Proc. Indian Acad. ScL(Earth Planet. Sci.), Vol. 105, No. 2, June 1996, pp. 101-117.
9 Printed in India.
Origin of a trend in ECMWF wind analysis and its objective removal
B N G O S W A M I and H A N N A M A L A I
Centre for Atmospheric Sciences, Indian Institute of Science, Bangalore 560012, India
MS received 7 September 1995; revised 29 March 1996
Abstract. The suitability of the European Centre for Medium Range Weather Forecasting
(ECMWF) operational wind analysis for the period 1980-1991 for studying interannual
variabilityis examined.The changes in the model and the analysis procedure are shown to give
rise to a systematicand significanttrend in the large scale circulation features. A new method
of removing the systematic errors at all levelsis presented using multivariate EOF analysis.
Objectivelydetrended analysis of the three-dimensionalwind field agrees well with independent Florida State University (FSU) wind analysis at the surface. It is shown that the
interannual variations in the detrended surface analysis agree well in amplitude as well as
spatial patterns with those of the FSU analysis. Therefore, the detrended analyses at other
levels as well are expected to he useful for studies of variability and predictability at
interannual time scales. It is demonstrated that this trend in the wind fieldis due to the shift in
the climatologiesfrom the period 1980-1985 to the period 1986-1991.
Keywords. ECMWF objective analysis; interannual variability; combined EOF analysis;
removal of unphysical trend.
1. Introduction
The need for a comprehensive, internally consistent, homogeneous, multivariate data
set for studies of the mean climate system and its interannual variations has been
recognized for sometime (Bengtsson and Shukla 1988). Model-based operational
analyses from numerical weather prediction (NWP) centres, such as the National
Meteorological Center (NMC) of the United States and the European Center for
Medium range Weather Forecasts (ECMWF), have been used for this purpose.
Although these data sets have been used in quite a large number of diagnostic studies of
the general circulation, they have some serious deficiencies (Trenberth and Olson
1988), specially for studying variability and predictability at interannual time scales.
Three major reasons are: (a) gradual improvement of the models' resolution and
physical parameterizations, (b) frequent changes in the data assimilation system and (9
unavailability of delayed m o d e data. F o r example, the N M C model has evolved from
a 12-layer R24 resolution version in 1980 to an 18-layer T106 resolution version in
rec~nt times. Similarly, the E C M W F model has grown from a 15-level grid point
(1.875 ~ version in 1980 to a 19-level T213 spectral model in recent times. The analysis
system has also gone through gradual changes over the last 15 years. The most
significant change is the introduction of multivariate o p t i m u m interpolation and
diabatic nonlinear normal m o d e initialization in the mid-eighties. Trenberth and
Olson (1988) note that the introduction of a multivariate o p t i m u m interpolation
scheme to the E C M W F analyses on M a y 1st, 1985, increased the divergent part of the
tropical wind. Trenberth et al (1990) noted that the 1000rob monthly mean wind
101
102
B N Goswami and H Annamalai
convergence near the equator in the ECMWF analysis during 1980-1986 does not agree
with that expected from satellite-observed longwave radiation. Hurrell and Trenberth
(1992) applied a weighting function for the ECMWF temperature fields for the period
1982-1989. They observed that over the tropics, the analyses have improved, especially
after May 1985, when the 1000-200mb temperatures show more coherence. However,
Trenberth and Guillemot (1994) note that only during the most recent period, 1990-1993,
the ECMWF analysed mean annual surface pressure is reasonably stable.
As a result of these changes, the archived analysis includes unphysical jumps at
different points of time. Therefore, it has become necessary to reanalyse the past
data using a recent model and a more recent analysis system (Bengtsson and Shukla
1988). Major reanalysis efforts have been already started in NMC (Kalnay and Jenne
1991), ECMWF and NASA (Schubert et al 1993). In spite of the changes noted in the
ECMWF analysis, the present analysed circulation data sets have been used increasingly in recent times to study interannual variability (Wang 1992; Murakami and
Matsumoto 1994) and predictability (Barnett etal 1994). These studies utilize
anomalies defined as deviation from a climatology based on the entire time span of the
data. The changes in the model-based analysis system give rise to some unphysical
interannual variability. The studies that do not attempt to separate this analysis system
induced interannual variability from the real interannual variability are expected to
arrive at erroneous conclusions. In this study, we investigate whether the analysis
system induced interannual variability is systematic. If it is so, could it be separated
from real interannual variability? After removal of the analysis system induced
variability, is the analysis still useful for interannual variability studies?
While examining the ECMWF analysis from 1980 to 1991 to study interannual
variability of the Indian monsoon, we found that the circulation anomalies calculated.
traditionally have a clear trend, specially over the tropical region. In an attempt to
bring out the patterns of the variability related to sea surface temperature variations,
we carried out a combined empirical orthogonal function (EOF) analysis of sea surface
temperature (SST), outgoing longwave radiation (OLR) and winds at 1000, 850 and
200mb from ECMWF analysis. We found that some of the leading principal components have a significant trend. As the physical origin of the trend was unclear, we
conducted an EOF analysis of the SST, OLR and winds separately. The SST and OLR
data do not show any trend. However, EOF analysis of the ECMWF winds at the
different levels carried out one at a time or a combined EOF analysis of the winds at all
the levels taken together shows a significant trend in the dominant modes. Thus, it was
clear that the trend in the combined EOF analysis arose from the trend in the ECMWF
wind data. The wind anomalies prior to the mid-eighties tend to be of one sign, while
those in the later period tend to be of the opposite sign. Here, we show that this trend
arises mainly due to a shift of the climatologies from a state before the mid-eighties to
another state after the mid-eighties. This shift in the climatology is due to major
changes in the model and analysis system in the mid-eighties. Previous attempts to
correct the analysis, compared it wilh observation at certain points and tried to develop
weighting functions to correct the difference between observations and analysis. The
limitations of this method are that the weighting functions developed over a particular
region may not be applicable over other regions. Moreover, upper air data over the oceanic
regions are unavailable to construct these correcting functions. Here, we present an
alternative method to remove the systematic errors globally at all levels by comparing with
an independent analysis at the surface and using a multivariate EOF technique.
E C M W F wind analysis
103
2. D a ~
The main data set here is the monthly mean operational analysis oftbe ECMWF from
January 1980 to December 1991. Zonal and meridional winds at 1000, 850 and 200 mb
are used. The horizontal resolution is 2.5 ~ x 2.5 ~ We have confined our study to the
global tropics between 30~ and 30~ To compare the anomalies with an independent
analysis we use the Florida State University (FSU) analysis of the surface wind
(Goldenberg and O'Brien 1981). This data set is based on available in situ observations
and a subjective analysis. This data set covers only the Pacific domain. Updated FSU
winds are available with us up to 1991.
3. Results and discussion
3.1 Anomalies relative to climatolo#y based on the entire period
First, we constructed the climatology of each field at every grid point for each calendar
month averaged over all the 12 years. Anomalies are then calculated by subtracting the
climatology corresponding to the given calendar month from the analysis for the
particular month. In figure I (a) we show the time series of anomalies of zonal wind
shear (Uaso-U2oo) averaged over the NINO-4 (5~176
160~176
region of the
Pacific. This index is related to the strength of the Walker circulation. Figure 1(b)
shows the time series of the broad scale monsoon index (Webster and Yang 1992). This
is the zonal wind shear (Usso-U2oo) averaged between 5~176
and 40~176
Similarly, in figure 1(c), we show the vertical shear of the meridional wind anomalies
averaged over the equatorial Indian Ocean (10~176
40~176
This represents
the strength of the regional Hadley circulation. A clear trend is evident in figures 1(a)
and (c). We note that the trend is more pronounced on the time series representing
equatorial fields (e.g., figures la and c), as compared to fields away from the equator
(e.g., figure lb). This is consistent with the findings of Trenberth et al (1990).
It is recognized that the ECMWF analyses have certain inconsistencies in all the
fields and at all the levels (Hurrell and Trenberth 1992). However, to obtain a homogeneous data set, it is a formidable task to identify the bias and apply a correction at all
grid points and at all levels. For studying variability and predictability over the tropics
at interannual time scales, one is interested in examining the large scale patterns of
spatial and temporal variability. EOF analysis helps us to identify large scale variability and describes most of the variability in a few leading EOFs. Therefore, systematic
errors in these patterns can be corrected more easily by correcting the principal
components and then reconstructing the data. To examine how this trend is manifest in
the large scale part of the circulation, we carried out a multivariate or combined EOF
analysis (Wang 1992; Nigam and Shen 1993). The choice of combined EOF analysis has
been adopted for the following reasons. Firstly, a combined EOF analysis provides
a more efficient way of compacting a large volume of multifield data. As geophysical
fields are not only spatially correlated but also correlated with one another, the
combined EOF analysis brings out the patterns that are highly interrelated. This helps
us to bring out the three-dimensional patterns of anomalies associated with certain
types of variability. We thus get an insight into the dynamics of such variability. More
importantly, in the context of our analysis, it is expected to bring out the systematic
104
B N Goswami and H Annamalai
Shear
Index
NINO4
15
10
u~
5
~0
E
-5
-10
-15
-2o
6
Broadscale
Monsoon S h e a r
Index
r
f~
2
-2-
-6
3
Meridionol
Shear
Index
C
2
1981
1982
1983
1984
1985
1986
1987
1988
1969
1990
1991
Figure 1. Anomaly time series, based on the 1980-1991 data. (a) The shear index,
(Uss o- U2oo) averaged over the NINO-4 (5~176
160~176
region of the Pacific; (b)
Broadscale monsoon index (U85o-U2o o) averaged between 40~176
and 5~176
(e)
The meridional shear index, (Vsso-V2o o) averaged over the equatorial Indian Ocean
( 10~ 10~ 40~ 110~
errors that are common to all variables in a limited number of EOFs. As we shall show
later, this will not only allow us to apply corrective procedures to the data with least
amount of effort, it will also provide information about the origin of the errors. As this
brings out the patterns that are interrelated amongst different variables, for verification
purposes, all that we need is one variable from an independent analysis. This is the most
useful feature of this analysis, as we shall show later.
As we are interested in interannual time scales, wind anomalies at all levels are passed
through a five-month running mean filter. Unless otherwise stated, all the analyses
described below are based on these filtered winds. A combined E O F analysis of winds
at surface, 850 and 200 mb was carried out. Figure 2 shows the time series of the first six
principal components of the combined EOFs. Figures 3 and 4 show the vector
representation of the first four EOFs at the surface and 200mb respectively. It is clear
from figure 2 that the PC 1, PC2 and PC5 have significant trends. The anomalies tend to
change sign in the mid-eighties. Following Bethea et al (1975) the significance of the
EC M WF wind analysis
1264~t
~
105
PC1
60-1 /N
PC 2
-3
o-I
_
-90
3O
0
,".
-so "1
.
.
.
.
"
I
I
19'81 19'82 19'83 19~1. 19'85 19'86 19'87 19'88 19'89 1990 19'91
Figure 2. First six principal components (PCs) of the combined EOFs of zonal and
meridional winds at 1000, 850 and 200 mb.
slope of the linear regression line was tested by calculating the 't-statistic' of the
regression coefficients for all the PCs. It was found that the regression coefficients for
the PC1, PC2 and PC5 are significant at 0.005 level. To test if the five-month running
mean introduced any spurious low frequency signal, we carried out a similar combined
EOF analysis with the unfiltered wind anomalies. It is seen (not shown) that the
structure of all the leading EOFs remains unchanged. The PCs also contain exactly the
same low frequency variability. The only difference is that some weak high frequency
noise is superimposed on the low frequency variability. The PCs have the same trends
seen in figure 2. Thus, no additional trend or low frequency variability is introduced by
thefilter.
This trend in the ECMWF data is related to the change in the 'analysis system'
(including model changes and analysis-initialization changes). However, to establish
whether this trend is due to model changes or not, we must compare this data with an
independent analysis that does not depend on the model. As we have mentioned earlier,
because the multivariate EOFs are interrelated, if we have an independent analysis on
one variable, comparison between patterns of variability of this variable with those of
106
B N Goswami and H Annamalai
E1 1000
9o(:
~v~
~P
/9
."
" 9
9
)
VAR 32.8~,
9
I'
9
9
)
9
~-/" "= "v \,,,
9 e-v.
9 4), ' ~
)
w
.~
',it
.it
-l.
r
t
9
~
~ 9
~,
'F9 ,q11 9
9
9
4I9
9149
E2 1000
.ua I
,,~ ~
3o~)-,
I
~
-~,
~
~
9
~
-= "=
~ 9
,,~
9
9
9 .~
11
VAR 1 2 . 7 ~
9
4- \~ N - . I ) ~ k..._,~, -.)
.
,
9
s
~,,,,
t,
~
--)
% ~, ~ - , 7
:
E3 1000
.
-4. ~
9
:
.
+ ;
,
=
9
~
.
o
=
VAR 8 . 5 ~
3014
20)4. .
t~
~"fl,<'9
9
I;
,,g
r
~
9
4
9
9
"> 9
9
9
~
')"-J'--~'
10N.
[Q,
lOS
20S,
3O(3
E4 1000
VAR 6 . 7 ~
3ON
!
R
10N , r..,.~ "-,,
%a
"
,.4~ 9
~
9
,e
F
9
9
-o ([ .'-a~;:-e. '%-o.'.~ .....e-...o-.-e.~ 9
~
.,' 4
" }~p~_lL_
E:Q
.
10~
~0~
,30~_
Ip
9
tl'~
~
Io
?~
9
"l
9
15
~
a
s
'm
"4'
',,qu'~
.u , " ) l , . ~ . . q 5 . . . 4 ,
A ~7
d-'~ tl.
q~,r a, " ~ . l ,
~
...,.~,~lp
~f
p
4)
,ll. j
9
~
9
~
m/
qb
q~
.8
9
9
9
II
9
.a
~
.=
9
9
9
Ib
,a,. 9
I.
11
9
var
9
9
9
*
9
9
I.
9
4
v
9
i.
A
,I,
e
1~0~
140~
lllO~
1110
1 6 Q W 141(11 1 2 ( I I
IO(M
0.03
F i g u r e 3.
First four vector E O F s of the winds at I000 mb. The combined per cent variance
explained is also given.
the corresponding variable in multivariate analysis will help us to identify the systematic errors that are common to all variables. FSU analysis of the surface wind provides
an alternative to compare the surface wind. Therefore, we carried out a similar EOF
analysis of the FSU winds (zonal and meridional winds combined) over the Pacific
domain. Figures 5 and 6 show the first four principal components and vector
representation of the corresponding EOFs of the FSU winds. We note that the
dominant PCs do not show any trend in the FSU winds. The significance of the
regression coefficients was carried out as before. It was found that the regression
EC M W F wind analysis
107
E1 200
VAR 32. 87.
1014.
94b
9
9
9 ..~,
.~
o
9
9
9 .-',Jb~tT
9
9
4'. 4 k . . . ~ q , . . . . .
qu,,..
4"
4
9
9
o,.
t,
t~
(Q.
b
9
9 %.,,!~ 9
'v , ~ t J . _ ' ~ -
"m
=r
..,. ~
e,.--~...~.,..~
~
10S.
20S.
3~
E2 200
VAR 1 2 . 7 ~
~,
I(~.
, /,,,,,~...,~,a.
lk.
~,V"
9
4
qP
Ill
"1
at'
al"
v ~ete"" ~','"4'- e - e . , , ' =,,' I,, ~. ~
v
~
~ 4
I,
.''*41-"
~
.~
9
=t
(O,
/~'
"
e--
e"
o'r6"
e"
~,
"
~
~"
~
u"
E3 200
VAR 8 . 5 ~
E4 200
VAR 6 . 7 ~
1QN
(0
lOS
2m~
'.-~b.
"~
4
Ir
u
9
9
9
,,4,,.,.41/o
9
9
4
'~
4
~
...~....~u,.~
9
9
l
4
9
9
9
~9
9
~
9
9
4,
.4'
.41,
JI
(Q.
lOS,
.
.
~
.
80[
.
~_
I0~
,,-,,.-r-~--~__~
,,.-
'120[ 14~t[ 1~1~ 1110 I~M
_
, ~
_
_
1,~1 12~W IOOW 801f
0.1
F i g u r e 4.
S a m e a s f i g u r e 3 b u t for 200 rob.
coefficients of none of the six PCs are significant. Moreover, we note that there is no
equivalent of EOF2 of E C M W F winds in the FSU wind. EOF2 of FSU wind
qualitatively agrees with EOF3 of the E C M W F surface winds. Thus, we infer that the
observed trend in the E C M W F wind analysis is not related to any physical mechanism
and the changes in the 'analysis system' seem to have introduced some spurious
patterns of variability. Therefore, the use of this data set for interannaal variability
studies may lead to erroneous conclusions.
108
B N Goswami and H Annamalai
:l
A
'~
~1
,c,
~'1
~c 3
:~]
-
,o] ,---,_
--~,4~
A
v
_
~
~.,
/x "~A_
I
~
-
1
-
v
I
/
A
~
20"I
I
PC 6
" ~
.v..
1981 1982 198,1 I W
II
1
\
. . I
1986 1987 1988 1989 1990 1991
Figure 5. First six principal components of the FSU winds.
3.2 Origin of the trend
As documented by Trenberth and Olson (1988) the ECMWF analysis has gone
through a number of significant changes (refer to their figure 3). Major changes were
introduced in May 1985 when the resolution of the model was increased to T106
(triangular wave 106), and when substantial changes in the physical parameterization
affecting the moisture field (clouds, shallow convection and condensation criteria) were
introduced. In March 1986 changes in the humidity analyses that had a positive impact
in the simulation of tropical rainfall and on latent heat release were also introduced.
Prior to this period, the tropical divergent winds were weak, while after this period they
were realistic. Thus the mean climatology changed from the period 1980-1985 to the
period 1986-1991. Trenberth and Olson (1988) show how the mean Hadley circulation
changed around this time. We suggest that the trend seen in the E C M W F analysis is
due to a change in the climatological mean from the period before 1985 to the period
after 1985. To support this hypothesis, we conduct the two following analyses.
First, we remove the trend from the analysis by objective means. To achieve this we
detrended PC1, PC2 and PC5 by fitting a straight line using the least squares method.
ECMWF wind analysis
109
E1 FSU
VAR 21.0~.
20N-lr~>~a~
~ v . ~ , ; I , i . i a.l.l,l.a, al..,l~, a.',..'i
b441p
~ u r
~ ~" & & L b
r
~* * i * # # , l , l / / / / / I
'~14 ~1V~44411
D p # # !, I, I!u
1. ~
44
A ~4
~,,
I////&.llddli
"~
r162162
tl
E2 FSU
~0~
V ~u
VAR 1 8 . 8 =
l ~lF ] J 2 1 [ . l . l_ l, T l q ll i 1 4
~os.1~'~' ~.'~[~i92r/~.~',&,
,~ ) 4 4 & ~
, . . f.- i',
;, ~ ,~ "- ~'~ ', "-%%* " ' , 9 9 9 ", " , 9
'-"~
2os I :4._.. tr176
~OS
' " "=----
,,
' . . . . .
;Jr"
. . . . . . . . . . . . . . . .
E3 FSU
9O d l O d l '
VAR 7 . 9 =
#1~1~1~ # ' 9
20S" 9 ~ L - 305
. .l~e-
'1114 , K & 6 A A A q~qP~l'IF'41" II. V q l ' O 0
~I ~ 4 &
~
,,...,
9
. . . . . .
~
--
._t
E4 FSU
.~r
* . ~ . 1
I 1 ~ % ~ , ~
~
-
VAR 7.17.
r
....
,'~'l~u~,'~,~'~,'~p~w,o.~*o.~-,~.~*k..I.
;- _ .
9 9
tO-,E4/r~m"(~,,,. ~ h , , , , ~ , . . . - . ,.,,
. ."~,-.,t.~#.~,~ a.s,~.~4~. ~, 9 9 I ~
3os :'"
~40E
.r" . - - -
160s
:..,.t~,,._
180
i, | l"J~llllllllllr.,~ t,,
: ~1,#,
1GOW 14011
l, r t u t ~ t r ,,_-
120#1
10011
0.1
Figure 6. First four vector EOF, s of FSU winds. The combined per cent variance explained is
also given.
This is shown in figure 7. As mentioned earlier, statistical significance tests indicate that
these trends are highly significant. The fields are then reconstructed using the three
detrended PCs together with the other PCs without trends. We then conducted an
110
B N Goswami and H Annamalai
tGO
120
90
60
30
0
-30
~
P.C1
/x
-120
1~1
1de2
'ta~
1de4
Im
IdN
~S7
1~ta
'teTee 'i~o
'toe't
1961
IW2
ISi43
IN4.
IWS
19U
117
I ~ 8 8 191M
I~90
lIPg!
1MI
19~2
Igt3
lg64
1~B5
Im
1S17
11
IS~O
1S~1
80
6O
4O
2O
0
-20
-40
-6O
20
10
0
-I0
-20
-40.
lm
Figure 7. The linear trends obtained for the dominant PCs of figure 2. Shown also is the slope
obtained from least squares fit.
EOF analysis as before with the detrended wind analysis. As expected no trend is noted
in the PCs of the leading EOFs in this case. The first four vector EOFs of the surface
and 200 mb wind of the detrended winds are shown in figures 8 and 9. It is now seen that
the first two EOFs of the analysed surface winds agree with the corresponding EOFs of
the FSU wind (figure 6). It is interesting to note that the detrending has removed the
spurious EOF2, explaining a large amount of variance in the original data (figure 3).
Reasonably good agreement between patterns in the detrended surface wind and those
of the independent FSU data set over the Pacific indicates that, even though the
climatology may have been wrong during the period prior to 1985, the large-scale part
of the interannual anomalies during that period may still be reasonable. The time
evolution of the first four PCs of the detrended analysis also agrees reasonably well with
that of the PCs of the FSU winds. This indicates that major interannual variations are
captured in amplitudes as well as in patterns by the detrended analysis. It captures the
anomalous surface westerlies and their eastward propagation in the Pacific during
1982/83 and 1987/88 episodes. The first two EOFs represent the low frequency
interannuai variabilities associated with the El Nino and Southern Oscillation
(ENSO). The two anticyclones at 200 mb straddling the equator in the Pacific and their
E C M W F wind analysis
111
E1 1000
VAR 3 4 . 8 ~
2ON.
~. 9
loll
o
~It.p
9
o
*'
9
,~l['~ 9
'qK~
.?
s, I~
9
9
9
9
9
9
4
9
9
u
" "~
9
~,
"8
u
t
9
"
9
9
~"
9
9
In
t
"'~.~..'~t._
9
"
~
9
s
9
~
4
9
9 ~
~
9
,e
9
9
10S.
20S
'~
~.''qF'*r'b
9
4,," 9
~',a[
"u
~A
e.
~
9
"5
II
.It
JI
3
..~ - -
=.1) --.
r
~
9
,~
,.~
E2 1000
VAR 1 0 . 1 ~
,,~
~.9
i~,, ~-
,v
~1[
'T.
_"11~,.."
~
4-. 9
Lr
9
9
~
~, " ) . L . ) 4 ~ . - . ' O ' ~ ' . ' / 3 ~ . , . e t ' m ' m ~
o...
m'
4..
,*,.."
%.
-~
m'
ir
9
ii,
9
4
,4,%~r
,,~
~
-4,
~
. . ~ ".,,~
l(Xq.
[0.
9
9
9
9
o
)
9
~
9
~;.,'a~_a.~,= -~C-~-
)'
,.
~
.......,.-~....~
9
!0~.
2mS
E3 1000
VAR 7.8X
E4 1000
VAR 7 . 3 Z
,.~
20N.
1ON(O"
10S"
~...~
ZO~ d,../u
30S4~
.~
q\
&
9
"~'lt
~
9
v
.
4
6~
8~
10~
W
~
.~.
12~
9
t
9
9
t
~
T~
9
4
9
.=.....~ . ~
11~
'~
9
,L
,I,
..= . . . . 9
,,
11~10 1~11
9
14~/
"l
~
~'qP
~
~,
=.
12~/
,)
10011 8011
0.03
Figmre 8.
S a m e as figure 3 b u t for t h e d e t r e n d e d d a t a o f E C M W F
analysis.
eastward migration during major ENSO episodes are also seen. The third EOF has
major loading in the Indian Ocean and represents a regional mode of oscillation
associated with the Indian monsoon. The actual anomalies are somewhat weaker
(about 70% of observed values) during major warm episodes associated with ENSO.
This is a general problem with many AGCMs (Goswami et al 1995) forced by the
observed SST.
Is the detrending of the analysis following the above method equivalent to calculating anomalies with respect to two climatologies; one for the period 1980-1985 and
112
B N Goswami and H Annamalai
E1 200
VAR 3 4 . 8 ~
30N
2ON.
10N.
~i
9
9
II
~',,~I~
*,'r'~i~t '
4
9
~" ~ ' 4 - - ' ~ " . ~
~,.
~
9
9
(O"
10S,
20'3'
30S
E2 200
VAR 10.1~
It_,-,.,-;" ; ~ > . d ; ; ; ; ~ ~ t i l ' - - : ~,~t :~-=,
30N
20N,
10N,
10S,
9
~
..
9
*
~..,,.~,.,,,~..~ ,,~ .~
,= ~ - ~ , . i ,
,.,= ~
~.S ,,,= 9 ~'. =,..~.~=,,,,.. ~'
20S
30S
E3 200
VAR 7 . 8 ~
30N
20N
10N
9~ , - - , 6 ~ '
o.
,.,~:' 9
EQ I~ . . . ~ _ ; ; . . 2 . : ~ , ~ . ~ . ~ .
10S
.
20S
.
.
9
4.
4.
.=.
-.~o~.';
.
.
.
.
9
~
9
*
,
,
.
, ,
.
4
30S
E4 2 0 0
VAR 7 . 3 ~ t
2OilION-
EQ~t
10'3-
~-
",. ~
20r
'l.
~
~.
9
,.
t.
-.
*-
,.
~/'~V'=~,~,"~..
~.
~. ~ / t . ~
~.,
,i,-
i,..
~
, i ~
I-.
i~.
i i
~.
."11%1
~'.
~-
.*-
I'
o
~.
~.
~, * .
~
9
It
I
l~
,I.
-
,
~
9
t
-
~
-
~
4-,-~,,.
t
,i
9
9
6T5
Figure
9.
Same
as figure
8 but for 200 mb.
another for the period 1986-1991? The delineation of the 12-year period into these two
six-year periods is not arbitrary. Even though the changes in the model have been
continuous, it is felt that the major model and analysis-initialization changes introduced
in mid 1985 had the most significant impact on the tropical analysis. This is why the
analyses before 1985 appears systematically different from that after 1985. Hence the
separation into the two periods. To test this, we calculate climatologies of each field at
each grid point separately for the period 1980-1985 and for the period 1986-1991. To
illustrate the differences in the two climatologies, we plot in figure 10, the annual cycle
E C M W F wind analysis
1 t3
of the climatological mean zonal wind averaged around the equator (5~176
corresponding to the two different periods, and their differences at 850 and 200 mb
respectively. We note that particularly large changes took place in zonal winds at
850 mb over both the Indian Ocean and Central Pacific sectors. The percentage change
from earlier climatology to the later ranges from 75% to 100%. There are also
significant changes in the 200 mb zonal wind climatologies for the two periods (figure
10b). However, the mean at this level is higher and the percentage change from the
earlier to the later climatologies range from 25 % to 50%. The largest changes are in the
equatorial region. The anomalies are then calculated with reference to the corresponding climatologies in the two periods. A combined EOF analysis is made with the
anomalies of zonal and meridional winds at the surface (1000mb), 850 and 200 mb.
Figure 11 shows the first 4 principal components of the analysis (solid curve) and this is
compared with the PCs of the objcctively-detrended anomalies (dotted curve). The
close matching of the two curves in all the PCs shows that the anomalies calculated by
these two different ways are nearly identical. Figure 12 shows the first 4 vector EOFs of
the surface wind from the combined EOF analysis of the newly constructed anomalies
MEAN 1 9 8 0 - 8 5
DEC"
NOV"
U 850
'? :.
OCT"
'::.
.."...::'
..:'.:
// f-?-,~--~
,~[p"
AUG"
I
I
(a)
/ .Z"-i--",."
: ,:":=:..
"
//;-.'-;
..,,,".:, ,:...:. ,
' ," ."
/ i
/
.:, : f: :"
', "."',',
i
; '..".'"
"
,"
Jt,A."
JUN"
..-"
...........
.. ;-:
,.,,
APR"
MAR
[rED.
Jii~l"
,'~":--.. --.....,
.......=.."..'..-
)
~Y
,) \ " - f
~'~,5
40E
S0E
80E
",:'---- 1"-'--" . . . . . . .
,'.:: ....
"<.:
9 " .'-':-""-~. "-':;:. "",
10OE 120[
140[
160E
X
180
"
--"
:
.........
k
".
. ......... ,,,,
.
", .
"
160W 14,0W 120W IOQW
MEAN 1986-91
"
80W
U 850
DEC
NOV
,.;:'.;
9
1-
~
OCT
SEP
-
...~::.::.., .,,", ....,]
AUG
JUL
Jr,JaN
MAY
APR
MAR
FEB
JAN
..::~-_~......
,mE
66E
- ....,....
86E
I0~
,::.
i2"0E 14OC 1601[ I1~0
DIFFERENCE
16"0'N 14~/ 12'0W I0~I
80W
U 850
DEC
NOV'
OCT
SEP
AUG'
JUt
JUN
HAY"
APR"
FEB"
JAN"
4.0[
G0[
6QE
Figure 10. (Continued)
100[
120([ 140[
160E
180
160W 14.0W 120W IOOW
8ffP/
114
B N Goswami and H Annamalai
MEAN
DEC
NOV'
OCT
SEP"
AUG
JUL
JUN
MAY
APR
MAR
FEB
JAN
1980-85
~d...----.<;;:......
U
~:...........
..-- .~ . . . . .
. . ~ .... - .................
": ....... . . - . . . . . . . . . . . . . . . . .
o-
.
,,-.......
:"-.
.-..-:-~---.:==:.
" ......
? ,"
MEAN
DEC
NOV
OCT
SEP
AUG
JUL
JUN
),lAY'
APR
MAR"
FEB"
JAN
,,.
..
.........
F
..,
,
(
..... -i;"
":.'
U
..................
..-,:............. )
~
':9.....
....
..... "-~
9......
.....
.........
40E
60E
80E
1;P"'/)
3
180
_"<--W-~--', " : ' 5
".
......
66E
8(7#
200
",",,.
.-~=-~---:
I
", - - ~ ~
,'-'-J:,":,'~ -
a0E 100E fl0E 14'0E I~E
~
.;:~/:...~'....:-=:..."--~...--"~
':: ..... . .....
,.':-:.- ..o, -.:~-.:.:
46E
.
160W 140W 120W IOOW
U
~
~
...=,- , ~ - - - . . ~ ~ . . .
,..::.--. . . . . .
IOOE 120E 140E 160s
,..-o~-
:
,._
DIFFERENCE
~ . , ~
-
200
....
~.::~::.._, ...... :;,::::.:~.----;...~
~-:-q-:.:::.----:,
"....
"......
. . . . . . . . . . . . . --'?"'-. ",.
1986-91
~.....i
.... :: . . . . . -: ..........
. . . . . . .-3
.-
.... -.
"~-1 .... '='[. . . . . . . .
3----""_
..:::- .........
.:---~ . . . . . . . . . . . . . . . . . . . ~. . . . . . . . .
",,':,.. (
DEC'
NOV
OCT'
SEP
AUG
JUL
JUN"
MAY
APR
MAR
FEB
JAN
Ib!
:..
.. . . . . . - s
. . . . . . ...
. . . . . . . . . . . . . . . . . .?'--I
....... ~'::-'-'-2
.......
,,
%
~---""
1::::----::::
. . . . . . . . . . .......... . . .
......... ::-::.....::
200
/ ..: _ _ ~ , ~ .
"-"
1~0 16~ 1 ~
.::-~-=q-i/'.,Li
1~
lo~
a~
Figure 10(a and b). Annual cycle of climatological zonal wind averaged over 5~176 for the
two periods and the difference between these two climatologies at (a) 850 mb and (b) 200 mb.
based on two climatologies. Again we note that they are nearly identical to corresponding EOFs of the objectively-detrended analysis (figure 8). This indicates that the origin
of the trend is related to differences in the mean conditions of the model-based analysis
system prior to 1985 and in the period after 1985. We also infer that a step-wise or linear
way of removing the trends yields the same result.
4. C o n c l u d i n g
remark
It has been recogriized for some time (Shaw et a11987; Trenberth and Olson 1988) that
operational analyses (e.g., the E C M W F analysis) contain discontinuities in time
introduced by the changes in the model and the analysis procedure. Efforts are now
underway for a reanalysis of the past data (e.g., Kinter and Shukla 1989; Kalnay and
Jenne 1991; Schubert et al 1993) using a 'frozen' analysis system. As it would still take
several years before such a set of homogeneous, consistent data set is available for
ECMWF
wind analysis
|~
115
,,,
80
0
_wi
-60
-40
-120
W
0
-.m
O,
-4O
~I
0
~4
.~
"
...
""
Figure 11. The firstfour PCs of the combinedEOFs of zonal and meridionalwindsat 1000,
850 and 200 mb but for anomalies formed from the detrended analysis (solid curve) and
anomalies formedwith respect to the two climatologies(dashed line).
a sufficiently long enough time, some researchers have made use of the existing
operational analyses for interannual variability and predictability studies (e.g., Wang
1992; Murakami and Matsumoto 1994; Barnett et al 1994). Some other studies
(Vincent et al 1991; Webster and Yang 1992) have exercised caution and used only the
data after 1985.
We show that the changes in the ECMWF analysis system for the period 1980-1991
give rise to a significant trend in the wind analysis. By comparing it with an
independent analysis of the surface wind (FSU analysis), we show that this trend is not
due to any natural processes. If this trend is not removed, the interannual anomalies
prior to 1985 tend to be of one sign while those after it tend to be of the opposite sign.
Therefore, erroneous conclusions may be drawn if the trend is not removed. Here we
provide a technique of filtering this trend using a multivariate EOF method. The
multivariate EOF technique provides a means of identifying the systematic errors that
are interrelated at all levels. Then by comparing with an independent analysis at one
level (e.g., surface), the systematic errors at all levels may also be corrected. We
demonstrated that this trend arises due to changes in the climatology before 1985 and
after 1985. We also strow that, although the climatology for the period prior to 1985 was
weaker and unrealistic, atleast in the tropical region, the interannual anomalies in the
116
B N Goswami and H Annamalai
E1 1000
3o.
6
told
,~ ~,~ 4
.)
9
I}'
*
!'
'~ e'
lOS
9
9
9
VAR 3 4 . 9 g
,e
Lf~
9
m~ 9
9
9 ,9
,-~
q."
9
9
,L \ , 9
nu gu 4
9
~
,mj,,,* 9
~'tI 9
~.~-,'~
4
-st-
9
/~
9
9
b
v
9
9
t
.
,,
9
9
9
9
.~
o'
~
-4,"~,,j",,,o",,,,,~",,,e,'~
ajr.~. lp-~ ~
9
.
"~
EO
T
9
wL~
~
~
~'~"
9
,,
lOS,
9
20S
~' ~ .
,
(/r.'.~
/',~,~
'.,
~'."~'-'-~-,-0,,,,-~...,.--~
~
.,,.%
"~'~.'~- ,,~ ? ~ ' ~ - . . ~ . ~ 4 . - - - [ , , ~
~,,~
..~.'-~;
,
"
"
.
,/-"~t..~"~"
;
#
.
9
.~..T
ION"
lOS.
"~
/*
20S' ~
~
f~
.
~10 6
..
~.
,,
..,,,~.,,~
@ , ~.i .,,,.~
~
9
~
i
4
v
"-e. ~
~
~
,.~
9
9
,~
9
9
v
.~
,.,, ..4. ~
-9
-.~ ."4' - 9 -,.,= '%,,
~,
9
.
=
"f
-*
-*
-~
-~
w,
-,
r
r ~
.
9
.~..
~'
-
~
-~
-"
-~'~--~-'~
9
",,
-
"~
=
i
"
.
,,
.
v ,,.. ~
~
.'~
VAR 6.77.
.
.
~-.'~...
,a \ l
9
4,-
E4 1000
2ON.
v
VAR 7 . 6 g
-
.
4
VAR 9.7%
E3 1000
ION,
4.
.,wm.., 9
E2 1000
ION'
.,,.
9
.
i
~.
,,~.,
..,~..,-..,, .,~,.,,
.,,
o,.
.~. '~. ~,
~,~'.. ~
9
~.
9
~,,"~'-""
9
. . . . .
~
9
"
"
qp
9
9
*~.dJ
9
+'~--~'~k~
~J
IL
'QN
4
.~
,,
,~
,,
~
~
9
~p
~'
9
-
"
~
~"
"~
-,,St,,
&
9
.
~
"
9
.;
0.03
Figm-e 12.
Same
as figure 3 but for the anomalies
formed
with respect to the two dimatologies.
detrended analysis are still reasonable. Before the actual reanalysed data are available,
the method presented here provides an objective and convenient way of making the
data useful for interannual variability studies.
Acknowledgements
The authors would like to thank V Krishnamurthy for helping us copy the ECMWF
analysis. We thank the Department of Science and Technology, Government of India
ECMWF wind analysis
117
for partial support for this work. We also thank Sulochana Gadgil for some comments
on the draft version of the manuscript.
References
Barnett T P, Bengtsson L, Arpe K, Flugel M, Graham N, Latif M, Ritche J, Roeckner E, Schlese U,
Schulzweda U and Tyrer M 1994 Forecasting global ENSO-related climate anomalies; Tellus A46
381-397
Bengtsson L and Shukla J 1988 Integration of space and in situ observations to study global climate change;
Bull. Am. MeteoroL Soc. 69 1130-1143
Bethea R M, Duron B S and Boullio T L 1975 Statistical methods for engineers and scientists;, (Marcel Pekker
Inc.)
Goldenberg S O and O'Brien J J 1981 Time and space variability of tropical Pacific wind stress; Mon.
Weather Rev. 109 1190-1207
Goswami B N, Krishnamurthy V and Saji N H 1995 Simulation of ENSO related surface winds in the
tropical Pacific by an AGCM forced by observed SST; Mort. Weather Rev. 123 1677-1694
Hurrell J W and Trenberth K E 1992 An evaluation of monthly mean MSU and ECMWF global
atmospheric temperatures for monitoring climate; J. Climate 5 1424-1440
Kalnay E and Jenne R 1991 Summary of the NMC/NCAR reanalysis workshop of April 1991; Bull. Am.
Meteorol. Soc. 72 1897-1904
Kinter J L and Shukla J 1989 Reanalysis for TOGA; Bull. Am. Meteorol. Soc. 70 1422-1427
Murakami T and Matsumoto J 1994 Summer monsoon over the Asian continent and western north Pacific;
J. Meteorol. Soc. Jpn. 72 719-745
Nigam S and Shen H S 1993 Structure of oceanic and atmospheric low-frequency variability over the tropical
Pacific and Indian Oceans. Part I: COADS observations; J. Climate. 6 657-676
Schubert S D, Road R B and Pfaendtner J 1993 An assimilated data set for earth science application; Bull
Am. MeteoroL Soc. 74 2331-2342
Shaw D B, Lonnberg P, Hollingsworth A and Unden P 1987 Data assimilation: The 1984/85 revisions of the
ECMWF mass and wind analysis; Q. J. R. Meteorol. Soc. 113 533-566
Trenberth K E and Olson J G 1988 An evaluation and intercomparison of global analyses from the National
Meteorological Center and the European Center for Medium Range Weather Forecasts; Bull. Am.
Meteorol. Soc. 69 1047-1058
Trenberth K E, Large W G and Olson J G 1990 The mean annual cycle in global ocean wind stress; J. Phys.
Oceanogr. 20 1742-1760
Trenberth K E and Guillemot C J 1994 The total mass of the atmosphere; J. Geophys. Res. 99 23,079-23,088
Vincent D G, North K H, Velasco R A and Ramsey P G 1991 Precipitation rates in the tropics based on the
Q~-Budget method: 1st June 1984-31st May 1987; J. Climate 4 1070-1086
Wang B 1992 The vertical structure and development of the ENSO anomaly mode during 1979-1989; J.
Atmos. Sci. 49 698-712
Webster P J and Yang S 1992 Monsoon and ENSO: Selectively interactive systems; Q. J. R. Meteorol. Soc.
118 877-926