Downscaling Indonesian pr ecipitation using large-scale ...david/Vimont_etal_2009.pdf · (paceklik , the season of scarcity) before the main rice harvest; and delayed planting of

INTERNATIONAL JOURNAL OF CLIMATOLOGYInt. J. Climatol. (2009)Published online in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/joc.2010

Downscaling Indonesian precipitation using large-scalemeteorological fields

Daniel J. Vimont,a* David S. Battistib and Rosamond L. Naylorca Department of Atmospheric and Oceanic Sciences and Center for Climatic Research, University of Wisconsin – Madison, Madison, WI, USA

b Atmospheric Sciences Department, University of Washington, Seattle, WA, USAc Program on Food Security and the Environment, Stanford University, Stanford, CA, USA

ABSTRACT: This study investigates the skill of linear methods for downscaling provincial-scale precipitation overIndonesia from fields that describe the large-scale circulation and hydrological cycle. The study is motivated by thestrong link between large-scale variations in the monsoon and the El Nino – Southern Oscillation (ENSO) phenomenonand regional precipitation, and the subsequent impact of regional precipitation on rice production in Indonesia. Threedifferent downscaling methods are tested across five different combinations of large-scale predictor fields and two differentestimates of regional precipitation for Indonesia.

Downscaling techniques are most skillful over the southern islands (Java and Bali) during the monsoon onset or transitionseason (Sep-Dec). The methods are moderately skillful in the southern islands during the dry season (May-Aug), and exhibitpoor skill during the wet season (Jan-Apr). In northern Sumatra downscaling methods are most skillful during January toApril with little skill during other times of the year. There is little difference between the three different linear methodsused to downscale precipitation over Indonesia. Additional analysis indicates that downscaling methods that are trainedon the annual cycle of precipitation produce less-biased estimates of the annual cycle of regional precipitation than rawmodel output, and also show some skill at reconstructing interannual variations in regional precipitation. Most of the skillof downscaling methods is attributed to year-to-year ENSO variations and to the long-term trend in precipitation andlarge-scale fields.

While the goal of the present study is to investigate the skill of downscaling methods specifically for Indonesia, resultsare expected to be more generally applicable. In particular, the downscaling models derived from observations have beeneffectively used to debias the annual cycle of regional precipitation from global climate models. It is expected that themethods will be generally applicable in other regions where regional precipitation is strongly affected by the large-scalecirculation. Copyright 2009 Royal Meteorological Society

KEY WORDS downscaling; Indonesia; precipitation; ENSO; monsoon

Received 8 December 2008; Revised 24 July 2009; Accepted 25 July 2009

1. Introduction

Indonesia’s geographical location (situated at equato-rial latitudes between the Southeast Asian and Aus-tralian land masses, and at longitudes between the Pacificand Indian oceans; Figure 1) places it in a positionto be strongly influenced by the annual march of theSoutheast Asian/Austral monsoon system, and by year-to-year fluctuations in the El Nino – Southern Oscil-lation (ENSO) phenomenon. The climatology of pre-cipitation over Indonesia has been described in detailduring the colonial period (Braak, 1929; this extensivework also documents the weather and climate phenom-ena of the region); with a focus on agricultural pro-duction – especially wet season rice (Schmidt and Fer-ugson, 1951; Oldeman, 1975; Oldeman et al., 1979);

* Correspondence to: Daniel J. Vimont, Department of Atmosphericand Oceanic Sciences and Center for Climatic Research, Universityof Wisconsin – Madison, 1225 W. Dayton St., Madison, WI 53706.E-mail: [email protected]

and more recently using objective criteria based on theamplitudes of the annual and semi-annual cycle in pre-cipitation (Hamada et al. 2002; Aldrian and Susanto2003; see also McBride, 1998, Chang et al. 2004, forreviews). In general, the latter classification schemesidentify regions that are wet all year round (northernSumatra, equatorial Kalimantan and Sulawesi, Papua);and regions that are strongly affected by the Australmonsoon, with rainfall maxima during October to Febru-ary (central and eastern Java, and the eastern islands inIndonesia’s archipelago) (McBride, 1998). These classi-fications closely follow the meridional variations of themonsoon system (Tanaka, 1994).

During some years, the annual march of the mon-soon is altered by large-scale conditions associated withconcurrent ENSO events (Haylock and McBride, 2003;Aldrian and Sustanto, 2003; Hendon, 2003; McBrideet al., 2003; Naylor et al., 2007). ENSO is most stronglyrelated to Indonesian precipitation (particularly over thesouthern islands) during the dry (Jun–Aug) and transition

Copyright 2009 Royal Meteorological Society

VIMONT ET AL.

Indonesia Provinces

1

2

34

5678

9 10 11 12 13 14

1516

17

1819

20

21 2223

24

100° E 110° E 120° E 130° E 140° E

100° E 110° E 120° E 130° E 140° E

10° S

5° S

0°

5° N

10° S

5° S

0°

5° N

Data Coverage

(a)

(b)

Figure 1. (a) Provincial scale over which precipitation observations areaggregated. Each number corresponds to the corresponding numberedregion in Table I. (b) Distribution of station precipitation observations.Grey dots indicate all station locations (after combination of the GHCN,GSOD, and EVE inventories) and dark black circles correspond tothe locations of stations that contribute to the station-based provincialprecipitation estimate [these stations have more than 30 years ofreported monthly precipitation (not necessarily contiguous)]. Gridboxes in (b) reflect the 0.5° resolution of the UDel precipitation

product.

(Sep–Nov) seasons, and is largely unrelated to precip-itation variations during the height of the wet season(Dec–Feb) (McBride and Nicholls, 1983; Haylock andMcBride, 2003; Hendon, 2003). During ENSO warmevents, the strength of the ascending branch of the Walkercirculation over the Maritime Continent is reduced,resulting in decreased precipitation and lower-level con-vergence over the region. These large-scale influencesare augmented by local air–sea interaction during thedry season, when increased easterly trade winds generateanomalously cool sea surface temperature (SST) in theregion through increased latent heat flux. The cooled SSTincreases anomalous subsidence in the region throughlarge-scale feedbacks and local hydrostatic arguments(Hackert and Hastenrth,1986; Hendon, 2003). Duringthe wet season the reverse occurs: anomalous easterliescounteract the climatological westerlies over the region,increasing SST through reduced latent heat flux and hencecounteracting the effects of large-scale subsidence overthe region. Alternatively, it has been argued that oncethe wet season commences precipitation variations aredominated by synoptic and mesoscale variations that areless sensitive to the large-scale circulation (Haylock andMcBride, 2003)

Despite the strong relationship between the large-scalecirculation and monsoonal precipitation, it is importantto note that precipitation over Indonesia is heavily

influenced by sub-seasonal variability and by small-scale variations. In particular, Qian (2008) shows thatthe regional geography of Indonesia leads to small-scaleland-sea breezes and mesoscale processes that influencethe diurnal cycle and mean intensity of convection overthe region. Houze et al. (1981) also points out theimportance of diurnal variations, and documents largesub-seasonal variations in rainfall over Borneo duringthe International Winter Monsoon Experiment that arerelated to synoptic scale wave activity over the region.Climate models that are capable of simulating presentand future large-scale climate variations are typicallyrun at a resolution that is far too coarse to resolvethe intricate topography and these small-scale variationsthat are known to influence monsoonal precipitation overIndonesia. These models also tend to have large biases intheir simulation of the hydrological cycle over Indonesia(Neale and Sligo, 2003; Naylor et al., 2007). In thepresent study, we test the ability of linear methods inboth downscaling and debiasing precipitation estimatesover Indonesia, given the large-scale circulation.

A motivation for the present work is the strong linkbetween year-to-year variations in Indonesian precipi-tation and rice production. Rice plantings follow themarked seasonality in rainfall, even in irrigated areassince most irrigation is categorized as ‘run of the river’(Naylor et al., 2001). In a typical year, most rice isplanted early in the ‘wet season’ (Oct–Dec), when thereis sufficient moisture to prepare the land for cultiva-tion and to facilitate the early rooting of transplantedseedlings. The main planting period occurs before thepeak of the monsoon, because excessive water at the veg-etative growth stage hampers rooting and decreases tillerproduction (De Datta, 1981). During the 90–120 daygrow-out period from transplanting to harvest, currentrice varieties require 600–1200 mm of water depend-ing on the agroecosystem and the timing of rainfall orirrigation (De Datta, 1981). Whether or not the regionis irrigated is important for gauging the critical amountof precipitation needed throughout the season. A smaller‘dry season’ planting (Apr–Jun) takes place in many, butnot all, rice-growing regions of Indonesia following themonsoon. Both the timing and amount of precipitationare thus critical for rice production systems.

Earlier work by Naylor et al. (2001) and Falconet al. (2004) shows that ENSO has been the primarydeterminant of year-to-year variation in Indonesian riceoutput over the past three decades, accounting for almosttwo-thirds of the total variation. During El Nino events,Indonesia’s production of rice – the country’s primaryfood staple – is affected in two important ways: delayedrainfall causes the rice crop to be planted later inthe monsoon season, thus extending the hungry season(paceklik, the season of scarcity) before the main riceharvest; and delayed planting of the main wet seasoncrop may not be compensated by increased planting laterin the crop year, leaving Indonesia with reduced rice areaand a larger than normal annual rice deficit (Naylor et al.,2001).

Copyright 2009 Royal Meteorological Society Int. J. Climatol. (2009)DOI: 10.1002/joc

DOWNSCALING INDONESIAN PRECIPITATION

The present study is motivated by the relationshipsbetween interannual climate variations, year-to-year fluc-tuations in regional precipitation, and rice production inIndonesia. The strong relationship between the large-scale circulation and precipitation over Indonesia leadsto the hypothesis that empirical downscaling techniquesmay be suitable for identifying precipitation variations ona regional scale over Indonesia. In the present study, wetest this hypothesis using linear methods to relate featuresin the large-scale circulation to regional precipitation.We also test the potential upper limit of predictabilityof regional rainfall in Indonesia from large-scale cli-mate variations. Although the study focuses on a specificregion (Indonesia), we expect the techniques to be gen-erally applicable to other tropical regions where precipi-tation variations have a signature in the large-scale circu-lation. We note that these models have also been used ina broader context of identifying whether natural climatevariability (including ENSO) will exert a greater impacton Indonesian rice agriculture and food security in 2050with changes in the mean climate Naylor et al., 2007).

The methodology herein is certainly not the only meansof downscaling precipitation over Indonesia. Numerousother methods for downscaling precipitation have beendeveloped by other research efforts; reviews can be foundin the work by Christensen et al. (2007) and Wilby et al.(2004). Other methods include regression approaches(under which our methodologies fall), weather genera-tors (e.g. Markov models), weather classification tech-niques, and analogue methods. Other researchers havefound that some techniques produce enhanced skill whendownscaling is applied to daily data, and then aggre-gated to a monthly time scale (Buishand et al., 2004),though the application to daily data does not guaranteea better estimate of monthly variations (Widmann et al.,2003). The techniques used in the present study have alsobeen employed by Widmann et al. (2003) over the PacificNorthwest region of North America, and have shown rea-sonable skill in estimating daily or monthly 50 km ×50 km resolution precipitation estimates from 1000 mbgeopotential height, 850 mb temperature or humidity, orcombinations thereof.

This paper is organized as follows. The data andmethodology used in this study are described in Section2, as well as in Appendices A and B. Results from thevarious downscaling methods, as well as the skill of thosemethods, are presented in Section 3. A discussion ofresults and attribution of skill is presented in Section 4,and a summary of results are presented in Section 5.

2. Data and methodology

The focus of the present study is to reconstruct provincial-scale precipitation over Indonesia using large-scale mete-orological fields. We investigate downscaling methodsusing two different estimates of provincial-scale precip-itation (described in Section 2.1 and Appendix A), fivedifferent combinations of large-scale meteorological vari-ables (described in Section 2.2 and Appendix B), and

three different statistical techniques (described in Section2.3 and Appendix B).

2.1. Precipitation data

In this study, precipitation data sets are constructed overIndonesia at a provincial scale that is appropriate forinvestigating relationships with agricultural production.The spatial scale for precipitation is defined using polit-ical and agricultural boundaries that correspond to thesame boundaries over which agricultural data are avail-able. Some provinces have been combined particularly toeliminate small delineations between regions (Table I).In particular, we aggregate precipitation over 24 dif-ferent political regions, listed in Table I, and shown inFigure 1(a). Data coverage over these 24 regions is farfrom uniform, with a large number of stations over Java,the most populated island in Indonesia (Figure 1(b)).Data are also averaged over three crop seasons roughlycorresponding to the peak of the Austral monsoon(Jan–Apr; JFMA), the dry season (May–Aug; MJJA),and monsoon ‘transition’ (Austral monsoon onset) sea-son (Sep–Dec; SOND). These seasons are chosen due totheir relationships with rice planting on Java as well asthe seasonality of the monsoon over Indonesia (Tanaka,1994; Naylor et al., 2007).

Two different regional precipitation data sets are con-structed and used in this analysis. The first is basedon the spatially interpolated, gridded data set from theUniversity of Delaware (Version 1.01 of the 1900–2006Terrestrial Precipitation product; Matsuura and Willmott(2007); hereafter referred to as UDel). Over Indonesia,that data set is primarily based on station observationsfrom the Global Historical Climate Network, version 2(GHCN; NCDC, 2008a), station observations from theGlobal Surface Summary of the Day (GSOD; NCDC,2008b), and a background climatology from Legatesand Willmott (1990). A provincial-scale data set is con-structed by simple averaging of the 0.5° × 0.5° UDeldata from all grid boxes centred in a given geographi-cal province (Figure 1b). We also tried aggregating pre-cipitation by averaging the standardized anomalies (thesame methodology as the aggregated station data; Hack-ert and Hastenrath, 1986 and Appendix A) and found nodifference in results.

Although the interpolation procedure for the UDel dataproduces a data set with fine-scale resolution (0.5° ×0.5°) the effective resolution is ultimately determined bythe original station density and interpolation algorithm.Note, e.g. the dense station network over Java (9–11)compared with Sulawesi (19–22) in Figure 1(b). In orderto interpolate to the original 0.5° × 0.5° grid resolution ofthe UDel data set, grid boxes in data-sparse regions mustbe interpolated from stations that are potentially far fromthe actual grid box, perhaps even across islands (one-point correlation maps for Java and for Papua, not shown,suggest this may be a problem). This may artificiallyaffect the skill of the downscaling technique, as data fora particular region is actually representative of a much


VIMONT ET AL.

Table I. Provinces in Indonesia considered in this analysis, and characteristics of data coverage. ‘Station’ lists the number ofstations with >30 years reported data from the combined station inventories; the total number of stations with any data at all inthe combined station inventories is listed in parentheses; the number of months with reported data is listed in brackets. ‘UDel’lists the number of 0.5 × 0.5° points in each province, from the UDel data. r(UD, Stat) is the correlation between the UDel and

station-based precipitation estimate (annual cycle has not been removed).

Province Name Station: >30 years(Total) [N mo]

UDel r(UD, Stat)

1. Aceh 3 (16) [563] 20 0.632. North Sumatra 2 (25) [597] 23 0.633. West Sumatra 0 (5) [0] 16 –4. Riau 3 (9) [585] 32 0.745. Jambi 2 (6) [588] 16 0.696. South Sumatra 2 (7) [600] 32 0.687. Bengkulu 0 (3) [0] 5 –8. Lampung 0 (0) [0] 11 –9. West Java (including Jakarta) 11 (77) [600] 16 0.85

10. Central Java (including Jogyakarta) 6 (75) [591] 12 0.8911. East Java 5 (80) [600] 16 0.8612. Bali 1 (20) [594] 2 0.7713. West Nusa Tenggara 2 (17) [578] 5 0.8314. East Nusa Tengara 4 (23) [598] 16 0.8515. West Kalimantan 3 (14) [600] 50 0.7516. Central Kalimantan 1 (13) [559] 51 0.6417. South Kalimantan 1 (11) [565] 15 0.6618. East Kalimantan 2 (19) [596] 64 0.5619. North Sulawesi 2 (13) [595] 8 0.5920. Central Sulawesi 1 (15) [367] 16 0.4321. South Sulawesi 1 (29) [584] 21 0.4722. Southeast Sulawesi 2 (6) [590] 11 0.6823. Maluku 5 (25) [596] 26 0.5924. Papua 12 (25) [599] 139 0.55

larger area (and hence more closely tied to the large-scaleflow used as a predictor). The UDel data set also uses datafrom stations that may have intermittent reporting overthe time period 1950–1999 (Figure 2), which may leadto spurious discontinuities in the estimated precipitation(Figure 3).

To provide an additional estimate of regional-scale pre-cipitation in Indonesia, a second, station-based regionalprecipitation data set is constructed. Station data forthe second regional precipitation data set is based onavailable Station data from the GHCN and GSODdata, and the NOAA/NCEP Climate Prediction Cen-ter reported precipitation [obtained under the name‘NOAA NCEP CPC EVE’ from the LDEO INGRID;IRI/LDEO (2007), (Ropelewski et al., 1985)]. Stationsfrom the three inventories were combined (see AppendixA for details) resulting in 71 stations with >30 yearsof data (not necessarily continuous) over Indonesia(open circles in Figure 1). Station data in each provincewere combined to produce an estimate of precip-itation that is representative of precipitation withinthe province, and that accounts for discontinuities inreported precipitation (Appendix A). This resulted ina station-based regional precipitation product with datafor 21 of the 24 provinces considered in this study(Table I).

Figure 2. Number of stations from all three data inventories reportingmonthly precipitation from 1950 to 1999. Light shading indicates thenumber of stations that reported precipitation for the given month;dark shading (thick black line) indicates the number of stations with>30 years of existing observations, that reported precipitation for agiven month; the dashed black line shows the number of stations withreported precipitation for a given month from the GHCN inventory(thus, the difference between the grey shaded curve and the dashedblack line indicates the number of stations with reported precipitation

from the GSOD and EVE inventories).

A comparison between various regional precipitationestimates over the island of Java is shown in Figure 3,including the UDel-based estimate, the station-based esti-mate, a satellite-based merged analysis of precipitationfrom the NOAA Climate Prediction Center (Xie and



1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 20000

5

10

0

5

10

0

5

10

0

5

10

NCEP

CPC

Station

UDel

Java and Bali Precipitation Estimates

Figure 3. Four different estimates of precipitation (mm day−1) over the islands of Java and Bali. The first and second curves are obtained asan average over regions 9–12 (Table I). The third curve is taken from a satellite-based estimate of precipitation from the CPC (Xie and Arkin1997). The bottom curve is from the NCEP reanalysis. For the bottom two curves, the average is taken over all grid points over the islands of

Java and Bali.

Arkin, 1997), and from the NCEP reanalysis (Kalnayand coauthors 1996). Of these, only the satellite-basedproduct is constructed from a homogeneous record. Thiscomparison also illustrates some of the challenges indefining a ‘true’ precipitation estimate for Indonesia. Inparticular, there are notable discrepancies between theUDel precipitation estimate and the Station precipita-tion estimate for various periods, especially from 1994to 1998. During this period, the station-based precipita-tion estimate shows better agreement with the satellite-derived precipitation product from the CPC (IRI/LDEO2007). During the 1990 wet season, though, the UDeland satellite-based product are in agreement, while theStation estimate is much lower. In general, the UDelestimate is slightly rainier than the station-based esti-mate (about 0.7 mm day−1) and has a more pronouncedseasonal cycle with up to 25% more rain during thewet season. Additional comparison between the UDeland station-based precipitation estimates is listed inTable I.

Although there are no clear winners in estimated pre-cipitation products, there is a clear loser over Indonesia.The estimate of precipitation from the NCEP reanal-ysis (Kalnay et al. 1996) shows large discontinuitiesin the early 1960s, late 1970s, and early 1990s (per-haps corresponding to the expansion of the radiosondenetwork and varying satellite coverage). The reanal-ysis demonstrates the large biases that can be intro-duced by a model-based precipitation parameterization(presumably due to the changes in data coverage andbiases in model parameterizations). We do note, how-ever, that inspection of the precipitation variability in thereanalysis estimate shows some similar variability as theobservational products suggesting that despite its prob-lems the reanalysis precipitation may be useful in someanalyses.

2.2. Large-scale data

Data from the National Center for Atmospheric Research/National Center for Environmental Prediction (NCEP/NCAR) Reanalysis project (Kalnay et al. 1996; here-after referred to as the reanalysis) is used to describethe large-scale circulation. We use three different vari-ables or combinations of those variables as predictors forprovincial precipitation: sea level pressure (SLP), 850-mbspecific humidity (SHUM), and 250 and 850-mb zonalwind (UWND). SLP is used due to the strong relationshipbetween precipitation and SLP over Indonesia and con-current ENSO events (Ropelewski and Halpert, 1987);SHUM is used to represent the large-scale hydrologicalcycle (including possible trends); and UWND is used tocharacterize the monsoon shear line, which is stronglyrelated to monsoon onset (Hendon and Liebmann, 1990;Tanaka, 1994). Five different predictor sets are used, anddescribed in Table II. When combinations of predictorsare used (e.g. SLP and SHUM), each field is scaled tohave equal variance. The predictor data are defined overthe region 20 °S–20°N, 90 °E–180° and are interpolatedfrom a 2.5° × 2.5° grid to a 5° × 5° grid (by spatial aver-aging). We interpolate so that fine-scale features (e.g.the monsoon shear line in UWND) do not unreason-ably affect estimates of downscaled precipitation whenlarge-scale GCM fields are projected onto the predictorstructures.

We also tried using the model-based large-scale pre-cipitation as a predictor for regional precipitation, butthe large biases in the modelled precipitation field(e.g. Figure 3) led to large errors in the downscaledprecipitation. As a result, we present results fromfive combinations of the large-scale predictor fields:(1) SLP, (2) SHUM, (3) SLP + SHUM, (4) UWND, and(5) UWND + SHUM. These variables, and combinationsof these variables, do exhibit some small non-stationary


VIMONT ET AL.

Table II. Large-scale variable sets used for downscalingprecipitation.

Variable set Physical process

1 Sea level pressure (SLP) Southern Oscillation,Large-scale monsoon

2 850-mb specific humidity(SHUM)

Hydrological cycle

3 SLP + SHUM See variable sets 1 and 24 850 and 200-mb zonal wind

(UWND)Large-scale monsoon shearline

5 UWND + SHUM See variable sets 2 and 4

behaviour over the time period 1950–1999, especiallyaround the introduction of the radiosonde network in theearly record, and in the transition to the satellite era inthe late 1970s. We do not attempt to adjust for thesebiases, except as described in the ‘annual cycle’ methodof downscaling below.

2.3. Methods

We test three different methods for estimating the rela-tionship between large-scale predictor fields and provin-cial precipitation: empirical orthogonal function/principalcomponent (EOF/PC) analysis used with linear regres-sion, maximum covariance analysis (MCA), and canoni-cal correlation analysis (CCA). These three linear meth-ods involve identifying spatial patterns in a predictorfield (the large-scale field) that relate to spatial patternsin a predictand (regional precipitation). These methodsare described in more detail in the work by Brethertonet al. (1992), and have been used to downscale precipita-tion over the Pacific Northwest (Widmann et al., 2003).We briefly describe each method in Appendix B. Eachmethod produces a set of large-scale predictor fields anda corresponding set of maps of regional precipitation.We refer to each methodology/predictor field combination(three different methodologies, five different predictorfields) as an ‘empirical downscaling model’ (EDM) thatcan be used to translate a new (independent) map of thelarge-scale circulation into an estimate of regional precip-itation. In the present study, we only apply the EDM toindependent data through our cross-validation method-ology (described below). However, each EDM can beapplied to general circulation model output to obtain anestimate of regional precipitation from the model’s large-scale fields. This will be discussed briefly in Section 4.

Regional precipitation for each crop season is predictedusing cross validation. The cross validation is performedas follows. First, one decade (1950–1959) is removedfrom a particular crop season, and the remaining 40seasons are used to train the statistical parameters forthe EDM. The resulting EDM is then used to predict thedecade that had been removed. The process is repeatedfor the remaining four decades of data, and then againfor the remaining two crop seasons. We also constructedEDMs using all seasons together (results not shown)and found poor performance, which is consistent with

findings that relationships between the large-scale flowand hydrology are seasonally dependent (McBride et al.,2003). Statistically significant correlations or regressionsare inferred when the correlation coefficient exceeds the95% level using a two-tailed t-test.

Finally, to address possible biases that may be intro-duced by non-stationarity in the large-scale fields, wealso train a set of EDMs on the 12-month annual cycle(with the grand mean removed) of large-scale fields andregional precipitation (e.g. the covariance matrix usedin MCA is defined only over the 12-month seasonalcycle). The 50 years of actual, monthly data (retainingthe annual cycle, but removing the grand mean) is thenfed into the EDM as a predictor, and the predicted pre-cipitation is averaged over each crop season. We thenevaluate how interannual variations of predicted precipi-tation relate to interannual variations in actual precipita-tion (the annual cycle is removed prior to evaluating theskill of the annual cycle EDMs). No cross-validation isnecessary here as the annual cycle that is used to trainthe downscaling technique is removed prior to evaluatingthe skill of the annual cycle EDMs in simulating inter-annual variability. These EDMs – trained only on theannual cycle – may show skill in predicting interannualvariations in precipitation if the interannual variations arerelated to changes in timing of the seasonal cycle (e.g.timing of the monsoon).

3. Results

3.1. Aggregated results

We begin by discussing results aggregated over all ofIndonesia. The average correlation (over all provinces)between the actual precipitation and the predicted (cross-validated) precipitation for the UDel precipitation esti-mate is shown in Figure 4. Results are shown as a func-tion of crop season, and as a function of the number ofmodes retained (Figure 4). Figure 4 demonstrates somecommon characteristics of the estimated precipitation. Ingeneral, the correlation increases from JFMA throughSOND, with the highest average correlation occurringduring SOND (the monsoon transition season). This isconsistent with (Haylock and McBride, 2003), who findthat precipitation is more coherent across stations duringthe dry season (JJA in their analysis) and transition sea-son (SON in their analysis). Despite the better skill duringthe transition season the average correlation is not high:in the best case, average correlations are near r = 0.6.These features are generally true for all three downscal-ing methods, and for both UDel (Figure 4) and Station(Figure 5) precipitation estimates.

In general, all three downscaling techniques havesimilar performance when at least two modes are used,indicating that the different large-scale predictor patternsshare similar information. This is also shown in Table III,which depicts the projection of the leading (unit length)spatial large-scale predictor pattern from each methodonto the leading and second predictor patterns from



JFMA MJJA SOND−0.2

0

0.2

0.4

0.6

0.8

Season

Mea

n C

orr

MCA

JFMA MJJA SONDSeason

CCA

1 Mode2 Modes3 Modes4 Modes

UDel Data; All–Indonesia


0

0.2

0.4

0.6

0.8

Season

EOF(a) (b) (c)

Figure 4. Mean correlation between predicted (cross-validated) and observed precipitation anomalies (as a function of method, crop season, andnumber of retained modes) over all of Indonesia, for the UDel data. Correlations are calculated on the provincial scale and then averaged overall provinces to create the all-Indonesia correlation. The three panels depict results for the MCA (left), CCA (middle), and EOF (right) methods.In each panel, results are shown for JFMA, MJJA, and SOND. Black circles, dark grey boxes, grey triangles, and light grey plus-signs referto reconstructions using 1, 2, 3, and 4 modes, respectively. The correlation is plotted for each of the five EDMs, based on the five differentpredictor fields (Table II). For example, the correlation using the MCA technique for SOND using one mode (panel (a), SOND, black circles)

ranged from about 0.18 to 0.53 for the five different predictor data sets.


0

0.2

0.4

0.6

0.8

Season

Mea

n C

orr

MCA


CCA


Station Data; All–Indonesia


0

0.2

0.4

0.6

0.8

Season

EOF(a) (b) (c)

Figure 5. As in Figure 4, except for the Station data.

Table III. Projection of the leading (unit length) UWNDpredictor pattern from each downscaling method (MCA, CCA,EOF) onto the leading and second predictor pattern from eachmethod, as a function of season. Spatial patterns are taken from

the full (not cross validated) 50-year data.

Season 1 (JFMA)

MCA1 CCA1 EOF1 MCA2 CCA2 EOF2

MCA1 1 0.97 0.85 0 0.57 0.4CCA1 0.97 1 0.88 0.03 0.6 0.39EOF1 0.85 0.88 1 0.49 0.9 0

Season 2 (MJJA)


MCA1 1 0.99 0.02 0 0.34 0.96CCA1 0.99 1 0.01 0.01 0.36 0.97EOF1 0.02 0.01 1 0.82 0.04 0

Season 3 (SOND)


MCA1 1 0.94 0.87 0 0.22 0.42CCA1 0.94 1 0.98 0.33 0.1 0.13EOF1 0.87 0.98 1 0.46 0.25 0

each method, as a function of season. Table III usespatterns derived from the entire 50-year data set (nocross validation is used in Table III). For brevity, weshow only the results from predictor set 4 (UWND). ForJFMA and SOND, the projections indicate that the spatialstructure of the leading mode from any one method isvery similar to the leading mode from any other method.The same is true for MCA and CCA during MJJA. Incontrast, for the UWND predictor set, the leading modefrom MCA and CCA has a larger projection onto EOF2than EOF1 for MJJA (further inspection shows that EOF1is dominated by the trend, while EOF2 is dominatedby ENSO). Inspection of Figures 4 and 5 shows thatfor three of the predictor sets (SHUM, UWND, andUWND + SHUM, from further analysis) the leading EOFmode (right panels) contributes very little to the meancorrelation, although the leading MCA and CCA modesdo (left and cenre panels, respectively). Because MCAand CCA are based on the covariance (or correlation)between the large-scale predictors and precipitation theunrelated large-scale variability captured by EOF1 isskipped over in these techniques.

The correlation between actual and estimated precipi-tation is mildly dependent on the number of modes used


VIMONT ET AL.

1 3 54−0.2

0

0.2

0.4

0.6

0.8

Variable Set

Mea

n C

orr

MCA

JFMAMJJASOND

Variable Set

CCA

Station Data; All–Indonesia

−0.2

0

0.2

0.4

0.6

0.8

Variable Set

EOF(a) (b) (c)

2 1 3 542 1 3 542

Figure 6. Mean correlation between predicted (cross-validated) and observed precipitation anomalies (as a function of method, variable, and cropseason) for the EDMs calculated using the Station data. Correlations are calculated on the provincial scale and then averaged over all provincesto create the all-Indonesia correlation. Precipitation estimates are made using 2 modes, except in cases where the prefilter for CCA retains onlyone mode. The three panels depict results for the MCA (left), CCA (middle) and EOF (right) methods. For each panel, variable set 1–5 referto large-scale predictions made with SLP, SHUM, SLP + SHUM, UWND, and UWND + SHUM, respectively. Black circles, dark grey boxes,

and grey triangles, depict results for JFMA, MJJA, and SOND, respectively.

to reconstruct precipitation. Note that for CCA, the num-ber of modes is limited by the number of EOFs retainedin the EOF prefilter (Appendix B), which is between 2and 3 for the UDel data, and between 1 and 6 for theStation data. Figure 4 shows that in general for MCAand CCA, the first mode dominates the correlation skill(with the exception of JFMA, where there is little skill tobegin with), with only mild improvements from remain-ing modes. Similar behaviour is seen in the Station data.As discussed above, 2 modes are necessary for the EOFmethod during MJJA and SOND.

Figure 6 highlights differences in results from differentlarge-scale predictors by stratifying results from thestation-based analysis as a function of method, predictorvariable, and season. In general, all large-scale variablesshow comparable skill during SOND, with almost nodistinction between the large-scale variables for the CCAtechnique during MJJA or SOND. It is also noteworthythat predictor sets that incorporate UWND (sets 4 and5) show higher skill during JFMA than other predictorsets, for all techniques. Still, the mean correlation duringJFMA for predictor sets 4 and 5 is not high (r ≈ 0.4).

3.2. Regional resultsSpatial maps of the cross-validated correlation coefficientbetween actual and predicted precipitation [using the EOFmethod with predictor set 4 (UWND) and 2 modes] areshown as a function of crop season for the UDel (inFigure 7) and Station (Figure 8) precipitation data. Thecorrelations tend to follow the seasonality of mean pre-cipitation. During JFMA, the Austral monsoon is at itspeak, and there are no significant correlations in the UDeldata over the southern-most islands in Indonesia. This isconsistent with earlier findings that precipitation is notcoherent during the height of the monsoon (Haylock andMcBride, 2003). Once the monsoon has begun, precipita-tion variability is likely dominated by meso- and synopticscale weather events that have a weaker relationship withthe large-scale, time-averaged flow. It is noteworthy thatFigure 8 shows statistically significant positive correla-tions in the Station data during JFMA over West and

JFMA Cross–validated Correlation

20 40 60 80

UDel Data

MJJA Cross–validated Correlation

20 40 60 80

SOND Cross–validated Correlation

20 40 60 80

Method: EOF(UWND), Modes 1–2

Figure 7. Correlation (radius of circles; see scale on figure) betweenthe predicted (cross-validated) and actual precipitation anomalies asa function of crop season, for the EDMs constructed from the UDelprecipitation data, and using the EOF methodology with predictor set4 (UWND) and 2 modes. The top, middle, and bottom panels showresults for JFMA, MJJA, and SOND. Dark (light) grey circles indicatepositive (negative) correlations. Statistically significant correlations are

outlined in black.



JFMA Cross–validated Correlation

20 40 60 80

Station Data

MJJA Cross–validated Correlation

20 40 60 80

SOND Cross–validated Correlation

20 40 60 80

Method: EOF(UWND), Modes 1–2

Figure 8. As in Figure 7, except for the station-based precipitation.

East Java, and over Bali, though these correlations areweak (r ≈ 0.4). In fact, statistically significant correla-tions during the height of the wet season originate frompredictor sets 4 and 5 (both incorporate UWND) in WestJava, predictor set 4 (UWND) over East Java, and frompredictor sets 1, 3, 4, and 5 (these incorporate SLP andUWND) in Bali. It is not clear whether the results fromthe UDel or station-based analysis are closer to ‘truth’.The selective sampling for the station-based precipitationproduct may bias the results to the more populated low-land areas, or it may provide a more homogeneous recordthat is less prone to discontinuities than the UDel data.

During MJJA, correlations using the UDel or station-based data are strongest in equatorial regions (southernSumatra, Kalimantan, Sulawesi, and Maluku), and weak-est in northern (northern Sumatra) and southern (Java,Bali) regions. In general, equatorial locations tend to havea weaker seasonal cycle in precipitation, and may also beless influenced by seasonality of the large-scale monsoon.During MJJA, northern Sumatra and the southern islands

(Java, Bali, Nusa Tenggara) are experiencing relative dryseasons (though dry season is much more pronounced inthe southern islands than in northern Sumatra). Recon-structions using other large-scale variables show similarfeatures.

The largest mean correlations between predicted andactual rainfall during SOND occur over southern Indone-sia (Java, Bali, Nusa Tenggara). At this time, correla-tion coefficients using the station-based precipitation arelarge (r = 0.7). Correlation coefficients with the UDeldata are not as large, as discussed above. In both cases,though, the large correlations suggest that there is a strongconnection between the large-scale circulation and pre-cipitation over southern Indonesia during the monsoontransition season. Mean monsoon onset dates range fromlate-October/early-November in western Java to mid-to late-December in Bali and Nusa Tenggara (Tanaka,1994; McBride, 1998; Hamada et al., 2002; Naylor et al.,2007). These regions also experience a large annual cyclein rainfall. Thus, rainfall variations during SOND are rep-resentative of variations in the timing of monsoon onset(Haylock and McBride, 2003; Naylor et al., 2007). Thelarge-scale circulation features that accompany monsoononset are well suited for predicting these variations in thetiming of the monsoon onset.

The islands of Java and Bali account for about 55% ofthe country’s total rice production and are thus watchedmost carefully from a food policy perspective. Agricul-tural data series for these regions are also longer andmore complete than for other provinces in Indonesia, andrainfall-rice relationships on Java and Bali have beenwell established Naylor et al. 2001). Thus, it is worthfurther examining the EDMs’ skill over these islands.The mean correlation coefficients between the predictedand actual precipitation over Java and Bali (provinces9–12 in Figure 1 and Table I) are shown in Figure 9(UDel precipitation) and Figure 10 (station precipitation).In general, the EDMs show similar characteristics overJava and Bali as they do over all of Indonesia (comparewith Figure 4 and 5). Still, notable differences exist. Inparticular, the UDel-based precipitation product exhibitsalmost no skill in estimating precipitation during JFMA(the peak of the Austral monsoon) for any method (forthe station-based precipitation, the average correlationcoefficient over Java and Bali during JFMA is compa-rable to the average correlation over all of Indonesia).Also, the average correlation over Java and Bali usingthe UDel-based precipitation during SOND (the transitionseason) is comparable to the average correlation over allof Indonesia, while the average correlation over Java andBali in the station-based precipitation estimation is muchhigher than the average over all of Indonesia. Resultsover Java and Bali suggest that the large-scale flow iscapable of predicting regional-scale precipitation varia-tions associated with the timing of the monsoon onset.This has special relevance to rice production, as timingof rice planting is tied to the timing of the monsoon onsetNaylor et al. 2001).


VIMONT ET AL.


0

0.2

0.4

0.6

0.8

Season

Mea

n C

orr

MCA


CCA


UDel Data; Java & Bali


0

0.2

0.4

0.6

0.8

Season

EOF(a) (b) (c)

Figure 9. Mean correlation between predicted (cross-validated) and observed precipitation anomalies (as a function of method, crop season, andnumber of retained modes) over Java and Bali, for the UDel data. Correlations are calculated on the provincial scale and then averaged over Javaand Bali (provinces 9–12 in Figure 1 and Table I). The three panels depict results for the MCA (left), CCA (middle) and EOF (right) methods.In each panel, results are shown for JFMA, MJJA, and SOND. Black circles, dark grey boxes, grey triangles, and light grey plus-signs refer toreconstructions using 1, 2, 3, and 4 modes, respectively. As in Figure 4, correlations are plotted for each of the five EDMs, based on the five

different predictor fields (Table II).


0

0.2

0.4

0.6

0.8

Season

Mea

n C

orr

MCA


CCA

Station Data; Java & Bali


0

0.2

0.4

0.6

0.8

Season

EOF(a) (b) (c)


Figure 10. As in Figure 9, except for the Station data.

3.3. Downscaling using the seasonal cycle

The downscaling methodologies described above useEDMs that are developed around anomalies from a par-ticular season. As such, these EDMs are unable to down-scale the climatological annual cycle of precipitation, andare subject to inhomogeneities in the data from whichthey are constructed. Motivated by the seasonality in themechanisms responsible for the variability in precipita-tion, and to test the impact of data inhomogeneity inEDM construction, we construct a series of EDMs thatare trained on the 12-month annual cycle of precipita-tion and large-scale fields (i.e. covariance matrices andregression coefficients are estimated across the 12-monthseasonal cycle with the grand mean removed). After theannual cycle is reconstructed at a regional scale, the grandmean for each region is added back to the downscaledprecipitation. For brevity, we restrict the presentation ofresults to the EOF methodology using 3 modes.

Figure 13 shows the annual cycle of actual and esti-mated precipitation averaged over Java and Bali. Theactual and estimated precipitations are taken from theUDel-based precipitation analysis. Also shown is the rawestimate of the annual cycle of precipitation for Java andBali taken from the NCEP reanalysis (estimated at thenearest 2.5° × 2.5° grid point). The latter is shown as

an example of the biases in model-based precipitation.Figure 13 suggests that the downscaling methodologymay provide a means for debiasing precipitation estimatesfrom large-scale models (because the large-scale predic-tor fields we use in the EDMs are also from the NCEPreanalysis). Indeed, Naylor et al. (2007) show this is thecase. Not surprisingly, percent errors (calculated as thedifference between predicted and observed precipitation,divided by observed precipitation for a given month) arelargest in magnitude during the dry season (when pre-cipitation rates are low), and are smallest during the wetseason. Given the importance of the timing of monsoononset, it is reassuring to see that the EDM estimates of theannual cycle are quite close to reality during the transitionseason (Oct–Dec). Results for the station-based precipi-tation data (not shown) are similar to those for the UDelprecipitation data. Inspection of other regions (not shown)indicates that the debiasing effect is slightly less effectivein regions with a strong semi-annual cycle (e.g. Aceh andWest Sumatra), but in all regions the downscaled annualcycle is in much better agreement with the original UDelestimate than is the NCEP reanalysis estimate. Nayloret al. (2007) have extended the use of the annual cycle-based EDMs to show that they are an effective meansof downscaling and debiasing climate projections from



the suite of models that contributed to the World ClimateResearch Programme’s Coupled Model IntercomparisonProject phase 3 (CMIP3) multi-model dataset.

Precipitation variability over Indonesia is stronglyrelated to the timing of the monsoon – a feature thatshould be captured in the annual cycle. Thus, we inves-tigate the skill with which the annual cycle-based EDMsare able to reproduce the seasonal variations of precipita-tion over Indonesia. In the case of the EOF methodology,we obtain the EOFs and PCs of the annual cycle of large-scale fields, and compute regression coefficients of theannual cycle of precipitation onto those PCs. Next, theseasonal anomalies are projected onto the annual cycleEOFs to produce a set of seasonal predictor expansioncoefficients, which are then multiplied by their associatedregional map of regression coefficients to produce an esti-mate of seasonal variations in regional precipitation. Themean correlation between the actual and estimated pre-cipitation for the EDM constructed using the annual cycleof the station-based precipitation is shown in Figure 14.For the most part, the EDMs perform much more poorlythan their counterparts based on seasonal anomalies, withsignificant skill coming only from predictor sets 1 and 3(both involve SLP) and only during MJJA and SOND(dry and transition seasons). Estimates that use UWNDas a predictor (sets 4 and 5) show quite poor skill forvariability, though these same EDMs perform quite wellwhen estimating the annual cycle. The skill in variablesets 1 and 3 in Figure 14 suggests that during the dryand transition seasons, year-to-year variations in precip-itation and SLP arise through variations in the timing oramplitude of the seasonal cycle.

The spatial distribution of annual cycle-based EDMskill is shown in Figure 15, which can be directlycompared to Figure 8. In general, the correlations areweaker using the annual cycle-based EDMs, especiallyduring JFMA (monsoon season) when there is almost noskill. However, there is still considerable skill over theequatorial regions and southern islands during MJJA andSOND, with correlations of r = 0.65 over Java. Again,the timing and regional dependence shown in Figure 15suggests that the physical processes that govern year-to-year precipitation variations produce variations in thetiming or amplitude of the seasonal cycle.

4. Attribution and discussion

In this section, we discuss the results presented above, aswell as possible applications of the EDMs to output fromglobal climate models.

Why the EDMs perform well over Indonesia: the annualcycle. The annual cycle-based EDMs are an effectivemeans of debiasing the modelled annual cycle of pre-cipitation over Indonesia because the annual cycle ofprecipitation is strongly related to large-scale circulationand humidity. These latter fields are generally better rep-resented in models than regional-scale or even large-scaleprecipitation. For regions where precipitation variations

are tightly connected to the annual cycle, the annualcycle-based EDMs are also capable of characterizing theinterannual variability in precipitation, suggesting thatthese interannual variations are related to changes in thetiming and amplitude of the seasonal cycle. Results fromthe annual cycle-based EDMs also confirm that thesemethods may be useful for debiasing model-based esti-mates of precipitation over Indonesia, or wherever thereis a strong relationship between the annual cycle of pre-cipitation and the large-scale circulation.

Why the EDMs perform well over Indonesia: the inter-annual variability. Figure 6 shows that the highest corre-lations between predicted and actual precipitation in thestation-based precipitation analysis occurs for predictorsets 3, 4, and 5 (SLP + SHUM, UWND, and UWND+ SHUM). Furthermore, Figure 5 shows that most ofthat predictability is due to the leading mode from anyanalysis. Thus, it is worth investigating the large-scaleconditions associated with that leading statistical modeof variability. Figure 11 shows the correlation map ofSST, SLP, 850 mb zonal wind, and regional precipitationwith the leading principal component of the UWND field(predictor set 4) for SOND. The spatial structure of thecorrelation maps bear a strong resemblance to the struc-ture of an El Nino event (a positive ENSO event) duringSOND. The SST pattern has been discussed in relationto Indonesian precipitation (Haylock and McBride, 2003;Aldrian and Sustanto, 2003; Hendon, 2003; McBrideet al. 2003). The SLP pattern bears the signature of theSouthern Oscillation Walker and Bliss, 1932) with lowpressure over the eastern equatorial Pacific ocean andhigh pressure over the western Pacific Ocean, Indonesia,and the Indian Ocean. The Southern Oscillation has beenshown to be a reasonable predictor of annual (Jul–Jun)precipitation by Hastenrath, 1987. The lower-level zonalwind depicts a characteristic relaxation of the climatolog-ical trades around the dateline, and anomalous divergence(at least in the zonal component of the flow) over Indone-sia. This divergence signifies a shift in the ascendingbranch of the Walker circulation, and as such it is notsurprising that the correlation coefficients between PC1and precipitation over Indonesia are large and negative.

Why the EDMs perform well over Indonesia: long-term trends. Figure 12 shows the relationship betweenthe large-scale flow and Indonesian precipitation duringMJJA. Plotted are the time series of mean station-basedprecipitation over all of Indonesia (PR), the Nino3.4index with polarity reversed (N34; this index is agood measure of ENSO variability, and is defined asSST averaged over 170–120 °W, 5 °S–5°N), and theleading two principal components of SLP (predictorset 1) during MJJA (PC1 and PC2). Visual inspectionshows a strong relationship between the time seriesof mean precipitation and ENSO events (warm ENSOevents – El Nino’s – are associated with dry conditionsduring this season), as discussed earlier. Figure 12 alsoshows a significant drying trend in precipitation overIndonesia. PC1 and PC2 are closely related to the long-term trend in SLP and variations in ENSO, both of


VIMONT ET AL.

0

2

4

6

8

10

12

Seasonal Cycle: Java and Balim

m d

ay−1

40

20

0

20

40

MA

Y

JUL

SEP

NO

V

JAN

MA

R

MA

Y

JUL

SEP

NO

V

JAN

MA

R

%

Percent Error in Seasonal Cycle

(a)

(b)

Figure 11. Reconstructed annual cycle (repeated) of UDel precipitationaveraged over Java and Bali (provinces 9–12 from Figure fig 1and Table I). (a) Annual cycle of precipitation from the UDel-basedprecipitation (thick solid black line), the NCEP reanalysis (thick dashedline), and individually downscaled estimates using EOF methodologyapplied to the annual cycle from predictor variable sets 1–5 (thingrey lines). (b) Percent difference between UDel-based annual cycleof precipitation, and NCEP estimate (thick dashed line) or downscaled

estimates (thin grey lines).

1 2 3 4 5

−0.2

0

0.2

0.4

Variable Set

Mea

n C

orr

Mean Correlation (All Indonesia)

Figure 12. Mean correlation between predicted and actual precipitationanomalies (as a function of variable and crop season) for the Stationdata. Correlations are calculated on the provincial scale and thenaveraged over all provinces to obtain the all-Indonesia correlation.Precipitation estimates are made using the EOF methodology appliedto the annual cycle, retaining 3 modes. Variable sets 1–5 referto large-scale predictions made with SLP, SHUM, SLP + SHUM,UWND, and UWND + SHUM, respectively. Black circles, dark greyboxes, and grey triangles, depict results for JFMA, MJJA, and SOND,

respectively.

which contribute to precipitation variations. The meancorrelation between the trend and station precipitationover Indonesia is r ≈ 0.3, while the mean correlation ofN34 with station precipitation is r = 0.4. Combined, thetrend and N34 produce a mean correlation of r ≈ 0.51(note that the trend is nearly uncorrelated with N34;

JFMA Correlation

20 40 60 80

Station DataAnnual Cycle EDM

MJJA Correlation

20 40 60 80

SOND Correlation

20 40 60 80

Method: EOF(SLP), Modes 1–3

Figure 13. Correlation (radius of circles; see scale on figure) betweenthe predicted and actual precipitation anomalies as a function of cropseason, for the Station precipitation data, using the EOF methodologyapplied to the annual cycle, with predictor set 1 (SLP) and 3 modes.The top, middle, and bottom panels show results for JFMA, MJJA,and SOND. Dark (light) grey circles indicate positive (negative)correlations. Statistically significant correlations are outlined in black.

r ≈ 0.1), and the combined PC1 and PC2 produce a meancorrelation of r ≈ 0.56. This highlights that both thelong-term trend and the interannual variability of ENSOare important for linking the large-scale circulation withregional precipitation variations. Similar arguments canbe made for other predictor sets.

Employing the EDMs for downscaling and debiasingoutput from the climate models (GCMs). An underlyingmotivation for the present study was to develop tech-niques to downscale and debias estimates of precipitationfrom the global climate models in the World ClimateResearch Programme’s Coupled Model IntercomparisonProject phase 3 (CMIP3) multi-model dataset (Meehlet al., 2007). Naylor et al. (2007) apply the annual cycle-based EDMs derived in this study (trained on the NCEPreanalysis) to GCM output from the CMIP3 archive and



SSTCorrelations with UWND PC1

SLP

850mb UWND

Station Precipitation

20 40 60 80

(a)

(b)

(c)

(d)

Figure 14. Contours (contour interval 0.2) of the correlation of UWNDPC1 (predictor set 4) for SOND with (a) SST, (b) SLP, (c) 850 mbUWND, and (d) regional station-based precipitation. In panels (a)–(c),solid contours denote positive correlations, dashed contours denotenegative correlations, and the zero line is shown as a thick black line.In panel (d) correlations are proportional to the radius of the circle (seescale on figure); dark (light) grey circles indicate positive (negative)correlations (all provinces except Aceh are negatively correlated withPC1), and statistically significant correlations are outlined in black.The UWND PC1 is taken from the full 50 years of data (not cross

validated).

find a considerable reduction in bias from the modelsimulated precipitation. The bias reduction indicates thatthe large-scale fields are better simulated than precipita-tion – not surprising given difficulties in parameterizingtropical convection in those models. The results also sug-gest that this technique may be generally applicable asa bias-reduction technique in other areas of the tropicswhere precipitation is closely tied to the large-scale cir-culation.

It is clear from Section 3 that the EDMs are skillfulin downscaling and debiasing interannual variations inregional precipitation from large-scale circulation fields.

−101

−101

−101

−101

PC2

PC1

−1×N34

PR

1950

1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

Season 2 (MJJA) Time Series

Figure 15. Standardized time series from MJJA: (PR) mean sta-tion-based precipitation over all of Indonesia, (−1 × N34) −1 timesthe Nino3.4 index, (PC1) the leading PC of SLP (predictor set 1) fromMJJA, and (PC2) the second PC of SLP (predictor set 1) from MJJA.For each time series, the trend over 1950–1999 is plotted as a dashed

line.

It is not clear, however, that the most skillful EDM inlinking the observed large-scale fields to the regionalprecipitation is also the best EDM for mapping the large-scale fields output from the GCMs to regional precipi-tation. For example, the UWND fields that comprise thepredictor fields from the observations tend to have fine-scale spatial structure that is difficult to capture in theGCMs. As a result, when fed output from the GCMS,EDMs based on the observed UWND produce largerbiases (for interannual variations) in downscaled precipi-tation than EDMs based on SLP (not shown). In contrast,the UWND-based EDMs do a good job in debiasing theseasonal cycle.

The strong relationship between ENSO and precipita-tion over Indonesia suggests that for practical purposes,a simple model that relates Indonesian precipitation tothe state of ENSO would suffice for producing a regionalestimate of year-to-year precipitation variations. We note,however, that the current generation of coupled generalcirculation models has serious biases in simulating detailsof ENSO variability, including biases in the structure,amplitude, and periodicity of ENSO events. In somecases, these biases will also translate to biases in thelarge-scale predictor fields used to train the EDMs inthis analysis. Naylor et al. (2007) addressed this issueby using the annual cycle-based EDMs to debias theseasonal cycle of precipitation from the CMIP3 mod-els. ENSO variability was incorporated into their estimatethrough simply adding observed ENSO variations to thereconstructed seasonal cycle. While that approach ignoresany possible changes in ENSO variability with climatechange, the serious biases in simulations of ENSO vari-ability in those models justify the methodology.


VIMONT ET AL.

5. Conclusions and discussion

Statistical techniques are used to downscale features ofthe large-scale circulation to regional precipitation overIndonesia. Three different techniques are tested (MCA,CCA, and EOF/regression analysis) using five differentcombinations of large-scale predictors. Results show thatthere is a skill in downscaling large-scale conditions toregional-scale precipitation, though the skill varies withregion, season and large-scale predictor. The followingmajor results are found:

• The strongest relationship between the large-scalecirculation and regional precipitation is found overJava, Bali, and the Indonesian archipelago during themonsoon ‘transition’ season (Sep–Dec). There is aweak to modest relationship in this region during thedry season (May–Aug), and during the wet (Jan–Apr)season results are conflicting across two differentprecipitation data sets. In the wet season, large-scalezonal wind variations show the strongest relationshipwith regional precipitation over this region.

• In northern provinces (especially northern Sumatra,Kalimantan, and Sulawesi) the strongest relationshipbetween the large-scale circulation and regional pre-cipitation occurs during the wet season. There is littleto no skill in northern Sumatra during the dry or transi-tion seasons. Skill in southern Sumatra and Kalimantantends to track skill in Java, Bali, and the Indonesianarchipelago.

• The skill of the downscaling method does not dependsensitively on the method tested. This is not overlysurprising as EOF, MCA, and CCA methodologies areall similar in their construction and methodologies.

• Downscaling methodologies that are trained on theannual cycle are an effective means of debiasing theseasonal cycle of precipitation on a regional scale inIndonesia. The annual cycle-based downscaling mod-els that incorporate SLP as a predictor are also capableof reconstructing interannual precipitation variationsover much of Indonesia during the dry and transitionseasons.

• Attempts to attribute skill in the downscaling methodsidentified ENSO and a long-term trend as importantpredictors for regional precipitation.

• A new, provincial scale, station-based precipitationdata set was developed for investigating relationshipsbetween precipitation, rice production, and the large-scale circulation over Indonesia (Appendix A).

Results from this study indicate that empirical down-scaling techniques are useful for producing downscaledand debiased estimates of precipitation variations, aswell as estimates of the annual cycle of precipitation,using features of the large-scale circulation over Indone-sia. These EDMs have also been used to downscaleand debias precipitation from large-scale fields output bythe current generation of global climate models (in theCMIP3 record; (Naylor et al., 2007). Results herein, and

in that study, suggest that these techniques may be use-ful wherever regional precipitation variations are stronglyrelated to the large-scale circulation.

The purpose of this study was to document the skill ofempirical downscaling techniques in relating the large-scale circulation to regional precipitation variations overIndonesia. The study emerged as part of a larger projectto assess the impacts of ENSO variations and global cli-mate change on rice production in Indonesia. The originalintent of the downscaling models in that context was todebias and downscale results from global general circula-tion models (especially the CMIP3 archive) to a regionalscale. This context has shaped many aspects of this study(e.g. the large simulation biases in the tropical hydrolog-ical cycle of those models have led us to ignore large-scale precipitation as a potential predictor of regional-scale precipitation – large-scale, satellite-based precipi-tation products may show more skill). Almost certainlyother efforts at downscaling precipitation over Indonesiaor other tropical regions will find other techniques andlarge-scale predictor variables that more closely suit theparticular purpose for those downscaling activities. How-ever, we do expect the general results of this study to bebroadly applicable and informative for those efforts.

Acknowledgements

This work was supported by the National ScienceFoundation under grant number SES-0433679. NCEPreanalysis data provided by the NOAA/OAR/ESRLPSD, Boulder, Colorado, USA, from their Web siteat http://www.cdc.noaa.gov/. Gridded precipitation datawere obtained from the University of Deleware. Sta-tion precipitation data were obtained from the IRI/LDEOClimate Data Library at http://iridl.ldeo.columbia.edu/.Thanks to two anonymous reviewers for their helpfulcomments on the original version of this manuscript.

Appendix A

Methodology for station-based precipitation

Station data for the second regional precipitation data setis based on available station data from the GHCN andGSOD data, and the NOAA/NCEP Climate PredictionCenter reported precipitation (obtained under the name“NOAA NCEP CPC EVE” from the LDEO INGRID;(Ropelewski et al. 1985). Over Indonesia, the GHCNinventory has observations from 1864 (Jakarta Observa-tory) to present; we use data from 1950 to 1999. TheGHCN data inventory contains observations from 484different stations, though station coverage is not con-sistent throughout the period considered in our analysis.In particular, GHCN station coverage drops markedly in1976 (Figure 2), and is almost completely absent by 1990(there are only a few stations with any data at all beyond1990). The GSOD data inventory contains data from 116



stations over Indonesia, and is available from 1979 to1996 (temporal coverage is variable); we use the ‘esti-mated precipitation’ rather than the ‘total precipitation’reported in the GSOD data, though differences betweenthe two are minor. The GSOD data set also reports thenumber of days each month in which a precipitationobservation was recorded. In general (and for the stationswe select) data coverage is good, though there are sometime periods with few observations per month, especiallyin the early 1990s. We do not attempt to control for thenumber of observations each month. The CPC data inven-tory contains observations from 129 stations from August1982 through the present, and we use data through 1999(temporal coverage is variable). We use the total precip-itation provided by this data set. All data were obtainedfrom the LDEO INGRID.

In order to construct a homogeneous data set forinvestigating precipitation variations over the period1950–1999, we combine data from the three precipitationdata inventories described above. We used the follow-ing methodology to identify common station observationsfrom the three different data inventories. First, we iden-tified stations that had the same five-digit WMO stationnumber. This resulted in a nearly seamless combinationof the GSOD and CPC data inventories. In the GHCNinventory, it was found that many stations had the sameWMO station number and were distinguished by a dif-ferent three-digit modifier (NCDC, 2007). In these cases,we classified two stations (from two different data inven-tories) as identical when any of the following criteriawere met (many stations met all of these criteria): thestations have the same station name (including differ-ent spellings), the same location (latitude and longitudewithin a tenth of a degree), and/or a high correlationbetween data points that were common to both records.Finally, all station records were visually examined, and afinal decision to combine records was made when cleardiscrepancies between data from different inventorieswere small (e.g. large changes in the mean, annual cycle,or variance were not obvious). Of the several hundredstations examined, there were on the order of 20 (or so)instances where the designation of identical stations wasnot clear; in these cases we did not designate the two sta-tions as identical. When multiple stations were classifiedas identical, the station data were averaged together (theaveraging was applied only to dates with reported data inmore than one station inventory; for dates with only oneinventory reporting, the available data were taken as is).Stations that were not classified as identical were retainedas independent station records.

The combined records from the three inventoriesresulted in a new data inventory containing 531 stationswith reported data. Of these stations, we retain only the 71stations with more than 30 years with reported data (thisdata need not be continuous). These station locations areplotted as open circles in Figure 1(b), and a timeline ofthe number of stations with reported data is shown as darkshading in Figure 2. The station coverage is relatively

sparse outside of Java and Bali, with some provincesretaining no station records at all (see also Table I).

Finally, a provincial-scale data set is constructed fromthe 71 retained station records. There are numerous waysto combine data from different stations in the sameprovince, each with its own set of disadvantages. It wasfound that a simple average of data from different stationsyielded a time series with non-stationary characteristicsas stations with missing data are introduced or eliminatedfrom the average. Thus, we (1) standardized all stationtime series within a given province by removing themean and dividing by the standard deviation, (2) averagedthe standardized time series together, (3) multiplied theresulting single time series for each province by the aver-age standard deviation of precipitation over all stationsin that province, and (4) added the mean precipitationover all original time series in that province to the singleprovincial time series. This resulted in provincial timeseries with much more stationary characteristics, but alsooccasionally resulted in negative precipitation reportedfor a province, especially during the dry season. As thisstudy focuses on explaining precipitation variance usingthe large-scale circulation as a predictor, we retain thenegative precipitation values.

Appendix B

Statistical downscaling methodology

The first technique, EOF analysis with linear regression,involves decomposing the predictor field into a setof spatial patterns (EOFs pi) and temporal expansioncoefficients (PCs zi). Linear regression of the predictandfield onto each principal component yields spatial maps ofregression coefficients (qi) that can be used to reconstructthe predictand. Given a new map x of the predictor atsome time t , a spatial map of the predictand (y) can beobtained via:

y =M∑

i=1

〈x, pi〉qi (B1)

where M is the number of retained EOF modes, and〈x, pi〉 is the projection of x onto the ith predictor EOFpattern. For the EOF-based EDM, each different field inthe predictor data set is scaled to have equal variance sothat all fields are weighted equally (i.e. for predictor set3, SLP and SHUM are scaled so that each field has thesame total variance).

MCA is a technique that identifies patterns in twodifferent fields that explain the maximum amount ofsquared covariance between the two fields. Again, in thistechnique each field in the predictor data set is scaledto have equal variance so that all fields are weightedequally. MCA is performed by applying singular valuedecomposition (SVD) to the temporal covariance matrixbetween the predictor and predictand fields, resulting ina set of paired spatial maps of the predictor (pi) and


VIMONT ET AL.

predictand (qi), and an associated singular value (σi).Each pair of patterns can be projected onto the originaldata to generate a predictor (ai) and predictand (bi)temporal expansion coefficient with covariance σi . Givena new map x of the predictor at some time t , a spatial mapof the predictand (y) can be obtained via linear regression:

y =M∑

i=1

σi〈x, pi〉var(ai(t))

qi (B2)

where M is the number of retained singular vectors,and 〈x, pi〉 is the projection of x onto the i’th predictorsingular vector.

The last technique considered here, CCA, is similarto MCA in that it identifies pairs of patterns betweentwo different data sets, but differs from MCA in that itmaximizes the squared correlation between the two datasets. As with MCA and EOF techniques, each field inthe predictor data set is scaled to have equal varianceso that all fields are initially weighted equally. An EOFprefilter is applied to the predictor and predictand fieldsprior to performing CCA to ensure that the analysisdoes not produce spuriously localized maps. In thisstudy, the EOF prefilter is used to retain the fewestPCs that explain at least 80% of the variance of eachfield (this amounted to between three to six principalcomponents for the precipitation fields). SVD is thenapplied to the correlation matrix between the retainedPCs, yielding pairs of canonical correlation patterns (inEOF space) for the predictor and predictand (pi and qi),an associated canonical correlation (Ci), and temporalexpansion coefficients (ai and bi). Given a spatial patternx from the predictor field, a prediction y is obtained via:

y =M∑

i=1

Nx∑

j=1

〈x, exj 〉

√λxj

pi(j )

Ci

Ny∑

k=1

√λ

ykey

kqi(k)

(B3)

where exj and ey

k are unit-length eigenvectors of thepredictor and predictand fields (from the EOF prefilter),λxj and λ

yk are the associated eigenvalues, and Nx and Ny

are the number of retained EOFs from the prefilter.

ReferencesAldrian E, Sustanto RD. 2003. Identification of three dominant

rainfall regions within Indonesia and their relationship to seasurface temperature. International Journal of Climatology 23(12):1435–1452.

Braak C. 1929. The climate of the Netherland Indies, Volumes I andII. Verhandelingen, No. 8. Koninklijk Magnetisch en MeteorologischObservatorium te Batavia. 1602 pp.

Bretherton CS, Smith C, Wallace JM. 1992. An intercomparison ofmethods for finding coupled patterns in climate data. Journal ofClimate 5(6): 541–560.

Buishand T, Shabalova MV, Brandsma T. 2004. On the choice ofthe temporal aggregation level for statistical downscaling ofprecipitation. Journal of Climate 17: 1816–1827.

Chang CP, Harr PA, McBride J, Hsu HH. 2004. Maritime continentmonsoon: Annual cycle and boreal winter variability. In East Asian

Monsoon, World Scientific Series on Meteorology of East Asia, vol.2, Chang CP (ed). World Scientific Publishing; 107–150.

Christensen JH, Hewitson B, Busuioc A, Chen A, Gao X, Held I,Jones R, Kolli RK, Kwon WT, Laprise R, na Rueda VM, mearns L,Menendez CG, Raisanen J, Rinke A, Sarr A, Whetton P. 2007.Regional climate projections. In Climate Change 2007: The PhysicalScience Basis, Contribution of Working Group I to the FourthAssessment Report of the Intergovernmental Panel on ClimateChange, Solomon S, Qin D, Manning M, Chen Z, Marquis M,Averyt KB, Tignor M, Miller HL (eds). Cambridge University Press:United Kingdom and New York; 847–940.

De Datta SK. 1981. Principles and Practices of Rice Production. JohnWiley and Sons: New York.

Falcon WP, Naylor RL, Smith WL, Burke MB, McCullough EB.2004. Using climate models to improve Indonesian food security.Bulletin of the Indonesian Economic Studies 40: 357–379.

Hackert EC, Hastenrath S. 1986. Mechanisms of Java rainfallanomalies. Monthly Weather Review 114: 745–757.

Hamada JI, Yamanaka MD, Matsumoto J, Fukao S, Winarso PA,Sribimawati T. 2002. Spatial and temporal variations of the rainyseason over Indonesia and their link to ENSO. Journal ofMeteorological Society of Japan 80(2): 285–310.

Hastenrath S. 1987. Predictability of Java monsoon rainfall anomalies:a case study. Journal of Climate and Applied Meteorology 26:133–141.

Haylock M, McBride JL. 2003. Spatial coherence and predictability ofIndonesian wet season rainfall. Journal of Climate 14: 3882–3887.

Hendon H. 2003. Indonesian rainfall variability: impacts of ENSO andlocal air-sea interaction. Journal of Climate 16: 1775–1790.

Hendon H, Liebmann B. 1990. A composite study of onset of theAustralian summer monsoon. Journal of Atmospheric Science 47:2227–2240.

Houze RA, Geotis SG, Marks FD, West AK. 1981. Winter monsoonconvection in the vicinity of North Borneo. part I: structure and timevariation of the clouds and precipitation. Monthly Weather Review109: 1595–1614.

IRI/LDEO. 2007. IRI/LDEO Climate Data Library. NOAA NCEPCPC EVE. Website accessed 25 November, 2008, http://iridl.ldeo.columbia.edu/SOURCES/.NOAA/.NCEP/.CPC/.EVE/.

Kalnay E and coauthors. 1996. The NCEP/NCAR 40-year reanalysisproject. Bulletin of the American Meteorological Society 77:437–471.

Legates DR, Willmott CJ. 1990. Mean seasonal and spatial variabilityin gauge-corrected, global precipitation. International Journal ofClimatology 10: 111–127.

Matsuura K, Willmott CJ. 2007. Terrestrial Precipitation: 1900–2006Gridded Monthly Time Series (Version 1.01). Website accessed 25November, 2008, University of Deleware. http://climate.geog.udel.edu/∼climate/html pages/Global ts 2007/README.%global.pts 2007.html.

McBride J, Haylock M, Nicholls N. 2003. Relationships betweenthe Maritime Continent heat source and the El Nino–SouthernOscillation phenomenon. Journal of Climate 16: 2905–2914.

McBride J, Nicholls N. 1983. Seasonal relationships betweenAustralian rainfall and the Southern Oscillation. Monthly WeatherReview 111: 1998–2004.

McBride JL. 1998. Indonesia, Papua New Guinea, and Tropical Aus-tralia: the Southern Hemisphere Monsoon, chap. 3A. In Meteorol-ogy of the Southern Hemisphere, Meteorological Monograph No. 49.American Meteorological Society: Boston, MA; 89–99.

Meehl GA, Covey C, Delworth T, Latif M, McAvaney B, MitchellJFB, Stouffer RJ, Taylor KE. 2007. The WCRP CMIP3 multi-modeldataset: a new era in climate change research. Bulletin of AmericanMeteorological Society 88: 1383–1394.

Naylor R, Battisti DS, Vimont DJ, Falcon WP, Burke MB. 2007.Assessing risks of climate variability and climate change forIndonesian rice agriculture. Proceedings of the National Academyof Sciences 104(19): 7752–7757.

Naylor RL, Falcon WP, Rochberg D, Wada N. 2001. Using ElNino/Southern Oscillation climate data to predict rice productionin Indonesia. Climatic Change 50: 255–265.

NCDC. 2007a. Global Historical Climate Network (GHCN)-monthlyversion 2. Website accessed 25 November, 2008, NOAA NationalClimatic Data Center. http://www.ncdc.noaa.gov/oa/climate/ghcn–monthly/index.php.

NCDC. 2007b. Global Surface Summary of the Day. Websiteaccessed 25 November, 2008, NOAA National Climatic Data Center.http://www.ncdc.noaa.gov/cgi–bin/res40.pl?page=gsod.html.



Neale R, Slingo J. 2003. The Maritime Continent and its role in theglobal climate: a G6CM study. Journal of Climate 16: 834–848.

Oldeman LR. 1975. An Agro-Climatic Map of Java. No. 17 in:Contributions form the Central Research Institute for Agriculture,Central Research Institute for Agriculture (Indonesia).

Oldeman LR, Las I, Darwis SN. 1979. An Agro-Climatic Map ofSumatra. No. 52 in: Contributions form the Central ResearchInstitute for Agriculture, Central Research Institute for Agriculture(Indonesia).

Qian JH. 2008. Why precipitation is mostly concentrated over islandsin the Maritime Continent. Journal of Atmospheric Sciences 65:1428–1441.

Ropelewski CF, Halpert MS. 1987. Global and regional scaleprecipitation patterns associated with the El Nino/SouthernOscillation. Monthly Weather Review 115: 1606–1626.

Ropelewski CF, Janowiak JE, Halpert MS. 1985. The analysis anddisplay of real time surface climate data. Monthly Weather Review113: 1101–1106.

Schmidt FH, Ferguson JHA. 1951. Rainfall Types Based on Wet andDry Period Ratios for Indonesia with Western New Guinea. No. 42

in: Verhandelingen, Kementerian Pehubungan Djawatan Meteorologidan Geofisik. Plus figures.

Tanaka M. 1994. The onset and retreat dates of the Austral summermonsoon over Indonesia, Australia, and New Guinea. Journal ofMeteorological Society of Japan 72: 255–267.

Walker GT, Bliss EW. Memorial Roy Meteorological Society. 1932.World weather V. 53–84.

Widmann MM, Bretherton CS, Salathe EP. 2003. Statistical precipita-tion downscaling over the Northwestern United States using numer-ically simulated precipitation as a predictor. Journal of Climate 16:799–816.

Wilby RL, Charles SP, Zorita E, Timbal B, Whetton P, Mearns LO.2004. Guidelines for Use of Climate Scenarios Developed fromStatistical Downscaling Methods. IPCC Task Group on Data andScenario Support for Impact and Climate analysis (TGICA).http://ipcc–ddc.cru.uea.ac.uk/guidelines/StatDown Guide.pdf.

Xie P, Arkin PA. 1997. Global precipitation: a 17-year monthlyanalysis based on gauge observations, satellite estimates, andnumerical model outputs. Bulletin of the American MeteorologicalSociety 78: 2539–2558.


Documents

Downscaling Indonesian pr ecipitation using large-scale ...david/Vimont_etal_2009.pdf · (paceklik , the season of scarcity) before the main rice harvest; and delayed planting of