on the observed annual mean mixed layer depth [Levitus, 1994]. The model has a standard resolution—triangular truncation at wavenumber 42 in horizontal and 19 vertical levels. A “q-flux” formulation is used to correct for the absence of ocean dynamics, so that the simulated SST annual cycle is in good agreement with observations.

[6] Two sets of forecast experiments are conducted. In the first set, the mixed layer ocean is initialized with the December monthly mean SST everywhere in the global ocean during the period 1981 to 1996. The coupled CCM3-ML model is then integrated forward for 9 months with ten slightly different initial conditions for the atmosphere, making for a total of 160 forecasts. These initial conditions are derived from the National Centers for Environmental Prediction (NCEP) reanalysis around December 15th of each of the 16 years. The second set is identical to the first except that observed SST values are used to initialize the mixed layer ocean only in the Atlantic Ocean between 30°S and 60°N. Outside the Atlantic the SST anomaly is set to zero everywhere.

3. Results

[7] Each set of the prediction experiments contains an ensemble of 10 × 16 monthly time series at a given lead time. The skill of the model is obtained by validating the ensemble mean of the prediction against the corresponding observations at each lead time. We focus on two important climate variables—SST and precipitation. The observed SST anomaly is derived from the reconstructed dataset of Smith et al. [1996] and the precipitation anomaly is derived from the dataset of Xie and Arkin [1996].

[8] Figures 1a and 1c show the correlation between the predicted (with the global SST initial condition) and the observed SST anomaly at a lead time of 4 months and the persistence of the observed SST anomaly in the tropical Atlantic domain. It is evident that the correlation skill of the coupled model is superior to the skill of the persistence forecast everywhere in the tropical Atlantic except along the equatorial cold tongue. The coupled model is particularly skillful in forecasting the SST anomaly in the North Tropical Atlantic (NTA) and the Caribbean region, with some correlation values exceeding 0.75 at the 95% significance level. Many earlier observational [Enfield and Mayer, 1997] and modeling [Saravanan and Chang, 2000; Alexander and Scott, 2002] studies have shown that the ENSO influence on the tropical Atlantic is strongest in this region, leading us to expect enhanced predictability in this region as well. To further examine the forecast skill in this area, we define an SST index by taking an area average in the western NTA region (indicated by the stippling in Figure 1a). We then calculated the correlation between the predicted and observed time series as a function of lead times (solid line in Figure 1d) and compared them with persistence (dotted line in Figure 1d). The coupled model retains a correlation skill of above 0.75 up to a lead time of 5 months, which is significantly higher than persistence.

Details are in the caption following the image
Correlations between the observed and predicted anomalies of SST and precipitation. (a) correlation between observed and predicted SST at 4-month lead time using global SST initial condition. Contour interval is 0.25, and shading indicates correlations significant at the 95% confidence level. (b) Same as (a) except Atlantic-only SST initial condition is used in the prediction. (c) Skill of persistence forecast at 4-month lead time. (d) correlation between the observed and predicted NTA SST indices: solid - prediction with global SST initial condition; gray - prediction with Atlantic-only SST initial condition; dotted - persistence forecast based on observed SST; dash-dotted - persistence forecast based on a 100-year coupled CCM3-ML simulation. The NTA SST index is derived by taking an area average over the region indicated in (a). (e) correlation between the observed and predicted Nordeste rainfall: solid - prediction with global SST initial condition; gray - prediction with Atlantic-only SST initial condition; dotted - persistence forecast based on Xie-Arkin observation; dashed - correlation between observed and simulated Nordeste rainfall based on an ensemble of CCM3 GOGA runs.

[9] To examine whether the model possesses any useful skill in predicting rainfall variability, we focus on rainfall anomalies in the Nordeste region. A precipitation index is derived by taking an area average of the precipitation anomaly in the Nordeste region as indicated in Figure 1a. First we calculated the potential predictability by correlating the observed Nordeste rainfall anomaly with the simulated anomaly taken from an ensemble of CCM3 GOGA runs where the model is forced by the observed SST everywhere in the global oceans. The correlation (dashed line in Figure 1e) retains an average value above 0.70 during the rainy season—March to May. Theoretically, it gives an upper limit of the model forecast skill because perfect SST values are used. The correlation skill of the coupled model (solid line in Figure 1e) is generally lower than the potential predictability. However, for March and April when the rainy season reaches its peak (which corresponds to lead time 3–4 months), the coupled model has a correlation skill of 0.6, which is statistically significant at the 95% level and is approaching the potential predictability limit. Note that the persistence forecast based upon the observed December rainfall anomaly (dotted line in Figure 1e) has essentially no skill at these lead times.

[10] To investigate the role of ENSO in seasonal prediction of TAV, we performed a similar analysis for the ensemble of predictions using Atlantic-only SST initial conditions, which excludes the direct influence of ENSO. Figure 1b shows the anomalous correlation for the predicted SST at a 4-month lead time (April). In comparison to the case with global SST initial condition, the model forecast skill is generally reduced, but is still significantly higher than the persistence forecast over much of the NTA region. The correlation of the NTA SST index (gray in Figure 1d) is above 0.5 up to a 5-month lead time, which is statistically significant at the 95% level. The decrease in model skill is particularly noticeable in the western NTA and Caribbean where the anomaly correlation drops from above 0.75 to about 0.5. However, in the deep tropics the model skill remains essentially unchanged. Perhaps it is for this reason that there is little decrease in the model's rainfall prediction skill. Even in the absence of an ENSO signal in the tropical Pacific, the coupled model can retain a useful correlation skill of above 0.5 at a 3–4 month lead time. This result suggests that Nordeste rainfall variability is determined, to a large extent, by the SST condition in the deep tropics of Atlantic Ocean.

[11] Figure 2 further illustrates the role of ENSO in predicting NTA SST variability. It shows the 4-month lead time predicted (solid) and observed April SST (dashed) time series for the NTA area as defined in Figure 1a. Superimposed on these time series are the SST time series of each ensemble member. Major El Niño events during this period are indicated by the shading. It is interesting to note that while the predicted SST time series agree well with the observations in the global initial prediction experiments, large discrepancies between the predictions and observations occur during all the major El Niño events, particularly the 82–83 and 86–87 events, when SST initial conditions are absent in the tropical Pacific. Without the remote influence from the Pacific ENSO, the predicted NTA SST values are consistently too weak and the model fails to capture the warming that takes place in the western NTA/Caribbean region in the Spring following major El Niño events. This result is entirely consistent with the well-known effect of ENSO on TAV [Enfield and Mayer, 1997; Penland and Matrosova, 1998].

Details are in the caption following the image
The observed (dashed) and 4-month lead time predicted (solid) NTA SST index (ensemble-average). (a) The ensemble of predictions with global SST initial condition. (b) The ensemble of predictions with Atlantic-only initial condition. Thin lines represent each ensemble member. Shaded vertical bar indicates major El Niño events during the time period.

4. Summary and Discussion

[12] The finding that the coupled CCM3-ML model shows high skill in the NTA region poses an intriguing question—what dynamic processes are responsible for the high predictability? An important contributor to the model forecast skill is the remote influence of ENSO [Enfield and Mayer, 1997], which manifests itself in SST anomalies in the NTA region with a 2–4 month delay associated with the mixed layer response to surface heat fluxes. Even if the slab ocean model were to simulate the evolution (or persistence) of the Pacific ENSO anomalies with some accuracy only over a 3 month period, it could still provide useful SST forecasts for up to 6 month lead times in the NTA region due to this delay. However, the remote influence of ENSO explains only part of the forecast skill, because significant skill is obtained even in the absence of ENSO anomalies in the Pacific, as seen in the Atlantic-only SST initial condition experiments. This means that we need to explain why the model is able to persist SST anomalies in the NTA region with realistic amplitudes even in the absence of the remote ENSO influence.

[13] In the simple coupled mixed layer model, the SST is entirely driven by atmospheric surface heat fluxes. The predictability of the system is determined by the persistence which is proportional to the mixed layer depth [Barsugli and Battisti, 1998; Bretherton and Battisti, 2000]. In the western NTA region, the average mixed layer depth is in the range of 10–20 meters, and is among the shallowest in the oceans (see Figure 2 of Saravanan and Chang [1999]). The shallowness of the mixed layer depth gives a relatively short decorrelation time scale for the SST anomaly. Indeed, we found, by computing the autocorrelation of the SST anomaly simulated by the coupled CCM3-ML in a 100-year control run, that the NTA is one of the regions in the global oceans where the SST persistence is very low. The simulated persistence from the 100-year CCM3-ML control run is shown in Figure 1d (dash-dotted), which is somewhat lower than the observed persistence (dotted line in Figure 1d) on timescales shorter than three months, but compares well with the observed persistence for longer time scales. The lower simulated persistence for short timescales could be attributed to the lack of ENSO and its remote influence in the CCM3-ML control run. Clearly, the model forecast skill is not tempered artificially by any exaggerated SST persistence in the simple coupled system.

[14] We therefore hypothesize that the high persistence of NTA SST anomalies is attributable, to a large extent, to the active air-sea feedback between surface heat flux and SST, in addition to the remote influence of ENSO. The active thermodynamic feedback, such as the Wind-Evaporation-SST feedback [Chang et al., 2000], can contribute to the enhanced predictability by 1) reducing the thermal damping and 2) introducing nonlocal effects into the coupled system so that the behavior of an AGCM-ML model can not be simply described by a 1-D mixed layer model.

[15] The correlation analysis shown in Figure 1 indicates that in the equatorial cold tongue and southeastern Atlantic region, the model's skill is generally worse than persistence. This seems to be a good indication that ocean dynamics, which is entirely absent from the coupled mixed layer system, may be an important contributing factor in the predictable dynamics in this region. Zebiak [1993] shows that a coupled mode akin to ENSO in the Pacific operates in the equatorial Atlantic. The dynamics of this coupled mode depend on subsurface ocean adjustment related to equatorial trapped waves. This coupled mode, albeit much weaker than its Pacific counterpart, is expected to contribute to the persistence and predictability of the SST in reality. Since these physics are absent in the coupled CCM3-ML model, there is little reason to expect that the model forecast would be superior to the persistence forecast.

[16] Potential advantages of using such a simple coupled system for seasonal climate prediction are: 1) Since it is a dynamically based system, it does not need to be trained using observational data and is therefore less vulnerable to statistical artifacts. 2) The only crucial variable for initializing the forecast is SST, which arguably is the most accurately observed climate variable. 3) Perhaps most importantly, the coupled AGCM-ML not only provides SST forecasts on seasonal timescale, but also provides forecasts of other important climate variables, such as precipitation, at mospheric temperature, etc. In this sense, the model can be viewed as the simplest tier-one global dynamical seasonal prediction (DSP) system whose skill can be used to benchmark a hierarchy of other, more sophisticated, DSP systems.

Acknowledgments

[17] This work is supported through NSF grant ATM-99007625 and NOAA grant NA16GP1572.