Potential predictability of tropical Indian Ocean SST anomalies

1. Introduction

[2] Prospects for forecasting SSTAs associated with the IDM, in which the zonal gradient in SST over the tropical Indian Ocean see–saws resulting in occasions of severe drought over Indonesia and flooding over east Africa [Saji et al., 1999], using a coupled dynamical forecast system appear hopeful at lead times of a season or so, as demonstrated by the NASA Seasonal–to–Interannual Prediction Project (NSIPP) system [Wajsowicz, 2004]. Herein, seasonal variations in forecast skill and potential predictability of interannual SSTAs are investigated using observations (NCEP analyses v.2, Reynolds et al. [2002], for 1982 to 2003 inclusive) and the recently completed two decades of retrospective forecasts from the NCEP CFS [S. Saha et al., The NCEP climate forecast system, submitted to Journal of Climate, 2005, hereinafter referred to as Saha et al., submitted manuscript, 2005]. Skill is simply measured by the anomaly correlation coefficient (ACC) between forecast and observed SSTAs averaged over the east (90°E–110°E, 10°S–0°) and west (50°W–70°W, 10°S–10°N) IDM poles defined by Saji et al. [1999].

[3] There is a marked drop in forecast skill of equatorial Pacific SSTAs in boreal spring in many coupled systems and a corresponding rapid decline in lagged correlations of El Nino-Southern Oscillation (ENSO) indices. Processes responsible for the decline (persistence barrier) are thought to create a ‘springtime predictability barrier’ [Webster and Yang, 1992]. For the eastern tropical Indian Ocean (Figure 1), a forecast of persistence is skillful (ACC > 0.6) for lead times of 3–4 months from late boreal winter to early fall. The ACC falls off rapidly in December. However, its magnitude increases again through boreal spring to give a peak ACC of −0.5 for persisted boreal fall anomalies. Further analysis (not shown) suggests that the sign change of the east IDM pole SSTA during boreal winter occurs irrespective of magnitude and source due to the march of the seasonal cycle and associated changes in feedbacks and upwelling and entrainment, as found for major IDM events [Vinayachandran et al., 2002].

[4] At the west IDM pole (Figure 1), forecasting persistence for 3–5 months is reasonable for all start months. ACCs decline further again around September for persisted boreal fall and winter SSTAs, as the observed SSTA changes sign. Skill returns, albeit of the opposite sign, in the following boreal spring. The 18–24 month timescale is suggestive of ENSO influence, see, e.g., Klein et al. [1999].

[5] Whether predictability barriers corresponding to the times of rapid decline in lagged correlations of observed SSTA over the tropical Indian Ocean may exist for dynamical CFSs, and whether intra-seasonal variability and the chaotic nature of the monsoon systems could severely limit the predictability of the system, are investigated below. The NCEP CFS is described briefly in 2. SSTA forecasts at the IDM poles for the last two decades are examined in 3, and the seasonal variation in skill assessed in 4. Potential predictability assuming a perfect model/ensemble is discussed in 5.

2. Brief Description of NCEP CFS

[6] The NCEP CFS consists of coupled almost-global ocean and atmosphere 3-D numerical general circulation models (GCMs) and a land model. The ocean and atmosphere GCMs are coupled once a day, and exchange daily–averaged quantities. The atmosphere is initialized from NCEP Reanalysis-2 (R-2) data [Kanamitsu et al., 2002]. Oceanic initial conditions are obtained by running the ocean GCM in stand–alone mode forced by weekly fluxes from R-2 with subsurface temperature data assimilated using a 3–D variational technique; subsurface salinity is adjusted using a climatological temperature–salinity relationship. Assimilated data include temperatures from expendable bathythermographs, the TAO array of moored buoys, and the ARGO network of floats. For the Indian Ocean, the amount of data available in real-time prior to 2003 is negligible. Sea–surface temperature is relaxed on a 90–day timescale to observed SST (NCEP OI data, [Reynolds and Smith, 1994]). The whole ocean system comprises the Global Ocean Data Assimilating System (GODAS). Fifteen-member ensembles are generated for each month from January 1981 to December 2003 by initiating runs from three sets of five consecutive days spanning each month. Daily atmospheric initial states are used with the 5–day averaged ocean initial conditions for each set. The system is then integrated forward in time for 9 months from each set of initial conditions.

[7] Further details on the CFS, and biases in its climatology and forecast skill over the Pacific and North Atlantic Oceans and North America, are given by Saha et al. (submitted manuscript, 2005). Over the tropical Indian Ocean, the CFS climatology drifts little from observed at a lead times of 2–3 months, but as the forecasts progress the zonal equatorial SSTA gradient weakens along with the equatorial westerlies (Figure 2). Evaporation is reduced over the southern tropics and Arabian Sea and Bay of Bengal, as the southeast trades and monsoon winds weaken respectively. The mean position of the Inter-Tropical Convergence Zone (ITCZ) shifts westwards with the warmer SSTs, as given by the lowering of outgoing longwave radiation (OLR). The zonal gradient in thermocline depth (20°C isotherm depth) also decreases resulting in a shallower (deeper) thermocline, and so stronger (weaker) air-sea coupling in the east (west). Associated with this bias is a westward shift in Ekman pumping (negative wind stress curl). Similar biases in position of the Warm Pool and thermocline tilt were found in the NSIPP free-running coupled system resulting in westward propagation of the eastern SSTA and associated convective anomaly during major IDM events [Wajsowicz, 2004].

3. Two Decades of Forecasts

[8] Time series constructed from monthly SSTAs averaged over the east and west IDM poles forecast at 2–3 months lead for 1981 to 2004 compare favorably with the observed (Figure 3); no smoothing has been applied. Cooling at the east IDM pole associated with major positive IDM events are captured with the exception of 1983 (Figure 3a). There is a slight delay in forecasting onset of major coolings. Warming during negative IDM events is not as intense, and the system captures this one-sidedness well. At the west IDM pole, the peak warming and decay associated with the major positive IDM events in 1982 and '83 and 1997/8 are well captured, but the onset is slightly delayed in '82 and '97 (Figure 3b). The 1989 and 1996 negative IDM events are reasonably well forecast, but the cooling associated with the 1992 and '93 events is not.

[9] As the coupled system drifts, the forecasts at the east IDM pole become poorer (Figure 3a). The magnitude and timing is only a little worse, but the 1990s are plagued with false alarms of a similar character to those in the NSIPP system [Wajsowicz, 2004]. Each boreal fall prior to the 1997/98 event, cooling is forecast, which turns out to be a false alarm. Similarly, post the 1997/98-event, warming is forecast, but these are also false alarms. For the 1980s, false alarms at 5–6 months lead are no more frequent than at 2–3 months lead. At the west IDM pole, there is only a slight deterioration in the forecasts at 5–6 months lead. Overall, these results are in-line with those for the NSIPP system described in Wajsowicz [2004], and compare favorably with those obtained by statistical methods, e.g., Collins et al. [2004].

4. Seasonal Variation in Forecast Skill

[10] Forecast skill is measured by calculating the ACC for the mean SSTA of the 15-member ensemble and the observed SSTA as given by GODAS. At the east IDM pole, the ACC is above 0.6 for a season or so for all start months (Figure 4a). The CFS simulates the observed quasi-biennial nature well for start months in the second half of the year. It skillfully captures the re–emergence of SSTAs with the opposite sign the following year; coincident with the December persistence barrier when SSTAs are weak, there is a temporary loss in skill. However, the return of skill is short-lived, as it falls off rapidly in late spring/early summer. The skill in forecasts initiated in January through April also declines rapidly at this time of year. These results suggest a late spring/early summer predictability barrier for forecasts initiated in boreal fall through early spring, which accounts for the tendency to forecast false alarms at 5–6 months lead in Figure 3a.

[11] At the west IDM pole, forecasts are typically skillful until boreal summer (Figure 4a). For forecasts initiated in boreal summer, there is a loss of skill in late summer/early fall, but then it returns, which is consistent with the delay in forecasting the onset of SSTAs associated with major IDM events found in Figure 3b, but with a good recovery to forecast the peak and decay. With the exceptions of forecasting boreal summer SSTAs at the west IDM pole and eastern SSTAs from late winter/early spring starts, the results from the NCEP CFS are encouraging, as the ACC of the ensemble mean typically beats that assuming persistence at lead times greater than 2–3 months.

[12] The results in Figure 4a compare favorably with the skill found using statistical methods, e.g., Collins et al. [2004].

5. Potential Predictability

[13] Potential predictability of a phenomenon may be examined by making the ‘perfect model/ensemble’ assumption, i.e. it is assumed that the coupled model and initial conditions are both perfect, and that the spread of ensemble members is sufficient. Each ensemble member is then taken in turn as truth, and compared using the ACC measure with the ensemble mean of the remaining members. High ACCs indicate that the climate system is predictable, or that there is a bias in the ensemble.

[14] At the east IDM pole the potential predictability of SSTAs appears limited (Figure 4b). The late boreal spring/early summer predictability barrier prevents successful forecasting of fall SSTAs at long lead times. Comparing the prediction-observed ACCs (Figure 4a) with Figure 4b suggests there is scope for improvement in the forecasts initiated in boreal spring and summer. Improvements could be in the GCMs, the initial conditions (almost no datum is assimilated into the Indian Ocean over the time frame considered), or in the construction of the ensembles.

[15] The potential predictability of SSTAs at the west IDM pole (Figure 4b) is high except for late summer and boreal fall anomalies forecast from start months in boreal winter and early spring. Comparing Figures 4b and 4a suggests that skillful forecast of boreal summer and fall SSTAs from April through August starts is feasible.

[16] Overall, these results are encouraging, and suggest that the quasi-biennial nature of the tropical Indian Ocean climate could be exploited to give skillful forecasts at longer lead times than currently published for ENSO prediction. However, overcoming the predictability barrier at the onset of the boreal summer monsoon at the east IDM pole, and at the height of the monsoon at the west IDM pole, which prevents skillful prediction of boreal fall SSTAs from start months early in the year, could be difficult if it is due to the noisiness of the climate system during the monsoons. At the east IDM pole, further difficulty could arise from trying to forecast from a time of year when the SSTAs are changing sign and the ocean-atmosphere coupling is weak (deep thermocline), especially as the subsurface ocean is not well initialized.