This study investigates whether air–sea interactions contribute to differences in the predictability of the boreal winter tropical intraseasonal oscillation (TISO) using the NCEP operational climate model. A series of coupled and uncoupled, “perfect” model predictability experiments are performed for 10 strong model intraseasonal events. The uncoupled experiments are forced by prescribed SST containing different types of variability. These experiments are specifically designed to be directly comparable to actual forecasts. Predictability estimates are calculated using three metrics, including one that does not require the use of time filtering. The estimates are compared between these experiments to determine the impact of coupled air–sea interactions on the predictability of the tropical intraseasonal oscillation and the sensitivity of the potential predictability estimates to the different SST forcings.
Results from all three metrics are surprisingly similar. They indicate that predictability estimates are longest for precipitation and outgoing longwave radiation (OLR) when the ensemble mean from the coupled model is used. Most importantly, the experiments that contain intraseasonally varying SST consistently predict the control events better than those that do not for precipitation, OLR, 200-hPa zonal wind, and 850-hPa zonal wind after the first 10 days. The uncoupled model is able to predict the TISO with similar skill to that of the coupled model, provided that an SST forecast that includes these intraseasonal variations is used to force the model. This indicates that the intraseasonally varying SSTs are a key factor for increased predictability and presumably better prediction of the TISO.
The potential and practical limits of predictability associated with the boreal winter tropical intraseasonal oscillation (TISO) remain relatively uncertain. A number of studies have attempted to estimate the predictability associated with the TISO using atmospheric general circulation models. Studies by Waliser et al. (2003a, b) estimate a theoretical or potential limit of predictability. This potential predictability represents an estimate of predictability for the hypothetical case of a “perfect model.” They use an atmospheric general circulation model (AGCM) to estimate the potential predictability for strong model events by performing twin predictability experiments forced with climatological sea surface temperature (SST). The predictability estimates were determined using a signal-to-noise measure defined by the authors. The limit of predictability was determined to be the time at which the mean error becomes as large as the mean signal. For the winter months (November–April), they estimated the predictability limit to be about 25–30 days for the 200-hPa velocity potential (χ200) and on the order of 10–15 days for rainfall. The limit of predictability for summer was determined to be 25 days for the 200-hPa velocity potential and 15 days for rainfall. However, the authors acknowledge that the predictability may be underestimated because they used an uncoupled atmospheric model forced by fixed climatological SST.
A study by Liess et al. (2005) estimated the potential predictability of the northern summer intraseasonal oscillation in the ECHAM5 model using the same signal-to-noise estimate of predictability. In Liess et al. (2005), the three strongest intraseasonal events were selected from a long simulation. Fourteen-member ensembles were run for each of the events and were forced by climatological SSTs. They estimated the upper limit of predictability for the 200-hPa velocity potential to be around 20 days. For the 200-hPa zonal wind (U200), the estimates were about 20–30 days, and about 15–20 days for precipitation. Like the studies described above, the study by Liess et al. (2005) does not use a CGCM, which may impact the limit of predictability.
Practical predictability of the TISO has also been investigated at some of the operational centers. These practical predictability estimates represent the actual skill in reforecasts of intraseasonal variability and are smaller than potential predictability estimates described above because model errors are included. For example, several studies use the dynamical extended-range forecast (DERF) experiments performed by the National Centers for Environmental Prediction (NCEP) using the Medium-Range Forecast (MRF) model to evaluate the predictability of the TISO when forced with observed climatological SST (Jones et al. 2000; Lau and Chang 1992; Chen and Alpert 1990). These studies estimate the predictability of the TISO based on the anomaly correlation coefficient. They estimate the limit of predictability to be about 5–10 days for outgoing longwave radiation (OLR) and the 200-hPa velocity potential. However, it is likely that the predictability is significantly underestimated in these studies because of large systematic errors over the tropics in the zonal wind, which result in a weak TISO signature (Jones et al. 2000). More recent DERF experiments have been performed with a new version of the NCEP model. The predictability of summer and winter intraseasonal variability was estimated by Seo et al. (2005) using the anomaly correlation coefficient. They found predictability of about 7 days for both summer and winter 200-hPa zonal wind. Predictability estimates were about 4–5 days for both summer and winter 850-hPa zonal wind (U850). The estimates for outgoing longwave radiation were the lowest at around 4 days for winter and 5 days for summer.
Additionally, two recent studies performed at the European Centre for Medium-Range Weather Forecasts (ECMWF) investigate the practical predictability of the TISO and its sensitivity to a range of factors, including air–sea interactions, physical parameterizations, resolution, and the representation of the diurnal cycle in the ocean (Woolnough et al. 2007; Vitart et al. 2007). The studies show practical predictability estimates for two intraseasonal events that occur during the Tropical Ocean and Global Atmosphere Coupled Ocean–Atmosphere Response Experiment (TOGA COARE) intensive observing period. The experiments consist of 32-day forecasts of five-member ensembles, initialized daily. The skill of the forecasts is measured by projecting the forecasts onto either the extended empirical orthogonal functions (EOFs) of χ200 (Vitart et al. 2007) or the combined EOFs of OLR, U850, and χ200 averaged in a tropical band from 10°S to 10°N. The anomaly correlation of the projected forecasts was calculated for the first two principal component time series (PCs) separately. Based on this metric, the skill of the ECMWF monthly forecast system is shown to be approximately 14 days in forecasting the evolution of the TISO. Vitart et al. (2007) demonstrate that the skill in forecasting the TISO is highly dependent on the representation of model physics, the quality of initial conditions, and the perturbation method used for initial conditions.
Because previous studies indicate that air–sea coupling has an organizing and intensifying role on tropical intraseasonal variability (Flatau et al. 1997; Waliser et al. 1999; Fu and Wang 2004; Fu et al. 2003; Zheng et al. 2004), it is possible that coupled air–sea feedbacks also have an impact on its predictability. Woolnough et al. (2007) demonstrate the importance of the feedback of the ocean to the atmosphere by performing a series of sensitivity experiments to test the importance of the diurnal cycle of SST and ocean mixing. They demonstrate that improved vertical resolution of the ocean and a high-frequency coupling interval (3 h), which improves the representation of the diurnal cycle of SST and ocean mixing, contributes to improved skill in forecasting the TISO for the PC that represents positive convection over the oceans (i.e., Indian Ocean and western Pacific) using the same experiment design as that in Vitart et al. (2007).
Two studies by Fu et al. (2007, 2008) specifically focus on the impact of air–sea coupling and SST variability on potential predictability estimates of the boreal summer intraseasonal oscillation. In one study (Fu et al. 2007), they perform a series of twin predictability experiments for 20 summer intraseasonal events selected from a long coupled run. They perform uncoupled experiments using the ECHAM4 AGCM forced with prescribed SST, and coupled experiments using ECHAM4 coupled to a hybrid ocean model. The coupled and uncoupled predictability experiments have the same initial conditions. However, the uncoupled experiments are forced by SSTs from the coupled control simulation in which the intraseasonal (20–90 days) variability of the SST has been removed. Using both the anomaly correlation coefficient and the signal-to-noise ratio defined by Waliser et al. (2003a, b), they find similar predictability estimates. Their estimates indicate that air–sea interaction extends the predictability of the northern summer rainfall by about 1 week, from 17 to 24 days. In the other study (Fu et al. 2008), they perform a series of perfect model predictability experiments with 10 ensemble members for the four phases of two model TISO events using both the coupled and uncoupled models. The uncoupled integrations are forced with different SST configurations derived from the control integration, including smoothed SSTs, where the intraseasonal variability is removed; damped persistent SST, where the SST from the coupled control run is linearly interpolated to the smoothed SST during the first 10 days of the forecast; SST anomalies forecast by coupling to a slab mixed layer model; and ensemble mean daily SST from the coupled forecasts. Comparison between the anomaly correlations and signal-to-noise ratios from these experiments are consistent with their previous results, which show that including intraseasonal variability in the SST extends the predictability of rainfall associated with the TISO by about 1 week.
Previously, Pegion and Kirtman (2008) demonstrated that coupled air–sea interactions are responsible for some differences in the simulation of winter (November–April) tropical intraseasonal variability between the coupled and uncoupled versions of the National Centers for Environmental Prediction (NCEP) operational climate model. In the current study, we investigate whether air–sea interactions contribute to differences in the predictability of the tropical intraseasonal oscillation during the extended winter season (November–April) for the same model. We do so in a perfect model framework (i.e., potential predictability), and thus we do not address the issue of how model error impacts the predictability estimates. A series of nine-member coupled and uncoupled predictability experiments are performed for 10 strong model intraseasonal events. The uncoupled experiments are forced by prescribed SST from the control simulation containing different types of variability (Table 1). Predictability estimates are compared between these experiments to determine the impact of coupled air–sea interactions on the predictability of the tropical intraseasonal oscillation and the sensitivity of the potential predictability estimates to the different SST forcing. Specifically, comparison of the coupled and uncoupled experiments addresses the potential impact of using tier-1 (fully coupled) versus tier-2 (uncoupled, with prescribed SST) methods to estimate the predictability of the TISO. This study is similar in focus and design to the work of Fu et al. (2008); however, it differs in several ways. First, we use an operational climate model to perform the predictability estimates as well as calculate predictability using a method similar to that of Woolnough et al. (2007) and Vitart et al. (2007) that can be applied to actual forecasts with the intention of providing predictability estimates that can be more readily compared with actual forecast skill. In fact, we present a series of prescribed SST experiments that are specifically designed to examine intraseasonal predictability in an operational setting. Second, we focus on the boreal winter intraseasonal oscillation. Third, we investigate a set of 10 events for a specific phase of the TISO with the intention of providing a more robust estimate of the impact of air–sea interactions and sensitivity to SST forcing.
2. Model description
This study investigates the impact of air–sea coupling and SST sensitivity of the potential predictability of the tropical intraseasonal oscillation using the NCEP Climate Forecast System (CFS; Saha et al. 2006). Here, we give a brief description of the model. A more extensive description of the CFS is given by Saha et al. (2006) and Wang et al. (2005). The CFS is the fully coupled atmosphere–ocean general circulation model used operationally by NCEP for climate forecasts. It is composed of the 2003 NCEP Global Forecast System (GFS) as the atmospheric component and the Geophysical Fluid Dynamics Laboratory (GFDL) Modular Ocean Model, version 3 (MOM3; Pacanowski and Griffies 1998) as the oceanic component. The GFS has a resolution of T62 in the horizontal and 64 layers in the vertical. Recent upgrades to the model physics include solar radiation following Hou et al. (1996), the cumulus convection scheme of Hong and Pan (1998), gravity wave drag (Kim and Arakawa 1995), and cloud water and ice (Zhao and Carr 1997). The ocean model has a quasi-global domain ranging from 74°S to 64°N; it has 40 layers in the vertical and a resolution of 1/3° × 1° in the tropics and 1° × 1° in the extratropics. The atmosphere and ocean exchange fluxes and sea surface temperatures once per day without flux correction. The sea ice extent is taken as climatology.
3. Experiment design
Because our goal is to estimate the impact of coupled air–sea feedbacks and forcing by different types of SST variability on the predictability of the tropical intraseasonal oscillation, we design a series of predictability experiments using both the coupled and uncoupled models. These experiments are designed so that we can determine the predictability in terms of how well each model “forecasts” the intraseasonal events from a coupled control simulation. The experiments are performed for 10 strong intraseasonal events. Atmospheric initial conditions are perturbed to produce nine-member ensembles and the experiments are run for 60 days. If coupled air–sea interactions are important for the predictability of the TISO, then we expect the coupled model to better predict the events than the uncoupled model. Additionally, we expect the uncoupled experiments that contain intraseasonal variability in the SST to better predict the events than the uncoupled simulations that do not contain the intraseasonal variations in SST.
a. Event selection
To conduct the predictability experiments, 10 strong (>2σ) intraseasonal events are selected from a 52-yr coupled control simulation, and the atmospheric initial conditions of these events are perturbed. The events are chosen according to the amplitude of the PC time series of an extended empirical orthogonal function (EEOF) analysis of precipitation from the control simulation (Pegion and Kirtman 2008). It is possible that the phase of El Niño–Southern Oscillation (ENSO) may affect the propagation of the intraseasonal oscillation (Tam and Lau 2005) and its predictability. Therefore, in an attempt to reduce this impact, we require that the selected events span the phases of ENSO. Potential events are categorized as occurring during warm, neutral, or cold phases of ENSO based on the value of the Niño-3.4 index from the control simulation during the extended winter season (April–November) in which the event occurs. The Niño-3.4 index is defined as the average SST anomaly in the 5°S–5°N, 170°–120°W region. If the Niño-3.4 index is above 0.5°C for four or more of the months during the extended winter season, then it is determined that the intraseasonal event occurs during the warm phase of ENSO. The same criteria are applied for the cold phase when Niño-3.4 is below −0.5°C. If events do not occur during either a warm or cold phase of ENSO, they are classified as occurring during a neutral phase. Four of the intraseasonal events occur during warm ENSO conditions, three during neutral conditions, and three during the cold phase. With this experimental design it is possible that the effects of ENSO on the predictability of the TISO are not entirely removed. However, while the interannual variability of the TISO and/or its relationship with ENSO is not the focus in this study, in practical applications (i.e., operational forecasting) the effects of ENSO influence the prediction of TISO.
We attempt to determine how well the evolution of the TISO events in the control simulation can be predicted by the coupled and uncoupled models when the oscillation is active but the initial conditions are perturbed. Because the observed precipitation anomalies for TISO begin in the Indian Ocean and propagate eastward, we choose to initialize the experiments when the precipitation is maximized in the Indian Ocean. By doing so, we are asking the following question: how well can the TISO be predicted when it is known that it has been initiated? The first two EEOFs of precipitation for day 0 are shown in Fig. 1. When the second EEOF is positive, the convection in the Indian Ocean is positive; therefore, we use the second EEOF (Fig. 1b) to identify initial conditions for events when they are in the Indian Ocean. Based on the second EEOF, the region of positive precipitation is located at 20°–5°S, 40°–80°E. Therefore, predictability experiments are initialized when positive precipitation anomalies are maximized in this region. As an example, we show the intraseasonally filtered (30–100 day) precipitation anomalies for one of the events (Fig. 1c). In this case, the initial condition date for the event is identified as 3 February.
This method of choosing events and identifying initial conditions is similar to the method used by Fu et al. (2007, 2008), but differs from the Wheeler and Hendon (2004, hereafter WH04) method of identifying events and their phase used by Vitart et al. (2007) and Woolnough et al. (2007). First, we note that the WH04 method is based on combined EOFs of OLR, U200, and U850. In the observations, the first two EOFs are in quadrature and are well separated from the others. However, in the CFS, as well as other models, the first two combined EOFs are in quadrature, but are not well separated from the other EOFs (not shown). This indicates that although the first two EOFs describe most of the TISO variability in observations, they are not sufficient to describe the TISO variability in the model. However, we note that all of the events chosen from the control run propagate eastward through the WH04 phase diagram (their Fig. 7) and have extended periods greater than one standard deviation during their evolution. The most notable difference between the events in our method and the WH04 method is that they do not always agree on the identification of the phase of the oscillation. While the WH04 method works well for identifying the phase in observations, deficiencies in model placement of maximum anomalies and failure of the models to properly represent the coherent relationship between the surface and dynamical variables (e.g., OLR, U850, and U200; Zhang et al. 2006) also make it difficult for the WH04 method to identify the phasing of the events in the model. For example, in the model used in this study, the precipitation and OLR anomalies are erroneously maximized in the southwestern Indian Ocean (Pegion and Kirtman 2008). Therefore, the application of the combined EOFs used by WH04 over a tropical belt (15°S–15°N) does not always identify the Indian Ocean phase well in this model.
To demonstrate the behavior of the events, we first show one of the events as an example. This example event occurs during the neutral phase of ENSO. The equatorial (10°S–10°N) anomalies of U200, precipitation, and SST for this event from the control simulation are calculated based on the climatology of the control simulation to remove the annual cycle and are shown in Fig. 2. Note that these values are not filtered. Despite this, the eastward propagation of anomalies of U200 and precipitation are apparent. Divergent (convergent) anomalies of U200 are associated with positive (negative) anomalies of precipitation. Consistent with the observed relationship between precipitation and SST, positive (negative) SST anomalies lead (lag) the region of enhanced precipitation. We also note that the SST anomalies are relatively large in the Indian and western Pacific Oceans in the first 50 days of this event.
b. Initialization strategy
The atmospheric initial conditions from the control simulation are perturbed to produce nine initial states for each event. In the example case shown above (Fig. 1c), the initial time is determined to be 0000 UTC 3 February. Perturbations are obtained by running the CFS in 1-h increments from the previous day. The 4 h prior to the initial conditions date (e.g., 2000, 2100, 2200, and 2300 UTC 2 February) and the 5 h subsequent to this date (e.g., 0100, 0200, 0300, 0400, and 0500 UTC 3 February) are chosen as the nine perturbed initial conditions. The calendar is then reset to the original date of the event from the control simulation. The coupled predictability experiments are initialized with the nine atmospheric states and the ocean initial conditions from the control simulation and evolve with their own SSTs. The coupled predictability experiments by design have different SST evolutions than that of the control. On the other hand, the uncoupled experiments are initialized with the same nine atmospheric states and are forced by the designated SSTs (Table 1). The different SST cases are described in detail in the next section.
c. SST cases
If coupled air–sea interactions are important to the predictability of the tropical intraseasonal oscillation, then the SST used to force the uncoupled predictability experiments may impact the predictability estimates. Therefore, we perform a series of experiments in which the uncoupled model is forced with different SST. In the first of the SST experiments, the uncoupled model is forced with the perfect daily SST from the control simulation. Because these SST are from the coupled simulation, they contain intraseasonal variations, which are due in part to the feedback of the atmosphere on the ocean. These intraseasonal variations in SST may then feed back to the atmosphere and impact the evolution of the TISO event (Fu et al. 2008). The perfect SST experiment represents a hypothetical case that is not realistic in an operational setting because the future SST cannot be known perfectly. Therefore, we perform a more realistic experiment in which the forecast SST from the coupled predictability experiments are used to force the uncoupled model (Fcst SST). However, the initial conditions are assigned to different ensemble members to avoid duplication of the coupled predictability experiments. This is similar to a “tier-two” approach and to the “ATMd-init” experiments performed by Fu et al. (2008). This set of experiments allows the uncoupled model to feel the forecast intraseasonal variations in the SST, but without being forced by the known, perfect SST. In an operational setting, forecast SSTs are lagged behind real-time SSTs because the ocean state cannot be initialized in real time because of the delay in receiving ocean observations. Therefore, it has been a common practice in seasonal prediction to maintain the observed SST anomalies at the initial time of the forecast. For example, NCEP uses a method called damped persistence in which the initial SST anomalies are damped to climatology (S. Saha 2007, personal communication). We perform a “persisted SSTA” experiment to provide a more realistic operational comparison. In this experiment, the initial SST anomaly is allowed to persist throughout the forecast. Two other SST sensitivity experiments are performed to demonstrate the impact of different SSTs on the predictability of the TISO. These experiments are ones in which the uncoupled model is forced by SST, which contains only monthly and climatological variations and therefore, do not contain intraseasonal variations.
4. Predictability metrics
We estimate the predictability in terms of the ability of each experiment to “forecast” the events from the control simulation. Three different methods are used to calculate estimates of predictability. The first two methods, anomaly correlation and signal-to-noise ratios, have been used in several other studies (Fu et al. 2007, 2008; Waliser et al. 2003a, b; Liess et al. 2005). These methods of calculating predictability require time filtering to extract the intraseasonal signal from high-frequency weather and are therefore not realistic for comparison with actual operational forecasting skill. To apply the time filtering, the 120 days prior to the event from the control simulation are appended to the 60 days of each of the predictability experiments following Waliser et al. (2003a, b) and Fu et al. (2007, 2008). The filtering is applied by calculating a 31-day centered running mean. The third metric uses spatial filtering of the TISO pattern and is being used at several operational centers (Woolnough et al. 2007; Vitart et al. 2007; J. Gottshalck 2007, personal communication).
a. Anomaly correlations of filtered data
For this calculation, we calculate the pattern correlation or anomaly correlation coefficient (ACC) of precipitation anomalies in the Indo-Pacific region (30°S–30°N, 32.5°E–92.5°W) over the nine ensemble members for the 10 events between the predictability experiments and the control. We also calculate correlations between the ensemble mean of the predictability experiments and the control over all 10 events. We use the ensemble mean in an attempt to reduce the “noise” and isolate any “signal” associated with the intraseasonal oscillation that is common among all of the ensemble members.
Over lead time, correlations will be reduced as the difference between the predictability experiment and the control simulation becomes larger. The limit of predictability for the correlations is subjectively defined as the time at which correlations fall below 0.5. We use the decorrelation time as the measure of predictability in order to mimic the way in which operational forecasts are verified. These predictability estimates are calculated for the coupled and uncoupled predictability experiments and their results are compared.
b. Signal-to-noise ratio
Traditionally, a signal-to-noise ratio has been used to estimate predictability for seasonal mean forecasts. This signal-to-noise ratio is usually defined as the ratio of the ensemble mean to the ensemble spread. However, this is not well suited for an oscillating phenomenon like the TISO because the amplitude of the oscillation inevitably passes through zero. Therefore, we use the signal-to-noise ratio defined by Waliser et al. (2003a, b). The amplitude of the TISO represents a measure of the signal. Waliser et al. (2003a, b) define this signal as the temporal variance in a sliding window that is large enough to encompass an entire event. Because the intraseasonal events in the CFS are much slower than those observed, we tested the sensitivity of the signal to windows of 51, 61, 81, and 101 days and found that the predictability estimates are relatively insensitive to the size of the window. Therefore, to be consistent with Waliser et al. (2003a, b) and Fu et al. (2007, 2008), we also use a 51-day window. The mean noise is estimated as the mean square forecast error relative to the control experiment and is determined by averaging over all nine ensembles and over all 10 events. We are interested in the time at which the forecast errors resulting from errors in the initial conditions grow to be as large as the signal of the TISO. Therefore, the limit of predictability is defined as the forecast lead time at which the mean signal is as large as the mean noise.
c. Anomaly correlations of projected PC time series
The ACC of the PCs calculated by projecting the forecasts onto the WH04 real-time multivariate Madden–Julian oscillation (MJO) index calculated for the control simulation is used to calculate predictability estimates that do not require time filtering. First, the combined EOF of intraseasonally filtered OLR, U850, and U200 averaged in a tropical band from 15°S to 15°N is calculated from the 52-yr control simulation. The first two EOFs are similar to those observed and explain about 12% and 10% of the variance, respectively. However, unlike the observations, the first two EOFs are not well separated from the subsequent EOFs, so some of the TISO variability in the model is not well represented by these first two EOFs. The predicted OLR, U200, and U850 from the predictability experiments are projected onto these two EOFs to obtain the projected PC time series. The ACCs are calculated for each of the ensemble means over all 10 events. We again estimate the predictability as the time at which it takes for the ACC to fall below 0.5.
5. Predictability estimates
As an example of the overall ability of the predictability experiments to predict the events from the control simulation, a composite of the precipitation anomalies over the 10 events from the control simulation is shown in Fig. 3, and composites of the ensemble mean equatorial (10°S–10°N) precipitation anomalies from the predictability experiments for each of the SST cases are shown in Fig. 4. The anomalies are calculated relative to the climatology from a corresponding coupled and uncoupled simulation experiment. All of the experiments have very similar precipitation anomalies that are consistent with the control for about the first 10 days. The evolution of the experiments is probably strongly influenced by the initial conditions during this period. It is apparent that the eastward propagation of the precipitation anomalies associated with the TISO can be simulated in the Coupled, Perfect, Fcst, and Persisted SSTA experiments, while the monthly and climatological SST experiments do not show eastward propagation. Clearly, the most coherent propagation is simulated by the Coupled and Perfect SST experiments. The Fcst SST experiment apparently has difficulty maintaining the amplitude of the precipitation anomalies over the Maritime Continent. The differences between the composite precipitation from the predictability experiments and the control simulation along the equator (10°S–10°N) as a function of lead time are shown in Fig. 5. The differences for the Coupled, Perfect, and Persisted SSTA experiments are clearly small for about the first 10 days, and then become larger. Generally, all of the experiments tend to have a large dry error (>8 mm day−1) at about 120°E at approximately days 20–25 and a large wet error (>8 mm day−1) at 120°E at approximately 35–40 days. These errors emphasize the difficulty that the model has in forecasting the evolution of the TISO across the Maritime Continent, even in perfect model predictability experiments. This is consistent with a predictability barrier at the Maritime Continent seen in NCEP (A. Vintzileos 2006, personal communication) and ECMWF (Vitart et al. 2007) experimental reforecasts. The dry error near 120°E in the Fcst SST experiment further highlights the fact that this experiment has particular difficulty in maintaining the amplitude of the oscillation over the Maritime Continent.
a. Comparison of estimates between weather and intraseasonal time scales
We now quantify the errors between the predictability experiments and the control events by calculating the predictability in terms of the metrics described in section 4. First, we show the anomaly correlations for unfiltered precipitation, OLR, U200, and U850 calculated over the ensemble members of the predictability experiments (Fig. 6). Clearly, for each of the variables, there is no significant difference between the different SST experiments, with the exception of the climatological SST experiment, which has clearly less predictability than the others. The predictability estimates with the unfiltered data are about 12–14 days for precipitation, 14–18 days for OLR, and about 16–20 days for U850 and U200. We believe that this predictability simply reflects the lack of long-term predictability associated with weather, so the predictability of the TISO is masked by the unpredictable weather. Therefore, it is necessary to filter the data to remove the day-to-day weather fluctuations and extract the variability associated with the TISO, in much the same way as in Waliser et al. (2003a), WH04, and Vitart et al. (2007).
b. Comparison of estimates using the ensemble mean versus ensemble members
The correlations calculated using filtered data are shown in Figs. 7 and 8. These correlations are calculated over all of the ensemble members (Fig. 7) and over the ensemble mean (Fig. 8). Correlations less than one at the initial time are due to the use of time filtering. Comparison of these two figures provides an estimate of the skill associated with using the ensemble mean rather than the individual ensemble members. Again, the assumption is that the ensemble mean is able to identify something common among the ensemble members that is more predictable than if the variability among the members is included. For the Coupled and Fcst SST experiments, using the ensemble mean increases the predictability to beyond the 45 days of available data for all of the variables. For the predictability estimates of precipitation in the Perfect and Fcst SST experiments, the ensemble mean provides increases in predictability of 25 and 16 days, respectively, while the increases are smaller for the Persisted SSTA (∼6 days) and monthly SST (∼11 days) experiments, and there is little to no change for the climatological SST experiment. Based on these predictability estimates, we find that the ensemble mean is useful in providing additional predictability for TISO-related precipitation, OLR, U200, and U850. These results are consistent with those of Fu et al. (2008) who found that the predictability estimates of precipitation are increased by 10 days for all their SST cases when the ensemble mean is used, although it appears that the ensemble mean has an even larger impact for the Coupled, Fcst, and Perfect SST cases in our experiments.
c. Impact of different SST cases on predictability
The primary focus of this research is to understand the role of the air–sea interactions and the intraseasonally varying SST on the predictability of the TISO. We now compare the predictability estimates for the different SST cases based on correlations of the ensemble mean with the events from the control run (Fig. 8) and those calculated using a signal-to-noise ratio (Fig. 9). Specifically, we point out some behavior of the predictability curves that is similar among the variables and predictability metrics. First, it is noted that the predictability curves are very similar in all SST cases for all variables during the first 10 days, with the exception of the climatological SST case, which loses predictability quickly, with a limit of about 10 days for precipitation for both the ACC and signal-to-noise estimates. As mentioned previously, this is likely due to the strong influence of the initial conditions. After 10 days, the curve for the coupled model appears to separate from the other cases in precipitation and OLR out to at least 45 days in both metrics, while the Perfect and Fcst SST experiments tend to stay tightly grouped with each other. For the dynamical variables, the coupled experiment does not appear to separate from the curves for the Perfect and Fcst experiments, but instead tends to be tightly grouped with these experiments throughout the forecast. For the Persisted SSTA case, the curve tends to be similar to the Perfect and Fcst SST cases in the first part of the forecast, but falls off more rapidly in the latter half of the forecast for all of the variables. All of these features are similar between the two predictability estimates. Based on this general behavior of the predictability curves, we conclude that the coupled model is able to better predict precipitation and OLR than the other SST experiments; however, this does not seem to be the case for the dynamical variables. Additionally, the experiments with intraseasonally varying SST (Coupled, Perfect, and Fcst) have longer predictability estimates than those without (Persisted, Monthly, and Clim) for all variables. It seems that the atmospheric model alone can generate skillful forecasts of the control events if the underlying intraseasonal variations in SST resulting from air–sea interactions are included. Clearly, because the coupled model has higher predictability for precipitation and OLR than the perfect SST experiment, a perfect SST forecast is not required, but it seems that it must be skillful enough to capture the intraseasonal variations in the SST.
d. Implications for real-time forecasts
In this section, we discuss the results of the predictability estimates above in terms of their application to real-time forecasts, including a third predictability metric that does not require the use of time filtering. The predictability estimates for all variables are highest for the Coupled, Perfect, and Fcst cases. In each of these cases, the SSTs contain intraseasonal variability resulting from air–sea coupling. Therefore, is seems that this intraseasonal variability in the SST is important and that potentially the best forecast should be obtained by either producing forecasts using a coupled model (tier 1) or by forcing the atmospheric component with the SSTs forecast by the coupled model (tier 2). Therefore, the fully coupled model is apparently needed to produce the most skillful forecasts of TISO-related precipitation and convection compared to the other cases. If the coupled model is not used, the best predictability estimates possible are those of the Persisted SST case. These estimates from the ACC, calculated using the ensemble means, are 26 days for precipitation and 30 days for OLR. For the dynamical variables, these estimates extend beyond 45 days.
There is an important caveat to the predictability estimates using the ACC and signal-to-noise predictability metrics. These metrics use time filtering to isolate intraseasonal time scales. However, this is not realistic in a forecast setting. Therefore, we also calculate the anomaly correlations of the first two PCs determined by projecting the ensemble mean forecast of OLR, U850, and U200 onto the combined EOFs of these variables from the control simulation. This EOF projection is designed to filter the data using spatial filtering rather than temporal filtering. The anomaly correlation of the PC time series with the control is shown in Fig. 10. First, it is clear that this metric produces predictability curves that are much noisier than the other metrics. The correlations vary largely throughout the lead times for all SST cases. It seems that more cases may be needed to produce stable statistics using this metric. For PC1 (PC2), all of the experiments have a large reduction in correlation at about day 18 (25), with the exception of the persisted SSTA case in PC2. This loss of skill occurs when the TISO-related convection is located over the Maritime Continent and further emphasizes the Maritime Continent predictability barrier. The loss of skill for PC1 around 18 days is less dramatic in the coupled model than in the other cases, with correlations remaining above 0.6. After this loss of skill, the Coupled, Perfect, and Fcst experiments rebound to correlations near 0.7 and remain above 0.5 throughout the forecast. Additionally, the monthly SST experiment drops to correlations of near zero at day 28, but then rebounds to have similar skill with the Coupled, Perfect, and Fcst experiments in the last 20 days of the forecast. For PC2, the persisted SST case has higher correlations than the other cases for days 20–25. This is likely due to the fact that PC2 is positive when TISO-related convection is near the Maritime Continent. Because the model tends to maintain convection over the Maritime Continent for about 10–15 days (Pegion and Kirtman 2008), the persisted SSTA case is more skillful for PC2 during these lead times. Overall, the experiments that contain intraseasonally varying SST have higher correlations from day 25 to 40 for PC1 and days 35–60 for PC2, further emphasizing the importance of the intraseasonally varying SST.
The impact of air–sea coupling on the predictability of the tropical intraseasonal oscillation and its sensitivity to SST forcing has been investigated by performing a series of “perfect” model predictability experiments with the coupled and uncoupled models. These estimates have been calculated using three different metrics. While these metrics in some cases provide different estimates of predictability, we caution against focusing on the specific number calculated, especially because the use of a 0.5 correlation coefficient is arbitrary. However, looking at the overall behavior of each of the different SST cases, the results are surprisingly similar for each of the metrics. Specifically, the experiments that contain intraseasonally varying SST (Coupled, Perfect, and Fcst) consistently predict the events from the control simulation better than those that do not after the first 10 days. Therefore, we conclude that intraseasonally varying SST that are in part due to coupled air–sea interactions are a key factor for the increased predictability of the TISO in these experiments. The uncoupled model is able to predict the TISO with similar skill as the coupled model, provided that a “skillful” SST forecast that includes these intraseasonal variations is used to force the model. It is clear that the SSTs must only be skillful; they do not need to be perfect because the Perfect SST experiment demonstrated similar skill as the Fcst SST experiment. These results are generally consistent with the results of Fu et al. (2008), in which a similar set of SST experiments was performed to investigate their impact on the boreal summer intraseasonal oscillation. For example, they also found that experiments either using both the coupled model and the uncoupled model forced with forecast SST from a slab mixed layer model, or daily mean SST from the coupled model have higher predictability estimates using both the ACC and signal-to-noise metrics in the region of Southeast Asia. One major difference between their results and ours is that they found that the uncoupled model has identical predictability as the coupled model for precipitation in Southeast Asia when forced with daily ensemble mean SST. This is attributed to the fact that the proper relationship between precipitation and SST is provided in the initial conditions and that it takes about 30 days for that relationship to be lost. In our Fcst and Perfect SST experiments the predictability is slightly lower for precipitation and OLR than the fully coupled experiments. We offer two possible reasons for these differences. First, we do not use ensemble mean SST to force the forecast SST experiment. Instead, we use the forecast SST from each ensemble member of the coupled experiments, and then shuffle the initial conditions associated with that SST. Perhaps the ensemble SST is more skillful or more consistent with the atmospheric state, thus providing a better forecast. Second, it is noted that even the Perfect SST has similar predictability to the Fcst SST experiments. Perhaps, the relationship between precipitation and SST is lost more quickly in this model or for these events.
We also wish to point out that the predictability experiments performed using climatological SST significantly underperform the other experiments in all metrics for precipitation, OLR, U850, and U200. Based on these results, it is likely that the potential predictability metrics estimated in previous studies using climatological SST (Waliser et al. 2003a, b; Liess et al. 2005) significantly underestimate the potential predictability of the TISO. Additionally, based on the results of Woolnough et al. (2007), our coupled model experiments and those of Fu et al. (2007, 2008) may also underestimate the predictability because they do not resolve the diurnal cycle.
Additionally, we find evidence of a Maritime Continent predictability barrier even in these perfect model predictability experiments, indicating that either this may be an actual predictability barrier or that current models must be improved in terms of their ability to represent the Maritime Continent, its topography, and the physics of convection in this region before this barrier can be overcome. Although studies have indicated that increased resolution does not solve this problem (Vitart et al. 2007), it may be that very high-resolution models are needed to predict the evolution of the TISO through this region.
The authors thank H. Pan and S. Saha at the Environmental Modeling Center of the National Centers for Environmental Prediction for providing the CFS model. The model experiments were performed on the National Center for Atmospheric Research supercomputing systems. This research was supported by grants from the National Science Foundation (ATM-0332910), the National Oceanic and Atmospheric Administration (NA04OAR4310034), and the National Aeronautics and Space Administration (NNG04GG46G).
* Current affiliation: Rosenstiel School for Marine and Atmospheric Sciences, University of Miami, Miami, Florida.
Corresponding author address: Dr. Kathy Pegion, Center for Ocean–Land–Atmosphere Studies, 4041 Powder Mill Road, Suite 302, Calverton, MD 20705. Email: email@example.com