Content uploaded by Vincent Moron
Author content
All content in this area was uploaded by Vincent Moron on Sep 01, 2014
Content may be subject to copyright.
NOTES AND CORRESPONDENCE
Spatial Coherence and Seasonal Predictability of Monsoon Onset over Indonesia
VINCENT MORON
CEREGE, UMR 6635 CNRS, Universite
´
d’Aix-Marseille, Aix en Provence, and Institut Universitaire de France, Paris, France, and
International Research Institute for Climate and Society, Columbia University, Palisades, New York
ANDREW W. ROBERTSON
International Research Institute for Climate and Society, Columbia University, Palisades, New York
RIZALDI BOER
Laboratory of Climatology, Bogor Agricultural University, Bogor, Indonesia
(Manuscript received 18 January 2008, in final form 30 May 2008)
ABSTRACT
The seasonal potential predictability of monsoon onset during the August–December season over Indonesia
is studied through analysis of the spatial coherence of daily station rainfall and gridded pentad precipitation
data from 1979 to 2005. The onset date, defined using a local agronomic definition, exhibits a seasonal northwest-
to-southeast progression from northern and central Sumatra (late August) to Timor (mid-December). South of
the equator, interannual variability of the onset date is shown to consist of a spatially coherent large-scale
component, together with local-scale noise. The high spatial coherence of onset is similar to that of the
September–December seasonal total, while postonset amounts averaged over 15–90 days and September–
December amount residuals from large-scale onset show much less spatial coherence, especially across the
main islands of monsoonal Indonesia. The cumulative rainfall anomalies exhibit also their largest amplitudes
before or near the onset date. This implies that seasonal potential predictability over monsoonal Indonesia
during the first part of the austral summer monsoon season is largely associated with monsoon onset, and that
there is much less predictability within the rainy season itself. A cross-validated canonical correlation analysis
using July sea surface temperatures over the tropical Pacific and Indian Oceans (208S–208N, 808–2808E) as
predictors of local-scale onset dates exhibits promising hindcast skill (anomaly correlation of ;0.80 for the
spatial average of standardized rain gauges and ;0.70 for standardized gridded pentad precipitation data).
1. Introduction
Rainfall over Indonesia is governed by the austral-
Asian monsoon, whose onset progresses from northwest
to southeast during the austral spring (Aldrian and
Susanto 2003; Naylor et al. 2007). This is also the season
when the El Nin
˜
o–Southern Oscillation (ENSO) exerts
its strongest influence on Indonesian rainfall, particu-
larly during the September–December monsoon onset
season (Hamada et al. 2002). The impact of ENSO
then diminishes during the core of the rainy season in
December–February (Haylock and McBride 2001; Hendon
2003; Aldrian et al. 2005, 2007; Giannini et al. 2007),
suggesting that the timing of monsoon onset may be
potentially predictable.
The date of onset of the rainy season is of particular
importance for the agriculture sector over Indonesia
(Naylor et al. 2002, 2007). It determines the suitable
time for planting crops, while delayed onset during El
Nin
˜
o years (Hamada et al. 2002; Boer and Wahab 2007)
can lead to crop failure. For irrigated rice farmers in
Java, information on onset timing is also important for
developing strategies (Boer and Subbiah 2005; Naylor
et al. 2007) to avoid exposure of the second rice crop to
higher drought risk at dry season planting (April–July),
Corresponding author address: Vincent Moron, CEREGE,
UMR 6635 CNRS, and Universite
´
d’Aix-Marseille, Europo
ˆ
le de
l’Arbois, BP 80, 13545 Aix en Provence, France.
E-mail: moron@cerege.fr
840 JOURNAL OF CLIMATE VOLUME 22
DOI: 10.1175/2008JCLI2435.1
Ó 2009 American Meteorological Society
particularly for farmers located at the tail end of the
irrigation system. Farmers in Indonesia often suffer
from ‘‘false rains’’ in which isolated rainfall events
around the expected onset date do not signal the sus-
tained onset of the monsoon. Such false starts occurring
in September prompt potato farmers in Pengalengan in
West Java to start planting. In the eastern part of
Indonesia, such as East Nuna Tenggara, multiple false
starts can cause multiple failures, with farmers some-
times planting up to four times in a season.
This paper discusses the seasonal potential predict-
ability of monsoon onset during the August–December
season over Indonesia. The approach taken is based on
quantifying the spatial coherence of specific rainfall
properties: the September–December (SOND hereaf-
ter) rainfall total, rainfall onset date, and postonset
rainfall totals following Haylock and McBride (2001)
and Moron et al. (2006, 2007). The seasonal predict-
ability of large-scale monsoon onset is then estimated
based on sea surface temperatures (SST) in July using a
cross-validated canonical correlation analysis (CCA).
The two precipitation datasets [rain gauge and Climate
Prediction Center (CPC) Merged Analysis of Precipi-
tation (CMAP)] are described in section 2, together
with the definition of onset. Results are presented in
section 3, with conclusions drawn in section 4.
2. Data and method
a. Global Summary of the Day (GSOD) station data
Daily rainfall at rain gauges for the period 1979–2004
was extracted from National Oceanic and Atmospheric
Administration (NOAA) CPC Global Summary of the
Day (GSOD) dataset, archived at the National Center for
Atmospheric Research (NCAR), and originating through
the World Meteorological Organization (WMO) Global
Telecommunication System (GTS). There are 91 avail-
able stations for Indonesia. The station years having at
least 50% of daily data are extracted and the 57 stations
having at least 10 available years are selected. Missing
entries (, 13%) were filled using a simple stochastic
weather generator (Wilks 1999), considering the wet-to-
wet and dry-to-wet persistence and a gamma distribution
for wet days, computed on a monthly basis at each sta-
tion. If a month is completely missing (, 3% of station
months for SOND), this method simulates a climatolog-
ical daily sequence for that month.
b. CMAP
GriddedpentadCMAPona2.58 latitude–longitude grid
was selected within a window (128S–68N, 90–1308E) over
the 1979–2005 period, based only on rain gauges and sat-
ellite estimates (Xie and Arkin 1996). Over this window,
there are typically 1–2 rain gauges per grid box including
land (P. Xie 2007, personal communication).
c. Definition of onset
Monsoon onset date can be defined in various ways.
We used an agronomical definition (e.g., Sivakumar
1988) based on local rainfall amounts using thresholds
to define the onset, requiring a certain amount of rain-
fall within a specified period of time, with no extended
dry spell occurring afterward. This local definition is
sensitive to small-scale processes but is used here in
order to be relevant to agricultural management, and to
prevent any a priori inflation of spatial coherence.
Onset date is defined to be the first wet day of the first
5-day sequence receiving at least 40 mm that is not fol-
lowed by a dry 10-day sequence receiving less than 5 mm
within the following 30 days from the onset date. Onset is
computed from 1 August because August–September
are the driest months over Indonesia (Aldrian and
Susanto 2003; Aldrian et al. 2007). The latter criterion
helps to avoid ‘‘false starts,’’ which could be defined, for
example, as the difference between the first 5-day wet
sequence receiving at least 40 mm and the onset as de-
fined above. The identification of false starts is sensitive
to the choice of the postonset dry spell length. In fact, the
sensitivity of crops to postonset dry spells varies. In trop-
ical countries, dry spells with a length of more than 7 days
would have serious impact on crop yields (Niewolt 1989).
Other studies found that 21 rice varieties being exposed
to dry spells with a length of 16 days during the vege-
tative stage will have a delayed harvesting time between
2 and 27 days and reduced yield between 10% and 91%
(Dikshit et al. 1987). Indeed, false starts defined with a
10-day dry spell in the following 30 days occur in 46% of
station years, ranging from less than 40% in northern
and central Sumatra and Kalimantan to a maximum
. 50% in western and central Java. These percentages
decrease by a factor of 2–3 when the length of the
postonset dry spell is chosen to be 15 days. The mean
onset date is also earlier (by one or two weeks in mean)
with a postonset dry spell lasting 15 days rather than 10
days. Nevertheless, this parameter (and the others en-
tering the onset definition) has only a very weak impact
on the large-scale and regional-scale interannual varia-
bility of onset dates (e.g., the spatial averages of CMAP
and GSOD onset-date anomalies computed with both
parameters are correlated at 0.99 and 0.97, respec-
tively). Increasing the length of the initial wet spell re-
duces the noise introduced by weather variability, but
the threshold of 5 days is used to facilitate comparison
between CMAP and GSOD datasets. The National Agency
for Meteorology and Geophysics of Indonesia (BMG)
defines the monsoon to start when, after 1 September,
1FEBRUARY 2009 N O T E S A N D C O R R E S P O N D E N C E 841
two consecutive 10-day sequences each receive at least
50 mm of rain. While changing the length and/or the
amount of rainfall of the initial wet spell modifies the
climatological mean onset date, its impact on interan-
nual variability is again found to be much smaller. The
onset date is undefined for two cases in CMAP and the
missing entries are filled with the latest available onset
dates for the corresponding grid points.
d. Spatial coherence estimates
The spatial coherence of interannual precipitation
anomalies is estimated empirically in terms of the in-
terannual variance of spatially averaged standardized
anomalies given by the Standardized Anomaly index
(SAI; Katz and Glantz 1986). Use of the SAI in the
context of tropical rainfall is discussed extensively in
Moron et al. (2006, 2007). The interannual variance of
the SAI (var[SAI]) measures the spatial coherence be-
tween M stations (or grid points) because it depends on
the interstation correlations; it ranges from var[SAI] 5
0 when two samples of equal size, perfectly covariant,
are perfectly out of phase; var[SAI] 5 1/M when all the
correlations are zero; and var[SAI] 5 1 when all stations
are perfectly correlated.
The SAI is an empirical estimate of the shared in-
phase ‘‘signal’’ across the network. The ‘‘noise’’ com-
ponent can be defined in terms of the (square rooted
spatial average) squared deviations relative to the SAI.
This definition of signal and noise is analogous to the
distinction between externally forced and internally
generated variance in ensembles of general circulation
model (GCM) simulations (i.e., Rowell 1998), with
stations or grid-points playing the role of GCM en-
semble members. A signal-to-noise ratio (SNR) can be
formed by dividing the SAI by the noise, but this sec-
ond-order statistic is more sensitive to sampling issues
than the SAI used here.
Statistical significance of interannual correlations is
assessed against 1000 synthetic time series of the same
length and spectral density as the observed pair, but
random phase (Janicot et al. 1996), with the two-sided
90%, 95%, and 99% significance levels indicated in the
following by one (*), two (**), and three (***) asterisks,
respectively.
3. Results
a. Onset date
The mean onset dates determined from CMAP and
GSOD, plotted in Fig. 1a, exhibit a northwest–southeast
progression from late August in northern-central
Sumatra and Kalimantan to mid-December in Timor. The
dates agree well between the two datasets, while there is
a large interstation variability over Java (Fig. 1a) that
could be related to small-scale topographic features.
Onset occurred before 1 November, 1 December, and
1 January in 67% (65%), 79% (86%), and 94% (95%)
of cases, respectively, in CMAP (GSOD). Mean onset
dates computed for subsets of GSOD stations averaged
by subregion (Table 1) are in good agreement with
Naylor et al. (2007; their Fig. 1). Using their definition
(i.e., the first day when accumulated rainfall from
1 August reaches 200 mm) leads to similar median dates
to those shown in Table 1, except in northern areas (not
shown). Moreover, the interannual variability is highly
consistent between both definition with cross correla-
tions . 0.85*** for all regions displayed in Table 1 ex-
cept for northern Sumatra (r 5 0.52***). Onset date is
less relevant in the northern regions because of the
differing seasonality of rainfall north of the equator
(Aldrian and Susanto 2003).
The interannual variability of onset date for the 14
stations over western and central Java is shown in Fig.
1b in terms of the individual standardized anomaly time
series (dotted). The signal that is common to the 14
stations, defined by the SAI (heavy solid), accounts for a
moderate fraction (var[SAI]50.41) of the total variance
at the individual stations, indicating substantial inter-
station noise. However, the SAI is correlated at 0.80***
(0.83***) with the large-scale SAI [leading principal
component (PC) time series] computed from CMAP
onset dates over all 128 grid points (heavy dashed blue
and red curves, respectively), suggesting that the signal
in onset over western and central Java is related to the
large scale despite considerable small-scale noise. This
is also seen in the other subregions (Table 1). The in-
fluence of ENSO is clearly visible in Fig. 1b, with big
delays in large-scale onset during the 1982 and 1997 El
Nin
˜
o events. In fact, the correlation between large-scale
SAI (leading PC time series) of CMAP is correlated at
0.84*** (0.84***) with the Nin
˜
o 3.4 SST anomalies in
October, corresponding to the mean onset date across
the domain. Some skewness is also visible with delayed
onsets exhibiting larger amplitudes than early onsets.
The leading EOFs of CMAP and GSOD onset dates
are plotted in Fig. 2. The leading CMAP EOF accounts
for 36% of total variance (EOF#2 accounts for 9% of
total variance), and consists of a large-scale monopolar
pattern with highest loadings over ‘‘monsoonal’’
Indonesia, (i.e., from southern Sumatra to the Timor Sea;
Aldrian and Susanto 2003, their Fig. 2). Loadings remain
substantial toward the southeast but fall off rapidly over
northern Sumatra, the Malay Peninsula, and northern
Kalimantan where they are generally close to zero. The
loadings of the leading EOF of GSOD onset dates (31%
of the variance) are generally similar to those of CMAP,
842 JOURNAL OF CLIMATE VOLUME 22
while their PC time series are correlated . 0.90***; there
is thus a high level of consistency at large scale between
these two contrasting datasets. Similarly, the cross cor-
relations between the SAIs of each region defined in
Table 1 are always positive and significant at the one-
sided 95% level or greater.
As discussed in section 2d, the station-scale noise can
be defined in terms of the (square rooted spatial aver-
age) squared deviations of the stations’ rainfall relative
to the SAI. The noise variance computed in this fashion
for each of the subregions (not shown) is fairly uniform
in space, though somewhat smaller in southern
Kalimantan, southern Sumatra and western and central
Java. However, differences in the spatial sampling be-
tween subregions do not allow for confidence in this
second-order statistic.
b. Seasonal rainfall total and postonset amounts
The temporal correlations between the leading PC of
onset (Fig. 1b) and the leading PC of SOND seasonal
FIG. 1. (a) Mean onset date computed in CMAP (shading) and GSOD (dot) as the first wet
day of a 5-day sequence receiving . 40 mm from 1 Aug without a dry 10-day sequence
receiving , 5 mm in the following 30 days from onset. (b) Standardized onset date for western
and central Java GSOD stations (dotted lines) with the average, i.e., SAI (solid black line),
together with the CMAP SAI (blue dashed line) and standardized leading PC time series (red
dashed line) computed from all 128 CMAP grid points. The dashed horizontal lines delineate
the 95% confidence interval of a set of 14 white noise time series. Note that one std dev
corresponds to an averaged deviation of ;20 days for western and central Java.
1F
EBRUARY 2009 N O T E S A N D C O R R E S P O N D E N C E 843
total exceed 20.90*** for both datasets. The variance
explained by the leading EOF of SOND total (51%
in CMAP and 32% in GSOD) is even larger than that
of onset, presumably because of the seasonal in-
tegration of rainfall that filters out some of the local-
scale noise inherent in the definition of onset date. In
fact, 46% (CMAP) and 67% (GSOD) of onsets oc-
curred between 1 September and 31 December. This
suggests that at least some of the spatially consistent
interannual variability of SOND amount is actually
conveyed by the anomalous timing of the monsoon
onset.
Three approaches are used to test this hypothesis, by
estimating the spatial coherence of rainfall amount be-
yond the onset date. (i) First, the spatial coherence of
the rainfall summed over the 15, 30, 60, and 90 days
following the local onset date is computed. Postonset
rainfall is a priori independent of the timing of the onset
of the monsoon, although both may be influenced by
ENSO and local-scale SST. The disadvantage of this
approach is that postonset amounts refer to different
temporal windows depending on the particular year
and station location. Nonetheless, 30-day amounts, for
example, refer to periods before 1 January in 69%
(CMAP) and 88% (GSOD) of cases. (ii) In the second
approach, the component of the SOND total accounted
for by the large-scale onset, defined as the leading PC of
each dataset (Fig. 2), is removed using a least squares
linear regression. The remaining residual is thus asso-
ciated only with postonset amounts, and all information
linearly related to large-scale onset is removed a priori.
(iii) The last method is to compare cumulative spatial-
FIG. 2. Leading EOF of CMAP (shading) and GSOD (dot) onset dates, plotted as cor-
relations with the principal component time series. The time series of onset date at each grid
point were standardized prior to EOF analysis.
TABLE 1. Statistics of GSOD station onset date by subregion (N is number of stations), computed from the SAI of each region. The
hindcast skill refers to the correlation between the observed and hindcast SAI with a cross-validated CCA using July SSTs as predictors.
One, two, and three asterisks indicate correlation significant at the two-sided 90%, 95%, and 99% level according to a random-phase test
(Janicot et al. 1996).
N
25%, 50%, and 75%
percentiles of the
spatial average Var [SAI]
Correlation with
large-scale SAI
of CMAP
Correlation
with PC#1
of CMAP
Hindcast
skill
Western and central Java (west of 1128E) 14 16 Oct, 28 Oct, 11 Nov 0.41 0.80*** 0.83*** 0.59***
Eastern Java (east of 1128E) 7 13 Nov, 21 Nov, 2 Dec 0.44 0.74*** 0.72*** 0.61***
Southern Sumatra (south of 18S) 6 5 Sep, 19 Sep, 17 Oct 0.60 0.86*** 0.88*** 0.74***
Central Sumatra (between 18S and 28N) 7 15 Aug, 24 Aug, 31 Aug 0.43 0.76*** 0.74*** 0.51**
Northern Sumatra (north of 28N) 6 1 Sep, 11 Sep, 15 Sep 0.23 0.46** 0.41** 0.22
Southern Kalimantan (south of 18S) 6 17 Sep, 22 Sep, 25 Oct 0.72 0.80*** 0.79*** 0.84***
Central Kalimantan (north of 18S) 5 11 Aug, 29 Aug, 12 Sep 0.48 0.70*** 0.69*** 0.46**
Eastern Indonesia (east of 1208E and
south of 88S)
5 30 Nov, 13 Dec, 25 Dec 0.57 0.63** 0.60** 0.49**
844 JOURNAL OF CLIMATE VOLUME 22
average rainfall anomalies computed from 1 August as
expressed as percentage of the long-term mean for early
and late onset years.
Estimates of var[SAI] for each quantity are given in
Table 2. The spatial coherence is high (i.e., large var
[SAI]) for both onset date and seasonal total but falls to
near zero for postonset rainfall and SOND residuals.
There is, nonetheless, a weak increase of spatial co-
herence as the length of the postonset averaging period
increases from 15 to 90 days, expected because of the
progressive cancellation of meteorological events as the
length of considered period grows. The difference be-
tween CMAP and GSOD results could come from the
area, mainly oceanic, that is not sampled in GSOD and/
or smoothing provided by gridbox-pentad averages in
CMAP.
Standardized anomaly time series of postonset 90-day
amounts for each CMAP grid box are shown in Fig. 3a,
together with the SAI. Spatial coherence is generally
low in most years, with the exceptions of the 1982 and
1997 large El Nin
˜
o events. The loadings of the leading
EOF of the postonset 90-day amount are displayed in
Fig. 3b. These are weak over monsoonal Indonesia,
especially between southern Sumatra to Sulawesi,
where those of the leading EOF of onset peak (Fig. 2),
and this mode explains less variance (22% in CMAP
and 11% in GSOD) than does the leading EOF of
CMAP onset date (36% in CMAP and 32% in GSOD).
The temporal behavior of the leading PC is nevertheless
consistent with that of onset date; that is, the postonset
season tends to be anomalously dry when onset is
anomalously late, and vice versa, at least for CMAP (r
between leading PC of onset and of postonset 90-day
amount is 20.87*** in CMAP and 20.19 in GSOD; the
postonset PCs being correlated at 0.37* between the
two datasets). Note that the second EOF of postonset
90-day amount in GSOD (not shown) explains 10% of
total variance and is correlated at 20.60*** (respec-
tively 0.51**) with the leading PC of onset date in
GSOD (respectively the leading PC of postonset 90-day
in CMAP). The fact that the loadings are rather large
over the eastern Indian Ocean and scattered patches of
the northern and eastern oceanic margins of the domain
(Fig. 3b) could be evidence of a deterministic signal and
warrants further study.
The leading EOF of SOND residuals (Fig. 3c) shares
some similarities with that of postonset 90-day rainfall
amounts (Fig. 3b), at least for CMAP (25% explained
variance); both have relatively high homogeneous
loadings over eastern Indonesia, and weak loadings
across monsoonal Indonesia. The leading EOF of
GSOD (16% explained variance) lacks similarity with
its CMAP counterpart, and their PCs are not signifi-
cantly correlated (r 5 0.23). Nearby stations often have
quite different loadings, such as over Java (Fig. 3c). By
construction, the leading PC of SOND residuals is or-
thogonal to the leading PC of onset date.
Figure 4 shows the spatial average of the cumulative
rainfall anomalies (averaged over the 57 stations across
Indonesia in the upper panel and the 14 stations of
western and central Java in the lower panel) computed
from 1 August and expressed as percentage relative to
long-term mean for the 6 latest and earliest mean onset
dates. A constant modulation of rainfall anomalies
would lead to a straight horizontal line at the mean
rainfall anomaly. The largest positive (negative) cu-
mulative anomalies occurred in both cases before or
around the early (late) onset dates while the curves
usually tend to zero thereafter (Fig. 4). The spatially
averaged rainfall anomalies at the end of the rainy
season, somewhere in March–April, are still consistent
with the phase of the onset date but the amplitude of
these anomalies is weak (Fig. 4). It suggests that the
strongest spatially coherent signal at large scale (Fig. 4a)
and for a particular subset of stations (Fig. 4b) is before
or near the onset date while it tends to cancel thereafter.
c. Seasonal predictability of onset
The substantial spatial coherence of onset date sug-
gests seasonal predictability. To provide a measure of
the latter, regression models are built using cross-
validated CCA between July SST over the tropical
Pacific and Indian Oceans (208N–208S, 808–2408 E) as pre-
dictors and GSOD or CMAP o nset dates as predictands.
Note that the 14% of missing entries in GSOD were first
filled with a simple linear regression using the closest
CMAP grid point as predictor. The models were built using
the Climate Predictability Tool (CPT) software developed
at the International Research Institute for Climate Pre-
diction (IRI; http://iri.columbia.edu/outreach/software/);
TABLE 2. Interannual variance of the SAI (Var[SAI]) of the 57
GSOD stations, and 128 grid points of CMAP for local onset date,
and postonset 15-, 30-, 60-, and 90-day rainfall totals. Var[SAI]
ranges between 0 (correlation of 21 between two equal-sized and
perfectly covarying samples), 1/ m (5 0.02 for m 5 57 and 0.008 for
m 5 128) where m is the number of locations for spatially inde-
pendent variations, and 1 (perfect correlation between stations)
(Moron et al. 2007).
Var(SAI) GSOD Var(SAI) CMAP
Onset 0.30 0.31
15 day 0.03 0.05
30 day 0.03 0.08
60 day 0.05 0.11
90 day 0.06 0.14
SOND 0.26 0.46
SOND residuals 0.10 0.16
1F
EBRUARY 2009 N O T E S A N D C O R R E S P O N D E N C E 845
the predictor and predictand fields were prefiltered using
EOFs, with the number of modes retained determined by
maximizing the model’s goodness of fit under cross vali-
dation, with 5 yr withheld at a time. The leading 5 and 2 (1)
EOF modes are retained in SST and CMAP (GSOD) and
most of the cross-validated skill is associated with the
leading CCA mode whose predictand pattern (i.e., SST
pattern) is almost identical for CMAP and GSOD.
Homogeneous correlation maps of the leading CCA
mode are shown in Figs. 5a and 5b for SST and onset
date, respectively. The SST anomaly structure (Fig. 5a)
exhibits a classical ENSO pattern, together with high
FIG. 3. (a) Individual standardized anomalies of rainfall total for the 90-day period after
the local onset date at the 128 CMAP grid points (dots) with the SAI (solid). The dashed
horizontal lines delineate the 95% confidence interval of a set of 128 white noise time series.
(b) Leading EOF of postonset 90-day amounts in CMAP (shading) and GSOD (dot). (c)
Leading EOF of SOND residuals in CMAP (shading) and GSOD (dot). Units in (b) and (c)
are correlations with the respective principal component time series.
846 JOURNAL OF CLIMATE VOLUME 22
correlations around Indonesia, such that warm ENSO
events are associated with delayed onset (Hamada et al.
2002; Hendon 2003). The corresponding structure
in onset dates (Fig. 5b) indicates that the delayed onsets
extend right across Indonesia, with high loadings over
monsoonal Indonesia, decreasing weakly (strongly)
toward eastern (northwestern) Indonesia. The regres-
sion model hindcast skill is plotted in Fig. 5c in terms of
FIG. 4. Spatial average of cumulative rainfall anomalies (a) for all 57 stations and (b) 14 stations from
western and central Java (Table 1) computed from 1 Aug and expressed as percentage from the long-
term mean for the six latest (in red) and earliest (in blue) onsets (computed from the spatial average of
onset dates). The dashed line indicates each year and the full bold line indicates the mean of the 6 yr.
The time series are low-pass filtered with a Butterworth filter (cutoff frequency 5 1/30 cpd). The
asterisks indicate the station-average onset date.
1F
EBRUARY 2009 N O T E S A N D C O R R E S P O N D E N C E 847
anomaly correlation, with regional averages given in
Table 1 (last column). Skill values are highest over
monsoonal Indonesia, exceeding 0.5** from southern
Sumatra to southern Kalimantan and Timor, reaching
0.80*** for the SAI computed over all stations (0.70***
for CMAP). The subisland subsets of stations in Table
1 achieve station-averaged skills ranging from 0.22
(northern Sumatra) to 0.84*** (southern Kalimantan).
The spatial variability of skill over Java could be due to
random sampling but also to deterministic signals as-
sociated with small-scale orographic features and/or
orientation relative to low-level winds.
4. Conclusions
The spatial coherence of onset date and postonset
rainfall is analyzed from GSOD rain gauges and the
CMAP dataset. The onset date is defined using an ag-
ronomic approach, that is, the first significant wet spell
(here 40 mm in 5 days) without any potentially damaging
FIG. 5. Homogeneous correlation maps of (a) SST, and (b) onset date from CMAP
(shading) and GSOD (circles), of the leading canonical correlation analysis (CCA) mode (c)
MOS skill (i.e., correlation between observed and hindcast onset date) associated with the
leading CCA mode between July SST and onset dates.
848 JOURNAL OF CLIMATE VOLUME 22
dry spell (here 10 days receiving less than 5 mm) there-
after (here in the 30 postonset days). This definition is
best suited for the end user’s purpose but suffers from the
subjective choice of the parameters. Nevertheless, these
parameters broadly reflect the needs and risks associated
with major crops of Indonesia, such as lowland rice. The
long-term mean onset dates, as well as the frequency of
false starts are sensitive to these subjective parameters
and future applications should carefully consider the
impact of these choices on specific crops. However, for
our main purpose of analyzing the spatial coherence of
anomalous onset dates, the sensitivity to these param-
eters largely vanishes.
The interannual variability of rainy season onset over
monsoonal Indonesia is shown from both gridded
pentad CMAP and daily station GSOD rainfall datasets
to be characterized by a large-scale coherent signal,
together with a moderate amount of local-scale noise
(Figs. 1b and 2). Considering small subsets of GSOD
stations recovers this signal, despite the complexity of
the island topography (Table 1). The interannual anom-
alies are dominated by delayed onsets (Fig. 1b). Con-
versely, the spatial coherence of interannual rainfall
anomalies beyond the onset date is weak, as revealed by
the amount of rainfall in the 15 to 90 days after the onset
and the SOND residuals from large-scale onset (Table 2
and Fig. 3a). The leading EOF of postonset 90-day
CMAP amounts exhibits weak and rather inconsistent
loadings over the main islands with high loadings re-
stricted to eastern Indian Ocean and scattered patches
of the northern and eastern margins (Fig. 3b). However,
this signal is strongly consistent in sign with onset date in
CMAP (i.e., late onset associated with smaller postonset
amount and vice versa). The leading EOF of SOND
residuals from large-scale onset lacks consistency be-
tween the GSOD and CMAP datasets but both none-
theless exhibit large spatially coherent loadings over
eastern Indonesia, but not over the eastern Indian
Ocean (Fig. 3c). The spatial average of cumulative
rainfall anomalies also exhibit their largest amplitudes
before and near the onset date (Fig. 4), while the post-
onset cumulative rainfall anomalies tend almost
monotonically toward zero. There may thus be some
predictability in postonset seasonal amounts, but most
of the spatially coherent signal in SOND seasonal total,
especially across islands, is merely related to the onset.
Our main finding is that most of the large-scale in-
terannual signal of SOND seasonal rainfall total is
conveyed by variations in the onset date of the rainy
season. This implies that (i) rainfall monitoring at a
small set of stations spread across Indonesia should be
sufficient to establish interannual anomalies of onset
date, and (ii) the scale of the interannual variability of
the onset suggests a large-scale forcing and potential
seasonal predictability. Indeed large-scale onset is found
to be highly correlated with an ENSO SST pattern
during July (Fig. 5a), that is, at least one month and half
before the mean local-scale onset date. A cross-
validated CCA using July SST in tropical Indian and
Pacific Oceans (208N–208S, 808–2408E) as a predictor
leads to promising skill values for the large-scale onset
date (r 5 0.80*** for GSOD; Fig. 5b,c). Further work is
needed to examine the associated circulation changes
and to investigate the roles of ENSO and Indian Ocean
climate variability (Hendon 2003).
The spatial variation of hindcast skill (Fig. 5c) and
onset EOF loadings (Fig. 2) warrants further study.
Both exhibit maxima from southern Sumatra to south-
ern Kalimantan—quite close to the equator—and de-
creases gradually southward across Java and Sonde
Islands and more rapidly northward (Figs. 2 and 5c,
Table 1). The latter decrease could be related to the
year-round rainfall there (Aldrian and Susanto 2003),
and the onset date should be viewed merely as an in-
crease of the rainfall rather than the transition between
a real dry and wet season. In that case, the onset date is
sensitive to the subjective choices used to define it and is
clearly less robust. This does not apply to monsoonal
Indonesia south of 58S. The highest EOF loadings and
SST-related skill over southern Sumatra to southern
Kalimantan coincide with the largest interquartile range
of interannual variability (Table 1). This subequatorial
band is perhaps the most sensitive to the spatial shift of
the ITCZ that probably triggers the onset of the rainy
season. The complex orography across Java could also
enhance the intraregional noise even between close
stations but we must also keep in mind that the spatial
sampling is highest over Java (Fig. 1a). Similarly, the
nature of spatial coherence for postonset rainfall and
SOND residuals over eastern Indonesia and the eastern
Indian Ocean (postonset rainfall only), as well as sea–
land contrast needs further investigation using better
sampled datasets and/or regional model simulations.
The large-scale signal in onset is still strongly present
in multistation small subisland regions (Table 1), indi-
cating the potential to downscale the large-scale onset
signal to the near-local scale. However, it is clear that
individual stations exhibit considerable noise (Table 1).
Thus, careful consideration needs to be given to the
trade-off between potentially more-accurate forecasts
at the aggregated scale versus local specificity for use in
climate risk management. The large-scale nature of
seasonal predictability of onset should enable improved
agricultural planning in the future, together with better
identification of false starts to the rainy season via real-
time monitoring and short-term forecasts of the large-
1FEBRUARY 2009 N O T E S A N D C O R R E S P O N D E N C E 849
scale evolving monsoon circulation. Forecasts of the
Madden–Julian oscillation may lend an additional source
of predictability at intraseasonal lead times (Wheeler
and McBride 2005).
Acknowledgments. We are grateful to P. Xie for in-
formation on the rain gauge measurements used in the
CMAP, and to two anonymous reviewers whose com-
ments helped us to clarify the paper. This research was
supported by grants from NOAA (NA050AR4311004),
the U.S. Agency for International Development’s
Office of Foreign Disaster Assistance (DFD-A-00-03-
00005-00), and the U.S. Department of Energy’s Climate
Change Prediction Program (DE-FG02-02ER63413).
REFERENCES
Aldrian, E., and R. D. Susanto, 2003: Identification of three dom-
inant rainfall regions within Indonesia and their relationship to
sea surface temperature. Int. J. Climatol., 23, 1435–1452.
——, D. Sein, D. Jacob, L. Du
¨
menil Gates, and R. Podzun, 2005:
Modeling Indonesian rainfall with a coupled regional model.
Climate Dyn., 25, 1–17.
——, L. Dı
¨
umenil Gates, and F. H. Widodo, 2007: Seasonal vari-
ability of Indonesian rainfall in ECHAM4 simulations and in
the reanalyses: The role of ENSO. Theor. Appl. Climatol., 87,
41–59.
Boer, R., and A. R. Subbiah, 2005: Agriculture drought in In-
donesia. Monitoring and Predicting Agricultural Drought:
A Global Study, V. S. Boken, A. P. Cracknell, and R. L.
Heathcote, Eds., Oxford University Press, 330–344.
——, and I. Wahab, 2007: Use of seasonal surface temperature for
predicting optimum planting window for potato at Penga-
lengan, West Java, Indonesia. Climate Prediction and Agri-
culture: Advance and Challenge, M. V. K. Sivakumar and J.
Hansen, Eds., Springer, 135–141.
Dikshit, U. N., D. Parida, and D. Satpathy, 1987: Geneticevaluation
and utilization: Drought tolerance. Int. Rice Res. News., 12, 6–7.
Giannini, A., A. W. Robertson, and J. H. Qian, 2007: A role
for tropical tropospheric temperature adjustment to ENSO
in the seasonality of monsoonal Indonesia precipitation
predictability. J. Geophys. Res., 112, D16110, doi:10.1029/
2007JD008519.
Hamada, J. I., M. D. Yamanaka, J. Matsumoto, S. Fukao, P. A.
Winarso, and T. Sribimawati, 2002: Spatial and temporal
variations of the rainy season over Indonesia and their link to
ENSO. J. Meteor. Soc. Japan, 80, 285–310.
Haylock, M., and J. L. McBride, 2001: Spatial coherence and
predictability of Indonesian wet season rainfall. J. Climate, 14,
3882–3887.
Hendon, H. H., 2003: Indonesian rainfall variability: Impacts of
ENSO and local air–sea interaction. J. Climate, 16, 1776–1790.
Janicot, S., B. Fontaine, and V. Moron, 1996: Sahel drought and
ENSO. Geophys. Res. Lett., 23, 515–518.
Katz, R. W., and M. H. Glantz, 1986: Anatomy of a rainfall index.
Mon. Wea. Rev., 114, 764–771.
Moron, V., A. W. Robertson, and M. N. Ward, 2006: Seasonal
predictability and spatial coherence of rainfall characteristics
in the tropical setting of Senegal. Mon. Wea. Rev., 134, 3468–
3482.
——, ——, ——, and P. Camberlin, 2007: Spatial coherence of
tropical rainfall at regional scale. J. Climate, 20, 5244–
5263.
Naylor, R. L., W. Falcon, N. Wada, and D. Rochberg, 2002: Using
El Nin
˜
o Southern Oscillation climate data to improve food
policy planning in Indonesia. Bull. Indones. Econ. Stud., 38,
75–91.
——, D. S. Battisti, D. J. Vimont, W. P. Falcon, and M. B. Burke,
2007: Assessing the risks of climate variability and climate
change for Indonesian rice agriculture. Proc. Natl. Acad. Sci.
USA, 104, 7752–7757.
Niewolt, S., 1989: Estimating of agricultural risks of tropical rain-
fall. Agric. For. Meteor.,
45, 251–263.
Rowell, D. P., 1998: Assessing potential seasonal predictability
with an ensemble of multi-decadal GCM simulations. J. Cli-
mate, 11, 1073–1093.
Sivakumar, M. V. K., 1988: Predicting rainy season potential from
the onset of rains in southern Sahelian and Sudanian climatic
zones of West Africa. Agric. For. Meteor., 42, 295–305.
Wheeler, M. C., and J. L. McBride, 2005: Intraseasonal variability
of the Australian–Indonesian monsoon region. Intraseasonal
Variability of the Atmosphere–Ocean System, W. K. M. Lau
and D. E. Walliser, Eds., Praxis Publishing, 125–173.
Wilks, D. S., 1999: Interannual variability and extreme-value
characteristics of several stochastic daily precipitation models.
Agric. For. Meteor., 93, 153–169.
Xie, P., and P. A. Arkin, 1996: Analyses of global monthly pre-
cipitation using gauge observations, satellite estimates, and
numerical model predictions. J. Climate, 9, 840–858.
850 JOURNAL OF CLIMATE VOLUME 22