The BMRC Coupled General Circulation Model ENSO Forecast System

APRIL 2002 WANG ET AL. 975

GUOMIN WANG,RICHARD KLEEMAN,NEVILLE SMITH, AND FAINA TSEITKIN Bureau of Meteorology Research Centre, Melbourne, Victoria, Australia

(Manuscript received 15 May 2000, in ®nal form 24 August 2001)

ABSTRACT An El NinÄo±Southern Oscillation (ENSO) prediction system with a coupled general circulation model and an ocean data assimilation scheme has been developed at the Australian Bureau of Meteorology Research Centre (BMRC). The coupled model consists of an R21L9 version of the BMRC climate model and a global version of the Geophysical Fluid Dynamics Laboratory modular ocean general circulation model with resolution focused in the tropical region and 25 vertical levels. A univariate statistical interpolation method, with 10-day data ingestion windows, is used to assimilate ocean temperature data and initialize the coupled model. The coupling procedure does not use any ¯ux corrections. Hindcasts have been carried out for the period 1981±95 for each season (60 in all), for up to a lead time of 12 months. This paper will describe these initial experiments and show that the skill of sea surface temperature (SST) hindcasts in the tropical Paci®c is comparable to other published coupled models. The skill of the model is strongest in the central Paci®c. SST skill tends to be lower during the earlier 1990s than during 1980s in the eastern Paci®c but not in the central Paci®c. Since the ENSO SST anomaly in the central Paci®c is the most important forcing of regional and global climate anomalies, the high SST prediction skill and its insensitivity over the hindcast period in this region in this model give grounds for optimism in the use of coupled general circulation models.

1. Introduction cycle and the anomaly deviation from it. The motivation The El NinÄo±Southern Oscillation (ENSO) is known for developing CGCMs is that, with increasing physical to be the strongest climate variation on seasonal to in- understanding of ENSO and the representation of all terannual timescales. The ENSO phenomenon, as a cou- relevant processes in the coupled model, comprehensive pled oscillation of the tropical Paci®c ocean and at- dynamical models have more potential for increased mosphere, has widespread and systematic in¯uence on ENSO forecast skill than purely statistical models and the ocean±atmosphere system (Trenberth et al. 1998). simpli®ed dynamical models. The embedded quasiperiodicity in this phenomenon The ®rst successful ENSO prediction using a CGCM suggests that ENSO may be predictable with a lead time was made by Latif et al. (1993). This CGCM, consisting of more than a year. Indeed the 10 years of the Tropical of a tropical ocean general circulation model (OGCM) Ocean Global Atmosphere (TOGA) program introduced and a global low-resolution atmospheric general cir- a hierarchy of ENSO prediction schemes, which out- culation model (AGCM), was initialized through sep- perform the persistence forecast in predicting gross in- arate spinup integrations of the ocean and atmospheric dices of ENSO on lead times of several months to a models in which only observed wind stress anomalies year (Latif et al. 1998). were needed and, then, integrated forward in coupled The ENSO forecast schemes include both statistical mode without using ¯ux correction. Another CGCM techniques and dynamical models. With regard to dy- ENSO forecasting system was developed by Kirtman et namical forecasting of ENSO, there are typically two al. (1997). This model consists of a high-resolution Pa- prototypes of model systems: one is the intermediate ci®c basin ocean model and a global AGCM. Ocean model, which is designed to simulate only the anomaly initial conditions were derived through an iterative wind part of the ocean±atmosphere system while treating the stress initialization procedure. An anomaly coupling mean state of the system as prescribed. The other is the strategy was employed, with the surface zonal wind comprehensive coupled ocean±atmosphere general cir- stress anomalies determined empirically from the culation model (CGCM), which is intended to simulate AGCM 850-hPa zonal wind anomalies. the full behavior of the system, that is, both the mean The above coupled models did not use any subsurface ocean measurements in initializing predictions. Results from both intermediate models (Kleeman et al. 1995) Corresponding author address: Dr. Guomin Wang, BMRC, GPO Box 1289K, Melbourne, VIC 3001, Australia. and general circulation models (Rosati et al. 1997; Ji E-mail: [email protected] and Leetmaa 1997; Ji et al. 1998) have demonstrated

᭧ 2002 American Meteorological Society 976 MONTHLY WEATHER REVIEW VOLUME 130 that subsurface ocean data have a positive impact on form is applied in the upper part of the atmosphere with the prediction skill. Forecasts for the period 1982±88 pressure Ͻ75 hPa. Gravity wave drag is determined from the Geophysical Fluid Dynamics Laboratory using the formulation of Palmer et al. (1986). Shortwave (GFDL) CGCM, which consists of a high-resolution and longwave radiation schemes follow the methods, global ocean model coupled to an AGCM, indicate that respectively, of Lacis and Hansen (1974) and Schwarz- data assimilation is crucial in obtaining skillful forecasts kopf and Fels (1991) with slight modi®cations. Pene- (Rosati et al. 1997). Several versions of the National trative convection is treated by the mass ¯ux scheme of Centers for Environmental Prediction (NCEP) CGCM, Tiedtke (1989), but without inclusion of momentum ef- which consist of a high-resolution Paci®c basin OGCM fects. Shallow convection is parameterized in terms of coupled to a modi®ed version of an operational weather the model's vertical diffusion scheme following Tiedtke forecasting model, treating the surface heat ¯ux in a (1988). Cloud fractions are calculated diagnostically in combination of full and anomaly coupling, have been three distinct layers. A simple sea ice model is included developed over the past few years (Ji and Leetmaa 1997; in the AGCM (Colman and McAvaney 1992). Further Ji et al. 1998; among others). All NCEP models use model details are described by Hart et al. (1988, 1990) ocean initial conditions derived from a full ocean data and McAvaney and Hess (1996) and references therein. assimilation system. Evaluation of ENSO prediction results, from the latest version of the model (CMP12), b. Ocean model show that the improved coupled model, along with more accurate ocean initial conditions, gives higher predictive The OGCM component of the coupled model system skill. is a global version of the GFDL Modular Ocean Model This paper presents results of the Australian Bureau (Pacanowski et al. 1991; Power et al. 1995). The res- of Meteorology Research Centre (BMRC) efforts to de- olution near the equator and near the surface has been velop an ENSO forecast system based on an ocean anal- enhanced to allow better representation of the seasonal ysis/assimilation component and a CGCM component. variation of the mixed layer and the equatorial wave We will show that our coupled model exhibits relatively guide, which is thought to be crucial to the dynamics small systematic biases in SST in the tropical Paci®c of ENSO. The grid spacing is 2Њ in the zonal direction, even though ¯ux corrections have not been adopted. We while meridional spacing varies from 0.5Њ within 7Њ of will also show that the use of ocean data assimilation the equator, smoothly changing to a maximum of 5.8Њ in deriving oceanic initial conditions can signi®cantly near the North Pole. There are 25 vertical levels, with improve ENSO prediction skill, consistent with results 12 concentrated in the top 185 m of the ocean. The described above. Section 2 describes the component bathymetry is represented by a smoothed approximation models used in the system. The BMRC ocean thermal to the high-resolution dataset of Gates and Nelson data analysis and assimilation scheme is presented in (1975) with a maximum depth of 5000 m. In order to section 3. Sections 4 and 5 show the hindcast forecast save computational cost, only Australia (combined with design and results emphasizing the coupled model's per- New Guinea), New Zealand, and Antarctica are sepa- formance in predicting sea surface and subsurface tem- rated from the remaining land points. No-slip boundary perature anomalies in the tropical Paci®c. Some sensi- conditions are used along all sidewalls. The vertical tivities of the forecast skill of the coupled model are diffusion is determined via a mixing scheme based on given in section 6 and are followed by conclusions in the scheme described by Chen et al. (1994). The scheme section 7. consists of two parts: a turbulent kinetic energy equation for the surface mixed layer, and a gradient Richardson number scheme for the ocean interior. Further details 2. The component models on the ocean model are described by Power et al. (1995). a. Atmospheric model A very similar OGCM has been used in studying dynamics and thermodynamics of the Indian Ocean (Schil- The global AGCM used in the coupled system is an ler et al. 1998). R21L9 (roughly 5.6Њ latitude ϫ 3.2Њ longitude in the horizontal and nine levels in the vertical) version of the c. Coupling strategy BMRC climate model. This model has been used in various climate and climate change studies. The simu- The coupling strategy is described in Power et al. lation of present-day climate by a closely related version (1998). The AGCM has a time step of 1350 s (64 time of the AGCM with observed sea surface temperature steps per day), while the OGCM time step is 1800 s (48 (SST) is given by Frederiksen et al. (1999). The present time steps per day). A full day is used as the time step version of the model uses boundary layer parameteri- of the ice model. The coupled model proceeds as a series zations and vertical diffusion based on the stability- of 1-day coupling intervals where the atmospheric mod- dependent bulk formulations of Louis et al. (1981). A el runs separately for 1 day with ®xed SSTs and sea ice linear sixth-order horizontal diffusion is applied on lev- extents, accumulating relevant time-mean surface ¯ux- els with pressure Ͼ75 hPa, while a linear second-order es. These are then used to drive the corresponding 1- APRIL 2002 WANG ET AL. 977 day integration of the ocean and sea ice models, fol- The top boundary condition of the OGCM during the lowing which the updated SSTs and sea ice extents are assimilation integration are speci®ed from the observed passed back to the atmosphere model for the next cou- wind stress derived from the Florida State University pling interval. (FSU) pseudostress (Goldenberg and O'Brien 1981) and The quantities needed to drive the ocean are wind observed SST (Reynolds and Smith 1994). stress (applied at both open ocean and sea ice points), freshwater ¯ux (i.e., precipitation minus evaporation), and downward energy ¯ux, composing net shortwave b. The data and longwave radiation, and turbulent ¯uxes of sensible The temperature data are a consolidated set from var- and latent heat. All components of the energy ¯ux are ious real-time and delayed-mode sources and from var- assumed to be absorbed exclusively in the top layer of ious special projects and data archives. the ocean with the exception of shortwave radiation, a portion (45%) of which penetrates to deeper layers ac- Real time: Includes all subsurface data received at cording to the scheme of Paulson and Simpson (1977). the Australian Bureau of Meteorology via the Glob- In addition, surface friction velocity (proportional to the al Telecommunication System (GTS) since around cube of the surface wind speed) is provided for use in 1987. the mixing scheme. No ¯ux corrections are made to Near±real time: Through the agencies of the U.S. these interfacial forcing ®elds. National Oceanographic Data Center, the Global A preliminary coupled simulation using the CGCM Temperature Salinity Pro®le (nee Pilot) Project, the (with slightly different atmospheric physics) exhibits Canadian Marine Environmental Data Service, and signi®cant interannual variability and a number of the the Tropical Atmosphere-Ocean array (TAO), we essential ingredients of the delayed action oscillator in have compiled further data that have been subject the tropical Paci®c (Power et al. 1998). to some quality control. In cases of duplication we give this stream precedence over the real-time 3. The ocean analysis and assimilation system stream. Data assembly centres: The World Ocean Circulation a. The assimilation method Experiment established a system of data assembly The ocean data assimilation method is based on the centres for upper ocean thermal data. As these data BMRC ocean analysis system (Smith et al. 1991; Smith have been made available they have been added to 1995a,b; Smith and Meyers 1996). The scheme is a the database and given precedence over the real- univariate implementation of the statistical interpolation time and near-real-time streams. In the case of method described by Lorenc (1986). The data are an- TAO, their quality controlled dataset is used. alyzed in 10-day bins ignoring the temporal distribution Levitus's World Ocean Atlas 1992: The data de- of the data within that bin and using the ocean model scribed in Levitus and Boyer (1994) have also been ®elds at the start of the 10-day period T model as an initial added to the database. These constitute the prin- estimate (®rst guess). The merging of the model ®eld cipal source of data for the 1980s. It was assumed and data is done according to they had been subject to the same levels of quality

anal model obs model control as the scienti®c data assembly centres. T ϭ T ϩ⌺wm(T Ϫ T ), where the sum is over all observations m, T obs is the observed value, wm are the optimum interpolation anal c. Quality control weights, and T is the analyzed value. The wm depend upon the distribution of data about the target model grid Quality control is handled through the objective pro- point and upon various statistical assumptions used in cedures described in Smith (1991). As noted in that the interpolation (forecast and observation error). This paper and in Smith (1995a,b), the ef®cacy of quality method is applied separately for each model level and control is strongly dependent on the quality of infor- only for temperature. No further adjustments are made mation provided and, in particular, on the accuracy of to the model state, which does lead to some initialization the forecast. When the ``model'' is climatology, we are shock subsequent to the assimilation step. faced with a poor estimate in extreme conditions (e.g., The statistical parameterizations are similar to those El NinÄo) and this does lead to bad decisions with respect adopted in the above-mentioned papers. The model er- to rejections. A dynamical model alleviates this problem rors are assumed to be anisotropic in the equatorial re- to some extent since the model is better able to capture gion with scales of 1500 km east±west and 250 km (predict) extreme events given accurate wind forcing. north±south. Outside Ϯ10Њ latitudes the scales are 500 However this advantage can be offset by the intrusion km and isotropic. The noise-to-signal ratio is kept at of systematic errors into the data assimilation process around 1. The observation errors are assumed to be in which case biased estimates may lead to erroneous isotropic with spatial and temporal scales of 150 km decisions. and 5 days, respectively, and amplitude 1ЊC. Our experience with the present system and within 978 MONTHLY WEATHER REVIEW VOLUME 130 the operational environment described in Smith (1995b) suggests the methodology is effective but imperfect. The European Centre for Medium-Range Weather Forecasts (ECMWF) experience with this scheme (Stockdale et al. 1998) has been similar.

4. Hindcast experiment design a. Initial conditions The initial conditions for the atmosphere are provided by driving the AGCM component model with observed surface forcing up to the start time of the forecast. The forcing consists of SSTs from Reynolds and Smith (1994) and sea ice extents inferred from the SST anal- FIG. 1. (a) SST NinÄo-3 index climatologies from observations (dashed line) and the hindcasts for each start month (solid lines). (b) ysis. The ocean initial conditions are provided by the SST NinÄo-3 biases, i.e., the differences between the hindcast and BMRC ocean data assimilation system as described in observational climatologies. section 3. In this paper results from 60 hindcasts of 1-yr integrations are presented. These experiments start from the same manner. We use the same 15 years of hindcast last day of the selected months, February, May, August, period to de®ne the observational climatologies. and November, for the years 1981 through 1995. In the subsequent sections, forecast lead time is de®ned as the time lag between the initial month and the target month. 5. Results For example an August forecast with lead time of 6 In this section we ®rst focus on the SST ®eld and months is the August monthly mean from a coupled look at its bias and forecast skill as a measure of per- model run that is initiated on the last day of the previous formance of the coupled model, followed by an ex- February. amination of subsurface temperature forecast skill. Then we show comparisons of model ENSO composites with b. Postprocessing observations to examine the model's ability to capture the major ENSO response characteristics. As both the AGCM and OGCM have de®ciencies in their respective stand-alone mode simulations and no ¯ux corrections are used in the coupling procedure, a a. SST bias systematic drift from reality is expected to occur during The model hindcast and observed SST NinÄo-3 index the course of the coupled model's forecasts. Hence the (SST averaged over the region 5ЊS±5ЊN and 90Њ± forecasts from the coupled model are bias corrected be- 150ЊW) climatologies and the corresponding bias for fore being validated against observation. Stockdale each start month are shown in Fig. 1. The overall cou- (1997) notes that removal of the coupled model bias pled model SST NinÄo-3 means show a very weak sea- should make the corrected forecasts a better estimate sonal cycle (Fig. 1a). The bias is dominantly positive for the expected variability. However it is likely that and peaks around September and October with ampli- only the linear part of the bias can be estimated with tude about 1.5ЊC, similar to the observed amplitude of con®dence in this manner (Stockdale 1997). The bias interannual SST variability (Fig. 1b). The bias is close is estimated from the mean drift of the forecasts, as a to the negative of the annual cycle for forecasts starting function of the initial months and the forecast lead times. in February and May. Those starting in August and The model drift is de®ned as the difference of model November follow the seasonal cycle reasonably well for climatology and observational climatology. The model the ®rst 6±8 months, but then the model remains in a SST climatology for a given start month, for example, warm state when reality cools. This means that the east- is obtained by averaging all predictions from the start ern Paci®c cold phase is not well simulated. Based on month of all 15 years and over all lead times as follows: these results we opted to remove the bias as an a pos- teriori correction to the forecast, following Stockdale 1 N (1997). k k SSTclimatology (␶) ϭ SSTj (␶), Shown in Fig. 2 is the spatial pattern of systematic N ͸ jϭ1 error of SST in the tropical Paci®c at lead times of 3, where the summation goes from 1981 to 1995 so that 6, 9, and 12 months for the February start case. The k N ϭ 15 andSSTj (␶) is the ␶ lead time monthly mean largest bias occurs in the eastern Paci®c and during the forecasts initialized in month k of year j. The model cold phase of the annual cycle (Figs. 2b and 2c, cor- climatologies for other variables are calculated in the responding to forecast months of August and November, APRIL 2002 WANG ET AL. 979

FIG. 2. Mean SST bias in the tropical Paci®c for all Feb start forecasts at lead times of (a) 3, (b) 6, (c) 9, and (d) 12 months. Contour interval is 1ЊC and areas with bias exceeding 1ЊC are shaded. respectively), while only a minor bias exists during the dictions track the observed anomaly well for the ®rst warm phase (see Figs. 2d and 2a, for forecast months half year of hindcasts, although there are some notable of February and May, respectively). In the subtropics a exceptions. The forecasts initialized before the peak general cooling in the North Paci®c and warming in the phase of the 1982/83 ENSO event produce colder than South Paci®c are seen to develop, which seems to be observed anomalies, while forecasts started just after an inherent bias of our coupled model (Power et al. the peak phase produce too warm anomalies. Several 1998). However these might be expected to have only forecasts, such as the ones initialized in November 1986 a small impact on the ENSO forecasts, which are the and May 1988, show very large deviations from the major focus of this paper. In the following analyses the observation even at short lead times. Forecasts initial- systematic errors have been removed from the original ized in the late 1980s and early 1990s tend to be biased hindcasts. toward warmer conditions and show larger case to case variability. During the 1992±95 period, the forecast SST anomalies are often of the wrong sign, although the b. SST forecast skill actual amplitude is relatively small. The 1-yr evolution of the SST anomalies in the NinÄo- Two commonly used metrics of forecast skill are the 3 region for all hindcasts is shown in Fig. 3. Most pre- anomaly correlation coef®cient (ACC) and the root-

FIG. 3. Time evolution of the NinÄo-3 SST anomalies of all hindcast cases (black lines, thick for Feb start forecasts) and the observation (gray line). Year ticks indicate Jan of each year. 980 MONTHLY WEATHER REVIEW VOLUME 130

maintained at a level of about 70% of the observed magnitude. Interestingly the level of interannual variability seems related to the level of forecast skill for the NinÄo-3 SST as shown in section 6b. It is dif®cult to quantitatively compare the skill scores from this model to those from other forecast systems because of the large diversity in the models' formulation and treatment of coupling, as well as in data initialization methods. However a general comparison can be made against the recent results presented by Kirtman et al. (2000) in which an assessment of the current status of NinÄo-3 SST anomaly forecast skill from six dynamical and two statistical systems is conducted in a uniform format. Here we only compare our model results with that of the Center for Ocean±Land±Atmosphere Studies (COLA) model and NCEP model, both of which are comprehensive coupled GCMs. To be consistent with Kirtman et al. (2000) the same common 8-yr subset (1982±84, 1986±89, and 1991) is used. Our results are based on 32 cases initialized 3 months apart while COLA and NCEP results are on 96 cases initialized 1 month apart. The ACC skills from COLA, NCEP, and the present models are, respectively, 0.70, 0.79, and 0.77 at lead times of 6 months, and 0.66, 0.69, and 0.68 at lead times of 9 months. This indicates that our model performance is comparable to that of the two other CGCMs in terms of predicting a gross SST anomaly index. The spatial distribution of the SST anomaly correlation coef®cient at a lead time of 6 months for the coupled model forecasts and persistence is shown in Fig. 5. The coupled model has the largest correlation skill over the central to eastern Paci®c and within 10Њ of the equator (Fig. 5a). Unlike many intermediate models, FIG.4.NinÄo-3 SST anomaly (a) correlation coef®cient, (b) rms such as Kleeman et al. (1995) and Chen et al. (1995), error, and (c) standard deviation (STD). The solid and dashed lines, the forecast skill in the far eastern Paci®c is relatively respectively, indicate results for the coupled model forecast and per- poor. This region in our model has the largest SST bias sistence in (a) and (b) and for the coupled model forecast and ob- (see Fig. 2). It is an area where both oceanic processes servation in (c). (such as horizontal advection, mixed layer entrainment, upwelling) and atmospheric processes (low-level stratus mean-square (rms) error between observed and pre- cloud formation) contribute to the SST variability. Un- dicted anomalies. Such metrics for the NinÄo-3 SST fortunately most general circulation models deal poorly anomaly are shown in Fig. 4 based on the same 60 cases with some of these processes, even in sand-alone mode used in Fig. 3. For comparison, skills for persistence with prescribed forcing (Schiller et al. 2000; Saji and forecast are also shown. It is seen that for all lead times Goswami 1997; Haskins et al. 1995). The skill of per- our coupled model forecast is more skillful than persistence, shown in Fig. 5b, is signi®cantly lower than sistence in terms of both ACC and rms error metrics. that achieved by the coupled model. The skillful area Using a correlation of 0.6 as a minimum value for useful just around and east of the date line is a sensitive area forecast skill, Fig. 4 indicates that the forecast is skillful having signi®cant regional and global impacts on cli- for lead times up to about 8 months. mate and weather (Hoerling and Kumar 1997; Nicholls As indicated in Fig. 1 the modeled seasonal cycle is 1989; Barsugli et al. 1999). We will further stress this too weak. What does the modeled interannual variability point in section 6a. look like? We produce in Fig. 4c the observed and the model-estimated NinÄo-3 SST standard deviations as a c. Heat content forecast skill measure of interannual variability. After the ®rst few months during which the coupled model seems to un- The existence of independent subsurface thermal data dergo a fast adjustment from a likely initial imbalance, analyses from the BMRC ocean analysis system (Smith the estimated model interannual variability has been 1995b; see section 3 of this paper) allows similar sub- APRIL 2002 WANG ET AL. 981

FIG. 5. Spatial pattern of SST anomaly correlation at a lead time of 6 months for (a) the coupled model forecast and (b) persistence. Contours are drawn from 0.5 with a contour interval of 0.1. surface temperature skill estimates as done for SST correlation mainly results from the fact that the warm above. We validate the forecast anomaly of heat content, pool SST is strongly in¯uenced by mixed layer pro- de®ned as the vertically integrated temperature of the cesses driven by surface ¯uxes associated with atmo- upper 400 m, against the BMRC analysis. Figure 6 shows spheric convective variability on a range of shorter time- the heat content ACC for both forecast and persistence scales (Fasullo and Webster 1999). But the existence of at lead time of 6 months. Interestingly the maximum skill a strong signal underneath the surface of the warm pool, for the heat content is mainly in the western Paci®c and with comparable forecast skill as the SST in the central off the equator, a pattern reminiscent of Rossby wave Paci®c, provides a promising perspective for enhancing incidence and Kelvin wave re¯ection on the western SST skill in this region. boundary, as implied by delayed action oscillator theory. The ocean temperature forecast skill tends to be sus- d. ENSO composite tained longer in the subsurface than at the surface. Fig- ure 7 shows a depth±longitude section of temperature The above analysis demonstrates that there is useful forecast and persistence ACCs along the equator at lead SST forecast skill in the central tropical Paci®c in our times of 6 and 12 months. The skillful region at 6-month coupled model. Observational studies have shown that lead time is near the top layer in the central Paci®c and SST anomalies in the Tropics can have a signi®cant extends into the western Paci®c at a depth of about 150 in¯uence on global climate. It is interesting to examine m. At 12-month lead time, the surface levels have vir- the ability of the CGCM to capture regional and global tually no skill but in the western Paci®c at a depth of responses associated with tropical Paci®c SST anoma- 100±200 m there is still a signi®cant area with ACC lies. Due to the presence of systematic errors in the greater than 0.5. The ACC skill from persistence is less model and unpredictable internal atmospheric noise, the than 0.5 almost everywhere at the given lead times. common skill measurements, such as the ones used in The subsurface skill in the subsurface of the western the previous section, are poor indicators of model skill. Paci®c does not show up at the surface because of the Instead, an ENSO composite analysis of various ®elds well-known weak correlation between the subsurface from both the forecasts and observations, which en- and the surface temperature in this area. Such a weak hances the signal-to-noise ratio, will be presented here,

FIG. 6. As in Fig. 5 but for ocean heat content, de®ned as the mean integration of temperature of upper 400 m. 982 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 7. Depth±longitude section along equator of anomaly correlation at a lead times of (left panels) 6 and (right panels) 12 months for the coupled model (upper panels) forecasts and (bottom panels) persistence. Contours are drawn from 0.5 with a contour interval of 0.1. in an attempt to understand some of the attributes of boreal winter season (December, January, and February) the coupled model in response to ENSO. from August start forecasts. The composites are thus the There are three warm (1982/83, 1986/87, and 1991/ two-season lead forecasts initialized a few months prior 92) and two cold (1984/85 and 1988/89) ENSO events to the mature phase of ENSO. during the hindcast period. We form composites for the Figure 8 shows the SST anomaly composites for

FIG. 8. ENSO SST anomaly composites for the (left panels) observed and (right panels) coupled model forecast (upper panels) warm events and (bottom panels) cold events. Event de®nitions are given in the text. The veri®cation is based on Reynolds reanalysis (Reynolds and Smith 1994). Contours are drawn from Ϯ0.5ЊC with contour interval of 0.5ЊC. Positive and negative values are shaded by heavy and light shading, respectively. APRIL 2002 WANG ET AL. 983

FIG. 9. As in Fig. 8 but for subsurface temperature anomaly composites. Contours are drawn from Ϯ0.5ЊC with contour interval of 1.0ЊC. The BMRC operational ocean analyses (Smith 1995b) are used to construct the veri®cation composite. Positive and negative values are denoted by heavy and light shading, respectively. warm and cold events. It indicates that the amplitude tions lack amplitude away from the equator particularly of the warm SST anomaly produced by the coupled in the Northern Hemisphere where the strongest re- model is about half of, but with spatial extent similar sponse is expected to occur during the winter season to, the observations. In contrast, the simulated cold (see panels a and c in Figs. 10 and 11). event composite is close to the observations in its Global composites for mean sea level pressure, 500- strength. Although we have only considered a small hPa height, and 200-hPa zonal wind are presented in number of cases in our composites, they show that the Figs. 12±14, respectively. Over the tropical Paci®c and model distribution of equatorial Paci®c SST anomalies Indian Oceans, a seesaw-type pattern in surface pressure has a different skewness to that observed; that is, the is evident in the forecast and in the veri®cation, but the model does not capture the full strength of the warm signals diminish quickly toward the Indian Ocean in the events. This is a feature commonly seen in the majority forecast (Fig. 12). In the upper troposphere the upper of coupled ocean±atmosphere ENSO forecast models branch of the Walker circulation weakens during warm [e.g., see Landsea and Knaff (2000) for forecasts of the periods (and strengthens during cold phases) in both the 1997/98 El NinÄo]. In the subtropical western Paci®c the forecast and veri®cation, but the magnitude in the fore- anomalies are of opposite sign in the observations and cast is only half of that in the veri®cation (Fig. 14). almost nonexistent in the prediction. The composite for This is consistent with de®ciencies in the precipitation the subsurface thermal anomaly is given in Fig. 9. The and zonal wind stress responses (Figs. 10 and 11), in- overall pattern of anomalous thermal structure and re- dicating too weak convection in the model atmosphere. versal between warm and cold event are well reproduced The anomalous upper-level zonal wind associated with by the model, but the strength is still too weak, espe- tropical convection forcing is seen to propagate away cially for the warm composite. In addition the subsurface anomaly does not extend as deep as the analyses. from the equator into midlatitudes of both hemispheres The two critical variables of the atmospheric response but the responses are weak poleward of about 30Њ.In to the tropical SST anomaly are the zonal wind stress the Paci®c±North American regions the responses and precipitation. Their composites are shown in Figs. shown in the veri®cation are strong, while in the forecast 10 and 11, respectively. Along the equator the predic- the responses are much weaker. The centers of major tions contain many features of ENSO: a reduced easterly maxima and minima are also located differently (Figs. zonal wind stress and enhanced precipitation accom- 12 and 13). Given the relatively weak tropical forcing panying the eastward shift of the intertropical conver- in our coupled model, particularly for warm events as gence zone (ITCZ) during warm periods in the central seen in the precipitation composites (Fig. 11), it is not Paci®c, and the reverse during cold phases, although the surprising that the overall extratropical response pat- amplitude and extent during the warm events is less terns are weak and shifted somewhat relative to their than observed, consistent with the de®ciencies in the observational counterparts, as they are largely depen- SST anomalies discussed above. However the predic- dent on both strength and location of the tropical forc- 984 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 10. As in Fig. 8 but for zonal wind stress anomaly composites. Contours are drawn from Ϯ1 ϫ 10Ϫ2 NmϪ2 with contour interval of 1 ϫ 10Ϫ2 NmϪ2. Positive and negative values are denoted by heavy and light shading, respectively. The veri®cations from this ®gure through Fig. 14 are derived from the NCEP±NCAR reanalyses (Kalnay et al. 1996). ing, as demonstrated by Hoerling and Kumar (1997) (®gures not shown). This implies that the ``spring pre- and Hoerling et al. (1997). diction barrier'' may not be a signi®cant factor in our The composite results indicate that our coupled model model. In the NCEP coupled model (Ji et al. 1996) the is capable of reproducing many basic features of ENSO seasonal skill dependence compared to the previous ver- in the tropical Paci®c but is unable to capture desirable sions tended to become less apparent with continuous responses in midlatitudes both in strength and pattern. improvement in various parts of the coupled model sys- In addition to weak SST and convection anomalies in tem. The introduction of a new initialization scheme in the coupled model, the weak midlatitude response to the Zebiak and Cane model almost eliminated the spring the tropical forcing is likely to be an inherent feature barrier from their model (Chen et al. 1995). These re- of our AGCM. A similar experiment using an earlier sults might suggest that the spring barrier to ENSO pre- version of the BMRC AGCM forced with observed SST diction may not be an intrinsic feature of the real climate exhibits a weak response as well (McAvaney and Col- system. However the present model does have signi®- man 1993). Thus improvements in atmospheric response cant bias and this might be masking the presence of to ENSO forcing in the AGCM are needed so as to take such effects. full advantage of the global CGCM framework. Fur- On the other hand, our coupled model appears to show thermore coupled model multimember ensemble runs a reduction in skill of SST anomaly forecast in the east- are also needed to enhance the signal-to-noise ratios in ern Paci®c from the 1980s to the early 1990s. Figure the extratropical regions as was shown by Stockdale et 15a compares NinÄo-3 SST anomaly correlation skill cal- al. (1998). culated for cases initialized during the earlier hindcast period (1981±88) with that from the later period (1989± 6. Skill sensitivities 95). In the eastern Paci®c both the forecast skill and the persistence skill in the earlier period (dark lines in Fig. a. Seasonal and decadal variations 15a) are better than that based on all cases (see Fig. 4). For lead times of 0±8 months the SST anomaly fore- However the skill for the later period (light lines in Fig. cast skill from our model does not show clear seasonal 15a) is much worse. Such skill variation is also observed dependence. After 8 months the skill decreases more in hindcast results from other dynamical models (e.g., rapidly for forecasts initialized at August and November Chen et al. 1995; Ji et al. 1996). There is no consensus APRIL 2002 WANG ET AL. 985

FIG. 11. As in Fig. 8 but for precipitation anomaly composites. Contours are drawn from Ϯ1 mm dayϪ1 with contour interval of 1 mm dayϪ1. Positive and negative values are denoted by heavy and light shading, respectively.

FIG. 12. As in Fig. 8 but for the global mean sea level pressure anomaly composites. Contours are drawn from Ϯ0.5 hPa with contour interval of 1.5 hPa. Positive and negative values are denoted by heavy and light shading, respectively. 986 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 13. As in Fig. 8 but for the global 500-hPa height anomaly composites. Contours are drawn from Ϯ5 m with contour interval of 15 m. Positive and negative values are denoted by heavy and light shading, respectively. on the origins of such skill variation over decadal time- variation between the earlier and the later periods. Fig- scales. It might be attributed to the decadal variation of ure 5 shows that the skill in the NinÄo-4 region is con- autocorrelation of tropical Paci®c SST anomalies (Bal- centrated in the eastern two-thirds of that region, with maseda et al. 1995) or to the presence of an SST decadal the peak skill in the central Paci®c. mode (Latif et al. 1997). Interestingly, we notice that As has been demonstrated by many studies the skill when the anomaly correlation skill for the NinÄo-4 region variation is closely related to the strength of signal and is examined, as shown in Fig. 15b, there is little skill noise in these indices. Figure 15c gives the observed

FIG. 14. As in Fig. 8 but for the global 200-hPa zonal wind anomaly composites. Contours are drawn from Ϯ1msϪ1 with contour interval of 2 m sϪ1. Positive and negative values are denoted by heavy and light shading, respectively. APRIL 2002 WANG ET AL. 987

(see Fig. 2). It may not be the decadal variability that causes such skill variation but a sensitivity to the systematic errors of the model. The relatively high skill in our model for the 1980s and 1990s in the central equatorial Paci®c has important implications for predictability in terms of tropical forcing of extratropical circulation anomalies (although our present AGCM only weakly captures tropical/extratropical interactions, as discussed in section 5d). The strong forcing of the atmosphere on seasonal and interannual timescales is closely related to latent heat release from tropical large-scale deep convection (Horel and Wallace 1981). In both warm and cold ENSO phases the convection anomalies, which are a function of both SST anomaly and total SST, mainly reside in the central Pa- ci®c (Hoerling et al. 1997; or see Figs. 11a and 11c in this paper). The convection anomalies in this region can create signi®cant impact in various regions of the world (Livezey et al. 1997; Trenberth et al. 1998; among others). It is also well established that there is a strong relationship between SST in the central Paci®c and win- tertime rainfall in eastern Australia (Nicholls 1989).

b. Sensitivity to ocean initial conditions We have demonstrated in preceding sections that our coupled model has skill in predicting SST anomalies in the central tropical Paci®c, comparable with that of other coupled models. It is interesting to examine how such skill is sensitive to changes to the physics of the model and to changes in the initial condition used to produce the forecasts. These aspects of sensitivity can tell us how well ENSO-related processes are represented and how much the skill might be enhanced by improving FIG. 15. (a) NinÄo-3 SST anomaly correlations of the forecast (solid lines) and the persistence (dashed lines) for separate periods of years physics and initial conditions. These issues are the sub- 1981±88 (dark lines) and years 1989±95 (light lines). (b) Same as in ject of future research and here only a brief discussion (a) but for NinÄo-4 SST anomaly correlations. (c) The observed NinÄo- on skill sensitivity to ocean data assimilation will be 3 (dark lines) and NinÄo-4 (light lines) SST standard deviations (ЊC) given. for separate periods of years 1981±88 (solid lines) and years 1989± We compare results from ``control'' coupled forecasts 95 (dashed lines) displayed as in (a) and (b). (CN), which are initialized by an ocean-only run that utilizes ocean subsurface data and SST, and which have NinÄo-3 and NinÄo-4 SST standard deviations calculated been described in previous sections, and forecasts ini- and displayed in the same manner as in Figs. 15a and tialized from an ocean-only run without assimilation of 15b. The NinÄo-3 region experiences a dramatic change subsurface thermal data (NA). These two ocean-only with the standard deviation in the earlier period being runs also differ in the strength of the relaxation of the more than twice that in the later period (dark solid and model SST to observed SST (relaxation timescale was dotted lines in Fig. 15c), whereas the NinÄo-4 region 3 days for CN experiments and 20 days for NA exper- undergoes less dramatic change (light solid and dotted iments). Figure 16 summarizes the hindcast skill mea- lines in Fig. 15c). This result is consistent with the ®nd- sured by the anomaly correlation of the SST NinÄo-3 ings of Kirtman and Schopf (1998). Extensive predic- index in the two sets of experiments. It is clear from tion experiments using a simple coupled model show Fig. 16 that the SST NinÄo-3 skill in the NA forecasts that during decades with high (low) amplitude of the is signi®cantly worse than that of the CN forecasts. The NinÄo-3 SST variance the forecast skill is relatively high use of different relaxation timescales in generating (low) and the delayed action oscillator (external noise ocean initial conditions for the CN and NA runs should forcing) is the dominant mechanism controlling the in- only have a small impact on forecast skill because the terannual SST variability (Kirtman and Schopf 1998). wind stress (which is the same in both cases) is the main Notice that in the present model the eastern Paci®c is forcing that affects the subsurface thermocline where also the region where there is the largest systematic error most of the skill resides (Fig. 7) and both timescales 988 MONTHLY WEATHER REVIEW VOLUME 130

FIG. 16. NinÄo-3 SST anomaly correlations of the assimilation control run (CN, black), no assimilation run (NA, dark gray), and persistence (dashed line). Thin dashed horizontal line represents the 0.5 correlation value.

are still relatively short compared to interannual timescales. Hence, the increased skill in the CN run is due to the assimilation of subsurface temperature observations. This conclusion is consistent with many other studies, such as Rosati et al. (1997) and Ji and Leetmaa (1997). It is worth pointing out that the interannual variability of the NinÄo-3 SST anomaly index for NA forecasts is weaker than that of the CN forecasts at all lead FIG. 17. Equatorial pro®les of (a) annual mean 20ЊC isothermal times. For example at 9-month lead time the variability depth from BMRC analysis (solid line), CN (dark dashed), and NA for NA is about 50% of observed while for CN it is (gray dashed) and (b) regression of zonal wind stress anomaly against about 70% (not shown). The ocean assimilation will SST anomaly from observations [wind stress from FSU and SST from have two impacts in the initial states: ®rst, to correct Reynolds and Smith (1994), solid line], CN (dark dashed), and NA the interannual temperature anomalies, and second to (gray dashed). correct the model's temperature mean state. Each of these can potentially contribute to the higher forecast skill and the stronger interannual variability as seen in the CN run (also see Fig. 17). From these experiments ocean in the NA run is systematic warmer than that in alone it is not possible to separate the relative contri- the CN run with a thermocline depth difference of about bution of each of these improvements due to the data 10 m in the east Paci®c and 20 m in the west Paci®c. assimilation. This is a topic for future research. The mean thermocline depth difference has a signi®cant A issue of concern is that the skill of the forecasts impact on the coupled system behavior as can be seen initialized without subsurface data is only comparable in Fig. 17b, which shows the local regression of zonal to persistence for the ®rst 8 months. This indicates that wind stress anomaly against normalized SST anomaly either the ocean model is unable to produce the correct along the equator. The observed regression uses FSU initial ocean state for a given wind forcing or there are pseudo stress and Reynolds SST datasets and shows a errors in the wind forcing. Certainly the model is not maximum wind stress regression of above 0.015 N mϪ2 perfect and observations are incomplete. By direct in- per positive unit of standard deviation of SST in the sertion of in situ observational oceanographic data in central Paci®c. The wind stress/SST regression in the the assimilation stage, at least part of these problems CN run is roughly the same shape as the observed but are overcome at initial time and the forecast skill is 30% weaker in the central Paci®c. On the other hand greatly improved. the regression relation for the NA run seems much As the CN and NA forecasts differ only in their initial weaker than both the observed and the CN results; its ocean conditions, it is worthwhile to examine how such maximum location is also biased westward. Detailed differences affect the coupled system. Figure 17a gives analysis of the relationship between the thermocline and an annual mean 20ЊC isothermal depth along the equator wind stress/SST association in our model is beyond the in the Paci®c from the BMRC analysis and from the scope of this paper. But we may infer that the larger two runs. These values are simply the averages of all warm error in the ocean's subsurface introduced at the hindcasts over all lead times and thus are approximately initial time in the NA run makes the SST less sensitive representative of model annual means. They are sys- to changes in the thermocline. One of the consequences tematically biased toward a reduced gradient; the upper is the lack of strong coupled activity in the key area of APRIL 2002 WANG ET AL. 989 the tropical central Paci®c, which in turn adversely af- dress this issue (Kleeman and Moore 1999). Ideally the fects SST evolution and skill. ensemble mean should reduce the uncertainty due to internal variability in the nonlinear ocean±atmosphere system and the ensemble spread should be an indication 7. Conclusions of reliability of the forecast. For the present model, en- This paper presents the Australian Bureau of Mete- semble forecasts produced by randomly perturbing ini- orology Research Centre coupled general circulation tial conditions fail to give such useable a priori estimates model ENSO forecast system and evaluates its perfor- of the uncertainty of a forecast (®gure not shown). Kirt- mance. The system includes an ocean thermal data as- man et al. (2000) reached the same conclusion in a similation component (Smith et al. 1991; Smith systematic evaluation of seven ENSO forecast schemes 1995a,b) and a CGCM component combining an R21L9 by looking at the same relationship of the ensemble version of the BMRC climate AGCM and a global ver- spread versus the ensemble mean derived from forecasts sion of the GFDL modular OGCM (Power et al. 1998). initialized 1 month apart. The failure suggests that either Ocean measurements are ingested through a window of the ensemble needs to be generated in a more sophis- 10 days in the assimilation system to provide ocean ticated way (Moore and Kleeman 1998) or other reli- initial conditions for the hindcasts. Hindcasts are per- ability measures should be utilized (Kleeman and Moore formed every season for the period 1981±95 (60 in all). 1999). The model is validated against observed surface and There are planned upgrades of the coupled model subsurface temperature datasets. forecast system focusing on improving the model res- The model exhibits skill comparable to, and in some olution and physics as well as the ocean data assim- cases exceeding that of, other published models. The ilation scheme. With better de®nition of the initial SST skill in the eastern Paci®c (NinÄo-3) region is sus- ocean states and better depiction of various atmo- tained out to lead times of around 8 months but higher spheric and oceanic physics in the new coupled model skill is concentrated in the central Paci®c (eastern two- system we hope that better model performance will thirds of the NinÄo-4) region from where potential sea- be achieved. sonal forecasts can be exploited for many regions. The model seems to show a skill variation from the period Acknowledgments. We wish to thank Scott Power, 1981±88 to the period 1989±95 in the eastern Paci®c Robert Colman, Jim Fraser, and Bryant McAvaney for but not in the central Paci®c judging from a comparison providing some detailed information of the coupled mod- between these two periods. Further analyses suggest that el. We would like to thank Oscar Alves and three anon- the skill variation may be related to the amplitude of ymous reviewers for their constructive comments on the the observed SST variance, which is consistent with the earlier versions of this manuscript. This work has been results of Kirtman and Schopf (1998). assisted by an LWRRDC grant since January 1999. This study is the ®rst to provide a systematic evaluation of subsurface skill. The results suggest that the model draws signi®cant bene®t from the ingestion of REFERENCES subsurface data, consistent with the results of the current Balmaseda, M. A., M. K. Davey, and D. L. T. Anderson, 1995: De- BMRC operational model (from Kleeman et al. 1995) cadal and seasonal dependence of ENSO prediction skill. J. Cli- and other CGCMs (Rosati et al. 1997; Ji and Leetmaa mate, 8, 2705±2715. 1997), and that this information tends to be retained Barsugli, J. J., J. S. Whitaker, A. F. Loughe, P. D. Sardeshmukh, and longer than at the surface. Z. Toth, 1999: The effect of the 1997/98 El NinÄo on individual The coupled model skill variation on longer timescales large-scale weather events. Bull. Amer. Meteor. Soc., 80, 1399± 1412. as demonstrated here and in many other papers is a chal- Chen, D., L. M. Rothstein, and A. J. Busalacchi, 1994: A hybrid lenge for ENSO prediction using coupled GCMs. The vertical mixing scheme and its application to tropical ocean mod- fact that NinÄo-3 SST anomaly skill from both coupled els. J. Phys. Oceanogr., 24, 2156±2179. models and persistence in the 1980s is better than that ÐÐ, S. E. Zebiak, A. J. Busalacchi, and M. A. Cane, 1995: An improved procedure for El NinÄo forecasting: Implications for in the early 1990s may re¯ect a possible climate regime predictability. Science, 269, 1699±1702. shift of the tropical Paci®c ocean±atmosphere system. In Colman, R. A., and B. J. McAvaney, 1992: Modelling of polar regions addition to the weakening signal, the larger systematic for climate change experiments. Modelling Weather and Climate: errors in this region in this model could be another factor The Third BMRC Modelling Workshop, Bureau of Meteorology affecting model skill. This may explain why SST anom- Research Centre Res. Rep. 33, 390±406. Fasullo, J., and P. J. Webster, 1999: Warm pool SST variability in aly skill in the NinÄo-4 region shows little sensitivity over relation to the surface energy balance. J. Climate, 12, 1292± different periods of time, because both the signal reduc- 1305. tion and the model errors are small here. Frederiksen, C. S., D. P. Rowell, R. C. Balgovind, and C. K. Folland, Another practical and outstanding question in ENSO 1999: Multidecadal simulations of Australian rainfall variability: The role of SSTs. J. Climate, 12, 357±379. forecasting is how to estimate the reliability of a forecast Gates, W. L., and A. B. Nelson, 1975: A new (revised) tabulation of made, say, by a dynamical model like the present one. the Scripps topography on a 1Њ global grid. Part I: Terrain heights. An ensemble forecasting technique is often used to ad- Rand Corporation Tech. Rep. R-1276-1-ARPA, 132 pp. 990 MONTHLY WEATHER REVIEW VOLUME 130

Goldenberg, S. B., and J. J. O'Brien, 1981: Time and space variability American region atmosphere to large central equatorial Paci®c of tropical Paci®c wind stress. Mon. Wea. Rev., 109, 1190±1207. SST anomalies. J. Climate, 10, 1787±1820. Hart, T. L., M. J. Gay, and W. Bourke, 1988: Sensitivity studies with Lorenc, A., 1986: Analysis methods for numerical weather prediction. the physical parameterizations in the BMRC global atmospheric Quart. J. Roy. Meteor. Soc., 112, 1177±1194. spectral model. Aust. Meteor. Mag., 36, 47±60. Louis, J.-F., M. Tiedtke, J.-F. Geleyn, 1981: A short history of the ÐÐ, W. Bourke, B. J. McAvaney, and B. W. Forgan, 1990: At- PBL parameterisation at ECMWF. Proc. ECMWF Workshop on mospheric general circulation simulations with the BMRC global Planetary Boundary Layer Parameterisation, Reading, United spectral model: The impact of revised physical parameteriza- Kingdom, European Centre for Medium-Range Weather Fore- tions. J. Climate, 3, 436±459. casts, 59±80. Haskins, R. D., T. P. Barnett, M. M. Tyree, and E. Roeckner, 1995: McAvaney, B. J., and R. A. Colman, 1993: The AMIP experiment: Comparison of cloud ®elds from atmospheric general circulation The BMRC AGCM con®guration. Bureau of Meteorology Re- model, in situ and satellite measurements. J. Geophys. Res., 100, search Centre Res. Rep. 38, 43 pp. 1367±1378. ÐÐ, and G. D. Hess, 1996: The revised surface ¯uxes parametris- Hoerling, M. P., and A. Kumar, 1997: Origins of extreme climate ation in BMRC: Formulation. Bureau of Meteorology Research states during the 1982±83 ENSO winter. J. Climate, 10, 2859± Centre Res. Rep. 56, 27 pp. 2870. Moore, A. M., and R. Kleeman, 1998: Skill assessment for ENSO ÐÐ, ÐÐ, and M. Zhong, 1997: El NinÄo,LaNinÄa, and the nonlin- using ensemble prediction. Quart. J. Roy. Meteor. Soc., 124, earity of their teleconnections. J. Climate, 10, 1769±1786. 557±584. Horel, J. D., and J. M. Wallace, 1981: Planetary scale atmospheric Nicholls, N., 1989: Sea surface temperature and Australian winter phenomena associated with the Southern Oscillation. Mon. Wea. rainfall. J. Climate, 2, 965±973. Rev., 109, 813±829. Pacanowski, R. C., K. Dixon, and A. Rosati, 1991: The GFDL mod- Ji, M., and A. Leetmaa, 1997: Impact of data assimilation on ocean ular ocean model users guide, version 1.0. Geophysical Fluid initialization and El NinÄo prediction. Mon. Wea. Rev., 125, 742± Dynamics Laboratory Ocean Group Tech. Rep. 2, 16 pp. 753. Palmer, T. N., G. J. Shutts, and R. Swinbank, 1986: Alleviation of a ÐÐ, ÐÐ, and V. E. Kousky, 1996: Coupled model forecasts of systematic westerly bias in general circulation and numerical ENSO during the 1980s and 1990s at the National Meteorolog- weather prediction models through an orographic gravity wave ical Center. J. Climate, 9, 3105±3120. drag parameterization. Quart. J. Roy. Meteor. Soc., 112, 1001± ÐÐ, D. W. Behringer, and A. Leetmaa, 1998: An improved coupled 1039. model for ENSO prediction and implications for ocean initial- Paulson, C. A., and J. J. Simpson, 1977: Irradiance measurements in ization. Part II: The coupled model. Mon. Wea. Rev., 126, 1022± the upper ocean. J. Phys. Oceanogr., 7, 952±956. 1034. Power, S. B., R. Kleeman, F. Tseitkin, and N. R. Smith, 1995: On a Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Re- global version of the GFDL modular ocean model for ENSO analysis Project. Bull. Amer. Meteor. Soc., 77, 437±471. studies. Bureau of Meteorology Research Centre Tech. Rep., 18 Kirtman, B. P., and P. S. Schopf, 1998: Decadal variability in ENSO pp. predictability and prediction. J. Climate, 11, 2804±2822. ÐÐ, F. Tseitkin, R. A. Colman, and A. Sulaiman, 1998: A coupled ÐÐ, J. Shukla, B. Huang, Z. Zhu, and E. K. Schneider, 1997: Mul- general circulation model for seasonal prediction and climate tiseasonal predictions with a coupled tropical ocean±global at- change research. Bureau of Meteorology Research Centre Res. Ref. 66, 52 pp. mosphere system. Mon. Wea. Rev., 125, 789±808. Reynolds, R. W., and T. M. Smith, 1994: Improved global sea surface ÐÐ, ÐÐ, M. Balmaseda, N. Graham, C. Penland, Y. Xue, and S. temperature analysis using optimum interpolation. J. Climate, 7, Zebiak, 2000: Current status of ENSO forecast skill. A Report 929±948. to the Climate Variability and Predictability (CLIVAR) Numer- Rosati, A., K. Miyakoda, and R. Gudgel, 1997: The impact of ocean ical Experimentation Group (NEG), CLIVAR Working Group initial conditions on ENSO forecasting with a coupled model. on Seasonal to Interannual Prediction. [Available online at http: Mon. Wea. Rev., 125, 754±772. //www.clivar.org/publications/wg࿞reports/wgsip/nino3/report.htm.] Saji, N. H., and B. N. Goswami, 1997: Intercomparison of the sea- Kleeman, R., and A. M. Moore, 1999: A new method for determining sonal cycle of tropical surface stress in 17 AMIP atmospheric the reliability of dynamical ENSO predictions. Mon. Wea. Rev., general circulation models. Climate Dyn., 13, 561±585. 127, 694±705. Schiller, A., J. S. Godfrey, P. C. McIntosh, G. Meyers, and S. E. ÐÐ, ÐÐ, and N. R. Smith, 1995: Assimilation of subsurface ther- Wijffels, 1998: Seasonal near-surface dynamics and thermody- mal data into an intermediate tropical coupled ocean±atmosphere namics of the Indian Ocean and Indonesian through¯ow in a model. Mon. Wea. Rev., 123, 3103±3113. global ocean general circulation model. J. Phys. Oceanogr., 28, Lacis, A. A., and J. F. Hansen, 1974: A parameterization for the 2288±2312. absorption of solar radiation in the earth's atmosphere. J. Atmos. ÐÐ, ÐÐ, ÐÐ, ÐÐ, and R. Fiedler, 2000: Interannual dynamics Sci., 31, 118±133. and thermodynamics of the Indo±Paci®c Oceans. J. Phys. Landsea, C. W., and J. A. Knaff, 2000: How much skill was there Oceangr. 30, 987±1012. in forecasting the very strong 1997±98 El NinÄo? Bull. Amer. Schwarzkopf, M. D., and S. B. Fels, 1991: The simpli®ed exchange Meteor. Soc., 81, 2107±2120. method revisited: An accurate, rapid method for computation of Latif, M., A. Sterl, E. Maier-Reimer, and M. M. Junge, 1993: Structure infrared of cooling rates and ¯uxes. J. Geophys. Res., 96, 9075± and predictability of the El NinÄo/Southern Oscillation phenom- 9096. enon in a coupled ocean±atmosphere general circulation model. Smith, N. R., 1991: Objective quality controls and performance di- J. Climate, 6, 700±708. agnostics of an oceanic subsurface thermal analysis scheme. J. ÐÐ, R. Kleeman, and C. Eckert, 1997: Greenhouse warming, de- Geophys. Res., 96, 3279±3287. cadal variability, or El NinÄo? An attempt to understand the anom- ÐÐ, 1995a: The BMRC ocean thermal analysis system. Aust. Me- alous 1990s. J. Climate, 10, 2221±2239. teor. Mag., 44, 93±110. ÐÐ, and Coauthors, 1998: A review of the predictability and pre- ÐÐ, 1995b: An improved system for tropical ocean subsurface tem- diction of ENSO. J. Geophys. Res., 103, 14 375±14 393. perature analyses. J. Atmos. Oceanic Technol., 12, 850±870. Levitus, S., and T. P. Boyer, 1994: Temperature. Vol. 4, World Ocean ÐÐ, and G. Meyers, 1996: An evaluation of XBT and TAO data Atlas 1994, NOAA Atlas NESDIS 4, 117 pp. for monitoring tropical ocean variability. J. Geophys. Res., 101, Livezey, R. E., M. Masutani, A. Leetmaa, H. Rui, M. Ji, and A. 28 489±28 502. Kumar, 1997: Teleconnective response of the Paci®c±North ÐÐ, J. E. Blomley, and G. Meyers, 1991: A univariate statistical APRIL 2002 WANG ET AL. 991

interpolation scheme for subsurface thermal analyses in the trop- scale models. Physically-Based Modelling and Simulation of Cli- ical oceans. Progress in Oceanography, Vol. 28, Pergamon, mate and Climatic Change, Part 1, M. E. Schlesinger Ed., Klu- 219±256. wer Academic, 375±431. Stockdale, T. N., 1997: Coupled ocean±atmosphere forecasts in the ÐÐ, 1989: A comprehensive mass ¯ux scheme for cumulus param- presence of climate drift. J. Climate, 10, 809±818. eterization in large-scale models. Mon. Wea. Rev., 117, 1779±1800. ÐÐ, D. L. T. Anderson, J. O. S. Alves, and M. A. Balmaseda, 1998: Trenberth, K. E., G. W. Branstator, D. Karoly, A. Kumar, N.-C. Lau, Global seasonal rainfall forecasts using a coupled ocean±at- and C. Ropelewski, 1998: Progress during TOGA in understand- mosphere mode. Nature, 392, 370±373. ing and modeling global teleconnections associated with tropical Tiedtke, M., 1988: Parameterization of cumulus convection in large- sea surface temperatures. J. Geophys. Res., 103, 14 291±14 324.