A Comparison of Tropical Precipitation Simulated by the ...people.tamu.edu/~k-bowman/pdf/Collier et al 2004.pdf · mania), Trigo and Palutikof (2001; Iberian peninsula), and the 32-model

1 SEPTEMBER 2004 3319C O L L I E R E T A L .

q 2004 American Meteorological Society

A Comparison of Tropical Precipitation Simulated by the Community Climate Modelwith That Measured by the Tropical Rainfall Measuring Mission

Satellite

J. CRAIG COLLIER, KENNETH P. BOWMAN, AND GERALD R. NORTH

Department of Atmospheric Sciences, Texas A&M University, College Station, Texas

(Manuscript received 29 August 2003, in final form 21 January 2004)

ABSTRACT

This study evaluates the simulation of tropical precipitation by the Community Climate Model, version 3,(CCM3) developed at the National Center for Atmospheric Research. Monthly mean precipitation rates from anensemble of CCM3 simulations are compared to those computed from observations of the Tropical RainfallMeasuring Mission (TRMM) satellite over a 44-month period. On regional and subregional scales, the comparisonfares well over much of the Eastern Hemisphere south of 108S and over South America. However, model–satellite differences are large in portions of Central America and the Caribbean, the southern tropical Atlantic,the northern Indian Ocean, and the western equatorial and southern tropical Pacific. Since precipitation in theTropics is the primary source of latent energy to the general circulation, such large model–satellite differencesimply large differences in the amount of latent energy released. Differences tend to be seasonally dependentnorth of 108N, where model wet biases occur in realistic wet seasons or model-generated artificial wet seasons.South of 108N, the model wet biases exist throughout the year or have no recognizable pattern.

1. Introduction

The purpose of this study is to measure the perfor-mance of a general circulation model in its simulationof precipitation in the Tropics. The motivations for amodel validation of this kind are numerous. Between308S and 308N, vast acreage is devoted to the productionof citrus, corn, cotton, rice, wheat, and sugar (Espen-shade 1995). According to figures obtained from thePopulation Reference Bureau, as of mid-2002, over 2.8billion people were living in the Tropics, 1.5 billionalone in Southeast Asia. Residents of the Tropics con-stitute approximately 45% of the earth’s population ofaround 6.2 billion (PRB 2003). Thus a large part of thehuman population benefits from the accurate predictionof long-term changes in precipitation within these lat-itudes.

Additionally, precipitation in the Tropics is the sourceof latent energy that comprises three-fourths of the en-ergy used to drive the general circulation (Kummerowet al. 1998). Therefore, variability in the horizontal dis-tribution and intensity of tropical convection has globaleffects, as evidenced in the teleconnections of the ElNino–Southern Oscillation (ENSO) phenomenon (seeRopelewski and Halpert (1987, 1989), Montroy (1997),

Corresponding author address: J. Craig Collier, Department ofAtmospheric Sciences, Texas A&M University, 3150 TAMU, CollegeStation, TX 77845.E-mail: [email protected]

Mo and Higgins (1998), Lau and Wu (2001), and Adleret al. (2000), among others).

General circulation models (GCMs) use Newton’sequations of motion and the laws of thermodynamicsalong with parameterizations for subgrid-scale process-es to simulate the atmospheric and oceanic circulationsand the land surface. Essentially, they are models of theentire climate system and can be used to assess theimpacts of variability in the mean state of the atmo-sphere. Long simulations (of several tens of years) canbe used to predict climate change. Validation studies ofthe precipitation simulated by GCMs are numerous,many concentrating on specific regions of the earth.Examples include the studies of Kirkyla and Hameed(1989) and Chen et al. (1996; United States), Chen andYen (1994; Indian monsoon), Busuioc et al. (1999; Ro-mania), Trigo and Palutikof (2001; Iberian peninsula),and the 32-model intercomparison of Sperber and Palm-er (1996) which examined the simulation of interannualrainfall variability in the Brazilian Nordeste, AfricanSahel, and the Indian subcontinent.

Since tropical precipitation is so important to the cir-culation of the atmosphere and thus to the climate sys-tem itself, its simulation by these models should beevaluated. This kind of evaluation may be performedby comparing the model with observations. In recentyears, an important new dataset has become available,that of the Tropical Rainfall Measuring Mission(TRMM) satellite, which is a joint venture between the

3320 VOLUME 17J O U R N A L O F C L I M A T E

National Aeronautics and Space Administration(NASA) and the Japan National Space DevelopmentAgency (NASDA). One of the primary goals of thismission, as stated by Simpson et al. (1988), is to helpmodelers of the general circulation know the locationand amount of latent energy released in the atmosphereto improve climate modeling (Kummerow et al. 2000).Launched 27 November 1997, the TRMM satellite or-bits at an altitude of approximately 401 km, boostedfrom an altitude of 350 km in 2001,1 and is equippedwith a microwave radiometer and a precipitation radarto infer rainfall rates between roughly 388N and 388Slatitude. The microwave radiometer (TRMM Micro-wave Imager, or TMI) measures the electromagnetic en-ergy emitted by the earth–atmosphere system in chan-nels between 10.65 and 85.5 GHz. The size of the in-strument’s field of view depends on the frequency, rang-ing from 63 km 3 37 km at 10.65 GHz to 7 km 3 5km at 85.5 GHz (Kummerow et al. 1998). Precipitationretrieval algorithms rely primarily on emission of ra-diation from hydrometeors (Kummerow et al. 2000).

The TRMM Precipitation Radar (PR), operating at a13.8-GHz frequency, can measure the three-dimensionalrainfall distribution. The energy backscattered by pre-cipitation and received by the radar is converted to anequivalent precipitation rate, as explained in Kumme-row et al. (2000). At present, the TRMM satellite re-mains in orbit, and fortuitously, it observed most of thestrong 1997/98 El Nino event, allowing modelers totake advantage of this rich dataset for evaluating theirsimulations of ENSO. Our research uses the TRMMsatellite data from both instruments, that is, the mergeddataset based on PR/TMI algorithm 2B31 (Kummerowet al. 2000), to evaluate the National Center for At-mospheric Research (NCAR) Community Climate Mod-el, version 3 (CCM3) in the simulation of tropical pre-cipitation.

2. Methods

a. TRMM data

This study uses the 3G68 dataset, which was obtainedfrom the TRMM Science Data and Information System(TSDIS). It consists of essentially instantaneous precip-itation rates derived from TMI, from PR, and from thecombination of both instruments averaged over 0.58 30.58 latitude–longitude boxes between 388S and 388N.For validating CCM3, we use the precipitation ratesfrom the combination of TMI and PR. Generally, theTRMM satellite is able to observe a given location inthe Tropics about once per day, at a different time eachday, with a cycle of 46 days, the period of its orbitalprecession (Negri and Bell 2002). Therefore, for the 44-month period considered here (January 1998–August

1 The data for this study are derived from observations at its pre-boost altitude.

2001), there are about 1320 observations of each of theaforementioned boxes. Precipitation is highly variablein both space and time, and the incomplete nature ofthe satellite’s sampling introduces an error in the re-trieval relative to actual ground truth. This error isknown as sampling error and has been the subject ofseveral studies (see Shin and North 1988; Bell and Kun-du 1996, 2000; Bell et al. 2001; and most recently Belland Kundu 2003, for example). According to Shin andNorth (1988), TRMM sampling errors in monthly meanrain rates in wet areas of the Tropics would be less than10% when averaging the observations on areas of 58 358. And the results of Bowman et al. (2003) show thatvery long-term averages of satellite-derived rain ratescompare remarkably well with those measured by raingauges on areas as small as 18 3 18.

b. GCM simulations

CCM3 is a three-dimensional global spectral model.For this study, simulations are carried out at T42 hor-izontal resolution (approximately 2.88 latitude 3 2.88longitude) with 18 vertical levels. There is a rigid lidat 2.9-hPa pressure. The model uses a hybrid terrain-following vertical coordinate with sigma coordinates atlower levels that transition to pure pressure coordinatesat upper levels. Physical tendency parameterizations in-clude those for clouds (Slingo 1987, 1989; Ebert andCurry 1992; Cess 1985; Liou 1992), radiative fluxes(Ramanathan 1976; Ramanathan and Downey 1986),surface fluxes (Holtslag and Boville 1993), boundarylayer height (Vogelzang and Holtslag 1996), and gravitywave drag (McFarlane 1987; Lindzen 1981). Adjust-ment physics consist of a convective parameterizationfollowing Zhang et al. (1998), Zhang and McFarlane(1995), and Hack (1994), large-scale stable condensa-tion, and dry convective adjustment. One of the inputsfor the model is a monthly mean sea surface temperatureboundary condition. For this study, we used sea surfacetemperatures provided to us by the Program for ClimateModel Diagnosis and Intercomparison (PCMDI) atLawrence Livermore National Laboratory (Taylor et al.2000). Additionally, the model requires a time-variantozone mixing ratio boundary dataset, and an initial con-ditions dataset that includes initial values of prognosticvariables (Kiehl et al. 1996).

Climate models may exhibit considerable internalvariability or noise, partly due to fluctuations on syn-optic time scales. Consequently, as has been shown byBarnett (1995) in his GCM simulations, a single modelsimulation of an interannual climate event or a climateforecast is woefully inadequate for the accurate evalu-ation of a model’s performance. We will show that thisfinding is also valid for CCM3. To distinguish the mod-el’s response to natural variations in the SST boundarycondition (external variability) from its response to itsown internal variability, it is helpful to look at statisticsfrom an ensemble of simulations. Therefore, for this


FIG. 1. (a) Map of the 48 regions of comparison and monthly mean precipitation rates in mm day21 for Jan 1998–Aug 2001 as simulatedby the eight members of the CCM3 ensemble for (b) 14W and (c) 15E. Gray curves represent the individual member means, and the heavyblack curve represents the ensemble mean.

project, we carried out eight separate CCM3 simula-tions, each forced by exactly the same sea surface tem-perature boundary condition and differing only in theirinitial conditions. While the sensitivity of extended-range forecast models to initial conditions is quite sig-nificant (Lorenz 1963; Tracton and Kalnay 1993), theactual initial conditions are largely irrelevant to climateforecasts (Barnett 1995), since a climate simulation‘‘forgets’’ its initial conditions after some limit of de-terministic predictability. Therefore, to generate our en-semble, we modified initial conditions among the mem-bers by adding random perturbations to the temperaturefield of a 4-month spinup run (1 September 1996–1January 1997). By the beginning of the TRMM ob-serving period, in late November of 1997, the realiza-tions have decorrelated from each other and can be treat-ed as statistically independent. Intermember correlationsare never identically zero since the observed sea surfacetemperature field, common to all the members, exerts acommon forcing.

The precipitation rates of each simulation, both con-

vective and large scale, are saved as hourly averages.Unless otherwise stated, CCM3 monthly, seasonal, orannual means shown here are ensemble monthly, sea-sonal, or annual means, and all TRMM results are eithermonthly, seasonal, or annual means averaged onto theCCM3 grid. However, since the sampling errors asso-ciated with TRMM are too large on this grid, both da-tasets will be spatially averaged over much larger re-gions (greater than 58 3 58) for the model validation.

3. Results

Before discussing the comparison, we briefly addressthe variability across the ensemble. For the purposes ofanalyzing the model’s internal variability and for thecomparison that follows, we divided the Tropics into 48separate, nonoverlapping regions with partitions restinghalfway between two adjacent grid points. See Fig. 1awhich is a map of their locations. Boxes are numbereddown the columns and then across the rows, each boxlabeled with a suffix of ‘‘W’’ or ‘‘E’’ depending on


whether it is west or east of the prime meridian. Timeseries are computed from the monthly means averagedover all grid points within a box. For example, the re-sults to be shown for box 1W are monthly means av-eraged over all grid points between longitudes 178.68Eand 136.48W, and between latitudes 19.58N and 30.78N.Longitudes and latitudes are rounded off to the firstdecimal place for brevity. Figure 1b shows area-aver-aged monthly mean precipitation rates for all eight en-semble members for box 14W. The ensemble mean isshown in heavy black. Figure 1c shows the results forbox 15E.

For box 14W, the model exhibits remarkably littlevariability across the ensemble. Each of the eight curvesfalls very close to the ensemble mean. This is not thecase for box 15E. Here, the model exhibits considerablymore spread among the simulations. Notice that a clearlydefined seasonal oscillation is evident, which is the mod-el’s response to SST variations.

The comparison of the simulated and observed pre-cipitation rates is now discussed. Figure 2 shows a mapof the climatological annual-mean precipitation rate inthe Tropics, as computed from the CCM3 simulationsand from the TRMM observations for the period 1 Jan-uary 1998–31 August 2001. These results are presentedon the model’s approximate 2.88 3 2.88 grid.

Features of the large-scale general circulation are ev-ident in the annual-mean precipitation intensity. Themodel clearly depicts the intertropical convergence zone(ITCZ) regions straddling the equator and the areas ofsubsidence in the subtropics. It also captures the ascentand descent regions of the Walker circulation in thePacific Ocean. Along the equator, more intense precip-itation occurs over the Indian Ocean and the MaritimeContinent, while less precipitation occurs in the easternPacific. The geographical distributions of mean precip-itation agree fairly well with the TRMM climatology,but there are regions of overestimation by the model(see Fig. 2c, which plots the absolute differences). Theseinclude the Caribbean and eastern Pacific, equatorialAfrica, and the Indian Ocean. There are also regionswhere the model underestimates precipitation. These aregenerally smaller in area and include parts of north-western and southeastern South America, the centralequatorial Atlantic, western equatorial Africa, extremesoutheastern Asia, and parts of the Maritime Continent.

Maps of the climatological seasonal means (Fig. 3)show the tendency for CCM3 to simulate larger time-mean precipitation rates than observed by TRMM. Atthis point, it should be noted that the TRMM seasonalmeans are the time averages of 12 months of TRMMdata (four 3-month seasons). CCM3 results, on the otherhand, are computed using 8 times as much data. Inaddition, the CCM3 rain fields are available much morefrequently than the TRMM observations. Therefore,when averaging over these shorter time periods, theCCM3 maps are generally smoother than the corre-sponding TRMM maps; this is particularly evident in

maps of monthly means (not shown). Despite the sub-stantially greater amount of averaging, the CCM3 datahave numerous localized regions with higher precipi-tation rates than seen by TRMM. For example, duringDecember–January–February (DJF) simulated precipi-tation rates are quite high over portions of eastern SouthAmerica and the western equatorial Pacific Ocean, com-pared to the TRMM observations. During March–April–May (MAM), large localized differences are evident inequatorial Africa and in the western equatorial Atlantic.In June–July–August (JJA), the largest differences areconcentrated in two regions: Central America (and theadjacent waters) and in the northern Indian Ocean. Thedifferences in these regions are large not only in mag-nitude but also in spatial scale. In September–October–November (SON), model–satellite differences are morelocalized. Relative to TRMM observations, CCM3 sim-ulates too much precipitation in portions of the Carib-bean, the central equatorial Atlantic, equatorial Africa,the central Indian Ocean, and the western equatorialPacific.

To provide a more detailed view of the time structureof the differences between the model and the data, wecompare the fields month by month in each of the 48boxes described earlier. Note that the 48 boxes range insize from 458 3 8.38 to 458 3 11.28 and thus are largeenough to compare monthly means without worryingabout the sampling errors associated with the TRMMobservations. However, it is important to observe twopoints regarding these time series. First, the monthlymeans associated with these boxes are averages over atleast 48 2.88 3 2.88 model grid boxes, while the earliermaps of the annual and seasonal means are producedon the model’s approximate 2.88 3 2.88 grid. Therefore,some of the details evident on the model’s grid can beblurred or hidden in the averages over these larger box-es. Second, due to its orbital precession, TRMM requiresat least 1.5 months to completely sample the diurnalcycle of rainfall over a given grid box. Therefore, in agiven month, boxes with a strong diurnal cycle mayexperience most of their rain in hours unobserved bythe satellite, creating a low bias in the satellite monthlymean. It has been noted by Lin et al. (2002) that thisphenomenon may introduce a spurious signal into amonthly time series. It is difficult to estimate the sizeof this effect without a thorough investigation of thediurnal cycle simulated by the model. Such an inves-tigation is currently underway.

Time series of monthly mean precipitation rates areshown in Fig. 4. In order to quantitatively measure howwell one series tracks the other, the Spearman rank cor-relation coefficient is computed for each box. It is anonparametric statistic that assumes no a priori knowl-edge of the distributions of the monthly means (Desh-pande et al. 1995), and its value may span the realnumbers from 21 to 1. In Fig. 4, the coefficient isprinted in the top left-hand corner of each region’s plot.A deficiency of the difference map of Fig. 2c is that it


FIG. 2. Climatological annual mean precipitation rate in mm day21 (a) as simulated by CCM3 and (b) as observed by TRMM (TMI 1 PR)for the period 1 Jan 1998–31 Aug 2001. (c) The climatological mean difference (CCM3 2 TRMM) in mm day21. Note that all results areon the model’s approximate 2.88 3 2.88 grid.


FIG

.3.

Pre

cipi

tati

onra

tes

(mm

day

21)

sim

ulat

edby

(top

)C

CM

3an

dth

ose

esti

mat

edby

(bot

tom

)T

RM

M(T

MI

1P

R)

data

aver

aged

over

the

four

met

eoro

logi

cal

seas

ons,

from

1998

–20

01.

Not

eth

atal

lre

sult

sar

eon

the

mod

el’s

appr

oxim

ate

2.88

32.

88gr

id.


FIG. 4. Time series of monthly mean precipitation rates in mm day21 as simulated by CCM3 and as observed byTRMM from Jan 1998 through Aug 2001 for 48 separate regions in the Tropics. The solid black curve is the CCM3ensemble-average mean, and the dashed black curve is the TRMM mean. The rank correlation coefficient is given inthe upper left-hand corner of each box while the absolute and relative differences between CCM3 and TRMM totalprecipitation are given in the upper right-hand corner of each box. Absolute differences are given in cm (per year),and relative differences are given in %.


FIG. 5. CCM3 (solid) and TRMM (dashed) monthly mean precip-itation rates in mm day21 for the southern tropical Pacific Ocean(boxes 5W, 11W, and 12W of Fig. 4a). Rank correlation coefficientsare given in the upper left-hand corner of each box, while the absoluteand relative differences between CCM3 and TRMM total precipita-tion are given in the upper right-hand corner of each box. Absolutedifferences are given in cm (per year), and relative differences aregiven in %.

FIG. 6. Same as in Fig. 5 except for the Central American–Carib-bean–far eastern Pacific region (boxes 7W, 8W, 9W, 13W, and 14Wof Fig. 4a).

is not useful for discerning the model’s underestimationsor overestimations relative to what TRMM observes.While the absolute differences are more important forevaluating the model’s simulation of latent heating, therelative differences are important for appraising themodel’s success with precipitation; relative differencescan be large in regions where TRMM observes little tono rainfall. Therefore, the absolute and relative differ-ences in total precipitation as simulated by the modeland as observed by the satellite are computed for eachof the 48 regions. Total precipitation refers to the sumover all 44 months of the monthly mean precipitationrates weighted by the number of days in each month.The absolute difference in total precipitation refers toCCM3 total precipitation 2 TRMM total precipitation,where the sign of the difference is retained. That is, anabsolute difference that is less than zero indicates thatwhen rainfall is totaled over the comparison period, themodel exhibits a dry bias. The relative difference is theratio of this absolute difference to the TRMM total pre-cipitation multiplied by 100. In Fig. 4, the absolute andrelative differences are printed in the top right-hand cor-ner of each region’s plot, the absolute differences ex-pressed in centimeters per year, and the relative differ-ences in %. We divided the Tropics into six geographicalregions, labeled A through F and discuss each regionin turn. Where absolute and relative differences in totalprecipitation are quite large, the region is further sub-divided into smaller boxes for closer inspection. How-ever, even these smaller boxes are large enough that weexpect TRMM sampling errors to be sufficiently small.Within these subregions, large model–satellite differ-ences are diagnosed as either wet season, dry season,or seasonally invariant, based on relative differencesbetween the monthly means.

a. Region A: The central and eastern Pacific

Over the central and eastern Pacific, CCM3 performsreasonably well in it simulation of precipitation. For

example, the model simulates well the annual cycle ofprecipitation, where it is discernible in the observations.The simulations of boxes 3W, 4W, and 10W are goodexamples. The simulation of box 4W in the central Pa-cific is particularly noteworthy in that CCM3 tracksTRMM closely during the relatively wet El Nino periodof 1998 and then continues to simulate the normal an-nual cycle thereafter. In boxes where the model’s phaseof the monthly precipitation is simulated well, the cor-relation coefficient is high, as would be expected. Evenwhere the annual cycle is somewhat less pronounced,the model results remain close to the observations, suchas in box 6W. However, where the model correctly cap-tures the phase of the annual cycle, its precipitation ratesdo not always agree so well with the observations inmagnitude. Absolute and relative differences are par-ticularly high in boxes 5W, 11W, and 12W. For a betterunderstanding of where and when the largest of thedifferences occur, these regions are further subdivided,as seen in Fig. 5.

Over this region of the southern tropical Pacific, dif-ferences in the monthly means are generally indepen-dent of season and are highest in boxes 1 and 2, where,compared to the satellite observations, the model sim-ulates at least 83 cm more precipitation per year duringthe comparison period. Figure 3 suggests that the largemodel–satellite differences here are caused by the mod-el’s generating a South Pacific convergence zone(SPCZ) which is both too large and too intense. Theabsolute differences in total precipitation taper towardthe South American coast, as do the monthly mean pre-cipitation rates in general. However, relative differencesare larger over this part of the region. For example, themodel simulates 328% too much precipitation in box 4and in excess of 100% too much precipitation in box3. In spite of its difficulties with simulating magnitudein these dry areas, the model is reasonably successfulwith capturing the annual cycle. The success is mostevident in boxes 1 and 2 where it correctly simulatesmaxima in DJF and minima in JJA. However, themodel’s maxima and minima are large relative to theobservations.


FIG. 7. Same as in Fig. 5 except for the southern tropical Atlanticregion (boxes 22W and 23W of Fig. 4a).

FIG. 8. Same as in Fig. 5 except for the equatorial Africa region[boxes 2E (southern half ) and 3E of Fig. 4b].

b. Region B: Central America, the Caribbean, theGulf of Mexico, and the far eastern Pacific

In contrast to region A, where the model–satellitedifferences are relatively small except in the southern-most boxes, differences are quite large in most of regionB, encompassing Central America, the Caribbean, theGulf of Mexico, and the far eastern Pacific. While thesimulation in box 7W is satisfactory, the model exhibitsproblems with magnitude elsewhere. In box 8W, themodel simulates two precipitation peaks per year, whichare not seen by the satellite. Overestimation by the mod-el appears both in JJA and in DJF. In boxes 13W and14W, it is largely confined to JJA. Differences in themonthly means are considerable in box 14W, where sim-ulated precipitation rates are some 3 times higher thanobserved by TRMM during this period, resulting in ex-cessive rates of near 14 mm day21. In box 9W, whereabsolute and relative differences in total precipitationare small, the model appears to be having trouble cap-turing the annual cycle. This region is further subdividedfor a more precise evaluation in Fig. 6.

Over this region, model–satellite differences aremuch more seasonally dependent than they are over thesouthern tropical Pacific. In box 5, the model produces3 times as much precipitation in the wet season despitefair agreement with the satellite observations during thedry season. In box 4, the model’s overestimation is aslarge as 100% in both the wet and dry seasons despitefair agreement during the transitional periods of the year.In box 7, the simulation is relatively satisfactory duringthe wet season, but significantly wet-biased during thedry season, estimating between 2 and 6 times the sat-ellite-measured precipitation during this time. Box 3 isdrier than the other boxes examined here. However, inthis box, the relative difference in total precipitation islarge. Relative to the observations, the annual maximumis simulated late in the year and is about twice as high.In contrast to boxes 3, 4, 5, and 7, where absolute and/or relative differences in total precipitation are signifi-cant, these differences are almost negligible in box 6.Here, the simulation disagrees with the observations inthe phase of the annual cycle. Annual maxima are placedreasonably correctly, but annual minima are consistentlyabout 1/3 of a year (4 months) behind their location in

the data. The model is comparatively successful withboth magnitude and phase in boxes 1 and 2, except fora model wet bias of around 100% in the August–Sep-tember time frame. In terms of the total precipitationover the comparison period, the absolute and relativedifferences in these two boxes are quite a bit lower thanthose of boxes 4 and 5 to the south.

c. Region C: South America and the tropical Atlantic

Compared to the rest of the tropical Western Hemi-sphere, the simulation over the continent of South Amer-ica (boxes 15W–18W) is superior. In general, the CCM3annual cycle and magnitudes of precipitation are in goodagreement with TRMM. Over the northern tropical At-lantic (boxes 19W, 20W, and 21W), agreement is alsoquite good. Boxes 19W and 20W are dry compared tothe rest of region C, and the model’s low monthly meanprecipitation rates closely follow those observed. Box21W is considerably wetter throughout the year, but theCCM3 monthly means are still quite close to those ob-served, and the annual cycle evident here is similarlywell simulated. As a whole, region C shares a similaritywith central and eastern Pacific region A: the simulationand observations are in good agreement except in thesouthern oceanic boxes, where simulated magnitudesare too large. The most problematic portion of the regionresides in southeastern boxes 22W and 23W, where themodel tends to be consistently too wet throughout theperiod. Absolute and relative differences in total pre-cipitation are quite large in these two boxes. See Fig.7, in which this area is further partitioned.

Like those of the southern tropical Pacific, the model–satellite differences in the subregions of the southerntropical Atlantic are largely independent of season. Alsolike those of the southern tropical Pacific, the differencesare exclusively ones of magnitude. Where an annualcycle is evident, such as in boxes 1, 2, and 4, CCM3simulates it well. In this region, the largest absolutedifference in total precipitation is in box 1, encom-passing extreme eastern South America and the adjacentocean. Here, the model’s monthly means are nearly con-sistently between 100% and 150% larger than the sat-ellite’s monthly means throughout the year, excepting


FIG. 9. Same as in Fig. 5 except for the south-central Asia–northernIndian Ocean region (boxes 7E–10E and western portions of boxes13E and 14E of Fig. 4b).

FIG. 10. Same as in Fig. 5 except for the northwestern tropicalPacific Ocean region (boxes 19E and 20E of Fig. 4b).

reasonable agreement in November. The same behaviorcan be seen in box 4, yet, as in box 1, there are singlemonths where the model output and satellite data agree.Relative differences are huge in arid boxes 5 and 6,where there is no annual cycle present. For example, inbox 5, over the comparison period, the model simulates262% too much rain, and in box 6, it simulates over700% too much, the largest relative difference in totalprecipitation found in the Tropics of the Western Hemi-sphere.

d. Region D: Northern Africa and southern Asia

In northern Africa and southern Asia, the comparisondoes not fare well. In box 3E, the model output andsatellite data is not well correlated. CCM3 simulates asemiannual maximum in precipitation which appearsunfounded in the TRMM data. Though the model’s sec-ond precipitation peak (in October) lies relatively closeto the observations, the first one greatly exaggeratesprecipitation in January through April, and the relativeminimum between the two peaks actually underesti-mates precipitation from May to October. A partitioningof this region into smaller boxes, as seen in Fig. 8, showsthat the model’s behavior with respect to the satelliteobservations varies considerably on smaller spatialscales.

In equatorial Africa, model–satellite differences arehighly spatially dependent. In box 1, while the modelagrees well with the satellite during the dry season, itis too dry during the wet season. In box 4, the modelis too dry during the wet season and too wet during thedry season, indicating that the monthly means are outof phase. CCM3 overestimates total precipitation inboxes 2, 3, 5, and 6. In box 2, the model’s monthlymeans are in good agreement with those of TRMM dur-ing the dry months of DJF, but when there is observedprecipitation, they are consistently higher. Exceptionsoccur during SON of 1998 and 1999. In box 3, themodel–satellite differences are seasonally invariant: themodel’s wet bias is present throughout the year, evenin the rainless months of DJF. The model–satellite dif-ference in total precipitation is largest in box 5, where

the contrast between the monthly means is dramatic.Here, CCM3 most prominently simulates the semian-nual cycle evident in the larger-scale average, but thisfeature is not well supported by the TRMM data. Themodel’s maxima are consistently higher than those ofthe observations, and its minima in DJF are far too high.

Returning to Fig. 4b, in south-central Asia and thenorthern Indian Ocean, the disparity between CCM3 andTRMM is also quite large. In this region, absolute dif-ferences in total precipitation are especially large in box-es 8E, 9E, and 10E, and with few exceptions throughoutthe 44-month period, the model consistently overesti-mates precipitation in these boxes. This region is sub-divided in Fig. 9 for a more detailed consideration.

One of the outstanding features of Fig. 9 is themodel’s excessive precipitation rates in box 1 (the south-ern Arabian peninsula) during JJA. Here, the model’soverestimation is on the order of 300%. According toTRMM observations, monthly mean precipitation ratesare less than 1 mm day21 every month of the year, instark contrast to the model’s simulation of an average6 mm day21 during the month of August. This model–satellite difference can hardly be classified as a wet-season model wet bias, since there is no seasonal var-iation evident in the TRMM observations. It appearsthat the model is creating a wet season that does notexist. According to the seasonal-means map (Fig. 3) thesource of the artificial wet season in JJA appears to bean extension of a rainfall belt stretching across equa-torial Africa, from Guinea on the Atlantic coast, throughthe Congo, and then northeastward into Saudi Arabia.However, in contrast to CCM3, TRMM data suggestthat this belt does not extend as far east and that thereis a clearly defined ‘‘dry’’ spot over Saudi Arabia, con-sistent with its small monthly means.

The precipitation of box 2 is simulated much moresatisfactorily. This box is dominated by a monsoon-typeclimate, wherein strong southwesterlies during the sum-mer months transport moisture from the Indian Oceaninto the Indian subcontinent (Pant and Kumar 1997;Kendrew 1953). Here, the model’s timing of the onsetof the monsoon is about right, according to the TRMMobservations. In this box, the model’s total wet bias isless than 10 cm yr21. In boxes 3 to 6, where there is


FIG. 11. Same as in Fig. 5 except for the western equatorial PacificOcean region (boxes 21E and 22E of Fig. 4b).

FIG. 12. Categorical (a) absolute and (b) relative differences be-tween CCM3 and TRMM total precipitation for the 44-month com-parison period. (c) Types of model–satellite differences are plotted,where a plus or a minus sign refers to the model’s wet or dry bias,respectively. Each location corresponds to one of the subregions ex-amined above.

little annual cycle evident, the comparison fares betterin boxes 4 and 6 than in boxes 3 and 5. Precipitationin box 6 is simulated remarkably well during all of 1998.By contrast, in boxes 3 and 5, the absolute differencein total precipitation is quite large (over 1 m yr21), yetit is difficult to discern any systematic pattern to themodel’s wet bias.

e. Region E: The Maritime Continent and the westerntropical Pacific

There appears to be little seasonal variation in theprecipitation over the Maritime Continent, both in theobservations and in the simulations. In boxes 15E and16E, monthly means are high throughout the year (.6mm day21). There are intermittent periods of modeloverestimation and underestimation, but the total pre-cipitation received over a course of a year is in goodagreement with the observations, relative to the rest ofregion E. Over the western tropical Pacific, CCM3 doesnot behave as well. Consider the partitioning of boxes19E and 20E in Fig. 10 for example.

While the model performs comparatively well in box1, its wet bias in boxes 2 to 6 lies between 50 and 80cm yr21. CCM3 appears to be simulating a strong sea-sonality to the precipitation in boxes 2 and 3, thoughsuch a seasonal variation is not apparent in the obser-vations. Thus, it is difficult to classify the model–sat-ellite differences here, except to conclude that the modelappears to behave as it does over the Arabian peninsula,simulating unrealistic wet seasons, during which pre-cipitation rates are too high compared to the observa-tions. In box 6, where an annual cycle appears slightlybetter supported, the model’s wet bias is over 100%during the wet season months.

The most remarkable feature of the model’s perfor-mance in boxes 21E and 22E of Fig. 4b (partitioned inFig. 11) is its close agreement with TRMM observationsover New Guinea and Papua New Guinea (box 4).

Absolute and relative differences in total precipitationin box 4 are particularly small. However, in the otherfive subdivisions, the comparison is quite poor. For ex-ample, in boxes 1 and 2, CCM3 overestimates DJF pre-cipitation by about 100%. Here, as over the Arabian

Peninsula and in subregions of the northwestern tropicalPacific, the model appears to be simulating an unrealisticwet season. The wet bias is of a similar magnitude inbox 3, though it resides in JJA, where a wet seasonactually appears supported by TRMM observations. Inboxes 5 and 6, the model appears to be tracking theoverall trend in the precipitation observed by the sat-ellite, but there are periods of excessive precipitation,particularly in the latter halves of 1999 and 2000.

f. Region F: The southern Eastern Hemisphere

In contrast to the previous regions discussed, there isvery little to be criticized of the model’s performancein the southern Eastern Hemisphere. Annual cycles andmagnitudes of precipitation are simulated quite wellfrom southern Africa to Australia. Absolute differencesin total precipitation are generally less than 0.5 m yr21

and relative differences are generally less than 60%.Where its measure is useful, the rank correlation be-tween CCM3 and TRMM is high, indicating the model’ssuccess in capturing the annual cycle. The largest dif-ferences in total precipitation are found in boxes 11Eand 12E of the southern Indian Ocean. These are causedby a small, consistent month-to-month model wet bias.


FIG. 13. Bar charts representing the magnitudes of 44-month total precipitation in cm yr21 as simulated by CCM3 and as observed byTRMM. The CCM3 precipitation bar (on the left) is partitioned into a convective part (black) and large-scale stable part (white). Thepercentage of the CCM3 precipitation that is convective is written above the CCM3 bar. The TRMM (TMI1PR) precipitation bar is representedin dark gray to the right.

Overall, the model’s simulation over region F is a tes-tament to the fact that, in certain areas, CCM3 can sim-ulate both wet and dry climates well, since they are bothpresent here; contrast box 23E with box 18E, for ex-ample.

4. Summary and conclusions

The results of this study illustrate that there are largeregions of the Tropics in which monthly mean precip-itation rates, computed from an ensemble of CCM3 sim-ulations, disagree with those computed from observa-tions made by the TRMM satellite. Similarly, there arelarge regions where CCM3 and TRMM monthly meansagree well. We summarize the differences on fairlycoarse regional and subregional scales in Fig. 12.

The top two panels (Figs. 12a,b) plot categorical ab-solute and relative differences in total (44-month) pre-cipitation for the previously examined subregions. Thebottom plot (Fig. 12c) labels the differences as occurringin the wet season (W) or in the dry season (D), as phasedifferences (P), or as differences which are consistentthroughout the period or have no discernible seasonalvariation to them (C), or which are associated with ar-tificial wet seasons created by the model (A). The mod-el’s bias is indicated by a plus sign or a minus sign forwet or dry, respectively.

Where there are differences, they are almost univer-

sally positive, indicating the model’s wet bias through-out the Tropics. Negative model–satellite differences ev-ident in the annual-difference plot (Fig. 2c) are not near-ly as large on the regional and subregional scales con-sidered here; the positive differences are far moreprominent. The only exception is in the interior Guinealands of western Africa (near the intersection of theprime meridian and 158N). Here, the model consistentlyunderestimates wet-season precipitation. Regions whereabsolute differences are large and positive correspondto regions where CCM3 simulates too much conden-sational heating, and unrealistically large amounts oflatent energy can have effects on the simulation of thegeneral circulation. In the central South Pacific, the largepositive absolute differences between the model and sat-ellite monthly means are most likely caused by the mod-el’s generation of a SPCZ which is too large and toointense. Outside this region, the largest positive absolutedifferences tend to be north of 108S and along the west-ern edges of the Atlantic, the Indian, and the PacificOceans’ basins. North of the equator, about half of thelargest model–satellite differences are wet biases in thewet season or wet biases in model-generated artificialwet seasons. South of the equator, they are all wet biaseswith no seasonal dependence.

In locations where absolute differences are large,CCM3 does a poor job of simulating both precipitationand latent heating. However, there are locations where


FIG. 14. Scatterplots of CCM3 and TRMM regional-mean monthly mean precipitation rates (mm day 21)for (a) Northern Hemisphere, (b) Southern Hemisphere, (c) ocean, and (d) land regions. Rank correlationcoefficients are printed in the lower right-hand corner of each plot. The solid black line represents the lineof unit correlation.

absolute differences are comparatively small but relativedifferences are extreme. Prime examples include por-tions of the Arabian peninsula and the southern tropicalAtlantic, where absolute differences in total precipita-tion are less than 75 cm yr21, but relative differencesare greater than 300% and 700%, respectively. In suchregions, the model’s overrelease of latent energy is notas important as its failure to simulate what is climato-logically reasonable ‘‘weather.’’

It should be noted that the previous analysis has notdistinguished between the convective and large-scalestable parts of the model’s precipitation. However, theprecipitation as simulated by CCM3 in the Tropics isoverwhelmingly convective in nature. See Fig. 13,which shows the magnitudes of model-generated pre-cipitation, partitioned into its convective and large-scalestable parts, and TRMM-observed precipitation, totaledover the comparison period. In all regions, the model’slarge-scale stable portion is smaller than not only itsconvective portion but also the portion observed byTRMM. Thus the model’s wet biases are wet biases inconvective precipitation.

In light of the problem regions described earlier, themodel’s simulation of the annual cycle of precipitationagrees quite well with the TRMM observations overSouth America and most of the southern Eastern Hemi-sphere (generally south of 108S), including Australia

and southern Africa. Over these regions, absolute andrelative differences between the model- and satellite-derived monthly means are small, and the rank corre-lations are large. In general, the model’s wet biasthroughout the Tropics tends to be smaller over theSouthern Hemisphere than over the Northern Hemi-sphere, and it tends to be smaller over land than overocean. This general conclusion is supported by scatter-plots of the regional-mean monthly mean CCM3 andTRMM precipitation rates, specific to hemisphere andland/ocean (see Fig. 14).

In the interest of improving the model, it would beuseful to explore how the model’s biases in precipitationare related to its biases in features of the general cir-culation. Unfortunately, describing the ‘‘true’’ generalcirculation in the Tropics is subject to its own errorssince such a description relies on tools like NationalCenters for Environmental Predictions (NCEP) and Eu-ropean Centre for Medium-Range Weather Forecasts(ECMWF) reanalyses, which are in part based on theirown models. Several studies have shown large differ-ences between the reanalyses and observational dataand/or large discrepancies between the NCEP andECMWF reanalyses themselves (see Newman et al.2000; Rouault et al. 2003; Wu and Xie 2003; and Roads2003 for examples). Therefore, such an analysis is a


complex undertaking and is reserved for a separatestudy.

Acknowledgments. This material is based upon worksupported by a NASA Earth System Science fellowshipand by NASA Grant NAG5-4753 to Texas A&M Uni-versity. The authors thank the Goddard Space FlightCenter for making the TRMM satellite data availableonline via the TRMM Science Data and InformationSystem; the Community Climate Model group at NCARfor developing and assistance with running CCM3; andthe Program for Climate Model Diagnosis and Inter-comparison at Lawrence Livermore National Labora-tory for providing the sea surface temperature datasetin CCM3 format. In addition, the authors thank theanonymous reviewers, whose insightful comments andsuggestions helped prepare this manuscript for publi-cation.

REFERENCES

Adler, R. F., G. J. Huffman, D. T. Bolvin, S. Curtis, and E. J. Nelkin,2000: Tropical rainfall distributions determined using TRMMcombined with other satellite and rain gauge information. J.Appl. Meteor., 39, 2007–2023.

Barnett, T. P., 1995: Monte Carlo climate forecasting. J. Climate, 8,1005–1022.

Bell, T. L., and P. K. Kundu, 1996: A study of the sampling error insatellite rainfall estimates using optimal averaging of data anda stochastic model. J. Climate, 9, 1251–1268.

——, and ——, 2000: Dependence of satellite sampling error onmonthly averaged rain rates: Comparison of simple models andrecent studies. J. Climate, 13, 449–462.

——, and ——, 2003: Comparing satellite rainfall estimates with raingauge data: Optimal strategies suggested by a spectral model.J. Geophys. Res., 108, 4121, doi:10.1029/2002JD002641.

——, ——, and C. D. Kummerow, 2001: Sampling errors of SSM/I and TRMM rainfall averages: Comparison with error estimatesfrom surface data and a simple model. J. Appl. Meteor., 40, 938–954.

Bowman, K. P., A. B. Phillips, and G. R. North, 2003: Comparisonof TRMM rainfall retrievals with rain gauge data from the TAO/TRMM buoy array. Geophys. Res. Lett., 30, 1757, doi:10.1029/2003GL017552.

Busuioc, A., H. von Storch, and R. Schnur, 1999: Verification ofGCM-generated regional seasonal precipitation for current cli-mate and of statistical downscaling estimates under changingclimate conditions. J. Climate, 12, 258–272.

Cess, R., 1985: Nuclear war: Illustrative effects of atmospheric smokeand dust upon solar radiation. Climatic Change, 7, 237–251.

Chen, M., R. E. Dickinson, X. Zeng, and A. N. Hahmann, 1996:Comparison of precipitation observed over the continental Unit-ed States to that simulated by a climate model. J. Climate, 9,2233–2249.

Chen, T.-C., and M.-C. Yen, 1994: Interannual variation of the Indianmonsoon simulated by the NCAR Community Climate Model:Effect of the tropical Pacific SST. J. Climate, 7, 1403–1415.

Deshpande, J. V., A. P. Gore, and E. A. Shanubhogue, 1995: StatisticalAnalysis of Nonnormal Data. John Wiley & Sons, 240 pp.

Ebert, E., and J. Curry, 1992: A parameterization of ice cloud opticalproperties for climate models. J. Geophys. Res., 97, 3831–3836.

Espenshade, E. B., Jr., Ed.,1995: Goode’s World Atlas. 19th ed. RandMcNally, 372 pp.

Hack, J., 1994: Parameterization of moist convection in the NationalCenter for Atmospheric Research Community Climate Model(CCM2). J. Geophys. Res., 99, 5551–5568.

Holtslag, A. A. M., and B. A. Boville, 1993: Local versus nonlocalboundary-layer diffusion in a global climate model. J. Climate,6, 1825–1842.

Kendrew, W. G., 1953: The Climates of the Continents. 4th ed. Clar-endon Press, 607 pp.

Kiehl, J., J. Hack, G. Bonan, B. Boville, B. Briegleb, D. Williamson,and P. Rasch, 1996: Description of the NCAR Community Cli-mate Model (CCM3). NCAR Tech. Rep. NCAR/TN-4201STR,Boulder, CO, 152 pp.

Kirkyla, K. I., and S. Hameed 1989: Harmonic analysis of the seasonalcycle in precipitation over the United States: A comparison be-tween observations and a general circulation model. J. Climate,2, 1463–1475.

Kummerow, C., W. Barnes, T. Kozo, J. Shiue, and J. Simpson, 1998:The Tropical Rainfall Measuring Mission (TRMM) sensor pack-age. J. Atmos. Oceanic Technol., 15, 809–817.

——, and Coauthors, 2000: The status of the Tropical Rainfall Mea-suring Mission (TRMM) after two years in orbit. J. Appl. Me-teor., 39, 1965–1982.

Lau, K.-M., and H. T. Wu, 2001: Principal modes of rainfall–SSTvariability of the Asian summer monsoon: A reassessment ofthe monsoon–ENSO relationship. J. Climate, 14, 2880–2895.

Lin, X., L. D. Fowler, and D. A. Randall, 2002: Flying the TRMMSatellite in a general circulation model. J. Geophys. Res., 107,4281, doi:10.1029/2001JD000619.

Lindzen, R. S., 1981: Turbulence and stress owing to gravity waveand tidal breakdown. J. Geophys. Res., 86, 9707–9714.

Liou, K., 1992: Radiation and Cloud Processes in the Atmosphere.Oxford University Press, 487 pp.

Lorenz, E. N., 1963: Deterministic nonperiodic flow. J. Atmos. Sci.,20, 130–141.

McFarlane, N., 1987: The effect of orographically excited wave dragon the general circulation of the lower stratosphere and tropo-sphere. J. Atmos. Sci., 44, 1775–1800.

Mo, K. C., and R. W. Higgins, 1998: Tropical influences on Californiaprecipitation. J. Climate, 11, 412–430.

Montroy, D. L., 1997: Linear relation of central and eastern NorthAmerican precipitation to tropical Pacific sea surface tempera-ture anomalies. J. Climate, 10, 541–558.

Negri, A. J., and T. L. Bell, 2002: Sampling of the diurnal cycle ofprecipitation using TRMM. J. Atmos. Oceanic Technol., 19,1333–1344.

Newman, M., P. D. Sardeshmukh, and J. W. Bergman, 2000: Anassessment of the NCEP, NASA, and ECMWF reanalyses overthe tropical west Pacific warm pool. Bull. Amer. Meteor. Soc.,81, 41–48.

Pant, G. B., and K. R. Kumar, 1997: Climates of South Asia. JohnWiley & Sons, 320 pp.

PRB, cited 2003: Population Reference Bureau: World populationdata sheet. [Available online at http://www.prb.org.]

Ramanathan, V., 1976: Radiative transfer within the earth’s tropo-sphere and stratosphere: A simplified radiative-convective mod-el. J. Atmos. Sci., 33, 1330–1346.

——, and P. Downey, 1986: A nonisothermal emissivity and absorp-tivity formulation for water vapor. J. Geophys. Res., 91, 8649–8666.

Roads, J., 2003: The NCEP–NCAR, NCEP–DOE, and TRMM trop-ical atmosphere hydrologic cycles. J. Hydrometeor., 4, 826–840.

Ropelewski, C., and M. Halpert, 1987: Global and regional scaleprecipitation patterns associated with the EI Nino/Southern Os-cillation. Mon. Wea. Rev., 115, 1606–1626.

——, and ——, 1989: Precipitation patterns associated with the highindex phase of the Southern Oscillation. J. Climate, 2, 268–284.

Rouault, M., C. J. C. Reason, J. Lutjeharms, and A. Beljaars, 2003:Underestimation of latent and sensible heat fluxes above theAgulhas Current in NCEP and ECMWF analyses. J. Climate,16, 776–782.

Shin, K.-S., and G. R. North, 1988: Sampling error study for rainfallestimate by satellite using a stochastic model. J. Appl. Meteor.,27, 1218–1231.


Simpson, J., R. F. Adler, and G. R. North, 1988: A proposed TropicalRainfall Measuring Mission (TRMM) satellite. Bull. Amer. Me-teor. Soc., 69, 278–294.

Slingo, J., 1987: The development and verification of a cloud pre-diction scheme for the ECMWF model. Quart. J. Roy. Meteor.Soc., 113, 899–927.

——, 1989: A GCM parameterization for the shortwave radiativeproperties of water clouds. J. Atmos. Sci., 46, 1419–1427.

Sperber, K., and T. Palmer, 1996: Interannual tropical rainfall vari-ability in general circulation model simulations associated withthe Atmospheric Model Intercomparison Project. J. Climate, 9,2727–2750.

Taylor, K. E., D. Williamson, and F. Zwiers, 2000: The sea surfacetemperature and sea-ice concentration boundary conditions forAMIP II simulations. Tech. Rep. 60, Program for Climate ModelDiagnosis and Intercomparison, Lawrence Livermore NationalLaboratory, 25 pp.

Tracton, M. S., and E. Kalnay, 1993: Operational ensemble prediction

at the National Meteorological Center: Practical aspects. Wea.Forecasting, 8, 379–398.

Trigo, R. M., and J. P. Palutikof, 2001: Precipitation scenarios overIberia: A comparison between direct GCM output and differentdownscaling techniques. J. Climate, 14, 4422–4446.

Vogelzang, D., and A. Holtslag, 1996: Evaluation and model impactsof alternative boundary-layer height formulations. Bound.-LayerMeteor., 81, 245–269.

Wu, R., and S.-P. Xie, 2003: On equatorial Pacific surface wind chang-es around 1977: NCEP–NCAR reanalysis versus COADS ob-servations. J. Climate, 16, 167–173.

Zhang, G. J., and N. A. McFarlane, 1995: Sensitivity of climatesimulations to the parameterization of cumulus convection in theCanadian Climate Centre general circulation model. Atmos.–Ocean, 33, 407–446.

——, J. T. Kiehl, and P. J. Rasch, 1998: Response of climate simu-lation to a new convective parameterization in the National Cen-ter for Atmospheric Research Community Climate Model(CCM3). J. Climate, 11, 2097–2115.

Documents

A Comparison of Tropical Precipitation Simulated by the ...people.tamu.edu/~k-bowman/pdf/Collier et al 2004.pdf · mania), Trigo and Palutikof (2001; Iberian peninsula), and the 32-model