modeling research on seasonal to interannual variability · improved, global gridded observational...

2001 CDC Science Review 7

CHAPTER 2

Modeling Research on Seasonalto Interannual Variability

2.1 Understanding and predicting sea-sonal tropical SST variations

CDC scientists have pioneered the devel-opment of Linear Inverse Modeling(LIM) as a diagnostic and forecastingtool. LIM is a method of extracting thedynamical parameters of a system fromdata. The assumption is made that the

dynamics can be modeled as a stable lin-ear multivariate process driven by geo-graphically coherent white noise. Theestimated dynamical parameters can thenbe used to make forecasts of the system.Real-time seasonal LIM forecasts oftropical Indo-Pacific SST anomalies arepublished monthly in the Climate Diag-nostics Bulletin and quarterly in the

The discovery in the 1980s of a significant extratropical impact of the tropical Pacific ENSO phenomenon,and the demonstration that several elements of the phenomenon could be predicted two or more seasons inadvance using relatively simple models, were important developments in climate research. They rekindledinterest in gaining a better understanding of ENSO dynamics, in the hope of making even better predictions.They also raised hopes that at least some aspects of interannual variability might be similarly predictable inother parts of the globe, not only from ENSO's remote impact but also through possibly predictable slow vari-ations in other ocean basins. International research programs such as TOGA and CLIVAR were launched inpursuit of these hopes. Largely from their impetus, the observing system in the tropical Pacific was vastlyimproved, global gridded observational datasets spanning several decades were generated through various'reanalysis' efforts, and extensive numerical simulations and predictions of interannual climate variationswere undertaken with both uncoupled and coupled global climate models. Our understanding of the variabil-ity and predictability of the climate system on this time scale has matured considerably as a result. CDC hascontributed significantly to the present store of knowledge in this area, and is at the forefront in addressingmany of the remaining questions.

CDC scientists have provided important evidence that the predictable evolution of ENSO and of the associ-ated remote teleconnections are governed largely by linear, low-dimensional dynamics. This may be one rea-son why simple empirical linear models remain competitive with sophisticated GCMs at making seasonalpredictions. Indeed the question of how much extratropical predictability exists on this scale beyond thatassociated with simple linear ENSO signals has become a recurring theme in CDC research. In the last fouryears, we have identified the situations in which the dynamics of ENSO are significantly nonlinear, clarifiedthe extent to which the remote impacts can be nonlinear, and explored the extent to which those impacts mightbe sensitive to the details of the tropical SST anomaly patterns, i.e. be “higher-dimensional”. We have alsoexplored how the nonlinear and higher-dimensional dynamics might affect the tails of the probability distri-butions of atmospheric variables on synoptic, intraseasonal, and seasonal scales, and thus the risk of extremeanomalies on those scales. We have investigated the predictability of SSTs in other oceans basins, boththrough “atmospheric bridge” ENSO teleconnections and through the year-long persistence and subsequentre-emergence of SST anomalies from below the seasonally-varying surface mixed layer. We have also contin-ued to assess the impact of such SSTs on the atmospheric circulation.

_____________________________

CHAPTER 2 Research on Seasonal to Interannual Variability

8 2001 CDC Science Review

Fig. 2.1 (a) The optimal initial structure for SST anom-aly growth. The pattern is normalized to unity. Thecontour interval is 0.025. Loadings greater than 0.025are colored red; dashed lines indicate negative con-tours. (b) The pattern predicted when the pattern in (a)is used as the initial condition.

Experimental Long-Lead Forecast Bulle-tin (ELLFB). They are also availablethrough CDC's website at http://www.cdc.noaa.gov/~lem/IndoPa-cific.frcst.html.

An important original result from ourLIM diagnosis of Indo-Pacific SST vari-ability was the identification, through asingular-vector analysis of the empiri-cally-determined system propagator, ofan optimal initial SST pattern for SSTanomaly growth over 7 months in thebasin. This structure evolves in 7 monthsinto a pattern resembling a mature ENSOevent (Fig. 2.1), and may therefore beviewed as a dynamically relevant precur-sor to ENSO. And indeed, as Fig. 2.2.shows, whenever an SST anomaly pat-tern projects appreciably on this struc-ture, there is a good chance of obtaininglarge SST anomalies in the Niño 3.4 area7 months later. Figure 2.2 also showsthat the projection statistics are similarwhether they are evaluated for a timeperiod including the training period

(COADS data: 1950–1990) or for inde-pendent data (NMC Real-time SurfaceMarine data: 1991–2001). Encouragedby this robust behavior, we now providereal-time monitoring of the projection ofthe SST anomaly field on this structurethrough our website at http://www.cdc.noaa.gov/~lem/opt/optstr.html.

We have also recently published proce-dures for estimating improved confi-dence intervals on our SST forecasts.Until recently, the confidence intervalswere those appropriate to our assumedstationary linear Markov process; that is,they showed the expected forecast errorof a stable linear model driven by sta-tionary white noise. The improved errorbars also include estimated contributionsfrom our neglect of the seasonal variationof the stochastic forcing, from estimatingthe model's parameters in a trainingperiod of finite length, and from initialcondition errors.

The actual forecast error normalized bythe rms of the total expected error esti-mated in this manner (Fig. 2.3) showshow much better the forecast skill is dur-ing La Niña than El Niño. This is consis-tent with our conclusion, stated in severalpapers, that nonlinear dynamics becomeimportant during the warmest phase ofwarm events. We have continued toinvestigate the failure of LIM duringsuch periods. Although the lack of skillis mostly due to the greater importanceof nonlinearity, an interesting recent find-ing was that LIM would neverthelesshave been useful during late 1994–early1995 had it been applied to weeklyinstead of seasonal SSTs. Evidence hasalso been found that unpredictable sto-

Research on Seasonal to Interannual Variability CHAPTER 2


chastic forcing accounted for a large por-tion of the observed warming during thestrong 1997–1998 event.

CDC also provides LIM forecasts of thetropical north Atlantic and CaribbeanSSTs in the ELLFB. Again, these fore-casts are available through our website athttp://www.cdc.noaa.gov/~lem/Atlan-tic.forecast.html. We have shown thatforecast skill in these Atlantic regions isrelated to the skill of predicting El Niño.We have also used LIM to diagnose thedynamical nature of tropical AtlanticSST variability, and published evidencethat the familiar “dipole” SST anomalypattern is a dynamically realizable struc-ture (as opposed to merely the dominantEOF of regional SST variability), butwhose development is often interruptedby ENSO influences from the eastPacific.

Our original claim that the basic dynam-ics of ENSO are stable, linear and sto-

chastically forced has beenindependently confirmed by others and isgradually forcing a paradigm shift in thisfield. We have continued investigating

-3

-2

-1

0

1

2

3

-1

-0.5

0

0.5

1

1950 1960 1970 1980 1990N

ino3

.4SS

T(0 C

)

Cor

r(S

ST,O

.S.)

-3

-2

-1

0

1

2

3

-0.8 -0.4 0 0.4 0.8

Nin

o3.4

SST

(0 C)

Corr (SSTA, O.S.)

R=0.68

-3

-2

-1

0

1

2

3

-1

-0.5

0

0.5

1

1992 1994 1996 1998 2000 2002

Nin

o3.4

SST

(0 C)

Cor

r(S

ST,O

.S.)

-3

-2

-1

0

1

2

3

-0.8 -0.4 0 0.4 0.8

Nin

o3.4

SST

(0 C)

Corr (SSTA, O.S.)

R=0.67

Fig. 2.2 Left panels: Time series of SST anomaly in Niño 3.4 (blue line), and the pattern correlation of the SSTanomaly pattern seven months earlier with the optimal structure (red line). The information in these left panels isalso displayed in the form of scatter plots in the right panels.

Fig. 2.3 Time series of actual LIM forecast errors nor-malized by one standard deviation of the total forecastuncertainty. The horizontal lines at + 1.96 indicate the95% confidence interval. The red and blue arrows indi-cate prominent El Niño and La Niña events during thisperiod.



the dynamical nature of ENSO, espe-cially in collaboration with scientists atTexas A&M University. A recently pub-lished study showed that an intermediatecoupled ENSO-prediction model of theCane-Zebiak type, tuned to be in a stablestochastically forced regime, generatedmore realistic SST variability than didthe same model tuned to give self-sus-tained oscillations.

2.2 Understanding and predicting theglobal impact of tropical SST varia-tions

CDC scientists are using a combinationof observational and general circulationmodeling approaches to address thisproblem. The observational studies relyheavily on the 50-year NCEP reanalysisdataset. The GCM studies are conductedby running various versions of the NCEPand GFDL global atmospheric modelswith prescribed SST forcing, in somecases by coupling to a mixed layer inparts of the world ocean. We have alsoanalyzed 10–12 member ensembles of50-year runs made by several GCMgroups (NCEP, GFDL, NCAR, ECHAM,IRI) with prescribed observed SST forc-ing, generally for the period 1950–1995,prescribed either globally or in the trop-ics.

2.2.1 Prediction skill and predictability

Outside the tropics, SST-forced signalsaccount for a relatively small (generallyless than 25%) portion of extratropicalvariability on seasonal to interannualscales. This fundamentally limits theaverage skill of a deterministic (asopposed to a probabilistic) seasonal fore-

cast, regardless of whether it representsthe mean of a large forecast ensemble oreven a multi-model ensemble. The limi-tation arises from a generally small sig-nal-to-noise ratio, and cannot beovercome by improving models. Thenoise is associated with chaotic (i.e.,unpredictable) nonlinear interactions andis intrinsic to the extratropical atmo-sphere. Still, in extreme individual cases,the signal can exceed the noise, makingrelatively skillful forecasts possible.

There are two other confounding factorsthat make it difficult for even sophisti-cated GCMs to improve upon the fore-cast skill of simple statistical modelsbased on linear correlations betweentropical SST and extratropical circulationanomalies. The first is the approximatelinearity of the remote response toENSO. The other is the relative insensi-tivity of that response to details of thetropical SST forcing; it appears thatknowledge of the area-averaged anomalyin Niño-3.4 alone is almost enough.

Figure 2.4 provides a good illustration ofthese points. It shows the correlation ofobserved JFM-mean 200 mb geopoten-tial height anomalies with those pre-dicted, over a 26-year period, using twoforecasting systems of vastly differentcomplexity. The top panel shows the skillof a 9-member ensemble-mean forecastby the NCEP atmospheric GCM forcedwith observed concurrent SST anomaliesin the tropical Pacific between 20°N and20°S. Consistent with many studies, thecorrelation of the observed and predictedheight anomalies is high in the tropics,and appreciable over North America andthe northeast and southeast Pacific



oceans. This is encouraging, although itshould be noted that the GCM forecastsare not true forecasts. Still, they give anidea of the potential predictability of sea-sonal anomalies around the globe if thetropical Pacific SSTs were to be pre-dicted accurately. The surprise in Fig. 2.4is the lower panel. It shows the skill ofthe simplest conceivable linear regres-sion forecasts for the same cases as in theupper panel, using the regression coeffi-cients of observed JFM 200 mb height

anomalies against the area-averagedobserved JFM SST anomaly in Niño-3.4.The forecasts themselves are made usingthe observed Niño-3.4 SST anomaly inthe previous 3-month period (OND) asthe predictor. These simple forecasts areclearly comparable in skill to the GCMforecasts. They also represent legitimate1-season ‘coupled model’ forecasts, inthat they incorporate a trivial persistenceforecast of the Niño-3.4 SST anomaliesfrom OND to JFM.

0.30.3

0.3

0.3

0.3 0.3 0.30.3

0.3

0.3

0.3

0.50.5

0.50.5 0.5

0.5

0.5

0.9

0.9

0.9

(a) Skill of atmospheric GCMforced with observed JFM tropical Pacific SST anomalies

0.3

0.3

0.30.3

0.3 0.3 0.3

0.3

0.3 0.3

0.3

0.3

0.50.50.5

0.50.5

0.5

0.5

0.5

0.9

0.9

(b) Skill of 1-dimensional statistical modelfrom regressions on observed Nino3.4 SST anomalies in the previous season (OND)

Local Anomaly Correlation of Predicted and ObservedJFM-mean 200 mb height anomalies (1969-94)

Fig. 2.4. Wintertime 200 mb height seasonal forecast skill of (a) the NCEP atmospheric GCM with specifiedobserved SSTs in the tropical Pacific basin and (b) a simple linear regression model based on seasonal 200 mbheight correlations with Niño 3.4 SSTs. See text for details.



The rough agreement between the twopanels in Fig. 2.4 may be interpreted asreflecting either true seasonal predict-ability limits or the need for furtherGCM improvement. There is room forboth interpretations, although we aremore inclined toward the former. GCMerror is probably not the main culprithere: several other GCMs analyzed by usyield skill patterns very similar to that inthe upper panel. Also, when the NCEPGCM is asked to predict its own behav-ior, such as when using an 8-memberensemble-mean to predict the 9th mem-ber's seasonal anomalies, its skill is againsimilar to that in the upper panel. Onecan thus make a case that the modestextratropical values in Fig. 2.4 aremainly a reflection of the limited intrin-sic predictability of extratropical sea-sonal averages associated with tropicalSST forcing. As mentioned earlier, thisin turn is mainly due to a modest signal-to-noise ratio.

CDC scientists have attempted to clarifythe relationship between the expectedanomaly correlation skill ρ of ensemble-mean forecasts and the signal-to-noiseratio s, defined as the ratio of the ensem-ble-mean anomaly to the ensemblespread. Figure 2.5 summarizes this gen-eral relationship, which is useful in inter-preting many GCM results. The and

curves show the expected skill of infi-nite-member and single-member ensem-ble-mean forecasts with a perfect model.The third (blue) curve depicts theexpected skill of infinite-member ensem-ble-mean forecasts with an imperfectmodel, whose systematic error se (i.e., itserror in determining the true s) is of thesame magnitude as s. Note that these

curves are applicable to any forecastvariable, in any forecasting situation, andto any forecasting method, including theregression method used in Fig. 2.4b.

The curve represents a hard predict-ability limit with a perfect model, andshows that to produce 'useful' forecastswith anomaly correlations greater than0.6, s needs to be greater than 0.75. Toproduce 'excellent' forecasts with anom-aly correlations greater than 0.9, s needsto be greater than 2. Given the evidencefrom several studies that s for ENSO-related 200 mb seasonal height anoma-lies is approximately between 0.5 and 1in the extratropical Pacific-Americansectors of both hemispheres, and greaterthan 2 in the tropics, the results in Fig.2.4 are not surprising. Figure 2.5 is alsouseful for assessing to what extent themodest skill in Fig. 2.4a might be due tomodel error or using only 9-memberensembles. The difference between

ρ∞ρ1

ρ∞

ρ∞

0 1 2 30.0

0.2

0.4

0.6

0.8

1.0

Ano

mal

y C

orre

latio

n ρ∞

s = ensemble mean anomaly ensemble spread

ρ1

ρ∞ for Se= S

Expected Forecast Skill as afunction of signal to noise ratio

Fig. 2.5 Expected anomaly correlation forecast skillof ensemble-mean forecasts as a function of signal tonoise ratio s. The red curve is the expected skill of aninfinite-member forecast, the black curve theexpected skill of a single-member forecast. The bluecurve shows the expected skill of an infinite-memberensemble forecast using an imperfect model whosestandardized systematic error is of the same magni-tude as s.



and ρ1 shows the potential gain in skillby using infinite-member ensemblesinstead of a single member. The maxi-mum gain is 0.25, for s ~ 0.6. However,most of this gain is attainable with about25 members, and even a ρ9 curve (notshown) is close to the curve. The lossof skill due to model error (blue curve) isprobably of greater concern in Fig. 2.4athan not having enough members. How-ever, model error could equally be affect-ing skill in Fig. 2.4b.

2.2.2 New research challenges

Figures 2.4 and 2.5 together suggest thatthe modest skill, on average, of determin-istic extratropical seasonal forecasts islargely consistent with the predictabilitylimits imposed by the modest local val-ues of s associated with the tropicallyforced signal. Given also the evidence inFig. 2.4b that similar skill can also beachieved with simple linear regressionmodels, the question naturally arises asto what further useful predictive informa-tion can be extracted by running GCMensembles.

CDC scientists have spent considerabletime pondering this issue, and have comeup with several encouraging possibilities.In one way or another, they all involvefocusing on the distributional aspects ofthe ENSO response rather than on theensemble-mean. For example, as Fig.2.6a shows, a modest shift in the mean of0.5 (in units of standard deviation), whilenot large enough to affect appreciably theexpected seasonal mean of an extratropi-cal variable, can still greatly affect theprobability of its extreme values. Therisk of obtaining an extreme positivevalue of greater than +1 increases from16% to 31%, and the risk of obtaining anextreme negative value decreases from16% to 7%. Thus without ENSO therisks of extreme positive and negativeanomalies are the same, but with ENSO,even for a modest s of 0.5, the risk of anextreme positive anomaly becomes 4.4(=31/7) times the risk of an extreme neg-ative anomaly. Figure 2.6b shows howthis risk ratio can be equally stronglyaffected by modest changes of noise. Inthis example, a 20% reduction of stan-dard deviation combined with a mean

ρ∞

-1 0 1

0.5

7% 31%

(a) shift of PDF only

-1 0 1

0.5

3%

27%

(b) shift with change of spread

Fig. 2.6. (a) An illustration of how a mean shift of 0.5 of a normal distribution increases the probability of extremepositive values from 16% to 31% and decreases the probability of extreme negative values from 16% to 7%. (b)Illustration of the altered probabilities of extreme values when the mean shift of 0.5 is also accompanied by areduction of the standard deviation from 1 to 0.8.



shift of 0.5 changes the risk ratio from 1(=16/16) to 9 (=27/3).

When even minor PDF shifts andchanges of variance imply large changesin the risks of extremes values, determin-ing them accurately becomes important.It is easy to see how a good GCM mighthave an advantage over the regressionmodel of Fig. 2.4b in this regard, whoseparameters cannot be estimated accu-rately enough from the limited observa-tional record to have confidence in itspredictions of extreme values. Further,one can run as many GCM ensemblemembers as necessary in individual fore-cast cases to predict the changes ofextreme risks within specified confidenceintervals. The regression-model alsoassumes, in effect, that ENSO-inducedmean PDF shifts are strictly linear withrespect to the SST forcing and that thereare no changes of noise, or variability.CDC scientists have spent considerableeffort on ascertaining the extent to whichsuch assumptions are valid, since theyhave a large bearing on the problem athand.

2.2.3 Understanding the sensitivity of theatmospheric response to details of theanomalous SST forcing

The regression-model used in Fig. 2.4balways predicts the same signal patternof the global atmospheric response; onlyits amplitude varies from forecast case tocase in direct proportion to the strengthof the Niño 3.4 SST anomaly. As wehave seen, this doesn't seem to affect itsdeterministic forecast skill overmuch,but the question is whether it limits pre-dictions of the risk of extreme anomalies.

To what extent does the remote SST-forced signal vary from case to case? Towhat extent are its variations determinedby the nonlinearity of the response to theamplitude and sign of the SST forcing inNiño 3.4, and to what extent by thedetails of the SST anomaly pattern in thewider tropical Indo-Pacific domain? Wehave conducted several studies to answerthese questions.

Figure 2.7 gives a sense of the signalvariation from case to case. Samplinguncertainty is an issue in this problem,given that the number of samplesrequired to establish that s is statisticallydifferent from zero is inversely propor-tional to the square of s. At the 5% level,the number of samples should be greaterthan 8/s2. To establish the significance ofs = 0.5 thus requires 32 samples; toestablish the significance of changes of sfrom case to case, say of order 0.25,would require many more, 128. With thisin mind, we ran a very large 180-memberseasonal ensemble with the NCEP atmo-spheric GCM with prescribed observedglobal SST forcing corresponding to theEl Niño of 1987, and another 180-mem-ber ensemble for the La Niña of 1989.The right-hand panels of Fig. 2.7 showthe ensemble-mean 500-mb heightanomalies obtained in these integrations(defined with respect to the ensemble-mean obtained in another set of 180 inte-grations with climatological-mean SSTforcing). To our knowledge this is themost statistically confident determinationof the global SST-forced signal evermade for two different observed SSTforcing patterns. Note that the sign of theresponse has been reversed in the lowerpanel for easier comparison with the



upper panel. For comparison we alsoshow in the left panels, in an identicalformat, observational 500-mb compositeanomaly patterns for 10 El Niño and 10La Niña events based on the Niño 3.4index, and defined with respect to 10“neutral” events. Note that the ampli-tudes of these composite patterns havebeen scaled by 0.73 and 1.36, in propor-tion to the observed Niño 3.4 SST anom-

aly magnitudes during the moderate1987 El Niño and the strong 1989 LaNiña events, respectively. Thus these leftpanels may be interpreted as the SST-forced 500-mb height signals during1987 and 1989 as predicted by an empiri-cal method. This method is superior, inprinciple, to that used in Fig. 2.4b in thatit predicts different response patterns in

Fig. 2.7 The SST-forced signal in 500 mb heights during the El Niño winter of JFM 1987 (top panels) and the LaNiña winter of JFM 1989 (bottom panels, with sign reversed) estimated by a GCM (right panels) and an empiri-cal method (left panels). The right panels show the ensemble-mean anomaly fields obtained in 180 seasonal inte-grations of the NCEP AGCM with prescribed global SST forcing during JFM 1987 and JFM 1989. The leftpanels show scaled historical composite 500 mb anomaly fields derived from 10 observed El Niño and 10observed La Niña events in the 1958–1999 period, defined with respect to a Niño 3.4 SST anomaly index. Thecomposite patterns were scaled by factors of 0.73 and 1.36 to account for the moderate Niño 3.4 index magnitudeduring 1987 and the larger magnitude during 1989, respectively. The contour interval is 10 m in all panels. Posi-tive values are indicated by red and negative by blue shading.

SST-forced 500 mb height signals in JFM

(a) Empirical model predicted signal in 1987

(b) Empirical model predicted signal in 1989 (times -1)

(c) GCM-estimated signal in 1987

(d) GCM-estimated signal in 1989 (times -1)



El Niño and La Niña cases, as shownhere.

Although the GCM's signal patterns forthese individual El Niño and La Niñawinters are generally similar to oneanother, there are some notable differ-ences. The El Niño response is strongerin the PNA sector, despite the weakerSST forcing. This is also true in theempirical forecast. There is little else tocompare between the GCM and empiri-cal forecasts because of sampling uncer-tainty in the empirical forecasts, i.e., thefact that the left panels are derived fromonly 10 cases each in the historicalrecord. In areas such as the North Atlan-tic where the left panels predict a strongasymmetric signal of the same sign in1987 and 1989, the significance of thatasymmetry is therefore questionable.The differences between the GCM's pre-dicted signal patterns for 1987 and 1989are much more reliable in this regard,and though more modest, are largeenough that they would have had impor-tant implications for predicting the risksof extreme anomalies during these win-ters.

There is thus evidence of significant sig-nal variation from case to case. To putthese results on a stronger footing, onemight ideally wish to generate similar180-member ensembles for each one ofthe past 50 or so winters. This has not yetbeen done. However, results from a 46-member multi-model ensemble of fourdifferent atmospheric GCMs (NCARCCM3, NCEP, GFDL, and ECHAM)offer additional evidence of the existenceof different response patterns. Between10 and 12 member ensembles were gen-

Fig. 2.8 The first three EOFs of SST-forced 500 mbheight signals in a multi-GCM ensemble of 50-yearintegrations made with prescribed evolving observedglobal SST boundary conditions. See text for furtherdetails. The values plotted are the correlations of theEOF's Principal Component time series with the local500 mb height time series of the SST-forced signal.Positive values are indicated by red and negative byblue shading.



erated for each model forced with identi-cal evolving observed global SSTs overthe past 50 years. For each one of the 50winters, the SST-forced signal at 500 mbwas defined as a weighted average of theensemble-mean responses of the 4GCMs. Finally, an EOF analysis of these50 signal patterns was performed. Figure2.8 shows the first three EOFs, togetherwith their fractions of the total signalvariance explained. The leading EOFalone accounts for 57% of the global sig-nal variance, and as much as 80% overthe PNA region. Most elements of thispattern are evident in all the four panelsof Fig. 2.7. The dominance of this signalpattern, its strong similarity to the classicobserved ENSO teleconnection pattern,and also to the unchanging forecast pat-tern of the regression model used in Fig.2.4b, explains why the upper and lowerpanels of Fig. 2.4 are so similar. Still,there is evidence in Fig. 2.8 of apparentlyminor but potentially important devia-tions from this dominant signal patternfrom winter to winter. The second EOF islargely zonally symmetric, with out-of-phase anomalies in polar and subtropicallatitudes. Locally, it explains a substan-tial fraction of the signal variability in thesubtropics, and its associated PrincipalComponent (PC) time series describes atropospheric warming trend of lower lati-tudes over 1950–1999 associated with along-term tropical SST warming trend.The third EOF resembles a tropicallyforced wavetrain with centers spatiallyshifted relative to those of the leadingEOF. Its PC time series regressed on thesimulated tropical precipitation fieldsyields a regression map with appreciablemagnitudes in the western and central

equatorial Pacific. This third EOF maythus reflect a genuine sensitivity of theSST-forced extratropical signal to varia-tions of the anomalous tropical SST pat-tern from winter to winter.

It should be stressed that the GCMresults in Figs. 2.7 and 2.8 are all for glo-bal SST forcing. To what extent can thesignal variation evident in these figuresbe attributed strictly to the tropicalPacific SST forcing, in particular toasymmetric responses to El Niño and LaNiña forcing? There is a strong sugges-tion of a weaker response to La Niñaforcing in the lower panels of Fig. 2.7.Such a weaker response (not shown) isalso clearly evident in the composite ElNiño and La Niña signals in the 4 differ-ent ensemble GCM simulations for1950–1995 described above. Are therealso significant differences in the pat-terns of the response? To what extent dothey contribute to the third EOF in Fig.2.8? Several CDC studies have attemptedto address such questions cleanly, byexamining the atmospheric response tothe first EOF pattern of tropical PacificSST (very similar to the Pacific portionof the lower panel of Fig. 2.1) with posi-tive (“El Niño”) and negative (“LaNiña”) signs. A weaker response to thenegative EOF forcing has indeed beenconfirmed, but is appreciable only forlarge amplitude forcing.

Figure 2.9 shows results from a GCMexperiment designed specifically toaddress this issue. A 9-member NCEPGCM ensemble was generated for 1963–1989 by forcing throughout the 27 yearswith the first SST EOF pattern, varying



in magnitude and sign as its associatedPC time series. The upper panels of Fig.2.9 show the ensemble-mean responsesduring the strongest warm and cold win-ters of 1983 and 1974, respectively, inthis period. The response is clearlyweaker for the 1974 event. This is sug-gestive but not conclusive, since themagnitude of the SST forcing was alsoweaker in 1974. To settle this, the entireexperiment was repeated with the sign ofthe PC time series reversed. The ensem-ble-mean responses for the sign-reversed

1983 and 1974 winters are shown in thelower panels. The 1974 response is nowstronger than the 1983 response, despitethe weaker magnitude forcing. We hadnoted this effect earlier in Fig. 2.7, butthe result here is cleaner. It confirms thatthe remote atmospheric response to thetropical Pacific SST forcing is apprecia-bly stronger for strong warm than for thestrong cold SST forcing. A top-to-bot-tom comparison in Fig. 2.9 compares theresponses to the same SST forcing but of

Fig. 2.9 GCM 200 mb height responses in northern winter (DJF) to the first EOF pattern of tropical Pacific SSTforcing, with the observed amplitude during DJF 1982/83 (upper left panel), the observed amplitude during DJF1973/74 (upper right), the negative of the observed amplitude during DJF 1982/83 (lower left), and the negative ofthe observed amplitude during DJF 1973/74 (lower right). The contour interval is 10 m. Positive values are indi-cated by red and negative by blue contours. See text for further details.



opposite sign, and further confirms thisresult.

2.2.4 ENSO-induced changes of variabil-ity

To what extent does ENSO affect theatmospheric noise (i.e., the variability) ashypothesized in Fig. 2.6b? We haveaddressed this issue in several recentpublications. In one study, we examinedthe standard deviation of seasonal-mean500 mb heights in our 180-memberGCM ensembles for the winters of 1987and 1989, and found a modest overallincrease in the warm (1987) and adecrease in the cold (1989) ensemblecompared to that in the neutral 180-member ensemble. This was speculatedto be forced partly by the increased vari-ability of seasonal precipitation in thewarm (and decreased variability in thecold) ensemble in the Niño-4 area of thecentral equatorial Pacific, which has beenshown to be sensitive region for forcing alarge global circulation response.Another study searched for ENSO-induced changes of seasonal noise in thesmaller AMIP-style GCM ensemblesused in Fig. 2.8, but found little impact inthe PNA region. While these findingsseem to conflict, a closer look at the pub-lished figures shows that the results arenot inconsistent in the PNA region. Tothe extent that the altered extratropicalnoise is due to the tropical precipitationnoise, the effect is also probably bothGCM-dependent and ENSO event-dependent. It should be mentioned thatsampling uncertainty is of even greaterconcern here than in Figs. 2.7 and 2.8.The number of samples required to

establish the significance at the 5% levelof a fractional change ∆ of an ensemble'sstandard deviation is close to 3/∆2. Toestablish the significance of the ∆ of 0.2(corresponding to a 20% change of stan-dard deviation) in Fig. 2.6b would thusrequire 75 samples from both the neutraland altered distributions. Any change ofsmaller than 17.5% would require morethan 100 samples.

The effect of ENSO on subseasonalextratropical variability is equally impor-tant, and somewhat easier to establish.These effects can be distinct from theeffects on seasonal mean quantities, andcan have important practical implica-tions. For instance, one may imagine asituation in which El Niño alters theoccurrence of both cold waves and warmspells in a winter. The effect is a mean-ingful change in the risk of extremeweather, even though little seasonalmean signal might be evident. The fewpublished studies on this topic, con-strained either by sampling requirementsor data availability, have formed compos-ites over several ENSO events to diag-nose the effect in limited regions. In anambitious recent study, we have esti-mated the effect globally from our largeAGCM ensembles for 1987 and 1989,and compared them with observationalcomposites based on 11 El Niño and 11La Niña events in the recent record. As inFig. 2.7, the purpose of this comparisonwas to gauge to robustness of thechanges of variability, their predictabil-ity, and their variation from event toevent.



OBS GCM

synoptic

monthly

El Niño-induced change of 500 mb height variance

Fig. 2.10 El Niño-induced changes of variance on three different subseasonal time scales. The quantity plotted isthe square root of the anomalous variance, with red shading for positive and blue for negative anomalous variance.The left panels are based on statistics averaged over 11 observed El Niño and 11 observed “neutral” JFM winters inthe NCEP reanalysis dataset. The right panels are derived from a large AGCM ensemble with observed SST forcingfor the El Niño winter of JFM 1987. Top panels: Synoptic scale (2 to 6 day periods), Middle panels: Intraseasonalscale (8 to 45 day periods.). Bottom panel: Monthly scale (30-day averages). Contours are drawn at 8 m intervalsstarting at 4 m in the top panels, and at 16 m intervals starting at 8 m in the middle and lower panels. See text for afuller explanation.

intraseasonal



The most important result from this anal-ysis, depicted in Fig. 2.10, is that the pat-terns of the SST-forced anomalous heightvariability are markedly different for thesynoptic (2 to 7 days), intraseasonal (8 to45 days), and monthly (30 -day average)time scales. In contrast, the patterns ofthe anomalous tropical rainfall variability(not shown) are nearly identical acrossthese time scales. Figure 2.10 showscontours of ∆σ , defined as the signedsquare root of the anomalous variancedifference of 500 mb heights on thesetime scales (where by “signed” we meanthat if the anomaly is negative, we depictits square root with a minus sign). Theresults for La Niña (not shown) are simi-lar and generally of opposite sign. Thecomparison between the GCM andobservational panels in Fig. 2.10 is notclean. Nevertheless, their gross similarityis reassuring, both for the robustness ofthe changes of variability and this GCM'sability to simulate them. To that extent,their dissimilarity can be attributed to thecomparison not being clean (i.e., event-to-event differences) and sampling error,especially in the observations (which iswhy the observational anomalousmonthly variability map has been omit-ted).

The main ENSO effect on the synopticscale is a southward shift of the stormtrack over the Pacific ocean and NorthAmerica. On the intraseasonal scale, it isa decrease of height variance over thenorth Pacific, consistent with a tendencyof reduced blocking activity during ElNiño. On monthly (and seasonal) scalesthere is a suggestion of an overallincrease of variance. Referring back to

Fig. 2.6b, it is evident that these differingENSO impacts on extratropical noise ondifferent time scales have very differentimplications for the risks of extremeanomalies on these scales. We believethat three quite distinct dynamical mech-anisms are responsible for these sharpdifferences, and are currently investigat-ing them in a hierarchy of dynamicalmodels.

2.3 Understanding the impact of extra-tropical SST variations

Despite many studies, the precise role ofextratropical SST variations in thedynamics of atmospheric low-frequencyvariability remains unclear. Many atmo-spheric GCM experiments have beenperformed with prescribed extratropicalSST anomalies to determine the atmo-spheric response to them. The resultshave been variable and confusing, notonly because the response is weak anddifficult to establish in short integrations,but also because it is sensitive to the pre-cise location of the prescribed SSTanomaly in relation to the atmosphericjet streams and their associated stormtracks. The largest extratropical SSTanomalies tend to occur near the jets, andmay be thought of as having a directeffect on the mean flow, and on the stormtracks through the change in the meanflow. The altered storm tracks can, how-ever, feed back on the mean flow. Thisindirect effect can be very important, butis difficult for GCMs to represent accu-rately.

We have investigated these issues in aseries of long NCEP AGCM runs with



prescribed extratropical SST forcing. In apreviously published study, an idealizedwarm SST anomaly was imposed nearthe Kuroshio region of the northwestPacific. The atmospheric response to thiswas radically different in perpetual Janu-ary and perpetual February integrationsof the GCM: a baroclinic response inJanuary and a nearly equivalent-barotro-pic ridge response in February. To under-stand this, we used a linear model toinvestigate the response to the SSTanomaly as the sum of the direct heating-induced component and the indirecteddy-driven component associated withthe altered storm tracks. (A novel featureof this calculation was that the alteredstorm-track forcing specified in the linearmodel was specifically that induced bythe heating altered mean flow, diagnosedusing Whitaker and Sardeshmukh'sstorm-track model). This diagnosis con-firmed the dominance of the eddy-drivencomponent, forced mainly by the alterededdy-momentum fluxes, in determiningthe barotropic ridge response. Withoutthe eddy forcing, the response to a shal-

low heat source associated with thewarm SST anomaly was always baro-clinic, with a trough near the surface anda ridge aloft and downstream. Thisexplains the baroclinic response to extra-tropical SST anomalies obtained in manylow-resolution GCMs, in which the sim-ulated storm tracks and eddy momen-tum-fluxes are generally too weak: insuch GCMs the SST-induced alterationsof the storm track are also too weak. Anequivalent-barotropic ridge responsedevelops only if the storm track is alteredsufficiently strongly that strong anoma-lous eddy-forcing is generated at the cor-rect locations to reinforce the heating-induced upper-level ridge and offset thelower-level trough. In this way, the atmo-spheric response becomes sensitive to thecharacteristics of the storm track and theposition of the SST anomaly relative tothe storm track.

As an example, Fig. 2.11 shows the lin-ear model results for idealized shallowheat sources associated with warm extra-tropical SST anomalies imposed on the

Fig. 2.11 A linear balance model's 250 mb geopotential height response in winter to a shallow elliptic heat sourcecentered (a) at 40°N,160°E and (b) at 40°N,160°W. The model's height response to the anomalous eddy vorticityfluxes induced by these heat sources is shown in (c) and (d), respectively. The contour interval is 5 m; positive val-ues are indicated by red and negative by blue contours.



observed winter-mean basic state. Withan elliptic source specified over the west-ern Pacific (left panels), the heating-induced upper-level ridge (upper panel)is reinforced by the eddy-driven equiva-lent-barotropic ridge (lower panel) overthe Pacific, similar to the GCM's Febru-ary response. No such reinforcement bythe transients occurs for a similar sourcespecified over the eastern Pacific (rightpanels).

2.4 Understanding and predicting SSTvariations outside the tropical Pacific

2.4.1 The atmospheric bridge

One example of global climate interac-tion is the “atmospheric bridge”, whereatmospheric teleconnections associatedwith ENSO drive anomalous ocean con-ditions outside of the equatorial Pacificthrough changes in the heat, momentum,and fresh water fluxes across the air-seainterface. The resulting SST anomaliescan also feed back on the initial atmo-spheric response to ENSO. As part of theGFDL-Universities Consortium project,we developed a coupled AGCM-mixedlayer ocean model and used it to conductexperiments to study the atmosphericbridge and other air-sea interaction pro-cesses. In the “MLM” experiment,observed SSTs were prescribed asboundary conditions in the tropicalPacific (15°N–15°S, 172°E–South Amer-ican Coast), and the remainder of the glo-bal oceans were simulated using thevariable-depth mixed-layer model. AsFig. 2.12 shows, the simulated SLP andSST anomalies associated with ENSOare fairly realistic, with stronger cyclonic

circulation and cold water over the cen-tral North and South Pacific and warmwater along the west coast of the Ameri-cas. ENSO-induced changes in theWalker circulation also lead to warmSSTs in the north tropical Atlantic andIndian Oceans.

How useful is this effect in actually pre-dicting the interannual variations of SSTsoutside the tropical Pacific basin? Figure2.13 provides one measure of the fore-cast skill. It shows the correlation of theobserved seasonal-mean SST anomalies,over the 50-year 1950–1999 period, withthose predicted by the MLM model. Thepredicted field for each season representsa 16-member ensemble-mean. Resultsare shown separately for the 50 winter(JFM) and 50 summer (JAS) forecast

Fig. 2.12. (a) Observed and (b) simulated El Niñominus La Niña composite of SLP (contour interval of1 mb) and SST (shading interval of 0.2 °C) for DJF(0/1), where 0 indicates the ENSO year and 1 the nextyear. The composite is based on 9 El Niño and 9 LaNiña events during 1950–1999. The observed valuesare from the NCEP reanalyses and the model resultsfrom an average of 16 MLM integrations.

Observed SLP (mb) / SST (°C)

MLM SLP (mb) / SST (°C)



cases over which the correlations werecalculated. The correlations are generallyhigher than 0.4 in the central north andsouth Pacific oceans, and also in the trop-ical Indian and tropical Atlantic oceans.This is encouraging, and also sheds somelight on why the LIM SST forecast mod-els described in section 2.1, that arebased solely on SST correlationsbetween different tropical locations, per-form as well as they do: the “bridge”effect is implicitly included in them. It isalso interesting to compare Fig. 2.13with Fig. 2.12 in areas such as the northAtlantic, where Fig. 2.12 suggests abridge effect but Fig. 2.13 shows it to beunimportant.

2.4.2 The re-emergence of long-livedsubsurface temperature anomalies

The atmospheric changes associated withENSO influence upper-ocean processesthat affect the subsurface temperaturestructure and mixed-layer depth (MLD)long after the ENSO signal decays. Ther-mal anomalies that form in the surfacewaters of the extratropics during winterpartially reemerge in the following win-ter, after being sequestered beneath themixed layer in the intervening summer.SST anomalies generated via the atmo-spheric bridge recur in the followingwinter in central North Pacific via thisreemergence mechanism (Fig. 2.14). TheMLD is substantially deeper in the cen-

Fig. 2.13. Predictability of seasonal SST anomalies worldwide via the “atmospheric bridge” from the tropicalPacific ocean, in winter (top panel) and summer (bottom panel). Values plotted are the correlations of observedSST anomalies with those predicted by an atmosphere–mixed layer ocean coupled model with specified observedSSTs in the central and eastern tropical Pacific ocean. See text for further details.



tral North Pacific during El Niño than LaNiña winters, but the reverse is true in thesubsequent winter. During El Niño win-

ters, enhanced buoyancy forcing (surfacecooling) and mechanical mixing createsa colder and deeper mixed layer. Afterthe MLD shoals in late spring, the coldwater stored beneath the surface layer aspart of the reemergence process increasesthe vertical stability of the water column,reducing the penetration of the mixedlayer in the following fall and winter.

The reemergence process is not confinedto the central North Pacific, nor does itoccur solely in conjunction with theatmospheric bridge. Figure 2.15 showsthe evolution of the leading pattern ofNorth Pacific SST variability in observa-tions and in two distinct AGCM – oceanmodel experiments. In the first experi-ment, the mixed layer ocean model isactive over the entire globe, including thetropical Pacific, and thus does notinclude ENSO. In the second experiment,

Fig. 2.14 The composite El Niño minus La Niña oceantemperature from Nov(0)–Jun(2) and the compositemixed layer depth (m) during El Niño (red line) andLa Niña (blue line) from the MLM experiment in thecentral North Pacific (180°–160°W, 28°N–42°N).

Fig. 2.15. The evolution of the leading pattern of SST variability over 20°N–60°N in the Pacific as indicated byextended EOF analyses of monthly SST anomalies from January through April of the following year. The resultsare presented as the correlation between the leading principal component (time series of EOF1) with SST anoma-lies at the individual grid points for March, September and March of the following year. (The other months, whichare not shown, indicate a similar evolution). Results are presented for (a) observations for 1950–1995, (b) anAGCM-global MLM simulation, and (c) 4 TOGA-50m slab simulations. The contour interval is 0.2 with values >0.4 shaded red and those < -0.4 shaded blue.



observed SSTs are specified in the tropi-cal Pacific, but the remainder of theworld oceans are simulated by a slabmodel without mixed-layer physics. Theobservational column in Fig. 2.15 showsthat the dominant large-scale SST anom-aly pattern that forms in the eastern two-thirds of the North Pacific during winterrecurs in the following winter withoutpersisting through the intervening sum-mer. Experiment 1 (middle column)reproduces this behavior, but Experiment2 (right column) does not. These resultssuggest 1) that the winter-to-winter SSTcorrelations are due to the reemergencemechanism and not due to similar atmo-spheric forcing of the ocean in consecu-tive winters, and 2) that the SSTanomalies in the tropical Pacific associ-ated with El Niño are not essential forreemergence to occur.

EPILOGUE

Climate variability on seasonal to inter-annual scales is dominated by the tropi-cal ENSO phenomenon and its globalimpacts. CDC scientists have been at theforefront in providing evidence thatmuch of the predictable evolution ofENSO and its remote atmospheric andoceanic impacts are governed by low-dimensional, linear dynamics. This clari-fication has proved extremely valuablefor basic understanding and for buildingsimple, useful, forecast models. How-ever, it has also raised important newquestions. Foremost among these is per-haps that raised by Figure 2.4. Howmuch predictability, especially extratrop-ical predictability, exists on this timescale beyond that associated with simplelinear ENSO signals? What additional

useful predictive information can beextracted by running large GCM ensem-bles? CDC scientists have addressedthese questions by exploring the nonlin-earity and sensitivity of the globalresponse to the details of anomaloustropical Pacific SST fields and by focus-ing on the distributional aspects of theresponse, especially the changes of vari-ance, rather than just shifts of the mean.We have also contributed to improvedunderstanding of the predictability ofSSTs in other ocean basins, through both“atmospheric bridge” and “re-emer-gence' mechanisms, and the impact ofsuch SSTs on atmospheric predictability.There is encouraging evidence that onewill be able to provide substantiallyimproved forecast guidance on this timescale, especially on the tails of the distri-butions, through better understandingand prediction of these formally second-order but still important effects. Theoverarching theme of new research inthis area must be a community-wide shiftfrom deterministic to probabilistic sea-sonal predictions. Ultimately, one canextract only so much information fromdeterministic predictions of a chaoticsystem. The utility of probabilistic pre-dictions, on the other hand, is unboundedin an important practical sense, in that itis determined to a large degree by theneeds of particular users. Future CDCresearch on this time scale will increas-ingly reflect this shift of focus to proba-bilistic predictions, with the climate-society interface and the needs of differ-ent categories of user groups in mind.

Contributed by: M. Alexander, J. Bar-sugli, G. Compo, M. Hoerling, S. Peng,C. Penland, and P. D. Sardeshmukh.

modeling research on seasonal to interannual variability · improved, global gridded observational...

Documents