Docstoc

An Improved Coupled Model for ENSO Prediction and Implications for

Document Sample
An Improved Coupled Model for ENSO Prediction and Implications for Powered By Docstoc
					1022                                          MONTHLY WEATHER REVIEW                                                                VOLUME 126




           An Improved Coupled Model for ENSO Prediction and Implications for
                    Ocean Initialization. Part II: The Coupled Model
                                   MING JI, DAVID W. BEHRINGER,                AND   ANTS LEETMAA
                             National Centers for Environmental Prediction, NWS/NOAA, Washington, D.C.
                                  (Manuscript received 1 October 1996, in final form 26 August 1997)

                                                                ABSTRACT
                An improved forecast system has been developed and implemented for ENSO prediction at the National
             Centers for Environmental Prediction (NCEP). This system consists of a new ocean data assimilation system
             and an improved coupled ocean–atmosphere forecast model (CMP12) for ENSO prediction. The new ocean data
             assimilation system is described in Part I of this two-part paper.
                The new coupled forecast model (CMP12) is a variation of the standard NCEP coupled model (CMP10).
             Major changes in the new coupled model are improved vertical mixing for the ocean model; relaxation of the
             model’s surface salinity to the climatological annual cycle; and incorporation of an anomalous freshwater flux
             forcing. Also, the domain in which the oceanic SST couples to the atmosphere is limited to the tropical Pacific.
                Evaluation of ENSO prediction results show that the new coupled model, using the more accurate ocean initial
             conditions, achieves higher prediction skill. However, two sets of hindcasting experiments (one using the more
             accurate ocean initial conditions but the old coupled model, the other using the new coupled model but the less
             accurate ocean initial conditions), result in no improvement in prediction skill. These results indicate that future
             improvement in ENSO prediction skill requires systematically improving both the coupled model and the ocean
             analysis system. The authors’ results also suggest that for the purpose of initializing the coupled model for
             ENSO prediction, care should be taken to give sufficient weight to the model dynamics during the ocean data
             assimilation. This can reduce the danger of aliasing large-scale model biases into the low-frequency variability
             in the ocean initial conditions, and also reduce the introduction of small-scale noise into the initial conditions
             caused by overfitting the model to sparse observations.




1. Introduction                                                         different ocean initialization approach (Ji et al. 1994;
                                                                        Kleeman et al. 1995; Rosati et al. 1997) used not only
   A major accomplishment of the recently completed
                                                                        the surface wind stress information but also observed
international research program, Tropical Ocean and
                                                                        surface and subsurface temperature data by way of
Global Atmosphere (TOGA, 1985–94), was the devel-
opment of the capability to predict the El Nino–Southern
                                             ˜                          ocean data assimilation. This ocean initialization meth-
Oscillation (ENSO) phenomenon. ENSO is the most                         od for a coupled prediction model is similar in spirit to
important known seasonal to interannual climate vari-                   the initialization of atmospheric models for numerical
ation and involves coupled interactions between the                     weather prediction.
tropical Pacific Ocean and the global atmosphere (Bjer-                     Chen et al. (1995) have developed an improved ocean
knes 1969; Wyrtki 1975, 1985). It has been linked to                    initialization scheme for the Zebiak and Cane (1987)
global climatic anomalies (Ropelewski and Halpert                       coupled model. By incorporating a new initialization
1987).                                                                  scheme into their prediction system without any mod-
   Cane et al. (1986) developed the first dynamical cou-                 ification to the coupled model itself, they showed sig-
pled model for the prediction of ENSO. This model and                   nificant improvement in ENSO prediction skill com-
a number of subsequent coupled models used similar                      pared to using the standard Zebiak and Cane (1987)
ocean initialization methods in which the observed sur-                 initialization method. Their result clearly shows that in-
face wind stress in the tropical Pacific was used to force               itialization improves ENSO prediction skill. Ji and Leet-
the ocean model prior to the start of a prediction (Latif               maa (1997) found that initialization of the ocean by the
et al. 1993; Kleeman 1993; Kirtman et al. 1997). A                      assimilation of observed subsurface temperature data
                                                                        generally leads to improved forecast skill. However,
                                                                        with one version of the coupled model, an improved
                                                                        oceanic initialization resulted in degradation of predic-
  Corresponding author address: Dr. Ming Ji, Climate Modeling
Branch, National Centers for Environmental Prediction, 5200 Auth
                                                                        tion skill for hindcasts in the winter season. The skill
Road, Rm. 807, Camp Springs, MD 20746.                                  for predictions started in the winter season was im-
E-mail: ming.ji@noaa.gov                                                proved when changes were made to the coupled model.
APRIL 1998                                            JI ET AL.                                                  1023

However, even with the improved model, data assimi-           also be the result of an improved coupled ocean–at-
lation did not have significant positive impact on the         mosphere forecast model. In this section, we describe
skill for these predictions. They concluded that im-          a new coupled forecast model, denoted as CMP12. Re-
proving the coupled model is at least as important as                       ˜
                                                              sults of El Nino forecast experiments using the im-
improving the initialization.                                 proved ocean initial conditions and the new coupled
   The research results presented here are Part II of a       model will be discussed in the next two sections.
two-part paper. In Part I (Behringer et al. 1998), an
improved ocean analysis system based on a three-di-
mensional variational ocean data assimilation scheme          a. The coupled model
(Derber and Rosati 1989) is described. We have shown             The new coupled model (CMP12) used for this study
that this improved ocean analysis system is capable of        is a modified version of the previous NCEP coupled
producing ocean initial conditions that can capture sea-      model denoted as CMP10 (Ji et al. 1996). The ocean
sonal to interannual variability in the tropical Pacific       model is the same as the one used in the ocean analysis
more accurately. This improvement is achieved by in-          system as described in Part I. It is known as the modular
corporating vertical variation in first-guess error vari-      ocean model developed at the Geophysical Fluid Dy-
ance and an overall reduction in magnitude of the es-         namical Laboratory (Bryan 1969; Cox 1984; Philander
timated first-guess error. The improvements in the ocean       et al. 1987), and is configured for the Pacific domain
data assimilation scheme lead to reduced instances of         of 45 S–55 N, 120 E–70 W. The atmosphere model is
locally overfitting the model to sparse observations;          a low-resolution version of the NCEP’s global medium-
hence, the result is reduced aliasing of model bias into      range forecast model (Kanamitsu et al. 1991) with a
low-frequency variability. The ocean initial conditions       horizontal spectral resolution of T40 and modified con-
produced are smoother, indicating less high-frequency         vection and cloud parameterizations designed to pro-
and small-scale noise, which is probably not relevant         duce a better climate simulation (Ji et al. 1994; Kumar
to ENSO. In Part II, we wish to demonstrate that the          et al. 1996). One-way anomaly coupling is used from
ocean initial conditions with more accurate low-fre-          the atmosphere to the ocean, that is, only anomalies of
quency variability and lower small-scale noise can sig-       surface stress, heat, and freshwater fluxes produced from
nificantly improve ENSO prediction skill.                      the atmosphere model are retained. This is because the
   In this paper, we will describe an improved coupled        AGCM is unable to produce realistic mean annual cycles
forecast model. We will show that the ocean initial con-      for these fluxes. The ocean is then driven by total fluxes
ditions produced with the improved ocean analysis sys-        made up of these anomalies and the climatological mean
tem can significantly improve ENSO prediction skill.           annual cycle wind stress of The Florida State University
However, this improvement is achieved only with an            (Goldenberg and O’Brien 1981) and the climatological
improved coupled model, therefore, the improvement in         heat flux components of Oberhuber (1988). Coupling
ENSO forecast will require improvement in both the            from the ocean to the atmosphere is accomplished by
ocean initialization and the coupled model.                   using total SST from the ocean model to force the at-
   This paper is organized as follows. The new coupled        mosphere model.
model is described in section 2. Hindcast experiments            Two changes were made to the ocean model for
using the new coupled model, and comparisons with the         CMP12: a relaxation of the OGCM’s surface salinity to
results from our previous coupled model are presented         the climatology of Levitus et al. (1994) is added, and
in section 3. In section 4, we discuss the reasons for        a modification to the parameter 0 in the Richardson
the improvement in the hindcasting skill by further an-       number–dependent vertical mixing formulation (Paca-
alyzing impacts from the improved ocean initialization        nowski and Philander 1981) for the eddy viscosity co-
and the improved coupled model. Our results are sum-          efficient. The eddy viscosity coefficient is given by
marized in section 5.
                                                                                         0
                                                                                                      b   ,         (1)
2. The coupled forecast model                                                      (1        Ri ) n
   The primary interest at NCEP is to improve the ca-         where R i is the Richardson number and , b , 0 , and
pability for predicting ENSO episodes using coupled           n, are parameters to be chosen empirically. Pacanowski
ocean–atmosphere models. A central part of this effort        and Philander (1981) suggested for 0 O(50 cm 2 s 1 ).
is improving the ocean data assimilation system to pro-       In tuning our earlier versions of coupled model, we had
duce a better ocean initialization for the coupled forecast   set 0 5 cm 2 s 1 . Over time, it has become clear that
model. In Part I, we demonstrated that the recent             this value is too small, resulting in an unrealistically
changes in the analysis system have improved the qual-        shallow equatorial undercurrent core in our earlier ocean
ity of the ocean analyses. This result indicates that the     analyses. In the new ocean analysis system, this param-
ocean initial conditions obtained from the new analysis       eter is changed to 50 cm 2 s 1 , as recommended by Pa-
system probably are better initial conditions for ENSO        canowski and Philander (1981). Therefore, the same
prediction. On the other hand, better forecast skill can      change is made for the new coupled model as well. Thus,
1024                                     MONTHLY WEATHER REVIEW                                               VOLUME 126


the ocean model in CMP12 is configured identically to            b. Postprocessing
the model used in the new ocean analysis system.                   Monthly mean total SSTs for the tropical Pacific were
   In addition, a number of changes in the coupling be-         saved during the forecasts. Monthly SST anomalies are
tween the ocean and atmosphere models have been                 obtained by removing the coupled model’s SST cli-
adopted. Specifically, the addition of anomalous fresh-          matology. The coupled model’s SST climatologies were
water flux from the atmosphere to the ocean and the              computed by averaging all predictions initiated from the
reduction of the active region of the Pacific basin, where       same month of all years. Therefore, the method for es-
SST from the ocean model is allowed to force the at-            timating the predicted monthly SST anomalies does not
mosphere model from the entire basin (45 S–55 N) to             involve observations of SST. This means that the av-
the equatorial band between 15 S and 15 N. However,             erage of the predicted SST anomalies is 0 over the hind-
basinwide stress, heat, and freshwater flux anomalies            casting period. However, the observed SST anomalies
produced from the AGCM are used to force the ocean              used at the Climate Prediction Center are estimated rel-
model. Additionally, a negative heat flux feedback of 5          ative to the climatology for the period of 1950–79
W m 2 C 1 has been introduced, which is in addition             (Reynolds and Smith 1995). Therefore, the observed
to the standard anomalous heat flux forcing used in the          SST anomalies, when averaged over the model’s hind-
CMP10 model.                                                    casting period of 1981–95, are not 0. This period is
   Since we do not have a realistic climatological fresh-       warmer than the 1950–79 period because of several
water flux to force the ocean to maintain a realistic            strong warm ENSO episodes that occurred during the
climatological sea surface salinity (SSS) field, relaxation      1980s and more frequent warm episodes that occurred
of the surface salinity to the Levitus et al. (1994) cli-       during the early 1990s. Hence, the predicted SST anom-
matology is a reasonable approach. Observations such            alies are adjusted so that when averaged over the hind-
as those of outgoing longwave radiation (OLR) indicate          casting period, they have the same mean as those for
that a shift of a significant portion of the tropical rainfall   the observed SST anomalies. This adjustment is made
from the western Pacific to the central Pacific occurs            by adding the mean observed monthly mean SST anom-
during a warm ENSO episode. AGCM simulations                    alies for the 1981–95 period to the predicted monthly
forced with the observed monthly SSTs show that our             mean SST anomalies. In essence, the average of ob-
model is able to reasonably simulate the tropical rainfall      served SST anomalies for 1981–95 period is treated as
anomalies associated with ENSO (Kumar et al. 1996).             a model bias. In addition, our postprocessing procedure
Hence a reasonable approach is to use the AGCM rain-            averages predicted monthly SST anomalies having the
fall anomaly to represent the anomalous freshwater flux          same lead time but produced from three predictions
while the effect of the climatological freshwater flux is        starting from three consecutive months. This signifi-
approximated by relaxation to the Levitus SSS clima-            cantly reduces month-to-month noise from different pre-
tology. Since the NCEP model has no demonstrated skill          dictions. In the subsequent sections, the forecast lead
for SST predictions beyond the immediate equatorial             time is defined as the average time lag between the initial
Pacific, we do not force the atmosphere model with the           conditions and the target month. For example, a 3-month
ocean model SSTs for the entire basin. The negative             lead prediction for July is the average of three monthly
heat flux feedback was included in order to damp the             mean predictions for June, July, and August, which are
predicted SST anomalies in the model, which at times            initiated in March, April, and May, respectively. How-
in the past appeared to be too large. However, the actual       ever, in real-time forecasting, this smoothing reduces
impact of this change on the model is unclear at present.       the effective lead time by 1 month, that is, the forecast
   Ideally, it would be desirable to study the model’s          for July is not available until the forecast made from
sensitivity to each of these changes. Unfortunately, the        May is completed. This procedure differs from that used
complexity of the coupled GCMs makes it nearly im-              for our previous models (Ji et al. 1994; Ji et al. 1996),
possible to explore the impact of each of these changes.        which averages three predicted monthly mean SST
A large number of hindcasts would be necessary to               anomalies for a given target month but with different
quantify the impact of each change by statistically ex-         lead times because they are produced from predictions
amining the model’s performance as measured by fore-            of three consecutive monthly starts. In the subsequent
cast skill. The computational cost to evaluate every sin-       sections, we will show comparisons of predictions made
                                                                from different versions of coupled models. For consis-
gle parameter change would be prohibitive. Many of the
                                                                tency, predictions made from our previous coupled mod-
changes implemented were based on making the model
                                                                el, that is, CMP10, are reprocessed using the same post-
more ‘‘realistic,’’ such as adding the relaxation to a
                                                                processing method for CMP12. We also postprocessed
surface salinity climatology and adding anomalous
                                                                CMP12 predictions using the previous postprocessing
freshwater flux by using the AGCM’s rainfall anomalies.          method, finding that it does not change our conclusion.
At present, we cannot isolate the impact of each of these
changes. However, as will be shown in the next section,
the combination of all these changes has had a positive         3. Results from hindcasting experiments
impact on the model performance, which has contrib-               To evaluate the performance of the new coupled fore-
uted to improved prediction skill.                              cast system, 1-yr hindcasts are carried out. Two sets of
APRIL 1998                                                JI ET AL.                                                         1025

                                             TABLE 1. List of prediction experiments.

Experiment                Model               Ocean IC                 Starting month                        Period
CMP10                     CMP10                 RA5                  Monthly, Jan–Dec                 Dec 1981–Dec 1995
CMP12                     CMP12                 RA6                  Monthly, Jan–Dec                 Dec 1980–Dec 1995
C10RA5*                   CMP10                 RA5                  Dec, Mar, Jun, Sep               1981–95
C12RA6**                  CMP12                 RA6                  Mar, Jun, Sep, Dec               1981–95
C10RA6                    CMP10                 RA6                  Mar, Jun, Sep, Dec               1981–95
C12CTL                    CMP12                 CTL                  Mar, Jun, Sep, Dec               1981–95

  * C10RA5 is a subset of experiment CMP10
 ** C12RA6 is a subset of experiment CMP12


ocean initial conditions, RA5 and RA6, are available.             time series of area-averaged SST anomalies between the
RA5 is produced with our previous ocean data assim-               predictions and the observations in an equatorial Pacific
ilation system (Ji et al. 1995); RA6 is produced with             region that is latitudinally bounded from 5 S to 5 N.
the improved ocean data assimilation system described             For this discussion, we choose the area between 170
in Part I. In this section, we describe hindcast results          and 120 W, which is also referred as the Nino3.4 region
using the CMP12 model and RA6 ocean initial condi-                (Barnston et al. 1994). Figure 1 shows time series of
tions. To evaluate improvements in prediction skill,              Nino3.4 SST anomalies predicted by CMP10 (dash) and
these hindcast results are compared with prediction re-           CMP12 (heavy) models for the 1981–96 period for lead
sults from our standard forecast model, that is, CMP10            times of 3, 6, and 9 months. The observed SST anom-
using RA5 ocean initial conditions. Note that the                 alies (Reynolds and Smith 1994) for Nino3.4 are shown
CMP10 results were actually produced previously and               in the thin solid lines. Both models predicted the low-
described in Ji et al. (1996). Major features of these two        frequency interannual evolution of the SST anomalies
sets of predictions are listed in Table 1.                        reasonably well, especially for the short lead time (3
   The common practice for evaluating a coupled                   months). At the longer lead times, CMP10 lags behind
model’s ENSO prediction capability is to correlate the            in predicting the 1982/83 warm ENSO episode, which




 FIG. 1. Predicted Nino3.4 SST anomalies ( C) for 1981–96 at 3-month, 6-month, and 9-month lead times. The prediction from CMP10
         (CMP12) models are shown in dash (solid) lines; the observed Nino3.4 SST anomalies are shown in the thin-solid lines.
1026                                     MONTHLY WEATHER REVIEW                                                    VOLUME 126




                    FIG. 2. Skill estimates as function of lead time for prediction of Nino3.4 SST anomalies for
                 CMP10 (dash) and CMP12 (solid) models for 1981–95 period (left panels) and 1992–95 period
                 (right panels). Shown in the top panels are anomaly correlation coefficients (ACC) between the
                 predictions and the observations; shown in the lower panels are rmse’s.


CMP12 predicted better. Both models lagged behind for             period, CMP12 tracks the observations much better than
the 1986/87 warm episode, especially for coming out               CMP10.
of the event, the rapid cooling during 1988. However,                The commonly used measures of skill for coupled
CMP12 performed somewhat better in this respect at                models are anomaly correlation coefficients (ACC) and
the 9-month lead. In addition, CMP12 exhibited a cool-            root-mean-square errors (rmse) between the predictions
ing during 1987, which is spurious. For the period of             and the observations for area-averaged SST anomalies.
the 1990s, CMP10 had trouble coming out of the                    These skill measures for the CMP10 and CMP12 models
1991/92 warm episode, which led to completely missing             for the Nino3.4 SST anomalies are shown in Fig. 2. For
the warming during late spring 1993. CMP10 also                   the entire common period from January 1982 to March
missed the warm episode during the late fall/winter of            1996, CMP12 clearly exhibited higher skill than CMP10
1994/95 at the longer lead time. On the other hand,               for lead times longer than about four months. The right
CMP12 fared much better during the 1990s. It was able             panels of Fig. 2 show ACC and rmse for both models
to come out of the 1991/92 warm event, and thus predict           based on a total of 48 predictions initiated from January
a small SST rise in 1993, although this had a different           1992 to December 1995. For this period, CMP12 out-
timing and amplitude when compared with the obser-                performed CMP10 by a bigger margin compared to the
vations. CMP12 also was able to predict the warm ep-              full 1982–95 period. The period since the 1991/92 warm
isode of 1994/95 and the cold episode of 1995/96 up               episode was a difficult one for ENSO prediction models.
to at least 6-months lead times. Overall, CMP12 tracks            Ji et al. (1996) showed that not only did the coupled
the observations better during 1981–91, mainly because            prediction models exhibit lower skill levels than during
it was better during the 1982/83 event and during 1988,           the 1980s (Barnston et al. 1994; Chen et al. 1995), the
especially at the 9-month lead time. For the 1992–96              persistence forecast also showed much lower skill, that
APRIL 1998                                                   JI ET AL.                                                         1027




                FIG. 3. Evolution of Nino3.4 SST anomalies ( C) from all individual predictions (thin solid lines) initiated
              monthly using the CMP10 model for 1982–95. Shown in each panel are the predictions grouped by three
              consecutive starting months. The observed Nino3.4 SST anomalies are shown in the heavy-dashed lines.


is, a shorter persistence time for anomalies, when com-               for useful prediction, Fig. 2 suggests that the predict-
paring the 1992–95 period to the 1982–92 period. The                  ability for the 1992–95 period for CMP10 model was
observations of sea level pressure, low-level zonal wind,             3–4 months. For the CMP12 model, the predictability
SST, and subsurface ocean heat content anomalies in                   for this period was about 7 months.
the Pacific during the two periods also show remarkable                   Individual 12-month predictions are shown in Fig. 3
contrasts in the characteristics of their interannual vari-           and Fig. 4. There are 168 1-yr predictions for CMP10
ability (Kleeman et al. 1996; Latif et al. 1997; Ji et al.            (Fig. 3) and 181 1-yr predictions for CMP12 (Fig. 4).
1996). The observational evidence suggests a possible                 The thin lines in each panel in the figure depict predic-
change in the nature of coupled interactions in the trop-             tions starting from three consecutive months for each
ical Pacific, thereby resulting in lower ENSO predict-                 year. For example, the top panels in both figures depict
ability for the 1990s. Thus it is a challenge to improve              all the predictions initiated in January, February, and
the skill of ENSO prediction for the early 1990s. If an               March (JFM) of each year for the entire hindcasting
ACC value of 0.6–0.7 is considered as a rough criterion               period. The three letters shown in the upper right of
1028                                   MONTHLY WEATHER REVIEW                                                VOLUME 126




                                FIG. 4. Same as Fig. 3 but for the CMP12 model for 1981–95.



each panel are the initials for each of the three starting     more consistent. However, CMP12 consistently pre-
months of the predictions depicted in the panel. Ob-           dicted a cooling during 1987 that did not actually occur.
served Nino3.4 SST anomalies are shown in dashed               The comparison of predictions from both models also
curves. We wish to point out several features in this          reveals that for the CMP10 model, the predictions ini-
figure. First, both models exhibit spread in the forecasts,     tiated prior to or during the onset of warm ENSO ep-
that is, hindcasts starting from consecutive months may        isodes appear to have difficulty coming out of the warm
result in very different predictions. Clear examples of        episodes. For example, many CMP10 predictions ini-
this difference can be seen in the top panel of Fig. 3         tiated in late 1982 to early 1993, in late 1987 and in
during 1992–96 period for CMP10 model. The spread              winter 1991 to spring 1992 were unable to predict the
of the predictions is the reason why averaging over sev-       cooling during the subsequent seasons. CMP12, on the
eral predictions reduces noise in the prediction results,      other hand, appears to be better in predicting the cooling
particularly for the 1992–96 period. The spread among          after warm ENSO events, particularly during 1983,
the predictions for CMP12 is smaller than for CMP10,           1988, and 1992.
indicating that the predictions from CMP12 model are              Figure 5 depicts differences between the observed
APRIL 1998                                                    JI ET AL.                                                              1029




           FIG. 5. Differences between the coupled model SST climatology (average of June and December starts) and the
         observed SST climatology for January, April, July, and October as representative for different seasons. Contour intervals
         are 0.5 C.


SST climatology (Reynolds and Smith 1995) and the                      thermore, in the equatorial region, biases for CMP12
coupled model SST climatology for CMP10 (left panels)                  tend to be negative (cold bias), except for the spring
and CMP12 (right panels). These difference (bias) fields                season (April) in the far eastern equatorial Pacific. Fig-
are obtained by subtracting the observed climatology                   ure 5 suggests that for CMP10 model, the coupled
from the model climatology. The model’s monthly mean                   ocean–atmosphere system is often in a ‘‘perpetual
annual cycle is obtained by averaging all the predictions                                                 ˜
                                                                       spring’’ or even ‘‘perpetual El Nino’’ state because SST
initiated in June and December. Monthly mean bias                      in the eastern tropical Pacific are strongly biased to be
fields for January, April, July, and October are shown                  warmer than the observed climatology. The northern
in the figure. These are the center months for each of                  spring season is characterized by warm SST throughout
the winter, spring, summer, and fall seasons. Comparing                the tropical Pacific and weak trades. Many previous
these systematic SST biases for the two models, we                     studies have found that the northern spring season is a
notice that the biases in SST climatology for CMP10                    period during which the coupled ocean–atmosphere sys-
are generally warm in the central to eastern tropical                  tem loses part of its ‘‘memory’’ (Zebiak and Cane 1987;
Pacific, whereas for CMP12, the magnitudes of the bias                  Battisti 1988) and the predictability for ENSO is the
fields are generally smaller than those of CMP10. Fur-                  lowest. The error growth rate for the coupled system
1030                                      MONTHLY WEATHER REVIEW                                                       VOLUME 126


during this time of the year tends to be the highest
(Blumenthal 1991; Xue et al. 1997). Hence, the per-
petual springlike state of the CMP10 model may facil-
itate random error growth, which results in the predic-
tions having a larger spread. The fact that CMP12 has
no tendency toward perpetual springlike conditions may
partially explain its better performance.
   One may ask if the difference in model bias char-
acteristics between CMP10 and CMP12 results from
different model characteristics or from different initial-
izations. We believe it is the former because we have
examined the CMP12 bias from predictions using a set
of different ocean initial conditions similar to RA5 (see
section 4). The bias fields for these (not shown) are
nearly identical to those shown in Fig. 5 for CMP12.
Therefore the difference in the bias appears to be a result
of changes in the model characteristics. Additionally, it
could be argued that the coupled model climatology is
referred to the 1981–95 base period during which the                FIG. 6. Spatial distribution (skill map) of the temporal anomaly
average SST in the tropical Pacific was warmer than the           correlation coefficients between the observed and the predicted SST
                                                                 anomalies. The skill map for CMP12 (CMP10) is shown in the upper
1950–79 base period. However, Reynolds and Smith                 (lower) panel. The anomaly correlation coefficients are computed
(1995) showed that the average warming during 1982–              based on 45 (42) 3-month averages centered in December, January,
93 for tropical Pacific region of 10 S–10 N and between           and February for the boreal winters of 1981/82 (1982/83) through
150 W and 90 W, is about 0.4 C. For many large areas             1995/96 for CMP12 (CMP10). Contour values are indicated in the
                                                                 figure.
in the equatorial Pacific, the CMP10 bias is above 1 C,
especially for the northern fall and winter seasons and
therefore, there is no doubt that the CMP10 has a sig-           western tropical Pacific near the date line. Hence im-
nificant warm SST bias.                                           proving SST forecast for this area is highly desirable.
   It is also of interest for potential users of the forecasts      Shown in Fig. 7 are skill comparisons for the pre-
to know the spatial distribution and the seasonality of          diction of Nino3.4 SST anomalies. In addition to overall
the prediction skill for this model. Spatial distribution        prediction skill, we grouped the predictions initiated
of the SST forecast skill for the six month lead predic-         during the northern warm seasons from May through
tions for the boreal winter seasons are shown in Fig. 6.         September, and during the early northern winter months
The anomaly correlation skill for the winter seasons is          (November, December, and January). We found from
defied as the temporal correlation of the observed and            this comparison that predictions starting in the warm
the predicted SST anomalies at each grid point for               seasons can sustain quite high skill (above or near 0.9)
3-month averages centered in December, January, and              for nearly three seasons while skill for the predictions
February. For CMP12, the predictions are verified for             initiated during early winter months starts to drop after
1981/82 through 1995/96, whereas for CMP10, predic-              only about one season. At about 1-yr lead time, the
tions are verified for 1982/83 through 1995/96. Hence,            predictions from all starting months have similar skill
there are 45 (42) seasons verified for CMP12 (CMP10)              of about 0.65, which is still considered useful.
models. Note that these sample sizes are quite small
because the actual degrees of freedom are much smaller
than the sample sizes, therefore the statistics could be         4. Discussion
unstable. Nevertheless, the skill maps can be indicative            In previous sections and in Part I, we have described
of the skillful regions of the forecasts for each model.         an improved ENSO prediction system that consists of
From Fig. 6, we found that the skillful forecasts are            an improved ocean data assimilation system and a new
expected only within a narrow region confined between             coupled forecast model. Hindcasting results show that
10 S and 10 N in the equatorial Pacific and east of the                                                          ˜
                                                                 this system produces more accurate El Nino forecasts.
date line. This is similar to our previous coupled models        It is of strong interest to understand if the improvement
and to many other coupled dynamical forecast models.             in prediction skill results from the more accurate ocean
Comparing CMP12 and CMP10, it is indicative that                 initial conditions or because of the improvements in the
CMP12 has somewhat higher skill for the northern win-            coupled model.
ter seasons, and more importantly, the skillful forecast            Figure 8 shows the 1-month rms spread for CMP10
region (area where the skill is above 0.8) for CMP12             (dash) and CMP12 (solid) for the individual predictions
extends further west of the date line. This is important         as a function of the prediction lead times. Symbolically,
because the significant region of coupled air–sea inter-          let A      {a i }, i  1, 2, . . . , 180 and B   {b i }, i 1,
actions that impact North America is in the central-             2, . . . , 180. If A represents a set of 6-month lead pre-
APRIL 1998                                                    JI ET AL.                                                   1031




                        FIG. 7. Anomaly correlation coefficients (left) and rmse’s (right) between the predicted and the
                     observed monthly mean SST anomalies as function of lead times for Nino3.4 SST anomalies. These
                     are for CMP12 for the period of 1981–95. Skills for the predictions starting in the northern warm/
                     cold/all seasons are shown in the dot–dashed/dashed/solid curves. The warm season is defined as
                     May through September; the cold season is defined as November, December, and January.


                                                                       dictions of monthly mean SST anomalies for the target
                                                                       period between July 1981 and June 1996 (a total of 180
                                                                       predictions), and B represents a set of predictions for
                                                                       the same target period but with a 5-month lead, then
                                                                       the value on each curve at the 6-month lead time is the
                                                                       rms difference between A and B. The difference between
                                                                       A and B arises because A has a lead time of 1 month
                                                                       longer than B. Note that this is not an estimate of the
                                                                       prediction error at a 6-month lead time; that is measured
                                                                       by the rms difference between predictions and obser-
                                                                       vations. This instead provides estimates of the error
                                                                       growth from one month to the next during the forecasts.
                                                                       The first value of each curve, that is, a lead time of 1
                                                                       month, represents the rms difference between all
                                                                       1-month predictions and the observations, which indi-
                                                                       cates the first-month error growth due to errors in the
                                                                       initial conditions. As the lead time becomes longer, the
                                                                       1-month rms separation becomes larger as indicated by
                                                                       the general increase of the 1-month rms spread. We
                                                                       notice that the two curves are approximately in parallel
                                                                       indicating that both models probably have a similar error
                                                                       growth rate. However, the fact that the curve for CMP12
                                                                       is below that for CMP10 reflects that RA6 initial con-
                                                                       ditions have a lower noise level in them, which con-
                                                                       tributes to improved consistency in the CMP12 forecast
                                                                       results (cf. Figs. 3 and 4). Lowering the noise level in
FIG. 8. One-month rms separation as an estimate of error growth for    the initial conditions can contribute to improved pre-
the CMP10 (dash) and CMP12 (solid) models (see text in section 7).     diction skill because some of the noises in the initial
1032                                      MONTHLY WEATHER REVIEW                                                       VOLUME 126




                    FIG. 9. Skill estimates (ACC, left and rmse, right) as a function of lead time for prediction of
                 Nino3.4 SST anomalies from four prediction experiments that used two coupled models and three
                 initial conditions. These predictions are designated as C12RA6 (solid), C12CTL (short dash),
                 C10RA6 (dot–dash), and C10RA5 (long dash). The predictions are initiated in March, June, Sep-
                 tember, and December of each year for 1981–95. The two coupled models, that is, CMP12 and
                 CMP10, are designated as C12 and C10; the three initial conditions are designated as RA5, RA6,
                 and CTL (see Table 1 and section 4).



conditions, when projected onto the growing modes, that             more accurate than CTL for the low-frequency vari-
is, the singular vectors (Xue et al. 1997), could have              ability. The CTL analyses are produced in order to fa-
nontrivial amplitudes, which can lead to bad predictions.           cilitate our understanding of the impact on analysis qual-
Therefore, a lower noise level in the initial conditions            ity by improving the first-guess error covariance func-
can extend the range of useful forecasts within the limit           tion. We use CTL as a proxy of less accurate ocean
of predictability given the same error growth rate for              initial conditions for additional hindcast experiments.
the model. Thus, the improved initialization for CMP12                 We denote the hindcasts using CMP12 model and
plays an important role in the improvement of the pre-              CTL initial conditions as C12CTL, which contains 1-yr
diction skill. Additional evidence comes from running               predictions initiated in March, June, September, and De-
the same forecast model with two different sets of vary-            cember of each year for the period of December 1980
ing quality ocean initial conditions.                               to December 1995. Furthermore, in order to quantify
   We have shown in Part I that our improved data as-               the impact of model improvement on prediction skill,
similation system results in analyses that are more ac-             we produced an identical set of 1-yr predictions as the
curate for the low-frequency variability. More accurate             C12CTL but using the CMP10 model and RA6 initial
analyses presumably lead to a better initialization for             conditions. We denote this dataset as C10RA6. For sim-
the coupled model. The obvious question, then, is what              plicity, we denote a subset of predictions using the
impact this actually has on prediction skill? To address            CMP12 model and RA6 initial conditions which are
this question, prediction experiments were carried out              common to the C12CTL experiment as C12RA6; and
using the CMP12 model and a third set of ocean initial              the same subset of predictions but from the CMP10
conditions denoted as CTL. The CTL analyses are pro-                model and RA5 initial conditions as C10RA5. Major
duced as part of the present research using an ocean                features of these prediction experiments are listed in
data assimilation scheme that adopted some improve-                 Table 1.
ments in the new system, but did not incorporate the                   Hindcast skills for Nino3.4 SST anomalies as mea-
two major changes of the new analysis system, that is,              sured by ACC and rmse, based on the predictions from
the vertical variation for the first-guess error variance            C12RA6, C12CTL, C10RA6, and C10RA5 are shown
and the overall reduction in the magnitude of the esti-             in Fig. 9. The skill estimates in Fig. 9 show that
mated first-guess error. Comparison with the indepen-                C10RA5, C10RA6, and C12CTL form a tight pack
dent observations of sea level data show that RA6 is                while C12RA6 clearly stands out as a set of more skillful
APRIL 1998                                            JI ET AL.                                                    1033

predictions. Note that there is a large random element        periments using the CMP12 model and the RA6 initial
inherent in coupled GCM predictions. Some indication          conditions showed a significant increase in the ENSO
of this can be seen in Figs. 3 and 4. Thus the sample         prediction skill over those using CMP12 and the less
size of 61–64 for each set of predictions may be still        accurate initial conditions (CTL) and over CMP10 using
too small to completely overcome the uncertainty, which       either set of initial conditions.
is difficult to estimate. However, the clear separation            An extended version of the NCEP ocean analysis sys-
between C12RA6 and the other three experiments that           tem that adopts stronger dynamical constraint than the
closely band together strongly suggest that C12RA6 is         previous analysis system is described in a companion
probably a better set of predictions.                         paper. The stronger dynamical constraint in the new data
   In Part I, we have shown that the RA6 analyses sim-        assimilation scheme not only resulted in analyses having
ulate the observed variability more accurately than the       better fit for low-frequency variability, but also proven
CTL analyses. The prediction results shown here sug-          to be a very important factor leading to the improved
gest that RA6 is a better set of initial conditions than      prediction skill. We believe the new analysis system
CTL (C12RA6 vs C12CTL). The ocean data assimila-              provides better ocean initialization because the more
tion system that produced RA6 uses stronger dynamical         accurate low-frequency variability and lower small-
constraint by overall reduction in the magnitude of the       scale noise level in the ocean initial conditions are de-
estimated first-guess error, that is, giving higher relative   sirable for ENSO prediction. This belief was quantified
weights to the model field and lower weights to the            by hindcasting experiments using the same coupled
observations. The hindcasting results indicate that for       model but different ocean initial conditions (C12RA6
the purpose of initializing the coupled model, it is de-      vs C12CTL). The higher relative weighting given to the
sirable for the analysis system to have stronger dynam-       ocean model field allows the ocean model data assim-
ical constraint because it gives better fits to low fre-       ilation system to act like a stronger filter. It reduces the
quency variability by reducing aliasing of large-scale        impact of the high-frequency, small-scale variations in-
model bias into the low-frequency variability. The RA6        adequately sampled by the observations, while retaining
ocean initial conditions are also smoother than CTL,          the low-frequency large-scale signals relevant to ENSO.
which indicates that RA6 contains less dynamically in-        Also, by avoiding the local overfitting to observations,
consistent features, probably caused by locally overfit-       it reduces errors that result from aliasing the mean mod-
ting the model to data. These results imply that the          el-forcing error into low-frequency variability, thus
C12RA6 predictions probably have lower initial error          leading to more accurate simulation of that variability.
and require less dynamical adjustment from initializa-        These improvements suggest that strong dynamical con-
tion as indicated by Fig. 8 and, therefore, help to achieve   straint in ocean data assimilation is desirable for the
the higher prediction skill.                                  purpose of ocean initialization for ENSO prediction.
   On the other hand, a comparison of skill levels in             Our improvement in the prediction skill resulted from
Fig. 9 between C12RA6 and C10RA6 suggests that im-            both more accurate ocean initial conditions and a better
provements in the prediction skill also result from the       coupled model. Our experiments (C10RA6) show that
improvements in the coupled model since both exper-           CMP10 model was unable to take advantage of the more
iments used the same more accurate ocean initial con-         accurate ocean initial conditions to produce better pre-
ditions (RA6). This indicates that both the model im-         dictions. On the other hand, the C12CTL experiment,
provement and the better ocean initialization are equally     which used the improved model (CMP12) with the less
important in improving prediction skill. The CMP10            accurate initial conditions, was also unable to achieve
model was unable to take advantage of the better ocean        better prediction skill. These reflect the complexity of
initial conditions from RA6 as illustrated by the com-        the coupled forecast system. Improving one component
parison of C10RA6, C10RA5, and C12RA6, suggesting             of the system does not necessarily lead to improvement
that the model error in CMP10 model plays a dominant          in prediction skill. Individual changes to the data as-
role in limiting further increases in the prediction skill.   similation system and the forecast model can some times
On the other hand, the increased skill level of C12RA6        lead to unexpected results. Future efforts in improving
over C12CTL, both of which used the CMP12 model,              forecast system must take a systematic approach to im-
suggests that improving the ocean initialization can have     proving both the ocean data assimilation system and the
a significant impact on the prediction skill, provided that    coupled model. This makes the task more challenging.
the coupled model is able to take advantage of the better
ocean initial conditions.                                       Acknowledgments. Support for this research is pro-
                                                              vided by NOAA’s Office of Global Program through the
                                                              Climate and Global Change Program. Dr. Y. Xue
5. Summary
                                                              (NCEP/UCAR) helped to calculate rms error spread of
  In this paper, we described a new version of the cou-       coupled model predictions. The authors wish to express
pled ENSO prediction model (CMP12). A number of               our gratitude to two anonymous reviewers. Their
changes in both the ocean model and in the coupling           thoughtful comments helped to improve this manuscript
scheme were incorporated into CMP12. Hindcasting ex-          greatly.
1034                                           MONTHLY WEATHER REVIEW                                                             VOLUME 126


                           REFERENCES                                         , A. M. Moore, and N. R. Smith, 1995: Assimilation of subsur-
                                                                              face thermal data into a simple ocean model for the initialization
Barnston, A. G., H. M. Van den Dool, S. E. Zebiak, T. P. Barnett,             of an intermediate tropical coupled ocean–atmosphere forecast
     M. Ji, D. R. Rodenhuis, M. A. Cane, A. Leetmaa, N. E. Graham,            model. Mon. Wea. Rev., 123, 3103–3113.
     C. F. Ropelewski, V. E. Kousky, E. A. O’Lenic, and R. E. Liv-            , R. A. Colman, N. R. Smith, and S. B. Power, 1996: A recent
     ezey, 1994: Long-lead seasonal forecasts—Where do we stand?              change in the mean state of the Pacific basin climate: Obser-
     Bull. Amer. Meteor. Soc., 75, 2079–2114.                                 vational evidence and atmospheric and oceanic responses. J.
Battisti, D. S., 1988: Dynamics and thermodynamics of a warming               Geophys. Res. (Oceans), 101, 20 483–20 499.
     event in a coupled tropical atmosphere-ocean model. J. Atmos.       Kumar, A., M. P. Hoerling, M. Ji, A. Leetmaa, and P. Sardeshmukh,
     Sci., 45, 2889–2919.                                                     1996: Assessing a GCM’s suitability for making seasonal pre-
Behringer, D. W., M. Ji, and A. Leetmaa, 1998: An improved coupled            dictions. J. Climate, 9, 115–129.
     model for ENSO prediction and implications for ocean initial-       Latif, M., A. Sterl, E. Maier-Reimer, and M. M. Junge, 1993: Structure
     ization. Part I: The ocean data assimilation system. Mon. Wea.                                           ˜
                                                                              and predictability of the El Nino/Southern Oscillation phenom-
     Rev., 126, 1013–1021.                                                    enon in a coupled ocean–atmosphere general circulation model.
Bjerknes, J., 1969: Atmospheric teleconnections from the equatorial           J. Climate, 6, 700–708.
     Pacific. Mon. Wea. Rev., 97, 163–172.                                     , R. Kleeman, and C. Eckert, 1997: Greenhouse warming, de-
Blumenthal, M. B., 1991: Predictability of a coupled ocean–atmo-                                        ˜
                                                                              cadal variability or El Nino? An attempt to understand the anom-
     sphere model. J. Climate, 4, 766–784.                                    alous 1990s. J. Climate, 10, 2221–2239.
Bryan, K., 1969: A numerical method for the study of the World           Levitus, S., R. Burgett, and T. P. Boyer, 1994: World Ocean Atlas
     Ocean. J. Comput. Phys., 4, 347–376.                                     1994. Vol. 3. Salinity. NOAA Atlas NESDIS 3, 99 pp.
Cane, M. A., S. E. Zebiak, and S. C. Dolan, 1986: Experimental           Oberhuber, J. M., 1988: An atlas based on the ‘‘COADS’’ data set:
     forecasts of El Nino. Nature, 321, 827–832.                              The budgets of heat, buoyancy and turbulent kinetic energy at
Chen, D., S. E. Zebiak, A. J. Busalacchi, and M. A. Cane, 1995: An            the surface of the global ocean. Rep. 15, Max-Planck-Institut
                                    ˜
     improved procedure for El Nino forecasting. Science, 269, 1699–            ¨
                                                                              fur Meteorologie, 20 pp. [Available from Max-Planck-Institut
     1702.                                                                      ¨
                                                                              fur Meteorologie, Bundesstrasse 55, Hamburg, Germany.]
Cox, M. D., 1984: A primitive, 3-dimensional model of the ocean.         Pacanowski, R., and S. G. H. Philander, 1981: Parameterization of
     GFDL Ocean Group Tech. Rep. 1, Geophysical Fluid Dynamics                vertical mixing in numerical models of tropical oceans. J. Phys.
     Laboratory/NOAA, Princeton University, 143 pp.                           Oceanogr., 11, 1443–1451.
Derber, J. D., and A. Rosati, 1989: A global oceanic data assimilation   Philander, S. G. H., W. J. Hurlin, and A. D. Seigel, 1987: A model
     system. J. Phys. Oceanogr., 19, 1333–1347.                               of the seasonal cycle in the tropical Pacific ocean. J. Phys.
Goldenberg, S. B., and J. J. O’Brien, 1981: Time and space variability        Oceanogr., 17, 1986–2002.
     of tropical Pacific wind stress. Mon. Wea. Rev., 109, 1190–1207.     Reynolds, R. W., and T. M. Smith, 1994: Improved global sea surface
Ji, M., and A. Leetmaa, 1997: Impact of data assimilation on ocean            temperature analysis using optimum interpolation. J. Climate, 7,
                             ˜
     initialization and El Nino prediction. Mon. Wea. Rev., 125, 742–         929–948.
     753.                                                                     , and       , 1995: A high resolution global sea surface temper-
     , A. Kumar, and A. Leetmaa, 1994: An experimental coupled
                                                                              ature climatology. J. Climate, 8, 1571–1583.
     forecast system at the national meteorological center: Some early
                                                                         Ropelewski, C. F., and M. S. Halpert, 1987: Global and regional scale
     results. Tellus, 46A, 398–418.
                                                                              precipitation patterns associated with El Nino/Southern Oscil-
     , A. Leetmaa, and J. Derber, 1995: An ocean analysis system
     for seasonal to interannual climate studies. Mon. Wea. Rev., 123,        lation. Mon. Wea. Rev., 115, 1606–1626.
     460–481.                                                            Rosati, A., K. Miyakoda, and R. Gudgel, 1997: The impact of ocean
     ,       , and V. E. Kousky, 1996: Coupled model prediction of            initial conditions on ENSO forecasting with a coupled model.
     ENSO during the 1980s and the 1990s at the National Centers              Mon. Wea. Rev., 125, 754–772.
     for Environmental Prediction. J. Climate, 9, 3105–3120.             Wyrtki, K., 1975: El Nino—The dynamical response of the equatorial
Kanamitsu, M., and Coauthors, 1991: Description of NMC global                 Pacific to atmospheric forcing. J. Phys. Oceanogr., 5, 572–584.
     data assimilation and forecast system. Wea. Forecasting, 6, 425–    —, 1985: Water displacements in the Pacific and the genesis of El
     435.                                                                     Nino cycles. J. Geophys. Res., 90, 7129–7132.
Kirtman, B. P., J. Shukla, B. Huang, Z. Zhu, and E. K. Schneider,        Xue, Y., M. A. Cane, and S. E. Zebiak, 1997: Predictability of a
     1997: Multiseasonal predictions with a coupled tropical ocean–           coupled model of ENSO using singular vector analysis. Part I:
     global atmosphere system. Mon. Wea. Rev., 125, 789–808.                  Optimal growth in seasonal background and ENSO cycles. Mon.
Kleeman, R., 1993: On the dependence of hindcast skill in a coupled           Wea. Rev, 125, 2043–2056.
     ocean-atmosphere model on ocean thermodynamics. J. Climate,         Zebiak, S. E., and M. A. Cane, 1987: A model El Nino–Southern
     6, 2012–2033.                                                            Oscillation. Mon. Wea. Rev., 115, 2262–2278.

				
DOCUMENT INFO