Modeling seasonal and interannual variability in ... - Wiley Online Library

3 downloads 0 Views 2MB Size Report
May 27, 2001 - vegetation land cover, fire counts, and smoke aerosol effects) reveals ... (normalized difference vegetation index, NDVI) from the advanced very high resolution ..... applied Fourier smoothing algorithms (FA) developed by Los.
JOURNAL OF GEOPHYSICAL

RESEARCH, VOL. 106, NO. D10, PAGES 10,423-10,446, MAY 27, 2001

Modeling seasonal and interannual variability carbon cycling for the Brazilian Amazon region

in ecosystem

Christopher Potter, • Steven Klooster, 2Claudio ReisdeCarvalho, 3 Vanessa Brooks Genovese, 4AliciaTorregrosa, 2Jennifer Dungan, 2 Matthew Bobo, 4andJoseph Coughlan • Abstract. Previousfield measurements have implied that undisturbedAmazon forestsmay representa substantial terrestrialsinkfor atmospheric carbondioxide. We investigatedthis hypothesis usinga regionalecosystem modelfor netprimaryproduction(NPP) andsoil biogeochemical cycling. Seasonalandinterannualcontrolson net ecosystem production (NEP) were studiedwith integrationof high-resolution (8-km) multiyearsatellitedatato characterizeAmazon land surfacepropertiesover time. Backgroundanalysisof temporaland spatialrelationships betweenregionalrainfallpatternsandsatelliteobservations (for vegetationlandcover,fire counts,andsmokeaerosoleffects)revealsseveralnotablepatterns in themodeldriverdata. Autocorrelation analysisfor monthlyvegetation "greenness" index (normalizeddifferencevegetationindex,NDVI) from the advancedvery highresolution radiometer(AVHRR) andmonthlyrainfall indicatesa significantlag time correlationof up to 12 months.At lag timesapproaching36 months,autocorrelation function(ACF) valuesdid notexceedthe95% confidenceintervalat locationswestof about47øW,whichis nearthe transitionzoneof seasonaltropicalforestand other(nonforest)vegetationtypes. Even at lag

timesof 12monthsor less,thelocationnearManaus(approximately 60øW)represents the farthestwesternpoint in the Amazonregionwhereseasonalityof rainfall accounts significantlyfor monthlyvariationsin forestphenology,as observedusingNDVI. Comparisons of NDVI seasonal profilesin areasof theeasternAmazonwidely affectedby fires(as observedfrom satellite)suggestthat our adjustedAVHRR-NDVI capturesyear-toyearvariationin land covergreenness with minimal interferencefrom smallfires andsmoke aerosols.Ecosystemmodelresultsusingthisnewly generatedcombinationof regional forcingdatafrom satellitesuggestthatundisturbed Amazonforestscanbe strongnet sinks for atmosphericcarbondioxide,particularlyduringwet (non E1Nifio) years. However, droughteffectsduringE1Nifio yearscanreduceNPP in primaryforestsof the eastern Amazonby 10-20%, comparedto long-termaverageestimatesof regionalproductivity.

AnnualNEP for theregionispredicted to rangefrom-0.4 Pg C yr -• (netCO2source)to 0.5 PgC yr -• (netCO2sink),with largeinterannual variabilityoverthestatesof Par/t,Maranhao, andAmazonas.As in the caseof predictedNPP, it appearsthatperiodsof relativelyhigh solarsurfaceirradiancecombinedwith severalmonthsof adequaterainfall are requiredto sustainthe forestcarbonsink for positiveyearly NEP estimates.

1. Introduction

In the Amazon basin, year-to-year variability in net ecosystem exchange(NEE) of carbonon a regionalbasisis

logging,andforestregrowth[Nepstadet al., 1999]. Both types of changecan generateuncertainties in the global carbonbalance[Tanset al., 1990; Ciais et al., 1995] andmay influencethe carbondioxide (CO2) levelsof the atmosphere,

influenced by (1) changes in thebalance betweennetprimary with importantimplications for climateforcingby so-called production (NPP) andsoilrespiration occurringasa resultof "greenhouse gases". variationsin climateand atmospheric composition, and (2) Since the 1970s, advances based on satellite remote changesin land use broughtaboutby fire, land clearing, sensinghave been made for estimatingthe total area deforestedin the Amazon [Skoleand Tucker, 1993; Stone et al., 1994; DeFries et al., 1998] and actualecosystemcarbon • NASAAmes Research Center, Moffett Field,California. 2 California State University Monterey Bay,Seaside, California. fluxes for the region [Potter et al., 1998]. However, there is still a large uncertainty surrounding the magnitude and 3 Laboratorio deEcofisiologia Vegetal, EMBRAPA-Amazonia Ocidental,Belfim, Parfi,Brazil. location of seasonaland interannual variability in CO2 flux 4 Johnson Controls WorldServices, NASAAmesOperations,betweenthe atmosphereand potentialland covertypesin the Moffett Field, California.

Amazon region as a whole. Recent estimates of NEE from tower-based

Copyright 2001by theAmericanGeophysical Union. Papernumber 2000JD900563. 0148-0227/01/2000JD900563

$09.00

flux

measurements

indicate

that uhdisturbed

(primary) Amazon forest ecosystemsmay representa large net carbonsink [Grace et al., 1995; Mahli et al., 1998]. This assumptionhas yet to be verified over relatively long time 10,423

10,424

POTTER

ET AL.:

MODELING

periods (> 5 years) and at a sufficiently large number of measurementsitesthat representwidely differingrainfall and soil conditions. Physical mechanismsused by trees to cope with interannualchangesin soil water supply may play an importantrole in sustainingthe productivityof Amazon rain forest ecosystems. In an analysis of the Amazon basin hydrologic cycle using observed rainfall, Zeng [1999] reportedthat the correlationbetween the E1 Nifio-Southern Oscillation (ENSO) index and observedprecipitation was high for the period 1979-1996. Changesin soil water storage were estimated to lag precipitation amounts by up to 3 months.

On the basisof previousglobal biospheremodelingresults

at coarseresolutions of 0.5ø to 1ø latitude/longitude [Tian et al., 1998; Potter and Klooster, 1998 and 1999a], we speculate that duringstrongE1 Nifio years, certaintropicalforestareas

arenetsources of carbon, withfluxratesof at least70 g C m-2 yr'• to theatmosphere, whereas thesameforestareasduring non E1 Nifio years can be net sinks of atmosphericCO2 of about twice that sourceflux magnitude. If this hypothesisis supportedby subsequentfield and mechanisticmodeling investigations,then the combinedevidencewould imply that assessmentsof "steady state" in carbon fluxes for moist tropicalecosystems mustbe made over at leasta decade,and possiblymuchlonger. For detailed regional investigations, the direct use of satellitesensordatato captureand integrateterrestrialsurface properties (e.g., canopy leaf area and light absorption properties)and land cover changesover large areashas been shownto meet the major requirementsfor accuratescalingof biosphere-atmosphere fluxes of water and carbon [Maisongrande et al., 1995; Goetz and Prince, 1996; Maimstrom et al., 1997]. Assuming that adjustmentsfor cloud cover, atmosphericsmoke and related aerosolscan be incorporated into the modeling of terrestrial ecosystem processesover a vast area like Amazonia, which commonly experiencesa large number of agricultural fires during the driest months, a model driver generated from the satellite

"greenness" index (called the normalized difference

AMAZON

CARBON

CYCLES

to come [LBA SciencePlanning Group, 1996]. In this paper we report on (pre-LBA field campaign) results from application of a regional model at 8-km resolution for simulation

of

seasonal

and

interannual

controls

on net

ecosystemproduction(NEP) in Amazonia,with integrationof multiyear satellite data to characterizedynamic land surface properties over space and time. Our regional model also includesgriddedsoil propertiesbasedon data specificto the Amazon region [Potter et al., 1998]. Additional model improvementsincludeuse of a validatedhigh-resolutiontime series of rainfall for the region, as well as equations for ecosystemwater balancedriven by variable surfaceradiation fluxes, topography,and deep rootingby vegetation. The use of multiyear satelliteNDVI to drive the model may uniquely capture certain patternsof large-scaleland use changeover the region, albeit limited in capacity to detect small-scale effects such as forest fires, selectivelogging, and secondary forestregrowth. Our main studyobjectiveis to determinethe temporaland spatial relationshipsamong predictedregional NEP, rainfall patterns, and satellite sensor estimates of vegetation land cover, fire counts,and smokeaerosols,using geo-rectifeddata sets for each of thesevariablesthat span at least a decadeof climate variability and land use change in Brazil. Another important objective is to use results from our simulation model to help formulate and refine hypothesesthat can be tested in tropical field studiesof temporal variability in NEP fluxes of CO2 with the atmosphere,suchas thoseplannedfor LBA.

2. Modeling Approach and Regional Drivers The NASA-CASA (Carnegie-Ames-StanfordApproach) model is a representationof major ecosystemwater, carbon, and nitrogencycles,which can be scaledup to regional and global levels [Potter and Klooster, 1997]. It includes interactionsof severalcontrolson net ecosystemproduction (NEP) of carbon:soil water balance,temperature,soil texture and fertility, and microbial activity affecting organic matter mineralization. The model is designedto simulatedaily and seasonalpatternsin carbonfixation, allocation, litterfall, and soil CO2 emissionsfrom microbial activity (Rh), as well as other biogenic trace gas emissions. Regional scaling is accomplishedby merging input data sets of NDVI from the advancedvery high resolutionradiometer(AVHRR), surface climate, radiation, vegetation,and soils with algorithmsfor moisture flow and nutrient transformation processes in terrestrial ecosystems. Use of satellite data to drive the CASA model's NEP (i.e., NPP - Rh) flux is an important feature, because (1) predicted carbon input fluxes are formulatedto be consistentwith the range of measuredrates from field studiesworldwide, and (2) actual regionalpatterns for land cover attributes, as capturedby the satellite time series, may differ substantially from potential vegetation maps. A completedescriptionof the previousregionalmodel design for Amazon ecosystemsis provided by Potter et al. [1998]. For applicationin this study,severalmodel componentsof the NASA-CASA versionfor the Amazon remainunchanged from thosedescribedby Potter et al. [1998]. For example,

vegetation index, NDVI) has the potential to provide extensive,relatively high resolution(e.g., 8-km) coverageof variable surfacepropertiesand land use patterns,currently over 10- to 15-yeartime periods. Parameters such as standing biomass, leaf area index (LAI), soil texture, water holding capacity, nutrient availability, root biomassallocation, and vegetationrooting depth have been shownto exert strong influenceson many land surface hydrologic and biogeochemical processesin Amazon ecosystems [Silver et al., 2000]. Generalized ecosystemmodels,which have been calibratedfrom a fairly small number of site-specific measurementsfor global biospheresimulations(see,for example,Melillo et al. [ 1993]; Kindermanet al, [1996]; and Tian et al. [1998]), typically are not designed for scaling up actual or dynamic effects of tropicalland usechanges,nor have they incorporatedregional data sets for soil propertiesunique to Amazon forest areas [Tomasella and Hodnett, 1998]. The long-term satellite NDVI recordcan provide a valuablemeansto overcomesome of theseinconsistencies in regionaldatasets. The Large Scale Biosphere-AtmosphereExperiment in Amazonia (LBA) is an internationalresearcheffort designed monthlyNPP} definedas net fixationof CO2by vegetation,is to augment the region-wide data set of climate, hydrology, computed on the basis of light use efficiency [Monteith, solar soils,and ecosystemmeasurements and modelingin the years 1972]. NPP is estimatedas a productof cloud-corrected

POTTER ET AL.: MODELING AMAZON CARBON CYCLES

(b) EcosystemProduction (a) Soil Moisture Balance Nutrient Mineralization and Plant Functional Types

10,425

(c) Biogenic Trace Gas Flux

FPAR PPT

PET

Soil Surface

Heat

Profile •-• R'øøt Litter I Layers •t•,Microbes [ SoilOrganic f(Temp) l Matter I

&

Water Flux

f(WFPS

f(WFPS) f(Lit q)

Grass

Shrub

Tree



CO 2•-•CHa NzO

f(Li•Mineral N/ NO

Figure1. Structure oftheNASA-CASA model,based onplantfunctional types(PFTs),forexample, grasses andwoodyplants (shrubs andtrees).(a) Soilwaterbalance, estimated asshown bytheshaded depthlevelsin soilprofilelayers (M•-M3).Fortheforestandsavanna classes, maximum rooting depthissetuniformly to 10 m, whereas fortheotherlandcoverclasses, it is setuniformly to 1 m. (b) Climatecontrols onnetprimary production (NPP),defined in equation (1), including fraction of absorbed photosynthetically activeradiation (FPAR), precipitation(PPT), and potentialevapotranspiration (PET). Controlson litter and soil C

decomposition aredefined assoiltemperature (Temp),water-filled porespace (WFPS),andnitrogen/lignin content (Lit q). (c) Biogenic emission fluxesof soiltracegases, including heterotrophic respiration (CO:), methane(CH4),andnitrous(N20) andnitric(NO) oxides.

irradiance (S),fractional intercepted photosynthetically active N mineralization fluxes was defined as an exponential radiation (FPAR)anda maximum lightuseefficiencyterm response using a Q10 (the multiplicative increase in soil (•), modified by normalized temperature (T) andmoisture (/4/) biologicalactivity for a.10øC rise in temperature),with a stressscalars(equation(1)): NPP

= S FPAR

•; T W

In the regional simulation mode, estimation of FPAR comesfrom a vegetationindex derived from AVHRR NDVI data [Sellerset al., 1994]. Our calibratedglobal value [Potter et al., 1993] of • basedon seasonalNDVI [Los et al., 1994] is

0.39 g C MJ-• PAR,whichis derivedfromcalibration to previous field estimates of NPP. The T stress term is computedwith referenceto an AVHRR derivationof optimal temperatures(ropt) for plant production[Potter et al., 1993].

The ropt settingrangesfrom near0øC in the Arctic to the middle

30s in low-latitude

deserts.

The

W stress term is

computed on the basis of the predicted monthly ratio of estimated evapotranspiration (EET) to potential evapotranspiration (PET). For the soil carbondecompositioncomponent(Figure 1), our design remains comparable to a somewhat simplified version of the CENTURY ecosystemmodel [Parton et al., 1992]. First-order equationssimulateRh loss of CO2 from decomposingplant residue(metabolicand structuralfractions of NPP) and microbesat the soil surface. Near-surface soil organicmatter (SOM) pools are presumedto vary in carbon residencetime and chemicalcomposition. Active (microbial biomassand labile substrates),slow (chemically protected), and passive(physically protected)fractionsof the SOM are represented.The effect of temperatureon litter and soil C and

value of 1.5 for surfacelitter decomposition and a value of 2.0 for soil decomposition. Regional data sets (8-km resolution) from a geographic information system (GIS) were used as model drivers and land surfaceparameterfiles. We assembleda completeset of GIS raster coveragesto serve as model-compatibleinputs, including monthly rainfall and surface air temperature, surfacesolarradiation,soil type and texture,land cover type, and satellite vegetation index (NDVI) for the country of Brazil and, in somecases,for the larger Amazon region. All original rastermaps (except solarirradiance)were griddedat 8-km spatial resolutionin an equal area projection. In terms

of single grid cell size (64 km2), this producesan improvement in spatial resolution of about 40 times,

compared to 0.5ø(---2500 km2cellsize)datadrivers[Tianet al., 1998]. The coastal boundary line file used as a base to geo-referencethe 8-km map set was taken from the Digital Chart of the World.

2.1. Satellite Vegetation Index

We obtainedmonthly compositesfor the years 1982-1991 of NDVI from the National Oceanic and Atmospheric Administration(NOAA) AVHRR, which is availablefrom the

NOAA/NASA PathfinderAVHRR Land (PAL) programat NASA GoddardSpaceFlight Center's(GSFC) Distributed ActiveArchiveCenter(DAAC). CompleteAVHRR datasets areproduced fromNOAA globalareacoverage(GAC) level

10,426

POTTER ET AL.: MODELING AMAZON CARBON CYCLES

lessthan 25% of the NDVI monthlyvaluesover the entire regionalland arearequiredmodificationwith this filtering

lB data, and consist of reflectances and brightness temperatures derived from the five-channel cross-track scanning AVHRR on board the NOAA Polar Orbiter 'afternoon'satellites(NOAA 7, 9, and 11). DAAC references by Agbu and James [ 1994] and Kidwell [ 1991] providemore informationon the derivationand potentialuseof theseNDVI products. Monthly compositedata setsare designedto removemuch of the contaminationdue to cloud cover presentin the daily AVHRR data sets[Holben, 1986]. To generatea composite data set, eight to eleven consecutive 'days of data are combined,taking the observationfor each8-km bin from the

prevalent at different times and locations in the Amazon

datewith the highestNDVI value. Only datawithin42ø of

region.Settings for thisFA correction includethreetemporal

step[Potteret al., 1998].

Although PAL composite datasetsareproduced expressly for studiesof temporaland interannualbehaviorof surface vegetation,subsequent processingis recommended if a more

completecloud-freesignalis required. Consequently, we appliedFouriersmoothing algorithms (FA) developed by Los et al. [1994] for AVHRR data sets to further remove

erroneousness NDVI signalsdue presumably to remaining cloud cover and smoke-aerosolinterference,which can be

nadir are usedin the compositeto minimize spatialdistortion and bidirectional effect biases at the edge of a scan. A Rayleigh correctionis calculatedand appliedusinga standard radiative transfer equation and methodology,which follows the work of Gordonet al. [ 1988]. Having obtained10 years of monthly compositePAL files for SouthAmerican NDVI (scaled in units (1-1000)), a lowpass spatial filter is run over the data to remove several narrowlines of anomaloushigh valueswhich are presumably a resultof compositing.This filter routinecomputesthe mean of six nearby grid cell values (located two rows and three columnsabove each cell location), and comparesthis average to the actualcell value. If the differencebetweenthe original cell value and the average of the values abovethat cell was greater than 200 NDVI units, then the value of that cell is replacedby the averageof the nearbycell values. Typically,

Jan 1980

July 1980

Jan 1981

July 1981

harmonicsand a weightedFouriertransform,in which values

which fall abovethe Fouriercurveare givenmoreweight thanvaluesbelowthe curve. Thisassumes thathigherNDVI values are more likely to be correctthan low NDVI values

which could occur during periods of cloud or smoke formation. Applicationof the FA algorithmmodifiedmean annualNDVI valuesby more than +10% of their original valuesin approximately four outof everytengridcellsin the region [Potter et al., 1998]. Adjustmentsof the NDVI for changes in surface inundation were not considered in this

analysis, although it is acknowledged thatseasonally flooded areas of the Amazon could alter light reflection over significant areasof theregion. From these adjustedmonthly PAL data sets for South

America, we appliedNASA-CASA empiricalalgorithms describedby Potter et al. [1993 and 1998] to compute

Jan 1986

Jul

1986

100.1-200 Jan 1987 200.1300

Jan 1902 J•l•• Jan 1983

3oo.-400

July 1983 500.1 - 600

•:

Jan 1984

July 1989

Jan 1990

July 1990

:-• •....

Ju[•3,•,.•_ • - 800

800.1

Jan 1985

Jan 1989

- 900

July 1985 900.1

- 1000

l

> lO0O

Figure2. Precipitation maps generated forNASA-CASA model simulations overtheAmazon region forthe period of 1980-1990. Records from50validation stations wereincluded inthefinalinterpolation procedure to minimizespatialerrors.

POTTER ET AL.'

MODEL1NG

AMAZON

CARBON CYCLES

10,427

!

!

!

!

!

i

Jan 80

Jan 82

Jan 84

Jan 86

Jan 88

Jan 90

Time

Figure 3. Root-mean-square error(RMSE; mm) for two rainfallinterpolation methods,basedon continuous monthlyrecordsfrom 50 rainfall gaugelocationsacrossthe Amazonregion and easternBrazil, selected randomlyas an error evaluationdata set. RMSE expressesthe averagesize of predictionerrorsfor the regression of monthlyobserved versusmonthlyinterpolated rainfallamountsat these50 validationstations. Kriging interpolation(solid line with squares)showsconsistently lower errorsthan the inversedistanceweightedinterpolation method(dashedlinewith circles),particularlyduringwet seasonmonths.

second-level model drivers for the fraction of intercepted photosyntheticallyactiveradiation (FPAR) and the fraction of yearly litterfall per month at each 8-km grid cell.

Amazon region and eastern Brazil were separatedfrom the remaining stationrecordsto serve as an error evaluation data set. Based on our kriging interpolation methods, average

prediction errors fallwithintherangeof 10-125mmmonth '•

dryversus wetseasons, compared toatypical range of50•

2.2. Monthly Rainfall

We generateda new set of monthly rainfall maps (19801990) for Brazil and the larger Amazon studyarea (Figure 2), based on two sourcesfor rainfall gauge stations. The main source for rainfall data from the Brazilian Legal Amazon statescomesfrom Brazil's DepartamentoNacional de Aguas e Energia Eletrica (DNAEE). Nearly 200 hydrometeorologic stations plus 400 stations of rainfall gauge locations are

1000mmmonth -• observed rainfall(Figure3). In addition, we find that the kriging interpolation method is more consistentin minimizing errors than an alternative inverse distance-weightedinterpolation method, particularly during wet periods and as the number of stations used for the interpolationchangeswith time. The region-wide averagefor

rainfalloverthenineyearsmodeled was201cmyr-• whichis

included in theDNAEEdataarchives, distributed through the within 5% of the long-term annual precipitationrate reported

NASA Earth Observing System (EOS) Amazon project [Richey et al., 1989b]. An additional set of meteorologic station data was obtained through the NOAA National Climatic

Data

Center's

Global

Historical

(GHCN) version 2 [Vose et al., 1992].

Climate

Network

On average, the

for each of six other Amazon basin rainfall data sets, all with

a spatialresolution of 0.5ø- 2.5ø[CostaandFoley,1998]. 2.3. Monthly Surface Irradiance Surface solar radiation

drivers for 1983-1991

are based on

computedsurface irradiance from the International Satellite (380 stations in 1980 to 50 stations in 1990). With the Cloud Climatology Project (ISCCP). These data sets are combined station records, we used spatial interpolation produced for the Sea-viewing Wide Field-of-view Sensor methodsbasedon kriging to generaterainfall maps on a 8-km (SeaWiFS) project through the implementationof algorithms (version resolution grid. The kriging method uses spatial distance from Bishopand Rossow[1991]. Second-generation calculations and covariance functions to calculate grid cell 2) algorithmsuse C1 data from ISCCP, which combinesdata values for each monthly rainfall map [Mathsoft, Inc., 1996]. from multiple geostationaryand polar-orbitingmeteorological The covariancefunctions are based on sphericalvariograms satellites to provide a global view of the occurrence and opticalpropertiesof clouds. Atmospheric,cloud, and surface modeledfor eachmonthly data set. Prior to the final interpolation procedure, continuous data from ISCCP are used as input with a fast schemefor monthly records from 50 rainfall gauge locations acrossthe computingclear-skyirradiancefrom the solarzenith angle, air GHCN

data set added 200 rainfall

stations over the time series

10,428

POTTER ET AL.' MODELING

AMAZON CARBON CYCLES Acre

Maranhao

Jan 84

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 84

Jan 85

Jan 86

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 84

Jan 85

Jan 86

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 84

Jan 85

Jan 86

Date

Jan 85

Jan 86

Jan 87

Jan 87

Jan 88

Jan 89

Jan 90

Jan 87

Jan 88

Jan 89

Jan 90

Jan 88

Jan 89

Jan 90

Date

Rondonia

Jan 84

Jan 90

Amazonas

Para

Jan 85

Jan 89

Date

Date

Jan 84

Jan 88

Arnapa

Mato Grosso

Jan 84

Jan 87 Date

Date

Roraima

Jan 88

Jan 89

Jan 90

Jan 91

Date

Jan 84

Jan 85

Jan 86

Jan 87

Jan 91

Date

Tocantins

Jan 84

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Date

Figure 4. Time seriesplotsof surfaceirradiancefor the nine statesof the Brazilian Legal Amazon [Instituto Brasileiro de Geografia e Estatistica(IBGE), 1991]. Seasonallydry statesare shown in the left plots, seasonallymoist statesin the right plots. Spatial mean valuesfor eachstatearea shownas the solid lines, with plusor minusonestandarddeviationasthe dashedlines.

properties, and surface reflectance. The schemethen uses Accuracyof the resultingclimatemapsis checkedby percent simple cloud properties (cloud fraction, cloud optical absolute difference analysis with comparison to original thickness,and diffuse albedo) to producesolar irradiance station values. Adjustmentsfor diurnal temperaturerange

fields. The dataaredocumented to havean accuracy of 9 W

anomaly are basedon GHCN records(1980-1991) with a

m'2ona dailybasisandlessthan4% overallbiasin the17- thin-platesplineinterpolation,griddedoriginallyto 0.5ø day meanrelativeto groundmeasurements. We regridded resolution [New et al., 2000]. Regional data for diurnal thesemonthlyfilesto 8-km spatialresolution usinga nearest temperaturerangeswere obtainedas monthly mean values neighbor algorithm. For use in this NASA-CASA Amazon from the Climate ImpactsProject(U. K. Departmentof the application, we smoothed theresultingdatasetsby assigning Environment)on behalfof the Climatic ResearchUnit (CRU), themeanof a 16 x 16 matrixaroundeachgridcell to thenew University of EastAnglia. 8-km grid cell value, to producea compatibletime series modeldriver(Figure4). Lackingsolarirradiancedatafor the 2.5. Land Cover and Land Use period 1980-1982,we usedaverageISCCP radiationdriv6rs For land cover characterization,we aggregated39 cover (1983-1991) asmodelinputsfor initialization. classesdevelopedfor SouthAmericanvegetationby Stoneet al. [1994] from analysisof 1-km AVHRR-NDVI patterns, 2.4. Monthly SurfaceAir Temperature into four generalclassesrepresented in the globalland cover Monthly meantemperaturemapsfor SouthAmericawere classificationschemeof DeFries and Townshenri[1994]. Of obtainedfrom ZedX, Inc. (Boalsburg,Pennsylvania).We the four generalland coverclassesrepresentedat 8-km grid regriddedthe originalfiles from 10-km spatialresolutionto cell resolution,broadleafforestis the mostcommon(48% of our nominal 8-km cell size using a nearest neighbor total area),followedby woodedgrasslandandsavanna(32%), algorithm.Thesemeanclimatedatasetsaregenerated based cultivatedlands(15%), and all other cover typescombined, on long-term(1961-1990)recordsfrom 98 weatherstationsin includingwetlands,river ways, and bare ground(5%). The Brazil,whicharepartof theGHCN. The spatialinterpolation vegetationmap of Stoneet al. [1994] is reportedto have an is performed by three-dimensionallinear regression. overallaccuracyof 90%, basedon groundtruthanalysisof all

POTTER

ET AL.'

MODELING

39 cover classes. It appears, nevertheless, that river and wetland areascouldbe underestimatedby Stoneet al. [ 1994], suggesting that improvements in certain categories are

AMAZON

CYCLES

10,429

Priestlyand Taylor[1972]methoddescribed by Campbell [ 1977]andBonan[ 1989].

PET = a (Ta + b) Rs

needed.

We appliedthe sameNPP algorithmcoefficientsdeveloped from the global version of the NASA-CASA model [Potter and Klooster, 1997] to the relatively undisturbed forests or savannasand to their respective"cleared or degraded"cover categories [Stone et al., 1994]. This means that any differencesin model resultsreportedfor different forest cover types (e.g., predominantlymoist evergreenversusseasonally deciduous forests; [Stone et al., 1994]) would not be attributableto NASA-CASA's internal model settingsin the four generalland cover groupsdefined above, but insteadto observed patterns in NDVI, climate, or soils inputs to the model. Although we have used a single land cover map for the multiyear simulationsreportedin this paper, large-scale changesattributableto major land use change(deforestation, pasture conversion,etc.) should be detected as significant alterationeffects on regionalNDVI values at 8-km resolution, thereby capturing certain land cover changesnot otherwise included in previous Amazon modeling studies [Tian et al., 1998]. As new regional land cover maps are developedover time for the Legal Amazon,theywill be addedto our dynamic modelingframework. The severalmodel parametersthat are assignedaccording to the four generalland covertypesincluderatio of leaf litter nitrogen:lignin content [Potter and Klooster, 1997], dec'omposition rates of soil carbonin cultivated soils, and vegetationrooting depth. For the forest and savannaclasses, maximum rooting depthis set uniformly to 10 m, whereasin the other three land cover classes,it is set uniformly to 1 m. Extraction of soil moisturefrom different soil layers to meet total computed PET demandscan be varied in the model accordingto datacollectedin localfield experiments.

CARBON

(2)

wherePETispotential evapotranspiration (calcm-• d-q),Ta is mean air temperature(øC), Rs is mean surface solar

radiation (calcm'• d-•),anda andb areempirical constants (set as functions of saturationvapor pressure) derived by densenand Haise [1963] and densen [1973]. Conversionof

PETto unitsof cmd-• ismadeby dividing by thelatentheat of vaporization. Estimated evapotranspiration(EET) flux is calculatedby comparing daily PET to the multilayer model estimate for daily soil moisture content (Figure 1). The soil profile is treated as a series of four layers: ponded surface water, surfaceorganicmatter, topsoil(0.3 m), and subsoilto rooting depth (1-10 m). These layers can differ in soil texture, moisture-holding capacity, and carbon cycling dynamics. Water balance in the soil is modeled as the difference

between

precipitation(PPT) or volumetricpercolationinputs,and PET and drainageoutputsfor each layer. All moistureinputs and outputs are assumedto progressfrom the surface layer downward. Inputsfrom rainfall canrechargethe soil layersto field capacity(FC). Excesswater percolatesthroughto lower layersand may eventuallyleave the systemasrunoff to rivers. Our simulation of ponded water dynamics on soil surfaces does not yet consider large-scale hydrologic processesin flooded wetland

areas.

3.2 Soil Moisture Holding Capacity and Retention Curves Moisture

retention

curves

for Amazon

soils have

been

derived by Tomasella and Hodnett [1998]. These water retentioncurvesare designedto reflect the hybrid characterof Amazon oxisols,which may act like sandsin terms of water movement at low tensions,while holding water like clays at higher tensions. We used the estimatedBrooks and Corey 2.6. Soil Texture [1964] parameters from these moisture retention curves to Soil attributemapsfor Brazil were createdby interpolation modify our model parametersfor a relative drying rate (RDR) of soilprofiledataavailablefor morethan 1000 Amazonsoil variable, describedby Potter et al. [1993]. A resultingfamily pits [Potter et al., 1998]. The procedureproduceda soil of Tomasellaand Hodnett [1998] logistic drying curvesfor an

attributemap for textureclasses, whichwere assignedon the basis of estimated clay content [Food and Agriculture

Amazon

soil RDR

scalar is based on a transformation

of the

relationship between soil water potential and volumetric

Organization,1971). We considerthesemaps to be most moisture content (equation (3)), such as the one reported by

accurate westward of 42øW and northward of 20øS.

3. Refinements

for Amazon

Saxtonet al. [1986].

Simulations

RDR= (1+a)/ (1+ a ©b)

(3)

In order to more accuratelyrepresentclimate controlsand where a and b are soil texture-dependent empirical soil processes for Amazonecosystemcarboncycling, several coefficients derived from Brooks and Corey [1964] modifications are introduced in this study for the Amazon version of the NASA-CASA model describedby Potter et al. [1998]. These changesincluderefinementof water balance Table 1. Soil MoistureHoldingCapacityfor AmazonSoils equations,and moistureholding and retention capacity for Derivedby TomasellaandHodnett[1998]for FAO Texture Amazon

soils.

Classes

3.1 Water BalanceEquations In the previous version of the model, we used

Thornthwaite's [1948] methodto estimatemonthlyPET from monthly mean surface air temperature. Thornthwaite's empirical forn•ulation may, however, underestimate PET when solar radiation is an important driver of latent heat fluxes[Rosenburget al., 1983]. Hence,the PET algorithms in this model are based on a modified formulation

of the

Class

% Clay WPa

FCb

PAWc

Coarse Coarse-medium Medium Medium-fine Fine

9 20 30 48 67

0.190 0.279 0.319 0.421 0.440

0.127 0.155 0.151 0.145 0.112

0.064 0.124 0.168 0.276 0.328

aWP iswiltingpointin unitsof cm2cm-2bedvolume bFCisfieldcapacity in unitsof cm2cm-2bedvolume

cPAW is plantavailable waterat FC in unitsof cm2 cm'2bed volume

10,430

POTTER

ET AL.:

MODELING

AMAZON

CARBON

CYCLES

Table 2. Allocationand ResidenceTime Paratnetersfor Major VegetationTypes,Following Global Cover ClassesDefined by DeFries and Townshend[ 1994] VegetationType

• Leaf

• Root • Wood

'r Leaf

'r Root

'r Wood

Annualgrasslandandcrops

0.45

0.55

--

1.5

5.0

--

Mixed deciduous forest

0.30

0.25

0.45

1.0

3.0

40

Desertandbare ground

0.25

0.25

0.50

1.5

3.0

50

Semi-arid shrub land

0.25

0.25

0.50

1.5

3.0

50

Savannaandwoodedgrassland

0.30

0.25

0.45

1.0

5.0

25

Tropicalevergreen rainforest

0.25

0.25

0.50

1.5

2.0

25

Here, ct is the proportionalallocationconstantof plant tissuepools,and 'r is the residencetime (in years)of carbonin plant tissuepools. Sourcesfor informationon parametersettingsinclude Cannel [1982], Abet and Melillo [1991], Running and Gower [1991], Lusk et al. [1997], and Terborghet al. [ 1997].

parametersand O is the predictedvolumetricmoisturecontent increasingroot biomassfor the acquisitionof soil nutrients [Wilson and Tilman, 1991]. On medium-to-highfertility (m3waterm-3bedvolume) of thesoillayer. Soil moistureholding capacityof Amazon soils was also soils,a similaradjustmentis favoredthat allocatesincreasing computed from the equations of Tomasella and Hodnett stem and leaf biomassfor light harvestingfunctionsin the [1998] for five texture classes[FAO, 1971]. Field capacity canopy[Gleesonand Tilman,1990;Lusket al., 1997]. was estimatedat matric potentialof- 10 kPa, whereaswilting point (WP) was estimatedat matric potential of-1500 kPa 3.4. Model Initialization (Table 1). The differenceof FC-WP is usedto assignplant Our ecosystemmodel must be initialized to obtain the availablewater (PAW) to textureclassesin the regionalmap. beginningbalancebetweenpredictedNPP and soil carbon 3.3 Biomass Allocation and Soil FertiliW

Carbonand nitrogenallocationamongleaf, fine root, and wood

tissues from

NPP

is defined

in terms of fractional

allocationconstantsof plant tissuepools (o0, and the mean residencetime ('r, in years) of carbon in the standingplant tissuepools (Table 2) [Potter and Klooster, 1999a]. Soil fertility effects are included in this model version to adjust allocationconstantsfor generalizednutrient limitations. We classifiedsoil typesin the Soil Map of Brazil [Minist•rio das Minas e Energia (MME), 1981] accordingto three relative levels of soil fertility (low, medium,and high). The resulting soil fertility classmap is shownin Figure 5. On low-fertility soils, an adjustment (+10%) is favored that allocates

poolsat all 8-km gridlocationsin theregion.We used3-year averagedclimateforcingconditions for representation of the period1980- 1982,withtheexception of solarirradiance data (described above), to initialize all biomass and soil carbon/nitrogen pools for monthly simulation results generated over a subsequent9-year period. Conditions of near steady state carbon pools can be reached following the equivalent of 1200 monthly time steps in the initialization procedure.

4. Analysis ofRegional ModelDrivers Exploratory analysis of spatial model drivers is a useful meansto understandthe inherenttrendsand possiblerelations among regional satellite observations,climate, and other land surface

variables

over

the

1980-1991

time

series

used

as

drivers in this modeling study of Amazon water and carbon fluxes. Results from a series of data summary and analysis approachesare presentedin this sectionto bettervisualize and compare the time-space patterns of NDVI, rainfall, fire counts, and smoke aerosol index maps that can influence seasonal to interannual forcing functions in a model like NASA-CASA. 4.1

Trends

in NDVI

We first examined patterns in the regional 8-km NDVI time seriesby averagingall pixel valuesfrom within eachof the areascoveredby the nine statesof the Brazilian Legal Amazon (as defined by Instituto Brasileiro de Geografia e Estatistica(IBGE) [ 1991]). NDVI time seriesplots(Figure6) for the raw (unprocessed)PAL data and our FA-corrected NDVI show the regular seasonalcycle of satellite-observed greenness, which generallypeaksin the monthsof April-May, near the end of the Amazon wet season, and reaches its

Figure 5. Soil fertility classificationbasedon the soilsmap of Brazil [MinistbriodasMinas e Energia(MME), 1981].

minimumvalue in October-November,generallynear the end of the Atnazondry season.The potentialeffectsof unusually dry seasonson vegetationcover during 1983 and 1987 are readily observedin the NDVI plots for all five southernand easternAmazon states(Figure 6, left plots).

POTTER ET AL.: MODELING

AMAZON

CARBON CYCLES

Maranhao

Jan 82

Jan 83

Jan 84

Jan 85

Jan 86

Jan 87

10,431 Acre

Jan 88

Jan 89

Jan 90

Jan 91

Jan 82

Jan 83

Jan 84

Jan 85

Jan 86

Time

Mato Grosso

Jan 82

Jan 83

Jan 84

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 82

Jan 83

Jan 84

Jan 85

Jan 86

Jan 84

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 82

Jan 83

Jan 84

Jan 85

Jan 84

Jan 85

Jan 86

Jan 87

Jan 86

Jan 91

Jan 88

Jan 89

Jan 90

Jan 91

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Jan 88

Jan 89

Jan 90

Jan 91

T•me

Rondonia

Jan 83

Jan 90

Amazonas

Time

Jan 82

Jan 89

Time

Para

Jan 83

Jan 88

Amapa

Time

Jan 82

Jan 87 Time

Jan 87

Roraima

Jan 88

Jan 89

Jan 90

Jan 91

Jan 82

Jan 83

Jan 84

Jan 85

Time

Jan 86

Jan 87 T•me

Tocantins

Mean Raw NDVI (dashed), FA-NDVI (dotted),and FAS-NDVI (sohd)for the states of the Legal Amazon

Jan 82

Jan 83

Jan 84

Jan 85

Jan 86

Jan 87

Jan 88

Jan 89

Jan 90

Jan 91

Time

Figure6. Timeseries plotsof monthlyNDVI for theninestates of theBrazilianLegalAmazon.Meanraw NDVI is shown bydashed lines,FA-NDVIis shown bydotted lines,andFAS-NDVIis shown by solidlines.

Seasonally drystates areshown intheleftplots, seasonally moist states intherightplots.

During the wet seasonmonths of all years, the difference Parfi,Rond6nia,and Tocantins),comparedto the wetterstates between

raw NDV!

values

and the FA-corrected

NDVI

is

relativelysmall, typically 5-10%. However, duringthe dry seasons of 1983, 1987, and 1988, notablylow valuesobserved for the raw NDVI are correctedupward substantiallyusing the FA processing,as seen, for example in the time series

(Acre, Amapa, Amazonas,and Roraima). In most years, spatialvariability appearsto be substantialover the statesof Maranhao,Mato Grosso,Parfi,Roraima,and Tocantins(data valuesnot shown),whereasthe statesof Acre, Amazonas,and

Rond6niashowmuchlower spatialvariabilitythroughoutthe time series,possiblydueto few effectsof periodicdrought. In the last threeyearsof the FA-NDVI time series(19891991), we note the possible artifacts of AVHRR sensor NDVI. This correctionpresumablyreducesthe contaminating changeover(from NOAA 9 to NOAA 11 in November1988), effects of smoke-aerosols or residual cloud cover on the andeffectsof drift in equatorialcrossingtimes(Figure6). An actual NDVI signal, to generate a more smoothedseasonal inconsistency in sensorcalibrationmay accountpartiallyfor plant phenology. However, it is not possibleto state with the trend upward in NDVI over about the last 36 months plots for the states of Mato Grosso, Parfi, Rond6nia, Tocantins, and Acre. Our maximum dry season FA correctionis typically between50% and 90% relative to raw

certaintyat this point in the analysishow much of the FA is

acrossthe entireregion. Calibrationcoefficientsto correctfor

correcting for smokeor cloud"noise"effects,versusaltering sensordifferences in NDVI (from AVHRR channels1 and2) the actual signal changesin NDVI, which could result from newly burned vegetation over a large portion of a 8-km resolutionpixel. This issueis addressed morethoroughlyin a subsequent section. There are noteworthy differences in the variability of NDVI amongstatesof the Legal Amazon. For example,the seasonalamplitude of NDVI appearshigher in the drier southern-easternAmazon states (Maranhao, Mato Grosso,

have been derived by Los [1998] based on time-invariant desert targetsfor a surface cover standard. Thus we used these monthly coefficients to correct for the NOAA sensor calibration

differences

over the entire NDVI

time series.

Gutman [1999] describes a similar sensor calibration

correctionfor the globalvegetationindex(GVI) dataset,plus an additionalpost-1988solarzenithangle(SZA) correction, calibratedfrom NOAA 9 as necessaryto reducethe influence

10,432

POTTER ET AL.' MODELING AMAZON CARBON CYCLES

of increasingSZA drift over the NDVI time series. Hence an additional adjustmentusing PAL-reported SZA values was

amount lessthan4.5 cmmonth '• [Potteret al., 1998],was consistentlythe longestduring 1983, typically with a duration

appliedto eachdatein thisNDVI time series,basedonthe of at least 3 dry months. Years of highest total rainfall in algorithmdescribedby Los et al. [ 1994]. The resulting AVHRR time series with SZA correction, which we call FAS-NDVI, shows maximum upward adjustmentsrelative to FA-NDVI of the order of 10-15% during the monthsof May through July in 1984, 1987, and 1988. As anticipated, the FAS-NDVI data set makes a correctionfor the drifts upward in NDVI over the 9-year time seriesby increasingseasonallyhigh NDVI valuesearlier (pre1989) in the sequenceto more closely resembleseasonally high NDVI valuespost-1988. 4.2. NDVI-

Rainfall Relationships

Seasonalchangesin vegetation canopy cover have been shown to be stronglyinfluencedby variation in precipitation patterns acrossthe Amazon region [Nepstad et al., 1994, 1999; Jipp et al., 1998]. Our 9-year time series of interpolatedmonthly rainfall totals show that 1983 and 1987 were relatively dry years in the states of Maranhao, Mate

most Legal Amazon states were 1985 and 1989, with the wettestperiodsoccurringduringthe monthsof January-May. To examinecovaryingpatternsof NDVI and rainfall more closely within and acrosspolitical (i.e., state) boundaries, extensiveregional transectscan be used to representmajor gradients in climate, coincident with varying ecosystem processesand land uses. For the following analysis we examined two ecoclimatic transects (ECT) in the region, which extendbeyondthe farthest west-eastborder points of the Brazilian Legal Amazon states(Figure 8). The selected ECTs pass through the approximate locations of several establishedfield study sites in Brazil (after LBA Science Planning Group [1996]). ECT 1 passesthoughthe 8-km cell o

locationsthat includethe coordinates at 2ø57'S, 59ø57'Wfor

the Institute Nacional de Pesquisasda Amazonia (INPA) tower site,Manaus,Amazonasand at 3ø53'S,54ø55'Wfor the TapajosNational Forest, Santarbm,Parfl. For comparison,a more southernECT 2 passesthough the 8-km cell locations that include the coordinatesat 10ø05'S, 6ø55'W for the Grosso, Parfi, Rond6nia, and Tocantins, in contrast to the Reserva Jaru, Rond6nia and at 15ø33'S,47ø36'W for the consistently wetter states Acre, Amapa, Amazonas, and ReservaEcologicaAguasEmendadas, Brasilia. Roraima (Figure 7). The average length of the dry season, Analysis of the FAS-NDVI and monthly rainfall (1982computedas the numberof monthsduring a year with rainfall 1990) totals acrossthe two transectsimplies that there are

Acre

Maranhao

Jan 80

Jan 80

Jan 82

Jan 82

Jan 84

Jan 86

Jan 88

Jan 90

Jan 80

Jan 82

Jan 84

Time

Mate G resso

Amapa

Jan 84

Jan 86

Jan 88

Jan 90

Jan 80

Jan 82

Jan 84

Jan 84

Jan 86

Jan 88

Jan 90

Jan 80

Jan 82

Jan 84

Jan 84

Jan 86

Jan 90

Jan 86

Jan 88

Jan 90

Jan 88

Jan 90

Roraima

Rondonia

Jan 82

Jan 88

T•me

Time

Jan 80

Jan 90

Amazonas

Para

Jan 82

Jan 86

Jan 88

Time

Time

Jan 80

Jan 86

Time

Jan 88

Jan 90

Jan 80

Jan 82

Jan 84

Jan 86 Time

Time

Tocantins

Jan 80

Jan 82

Jan 84

Jan 86

Jan 88

Jan 90

Time

Figure 7. Time seriesplots of rainfall (mm) for the nine statesof the BrazilianLegal Amazon.Seasonally dry statesare'shownin the left plots,seasonally moiststatesin therightplots. Spfitialmeanvaluesfor each stateareashownasthe solidlines,with plusor minusonestandarddeviationasthe dashedlines.

POTTER ET AL.: MODELING AMAZON CARBON CYCLES -80

8

: :

-78 -76



:

-74

-72

-70

-68

-66

-64

-62

-60

-58

-56 -54

-52

i... ::.....•.....:.....:•...i...::... •.....i .....i...• ......

-50 -48

-46

-.44 -42

-40

-38

10,433 -36

-34

•......i .....• ...... i...i... i...........•.....

i

:

:

:

.

:

E'dt'1 ' :

:

:

-18

,

-22

i

.... : .... , .. : ,

:

ß

,

:

:

,

.......

.

.

.

i

.:..

.:

:.......

i....

Figure 8. Ecoclimatictransects(ECT) acrossthe studyregion [after LBA SciencePlanning Group, 1996]. Site abbreviationsare MSN for the INPA tower site, Manaus, Amazonas,STM for the TapajosNational Forest, Santar6m,Par/t,RND for the ReservaJam, Rond6nia,and BSL for the ReservaEcologicaAguas Emendadas, Brasilia.

relateddeclinesin both vegetationgreennessand rainfall in 1982, 1983, and againin 1987. There also appearsto be a correspondingincreasein FAS-NDVI during or shortly followingunusuallyhigh rainfallperiodsof 1985, 1988, and 1989. Locationsfor a more detailedstatisticalanalysisof satellitegreenness-rainfall relationships were selectedalong the transects by sub-sampling 8-km pointsat 1.25ø longitude increments. Autocorrelationanalysis[Mathsoft,Inc., 1996] was performedfor monthlyFAS-NDVI laggedby monthly rainfall amountsat yearly time intervals for each selected analysispoint. The longestlag interval used (36 months)

9d-9f), which is a zone dominatedby savannavegetation types[Stoneet al., 1994]. Theseresultsimply that effectsof periodicdroughtson patternsin NDVI for seasonallydry Amazonforestsare readilydetectable up to about2.5 years later in the 1980s time seriesrecord. These temporal autocorrelation results are consistent with field studies in the

eastern Amazon that have linked seasonal changes in vegetationcanopycoverto precipitation andplant-available soilwaterconditions [Nepstad et al., 1994;Yippet al., 1998].

represents one third of the available time series data set 4.3. NDVI - Fire Count- SmokeAerosolRelationships The FA corrections for raw AVHRR-NDVI values are representedin the transects. ResultsacrossECT 1 show that a significantcorrelation intendedto remove the potentialcontaminatingeffects of

exists between monthly FAS-NDVI and monthly rainfall smoke aerosols and residual cloud cover on the actual NDVI amountwithin the LegalAmazonregionat pointslocatedeast signalthatcouldoccurin the atmosphere abovethe vegetated of about55øW,at lagtimesup to about24 months(Figures land surface. In the absenceof large-scaledeforestation 9a-9c). At lag timesof 12 monthsor less,the approximate lossesof plant cover,we have assumedthat the actualNDVI

locationof theManausforestsite(near60øW)represents the signalwill follow a fairly smoothseasonal phenologywhen farthestwesternpointin theAmazonregionwhereseasonality viewed as a monthly plot of FA-NDVI values [Los et al., of rainfall accountssignificantlyfor monthly variationsin 1994]. Typical dry seasondecreasesin NDVI have been forestphenology, asobserved usingFAS-NDVI. At lag times documented in fieldstudies at between10%and30% [Yippet of greater than 30 months across ECT 1, autocorrelation al., 1998]owingto progressive leaf shedding with decreasing function (ACF) values did not exceed the 95% confidence interval(CI) at pointslocatedwest of about47øW, which is near the transitionzone of seasonaltropical forestand other (nonforest)vegetationtypes[Stoneet al., 1994]. For ECT 2, ACF values were not significant (95% CI) at lag times approaching 36 monthsat pointslocatedwest of about53øW,

seasonalrainfall. It is well known,however,that fires in the Amazonregioncanrapidlyconsume largeareas(thousands of

squarekilometers) of forestandpasturelands[Nepstadet al., 1999],andcouldpotentiallygeneratetransientlow signalsin AVHRR-NDVI for 1 or 2 monthsduring and following burning.Hencethereis a needto evaluatethepossibilitythat but remainedsignificant between53øWand 42øW(Figures the empiricalFA corrections we make for tropicalregion

10,434

POTTER

ET AL.'

MODELING

AMAZON

CARBON

CYCLES

a. Lag = 2 months

I

[

I

I

-52

I

I

-48

'1

I

-44

I'

I

-40

I

-36

Longitude

b. Lag = 26 months

i

-80

r

i

-76

i

i

i

-72

i

-68

i

i

-64

i

i

i

-60

i

-56

i

i

-52

i

i

-48

i



-44

'

i

i

-40

i

i

-36

Longitude

Figure 9. Autocorrelationfunction(ACF) resultsfor monthlyFAS-NDVI, laggedby monthlyrainfall amountover the period 1982-1990.(a-c) Lag times for locationsacrossECT 1, and (d-f) lag times for locationsacrossECT 2, as shownin Figure 8. Dashedlines delineatethe 95% confidencelevel for significance of ACF valuesaboveandbelowtheseline levels.

NDVI

could confuse smoke aerosols or small-scale

cloud

cover interference effects with actual fire impacts on vegetation greenness. In the case of cells containinglarge fires, FA-NDVI values might be raised artificially during months of burning under the false assumptionof a smooth seasonalphenologyand low temporalvariability in leaf area

index(LAI) for tropicalvegetation canopies.

One possibleapproach,althoughnot oneattemptedfor this study,for evaluationof this confusionof noise(for example, smoke aerosols) and NDVI signal (fire depression of vegetationgreenness) is to estimateatmospheric opticaldepth and particulatepropertiesof aerosolsas inputsto a radiative transfer model, in an attempt to isolate and remove smoke aerosol interference from NDVI in a direct quantitative

POTTER ET AL.: MODELING AMAZON CARBON CYCLES

10,435

c. Lag = 36 months

-80

-76

-72

-68

-64

~60

-56

~52

-48

-44

-40

Longitude

d. Lag = 2 months

/'\ /

e•e•e

e-©

.

ß

i

,

-65

-60

-55

-50

i

-45

,

-40

'!

'

-35

Longitude

Figure 9. (continued) fashion. For aerosol correctionsover land, this method is

basedon the observation thatwatervaporandaerosolshavea much greaterimpact on AVHRR channel1 (0.5-0.68 gm; visible)reflectance(p) thanon channel2 (0.73-1.1 gm; nearIR) reflectance. Many land cover types, such as dense vegetationand somesoils,are darkin the blue (0.4-0.48 gm) andthered (0.60-0.65gm) visiblewavelengths.Thereforeit is appropriateto use the darkestpixels in the image to estimatethe aerosolopticalthicknessand its effect on NDVI

from AVHRR dataon a regionalscale,sincein orderto use a radiative transfermodel in this fashion,the surfacereflectance of these dark pixels must be estimated within a small

uncertainty(Ap = +0.005 to +0.01; Kaufmanet al. [1997]), to specifically accountfor all effects of atmosphericwater vapor. Moreover, the vegetated surface reflectance from channel

1 must be assumed to be constant over time for

multitemporalapplicationsof a direct aerosolcorrectionwith AVHRR. This assumptioncannotbe conclusivelyverified at thesewavelengths. Over the land, however, there is no for areasof the Amazonwhererainfallis highly seasonaland well-validatedmethodto accuratelyretrieveaerosolcontent forest fire effectsare prevalent. Thus it appearsthat there

10,436

POTTERET AL.' MODELFNGAMAZON CARBONCYCLES

e. Lag = 26 months

-80

-75

-70

-65

-60

-55

-50

-45

-40

-50

-45

-40

Longitude

f. Lag = 36 months

ß

e... // ß

f

-80

,

-75

-70

-65

-60

-55

Longitude

Figure 9. (continued) may not be sufficient spectralinformationin the AVHRR measurementdatabase,nor from ground-based measurements of vegetationoverthe vastAmazonregion,to makea reliable quantitativeestimateof aerosolinterferenceon the AVHRRNDVI signalover time. Our more qualitativeanalysisof the FA-NDVI correction

over the Amazonregionis basedinsteadon comparison of monthlyNDVI with coincidentsatellitedata setsfor AVHRR

datato evaluatethe accuracyof fire detectionandburnedarea estimatesin Brazil, and found that all AVHRR-detected fires

had corresponding Landsatthematicmapper(TM) fire scars, but thatfire sizescouldbe overestimated by 43%, on average. More recently,Kaufmanet al. [1998] reportedthat the typical sizeof thehottestareasof firesin Brazil is usuallysmall(e.g.,

0.005km2),witha fireareadistribution thatpeaks between 0.3 and 1 km2. Theseinvestigators conclude that1-km

fire counts,plussmokeaerosoldetection fromozone-mappingresolutionAVHRR analysiscan detectonly about25% of all satellitesoverthe sametime periods.As background for this firesburningin the Amazonregion,althoughthesefires may analysis,Pereira et al. [1991], for example,usedAVHRR include60-85% of the forestbiomassbeingburned.

POTTER ET AL.: MODELING

Therefore, while it may be unreasonableto consider AVHRR fire countsas a reliable sourcefor estimatingthe precise land area burned by fires in the Amazon, for the purposesof our anaylsis, 1-km AVHRR fire count data are intended mainly to detect the probability of severe fire impactson the NDVI signalwithin each 8-km pixel of the regionalcover image. We anticipatethat the higherthe count

AMAZON

CARBON

CYCLES

10,437

performedbefore applyingfire detectionalgorithms. This is accomplished using the added channel 4 temperature to eliminate false detection from cool clouds that are highly reflective in the 3.8-gm band. It is acknowledgedthen that the cloud-masking step leaves some actual ground fires undetected in the GFP data sets. On the other hand, wherever

the GFP indicates high fire counts, these pixels should

of firesfrom 1-kmAVHRR data,the greatertheprobability accuratelyrepresentcloud-freemeasurementsof surfacefire that anomalouslylow monthlyvaluesin the raw NDVI time seriescould be associatedwith extensivefire impactson actualvegetationgreenness, ratherthanappearingmerely as artifactsof atmospheric interference by smoke.Thispotential fire effecterror generatedby automaticFA correctionwould be most importantin areaswhere little or no smokeaerosol

activity. Therefore to compare AVHRR fire counts and the

FAS-NDVI data values,we first aggregatedthe 1-km GFP

firecountdatato theresolution of our8-kmAmazonregional grid. All firesdetected in thedailyGFPrecordduring19921993 were countedover the 15 daysprecedingthe date of recordedNDVI valuesfor eachmonth,in orderto capture

can be detected using satellite-borneozone mapping recentfire activityat eachpixel locationl Smoke from biomassburning and deforestationin the spectrometers [Hsu et al., 1996]. On the other hand, if there UV-absorbingtropospheric aerosols, which is not a strongrelationshipbetweenAVHRR fire countsand Amazonproduces the difference between raw NDVI and FA-corrected NDVI on can be measuredusing NASA total ozone mapping a regionallevel,thenourFourieradjustment (FA) for possible spectrometer (TOMS) instruments, from Nimbus 7 and smoke and cloud effects on NDVI should be making Meteor 3 usingmeasured340- and 380-nm radiances,and justifiableimprovements in thesurfacegreenness signal.This fromADEOSandEarthProbeTOMS using331- and360-nm channels [Hsu et al., 1996]. Daily globalmaps latterimplicationwouldbe thatwhilerelativelysmallfiresare wavelength widespread in the Amazonbasin,perhapsnumberingtensof over land and water have been generatedfor TOMS smoke thousandsper year, they are not extensiveenough to aerosolindex(SAI) values. The daily TOMS SAI is available at 1øresolution from1979to thepresent.Somelimitations of significantlyaltervegetationgreenness coverbasedon NDVI seasonal profiles [Ehrlich et al., 1997], nor are they large the SAI are that no absorbingaerosolscan detectedcloseto enoughto be detectedindependently of a possiblesmokeor the ground(below 1.5 km), and sensitivityis reducedunder cloud dampingeffect on raw NDVI values,which have been cloudyconditions.All firesdetectedin the daily GFP record aggregatedto coarser(8-km) resolution. during 1992-1993were countedfor the same 15 days The most readily availabledata set for our analysis,the previous to the recorded SAI and NDVI values for each International Geosphere-BiosphereProgramme (IGBP) month[Prinset al., 1998]. Sincesmoke"haze"createdby numerous smallfiresin the Global Fire Product(GFP), providesa consistentcount of daily fire eventsat a global scale[Stro_ppiana et al., 2000; Amazon does not generally rise to above 1.5 km altitude Dwyer et al., 1998; Justiceand Dowry, 1993]. At this time, [Reidet al., 1998],we do not expectthat SAI valueswill be the GFP productis availableonly for 1992- 1993,a period elevatedin the areaswhere, for instance,AVHRR detects10 which is fortunately also coveredby PAL data for South or fewer1-kmfire countsper 8-kmgridcell (i.e., covering American NDVI. The GFP fire detection method is based a areassmallerthanonefifth of an entire8-kmcell)for any 15selectionof 1-km pixels that couldpotentiallycontainfires, day GFP record. With this multiple satellite index regionaloverlaymaps(Figure10) canbe usedto with a confirmationprocedure for the fire pixel classification comparison, by comparison of the pixel with its immediategeographic identify small areaswhere the differencebetweenraw NDVI neighborhood [Flasseand Ceccato,1996]. The spectralbasis and FAS-NDVI during Septemberand October 1992 is for theAVHRR fire detectionalgorithmis channel3 (3.55 - relativelyhigh(> 50 NDVI units)andtheSAI is alsolow (< (> 10 counts 3.93gm;mid-IR),whichis locatedneartheoptimumfor high 1.5),butAVHRR fire countsarefairlynumerous radiative emittance from typical vegetationfires [e.g., per 15-dayGFP record). We find that all suchpixelsare in theextreme eastern andsouthern areasof theLegal Langaasand Muirhead, 1988]. It is possibleto detectfires located burning over areas much smaller than the nominal 1-km Amazon. Together,theseidentifiedpixelsconstituteonly AVHRR pixel size, becausethe total radiant emittancefrom a about2% of the regional fire pixel coverage. We can fire is disproportionallygreater than that from a cooler therefore marktheseassmallareasof highest uncertainty for applyinga FA correctionto the raw NDVI datasets. Overall, background [Kennedyet al., 1994]. There are, nevertheless, severalnoteworthyinterference however,we can concludethat small-scalefire activity problems with the AVHRR mid-IR signal that must be detectedfrom the satelliteAVHRR in the Legal Amazon, from overcomefor accuratefire detection, including surface specificallythat which can be detectedindependently reflectionof solarradiationin the 3.75-gm band,transmission possiblesmokeaerosolinterference,doesnot generatea depression effecton regionalNDVI valuesat 8-km effectsof atmospheric watervapor,subresolution clouds,and significant viewinggeometry(i.e., Sunglinteffect)of the AVHRR sensor

resolution.

[Gr•goire,1993]. Thereforeto distinguish betweensignal increasescaused by fires and those caused by warm 5. Carbon and Water Modeling Results (nonburning) surface reflection, the difference between

AVHRR channel 3 and channel 4 (10.3 - 11.3 gm) 5.1. Site Level Comparisons temperatures is used. Surfacefireswill generatemuchhigher To compare a range of model responsesfor surface emittance values for channel 3 than for channel 4. To remove radiationabsorption, waterbalance,andnet ecosystem fluxes most effects of cloud reflectances,maskingis normally of CO2, we selected locations of four transect-based LBA

10,438

POTTER ET AL.: MODELING

AMAZON

CARBON CYCLES

FA-correctedFire Counts(> 10) w/out Aerosols

All Fire Counts

Sept. 1992

22,018

385 Oct. 1992

20,102

368

Figure 10. Fire countsfrom the IGBP Global Fire Product(GFP) duringSeptemberand October1992, before(left maps)andafter(rightmaps)maskingby 8-km pixelsfor whichthe differencebetweenraw NDVI andFAS-NDVI is relativelyhigh(> 50 NDVI units),TOMS smokeaerosolindex(SAI) is low (< 1.5),and 1km AVHRR fire countsarenumerous (> 10 countsper 15-dayGFPrecord).

studysites,the INPA tower site, Manaus,Amazonas(MNS), Tapajos National Forest, Santarfim,Parfi (STM), Reserva Jaru, Rond6nia (RND), and the Reserva Ecologica Aguas Emendadas,Brasilia (BSL). Site characteristicsfor these single-pointresultsare summarizedin Table 3. The MNS forestsite typicallyreceivesthe highestrainfall annuallyand

seento increaseby lessthan 10% duringthesesameyears,a patternwhich is associated with only slightlyaboveaverage rainfall amounts,as indicatedby an averageanomalyof +1 cm observedover all of the state of Rond6nia during the combinedyearsof 1988 and 1989. The strongestdrought year in the modeledtime serieswas 1983, when the Legal

Amazon region receivedbelow average rainfall amounts,at typically 30 cm less than the long-termyearly mean. This effect can be seen clearly in the predictedFPAR at MNS, with intermediate rainfall amounts and surface radiation STM, and RND sites,which declinesby more than 25% by fluxesamongthe sites,is on averagethe mostproductivesite the end of 1983, comparedto the mean annualFPARs for the in terms of predictedannualNPP. Comparedto the STM decade. In addition to drought stress,we cannotrule out forest site, for example,the RND forest site is apparently entirely the possibilitythat forest fires negatively affected locatedon more fertile soils,which could explainthe higher FPAR values observed for these 8-km x 8-km model locations predictedFPAR, LAI, woodbiomasspool, and soil nitrogen during 1983 and 1987. pool at the RND site. Comparing EET predictionsacrossthe northern Amazon SeasonalFPAR is predictedto increaseconsistentl•over forest transectsites,it appearsthat althoughannual rainfall the period 1988-1990at the MNS and STM sites,possiblyas amountsat MNS are, on average,slightly higher than at STM a resultof 15-25%higherannualrainfall amounts(compared (238 cm rain per year at MNS versus213 cm rain per year at to averagerainfall totals)observedat theselocationsduring STM), highersurfaceirradiancefluxesat STM throughoutthe 1988 and 1989 (Figure 11). This higher FPAR effect is year result in predictionof higher averageEET fluxes from carried over into higher annual NPP predictedat the sites. forests(153 cm per year at MNS versus171 cm per year at Higher than average rainfall amounts could be observed STM). Owing to the model settingfor deep-rooted(10 m) during these years over the entire states of Amazonas and forest vegetationon heavy clay soils, it is rare that water Parfi (Figure 7). In contrast,FPAR at the RND forestsite is stressis predictedto limit EET at sites like MNS and STM. the lowest surface radiation fluxes, while the BSL savanna

site receives the lowest rainfall annually and the highest surfaceradiation fluxes (Figure 11). The RND forest site,

POTTER ET AL.'

MODELING

AMAZON

CARBON

CYCLES

10,439

Table3. Estimated NASA-CASAParameter Settings andSources forInitialization of AmazonSiteSimulations StudySites MNS

ModelParameter

RND

BSL

U•its

Geographic StudySite Name

Elevationb

STM

INPA tower site, Manaus m

Tapajos National Forest, Santar•m

55

88

ReservaJam, Ji-Parana 298

Reserva EcologicaAguas Emendad,Brasilia a 1052

Climate Drivers

Meanannual precipitation

cmyr4

238

213

189

147

Mean annualsurfaceradiationc

W m-2

217

231

236

251

1,084

1288

682

Vegetation

-1

Netprimary production (mean)

g C m-2yr

927

Maximum one-sidedLAI

m2 m-'-

3.9

Livewoodbiomass

g C m-2

Mineral

4.8

10,430

12,200

7.0

14,490

2.5

7670

Soils

Texture Class

Totalorganic carbon(to30cm) Totalorganic nitrogen(to30cm)

kgC m-2 gN m-2

medium

medium

coarse

coarse

4.2 227

5.2 276

6.0 301

3.8 204

a Unlessindicatedotherwise,data are from Potter et al. [1998] for 8-km grid cells and model initializationestimatesusedin this

studbY. INPA istheInstituto Nacional dePesquisas daAmaz6nia ofBrazil. LAIdenotes leafarea index. Data are from Geschand Larson [ 1996]. cDataarefromBishopandRossow[ 1991].

This is true evenduringthe notablydry year of 1983 (220 cm rain at MNS and 141 cm rain at STM), in part because Amazon forest soils are predicted to store water from previous(wetter)yearsthattreescanutilize duringoccasional droughtperiods. ComparingEET predictionsacrossthe southernAmazon transectsites, the RND forest on a more sandy soil type than eitherthe MNS or STM sitesmay experienceperiodsof water stress,as shownwhen predictedEET dropsbelow 5 cm per month in July-Septemberduring the notably dry years of 1983, 1987, and 1990 (with 189, 172, and 171 cm rain per year,respectively).This generatesstrongseasonalpatternsin soil respirationfluxes of carbon. In comparison,the BSL savannasite experiences regulardroughtperiodsof near zero rainfall, extendingfrom Juneto Augustin mostyears,which limits EET to an averageof 147 cm annually. With high mean surfaceirradiancefluxes at BSL, the model predictsthat only about 70% of total evaporativedemandcan be met by rainfall and stored soil moisture each year, which generates strongseasonalpatternsin both net plant productionand soil respirationfluxes of carbon,comparedto the three Amazon forest sites shown.

et al., 1999], but is somewhathigherthan the rangereported by Tian et al. [1998] in their interannualAmazonmodeling study,which estimatedNPP overthe regionfrom 660 to 830

g C m'2yr-l. Interannualpatterns in predictedNEP fluxes of CO2 suggestthat with the exceptionof a dry year like 1983, Amazonprimaryforestsiteslike MNS and STM canbe net

annual carbon sinksof 30-140g C m-2. However, during extreme drought years like 1983, both forest sites are

predicted to losemorethan150g C m-2yr-• in NEPflux, whichcan apparentlytake 3 or 4 yearsto recoverin termsof reversingto a positivemultiyearNEP for theseforestsites. The ecosystemmodel predictsthat both the RND and BSL sitesare more variablewith respectto predictedNEP fluxes over the modeled time series than are the MNS

and STM

sites. This is explainedby year-to-year changesin soil respirationlossesof CO2relatedto variablerainfall patterns,

which result inswings ofbetween-120 and+120g C m'2yr -• in NEP flux at the RND and BSL sites.

Over the entire modeled time series, the MNS, STM, and

BSL sitesarepredictedto be cumulativecarbonsinksof 265,

252,and357g C m-2respectively, whereas theRNDsitelost

All foursitesshow Model resultsof predictedmonthly NPP at the BSL site a totalof 197g C m-2totheatmosphere. positive carbon gains in NEP totaled over the final threeyears show an unusualpeak in early 1988, which coincideswith NEP fluxes,however,do atypicallyheavy rainsbeginningin late 1987 and extending of the simulation.Theseaggregate into March of 1988. Nevertheless,peak annual NPP at the not factorin any potentialbiomassburninglossesof CO2 to BSL site is estimated for the modeled time series in 1990, a the atmospherefrom ground fires within any of the site yearwhentherainfalltotalwaslowerthanaverage, butFPAR remainshigh as a carryovereffect from the previouswet

locationgrid cells. Standingbiomassestimatedfor all sites (Table 3), muchof which couldbe burnedoff during forest

monthsof late 1989, and solarradiationfluxes are increased fire events,would far exceedthesecumulativeNEP carbon gainsoverthe severalyearssimulated.It is importantto note comparedto precedingyears.

The modelpredictionfor meanannualNPP rangesfrom

that althoughthe aggregated FPAR and predictedNPP flux

950 to 1300g C m-2yr-• for the STM andRND forest over a modeled cell area may be estimatedto increaseover locations, 780 to 1030g C m-2yr-• for the MNS forest certainperiods,the deforestationloss of standingbiomass location, and560to 900g C m-2yr-• fortheBSLlocation.over evenrelativelysmall areaswithin a given 8-km x 8-km This overallrangein NPP is in closeagreementwith many locationmay be sufficientto offsetany multiyearNEP carbon areasof thegridcelllocation. otherterrestrialNPP modelsfor tropicalforestzones[Cramer gainsfromthelargerunburned

10,440

POTTER

a. MNS

ET AL.'

MODELING

ß450

350"

250

45O

350 •

0.8

•' •

• 0.6

250

0.4

0.4

[•

CYCLES

1.0

0.6

0.2

CARBON

b. STM

1.0 0.8

AMAZON

150

0.2

150

20

100

20

100

16

80

16

80

12

60

12

60

40

8

40

20

4

m 8

0

.20 •



0

180

180

180

180

150

150

150

150

120

120

120

120

Z 90

90

90

90

60

60

60

60

30

30

30

30

0

0

Figure 11. Monthly modelresultsfromNASA-CASA simulation(1982-1990)for four selectedsitesunder studyin the LBA project:(a) INPA towerSsite,Manaus,Amazonas(MNS), (b) TapajosNationalForest, Santar6m,Parfi(STM), (c) ReservaJaru,Rond6nia(RND), and(d) ReservaEcologicaAguasEmendadas,

Brasilia(BSL). (top) FPAR (thick line; 0-1) and surfacesolarirradiance(thin line; W m2). (middle) Estimatedevapotranspiration (thick line, cm) and precipitation(thin line; cm). (bottom)Net primary production (thickline, g C m'2)andsoilheterotrophic respiration (thinline;g C m'2). INIT standsfor model initializationresultsusingdrivervariablesaveragedfor monthsof the 1980s.

5.2. Amazon Regional Results

fluxes,increasingPET and EET are predictedon a gradient overthe region(Plate 1), mainly For the Legal Amazon regionas a whole, the driestyears from northwestto southeast for this simulationcorrespondto the E1 Nifio eventsin 1983 as a result of increasing surface radiation fluxes and air 'Duringthe wet season months(e.g.,January) and 1987, with averageyearly rainfall anomaliesof-30 cm temperatures. and -20 cm (1983 and 1987, respectively),comparedto the of mostyears,the modelpredictsthat practicallythe entire EET fluxesof greaterthan region-wideaverageof 201 cm rainfall per year over the nine Legal Amazonregionalsupports years modeled. The wettestyear overall was 1985, with an 10 cm per month. averageannualanomalyof +23 cm for the region. Therefore Predicted EET fluxesat the endof thedry season (August) 1983-1984 and .1985-1986 were selected as the primary in a droughtyearlike 1983 imply thatthe entirecerradozone comparisonperiodsfor presentationof our regionalmodeling (savanna and dry woodland) of the Legal Amazon has results,with 1987and 1990assecondary comparison periods. depleted soil moisture storage and near-zero EET flux. The model's estimatedrange for EET is consistentwith Despite the deep rooting settingsin the model for Amazon reported evapotranspiration measurements from Amazon forests, the area of depleted soil moisture extends far forests [Shuttleworth, 1988; Roberts and Cabral, 1993;

northward and westward into the states of Mato Grosso and

Nepstadet al., 1994; Costaand Foley, 1997;Jipp et al., Parfi,startingfromabout5ø16'S,50ø27'W(about1ø westof 1998]. Based on the model resultsfor ecosystemwater Maraba),northwardduringa yearlike 1983well into forested

POTTER ET AL.' MODELING

c. RND

450

AMAZON

CARBON CYCLES

10,441

d. BSL

1.0

450

1.0

350

0.8



0.6

250 • 0.4

0.8

350

0.4

250

0.6

0.2

150

0.2

150

20 I

lOO

20

100

16

80 m

16

80

4

8

40 20

0.

0

[. 12

60

e•



40



8

20 •

[ 120

180

180

180

150

150

150

120

120

120

90

90

60

60

60

60

30

30

30

30

z 90

90

Figure 11. (continued) zones (Selvas) of the Xingu and TocantinsRiver basins and estimated at about6.6x 1012m3 wateryr-I [Richey et al., 1989a]. This region-wide ratio of 1.4 for EET rate to river extending to areas of Parfi state north of the Amazon main channel. Scatteredareasof shallowrocky soils on plateaus discharge from runoff and subsurfaceoutflow is in close extendingacrossMato Grossoand Rond6nia also show near- agreement with Amazon regional water balance estimates zero EET fluxes in August 1983. In total, the model estimates made by Zeng [1999], and also with smaller-scaleestimates

thatmorethan1.24millionkm2oftheLegalAmazon region, by Franken and Leopoldo[1984] who measuredrainfall and mainly in the cerrado areas of the states of Maranhao, Tocantins,and Mato Grosso,suffer from severelydepleted soilmoistureconditionsduringa strongE1Nifio droughtyear. By contrast, in a much wetter year like 1985, the model predictsthat soils over much of the northern cerrado zone of Maranhao

and Tocantins

states can remain

moist well

into

August, supportingEET fluxes of greater than 10 cm per

streamflowtotalsfrom the BaciaModelo(TarumaA9u) and Barro Branco catchments near Manaus.

Different

results

were reported, however, in the water balance study by Forsberget al. [1988] andLesack[1993] at a catchmentnear Lake Calado, where the EET:river dischargeratio was estimatedat slightlybelow 1 duringa high rainfall year (> 287 cm).

month, as far southeast as the confluence of the Sono and

The modelestimates interannual variationin regional-level Tocantins Rivers(at 9ø00'S,47ø40'W).Evenin therelatively NPPof 4.8to 5.9PgC yr-•,withtheextreme lowandhigh wet forested zones of western Acre and Amazonas states, yer.rs for NPP being1983and 1988,respectively.This range EET fluxes are estimated by the model to be about 10% is towardthe upperboundsof the regionalNPP estimateof higherin 1985 thanin the dry E1Nifio monthsof 1983. Total annual EET fluxes of water for the entire Legal Amazon region (Plate 1) are estimatedto range from 8.8 x

3.8 to 5.7 PgC yr-• by Tianet al. [1998]for theLegal

Amazon. On an averageseasonal basis,NPP overthe Legal Amazonis predictedby the NASA-CASA modelto peak in 1012 m3wateryr-• (in 1983)to 10•3m3 wateryr-•(in 1988), March and declineto annuallow levelsby November. The which is somewhat higher than the combined annual annualamplitude(maximum- minimummonthly)NPP for dischargerates for the Amazon and Tocantins River basins, the regionis predictedto be lowestoverallduringthe high

¸ ¸

¸ o

i

z

POTTER ET AL.: MODELING

rainfall year of 1985, at < 0.05 Pg C, and to increasesteadily over the next 5 yearsto almostO.12 Pg C in 1990. We find, nevertheless, that this trend of increasingannualamplitudeof Amazon regional NPP is not simply explainedby trends in annual amplitude of rainfall (Figure 7), which does not increasenotably over the simulationtime period of 19851990. Rather,it appearsthat increasingannualamplitudeof solar surfaceirradiancemay accountfor the similar trendsin NPP, as indicated in the time seriesplots for Amazon site locations(Figure 11) andregionalaverages(Figure 4). For prediction of regional-level NEP for the Legal Amazon, our model estimatesinterannualvariability of-0.4

PgC yr'• (netCO2source) to 0.5 Pg C yr-I (netCO2sink),

AMAZON

CARBON CYCLES

10,443

3. Drought effects during E1 Nifio years can reduceNPP fluxes of CO2 in undisturbedforestsof the easternAmazon by 10-20%, comparedto long-termaverageestimatesof regional productivity. 4. Rates of soil respiration loss of CO2 in relation to variable rainfall patternsare significant sourcesof year-toyear variability in regionalAmazon NEP fluxes.

5. Annualdeforestationlossesof standin•g biomassfrom burning over even relatively small regional areas are sufficientto offsetmultiyearNEP carbongainsin undisturbed Amazon

forests of these areas.

In additionto ecosystemmodelsdriven by satelliteremote sensingdata,severalcomplementarymethodswill be brought to bear in testingtheseabove hypothesesas part of the LBA program. Numerous independent approachescan aid in defining limits on regional carbon stocksand narrowing the

with the extreme low and high years being 1983 and 1984, respectively (Plate 2). These predicted interannual NEP results are similar to those predicted by Tian et al. [1998], in net terrestrial sources of carbon in the even though regional NPP totals for the two models are uncertainties somewhatdifferent. The implication here is that because atmosphere. At the forest stand level, intensivesamplingof NEP is computedas the differencebetween two much larger plant, litter, and soil carbon pools can be combined with componentcarbon fluxes (productionand respiration),it can geographicinformation systemsfor extrapolationacrossthe be estimatedat nearly identical levels using widely differing landscape. A series of tower-based measurementsof net ecosystemcarbonexchange(NEE) over several seasonscan componentflux predictions. A regular seasonalpattern for NEP is predicted by our be made usingthe eddy correlationapproach[Gouldenet al., NASA-CASA model, with June-November net CO2 source 1996; Malhi et al., 1998]. This approachconsistsof fast fluxes from ecosystemsat the southernextreme of the region responsemeasurementsof the vertical componentof wind switchingto a net CO2 sink into ecosystems by Decemberand speed and humidity, which allows for calculation of forest spreading northward over the subsequent 12 months. moisture and carbon fluxes based on the covariance of these Particularly large interannualvariability in NEP is predicted meteorologicalvariables. Together with reliable images of over most of the states of Parfi, Maranhao, and Amazonas. dominantforest cover types from Landsator other relatively During a strong E1 Nifio year like 1983, almost the entire high resolutionmappingsources,theseeddy correlationtower flux measurementsof NEE might be extended, albeit with Legal Amazon region is predictedto be a net CO: sourcein excessof-100 g C m-2 yr-• NEP, whereasduringa caution,to regionalscalesfor comparisonto ecosystemmodel comparatively wet year like 1990, most of these same areas predictions.To work from regionalto global scales,a reliable arepredicted tobea netCO2sinkof50-200g C m-2yr-•NEP. network of gas-sampling stations is required to provide Areasin northcentralMato Grosso(approximately10ø- recordsover time for atmosphericCO2concentrations that can 1løS, 57ø-58øW),which drain into the lower TapajosRiver be used with models of global circulation to inversely track basin,are predictedto be consistentnet sinksof atmospheric terrestrial locations and fluxes of NEP. Subsequentrefinements of our NASA-CASA modeling CO2 throughthe relatively dry periodsof the 1983 and 1987 E1 Nifio events. It appears that near (long-term) average approachfor the Amazon region will includedaily resolution simulations [Potter et al., 1997], driven by a stochastic rainfall during the months of May-July, combined with periods of comparatively high solar surface irradiance, can rainfall algorithm. In addition to includingvariable datesof sustainthe forestcarbonsink in theseareasfor positiveyearly deforestationevents over the Amazon region, the daily time NEP estimates. This is in contrast to forest areas around step will facilitate estimations of soil nitrogen trace gas Santar6m and eastern Par/t, which received 15-35% lower emissions(e.g., N20 and NO) and methanefluxes in Amazon both of which appearto be controlledby rainfall comparedto long-term averagesduring the 1983 and wetland ecosystems, 1987 E1 Nifio eventsand are predictedto be net CO2 sources fairly rapid changesin water availability [Verchotet al., 1999; Wassmannet al., 1992]. Our revised model version is being duringtheseyears. tested also to include dynamic sequencesof deforestation, which can be scaledup to the Amazon region [Potter et al., 6. Future Research Directions 2001]. This approachshouldgeneratedetailedspatialmodel This ecosystemmodelingstudy raises severalresearch responsesfor biomasslossesthroughburning,decomposition hypothesesthat can be testedin future field studiesof of residualunburnedbiomass,and overall ecosystemchanges Amazonecosystemexchangeof CO2 with the atmosphere, in carbonand nitrogen pools following forest conversionto suchas thoseplannedfor LBA. agriculturaluses. 1. UndisturbedAmazon forestscan be strongnet sinksfor

atmospheric carbondioxide,particularlyduringwet (non E1 Nifio) years.

2.

Seasonalperiods of relatively high solar surface

irradiance combined with several months of near-average

rainfall are requiredto sustainthe forest carbonsink for positive yearly NEP estimatesin undisturbedAmazon ecosystems.

Acknowledgments.We thank three anonymous reviewersfor helpfulcomments onanearlierversionof themanuscript. Thiswork wassupported by grantsfromtheNASA LBA-EcologyProgramand the NASA Land Surface Hydrology Program. Eric Davidson providedguidanceon soil fertility classsettingsfor regionalmap inputs. SietseLosprovidedassistance with corrections for AVHRR sensor data sets.

10,444

POTTER ET AL.: MODELING

References

Aber,J.D., andJ.M. Melillo,Terrestrial Ecosystems, 429pp.,W. B. Saunders, Philadelphia, Pa., 1991. Agbu, P. A., and M. E. James, NOAA/NASA PathfinderAVHRR Land Data Set User's Manual, Goddard Distributed Active Arch.

AMAZON

CARBON

CYCLES

Wofsy, Exchangeof carbon dioxide by a deciduousforest: response to interannualclimatevariability,Science,271, 15761577, 1996.

Grace,J., J. Lloyd,J. Mcintyre,A. C. Miranda,P. Meir, H. Miranda, C. A. Nobre, J. M. Moncrieff, J. Massheder,Y. Malhi, I. R. Wright, and J. H. C. Gash, Carbon dioxide uptake by an undisturbedtropical rain forest in south-westAmazonia: 1992-

Cent,,NASA GoddardSpaceFlightCenter,Greenbelt,Md., i 994. Bishop,J. K. B., andW. B. Rossow,Spatialandtemporalvariability 1993, Science, 270, 778-780, 1995. of global surfacesolarirradiance,J. Geophys.Res.,96, 16,839- Gr6goire,J.-M., Use of AVHRR for the studyof vegetationfires in 16,858, 1991. Africa: Fire managementperspectives,paper presentedat Euro Courses: Advances in the Use of AVHRR Data for Land Bonan,G. B., A computermodelof the solarradiation,soil moisture' and soil thermalregimesin borealforests,Ecol. Modell., 45, 275Applications,Inst. RemoteSens.Appl., Ispra,Italy, 1993. 3O6, 1989.

Brooks, R. H., and A. T. Corey, Hydraulic propertiesof porous media, Hydrol. Pap. 3, 27 pp., Colo. State Univ., Fort Collins, 1964.

Gutman, G. G., On the use of long-term global data of land reflectancesand vegetation indices derived from the advanced very high resolution radiometer,J. Geophys.Res., 104, 62416255, 1999.

Campbell,G. S., An Introductionto EnvironmentalBiophysics,159 pp., Springer-Verlag, New York, 1977.

Holben, B. N., Characteristicsof maximum-valuecompositeimages from temporalAVHRR data,Int. J. RemoteSens.,7, 1417-1434,

Cannel, M. G. R., World Forest Biomassand Primary Production Data, Academic,SanDiego, Calif., 1982. Ciais, P., P. P. Tans, J. W. C. White, M. Trolier, R. J. Francey,J. A.

Hsu, N. C., J. R. Herman., P. K. Bhartia, C. J. Seftor, O. Torres, A.

Berry, D. R. Randall, P. J. Sellers, J. G. Collatz, and D. S. Schimel,Partitioning of oceanandlanduptakeof CO2asinferred

by •3Cmeasurements fromtheNOAAClimateMonitoring and DiagnosticsLaboratoryglobalair samplingnetwork,J. Geophys. Res., 100, 5051-5070, 1995.

1986.

M. Thompson,J. F. Gleason, T. F. Eck, and B. N. Holben, Detectionof biomassburning smokefrom TOMS measurements, Geophys.Res.Lett., 23,745-748, 1996. Instituto Brasileiro de Geografia e Estatistica (IBGE), Anuario Estatisticodo Brasil, vol. 51, pp. 1-1024,Rio de Janeiro,1991. Jensen,M. E., ConsumptiveUse of Water and Irrigation Water Requirements, 215 pp., Am. Soc.Civ. Eng.,New York, 1973. Jensen,M. E., and H. R. Haise, Estimatingevapotranspiration from solarradiation,J. Irrig. Drain. Div., 89, 15-41, 1963. Jipp, P. H., D.C. Nepstad,D. K. Cassel,and C. Reis de Cavalho, Deep soil moisture storage and transpiration in forests and pasturesof seasonallydry Amazonia,Clim. Change,39, 395-412,

Costa,M. H., andJ. A. Foley,A comparisonof precipitationdatasets for the Amazonbasin,Geophys.Res.Lett., 25, 155-158, 1998. Costa, M. H., and J. A. Foley, The water balance of the Amazon basin:Dependence on vegetationcoverand canopyconductance, J. Geophys.Res., 102, 23,973-23,990, 1997. Cramer, W., D. W. Kicklighter. A. Bondeau, B. Moore Ill, G. 1998. Churkina,B. Nemry, A. Ruimy, A. Schloss,and the Participants of the PotsdamNPP Model Intercomparison,Comparingglobal Jumikis, A. R., Thermal Soil Mechanics, 267 pp., Rutgers Univ. models of terrestrialnet primary productivity(NPP): Overview Press,New Brunswick, N.J., 1966. andkey results,Global ChangeBiol., 5, 1-15, 1999. Justice, C. O., and P. Dowty, IGBP-DIS satellite fire detection DeFries, R., and J. Townshend, NDVI-derived land cover algorithmworkshoptechnicalreport,IGBP-DIS WorkingPap. 9, classificationat global scales,Int. J. Remote Sens.,15, 356788 pp., 1993,NASA GoddardSpaceFlight Cent., Greenbelt,Md., 3586, 1994.

DeFries, R. S., M. Hansen, J.R.G. Townshend, and R. Sohlberg, Global land cover classificationsat 8 km spatialresolution:The useof trainingdataderivedfrom Landsatimageryin decisiontree classifiers,Int. J. RemoteSens., 19, 3141-3168, 1998.

Feb., 1993. Kaufman, Y. J., D. Tanr6, L. Remer, E. F. Vermote, A. Chu, and B.

N. Holben, Operationalremote sensingof troposphericaerosol over land from EOS moderate resolution imaging spectroradiometer(MODIS), J. Geophys. Res., 1 0 2, 17,051-

Dwyer, E., J.-M.Gr6goire,and J.P.Malingreau,A globalanalysisof 17,067, 1997. vegetationfires using satellite images: Spatial and temporal Kaufman, Y. J., et al., Smoke,Clouds,and Radiation-Brazil(SCARdynamics,Ambio,27, 175-181, 1998. B) experiment, J. Geophys.Res.,103, 31,783-31,808,1998. Ehrlich, D., E. Lambin,and J.-P. Malingreau,Biomassburningand Keller, M., W. A. Kaplan, and S.C. Wofsy, Emissionsof N20, CH4, broad scale land-coverchangesin western Africa, Remote Sens, Environ., 61,201-209,

1997.

Flasse,S. P., and P.S. Ceccato,A contextualalgorithmfor AVHRR fire detection,Int. J. RemoteSens.,17, 419-424, 1996. Food and AgriculturalOrganization(FAO)/UNESCO, Soil Map of the World, scale 1:5,000,000. U. N. Educat., Sci., and Cult. Org., Paris, 1971.

and CO2 from tropical soils,J. Geophys.Res., 91, 11,791-11,802, 1986.

Keller, M., E. Vledkamp,A.M. Weitz, and W. A. Reiners,Effect of pastureage on soil trace-gasemissionsfrom a deforestedarea of Costa Rica, Nature, 365, 244-246, 1993.

Kennedy, P. J., A. S. Belward, and J.-M. Gr6goire, An improved approachto fire monitoringin West Africa using AVHRR data,

Forsberg,B. R., A. H. Devol, J. E. Richey,L. A. Martinelli, andR. Int. J. RemoteSens., 15, 2235-2255, 1994. Dos Santos, Factors controlling nutrient concentrations in Kidwell, K., NOAA Polar Orbiter Data User's Guide, Nat. Clim. Amazonfloodplainlakes,Limnol. Oceanogr.,33, 41-56, 1988. Data Cent., Washington,D.C., 1991. Franken,W., andP. R. Leopoldo,Hydrologyof catchment areasof Kindermann, J., G. Wiirth, G. H. Kohlmaier, and F. W. Badeck, central-Amazonianrain forest streams,edited by H. Sioli, The Interannual variation of carbon exchange fluxes in terrestrial ecosystems, GlobalBiogeochem.Cycles,10, 737-755, 1996. Amazon:Limnology andLandscape Ecologyof a MightyTropical River and Its Basin,pp. 501-519, Dr. W. Junk,Norwell, Mass., Langaas,S., and K. Muirhead, Monitoringbushfiresin west Africa 1984. by weather satellite, paper presentedat the 22nd International Symposiumon RemoteSensingof Environment,Abidjan, ivory Gesch,D. B., and K. S. Larson,Techniquesfor developmentof

global1-kilometer digitalelevationmodels,paperpresented at Pecora 13, Human Interactions With the Environment -

Perspectives FromSpace,U.S. Geological Survey,SiouxFalls,S. D., Aug. 20-22, 1996.

Coast, October 20-26, 1988.

LBA Science Planning Group, The Large Scale BiosphereAtmosphere Experiment in Amazonia (LBA) : Concise ExperimentalPlan, 44 pp., CachoeiraPaulista,S•o Paulo,Brazil,

1996. Gleeson,S. K., andD. Tilman,Allocationandthetransient dynamics Lesack,L. F. W., Water balanceand hydrologiccharacteristics of a of succession onpoorsoils,Ecology,71, 1144-1155,1990. rain forest catchment in the central Amazon basin, Water Resour. Goetz, S. J., and S. D. Prince, Remote sensingof net primary

production in borealforeststands, Agric.For. Meteorol.,78, 149179, 1996.

Res., 29, 759-773, 1993.

Los, S., Estimation of the ratio of sensor degradation between NOAA-AVHRR channels 1 and 2 from monthly NDVI composites,IEEE Trans. Geosci. Remote Sens.,3 6, 206-213,

Gordon, H. R., J. W. Brown, and R. H. Evans, Exact Rayleigh scattering calculations for usewith the Nimbus 7 coastalzone 1998. colorscanner,Appl. Opt., 27, 2111-2122, 1988. Los, S. O., C. O. Justice,and C. J. Tucker, A global lxl NDVI data Goulden,M. J., J. W. Munger,S. M. Fan, B.C. Daube,and S.C.

POTTER ET AL.'

MODELING

set for climate studies derived from the GIMMS continental NDVI data,Int. d. RemoteSens.,15, 3493-3518,1994.

AMAZON

CARBON

CYCLES

10,445

flux and evaporationusinglarge-scaleparameters, Mont. I/Veather Rev., 100, 81-92, 1972.

Lusk, C. H., O. Contreras,and J. Figueroa,Growth,biomass Prins, E. M., J. M. Feltz, W. P. Menzel, and D. E. Ward, An overview of GOES-8 diurnal fire and smoke results for SCAR-B allocation andplantnitrogenconcentration in Chileantemperate and the 1995 fire seasonin SouthAmerica, J. Geophys.Res., 103, rainforesttree seedlings'Effects of nutrient availability, 31,821-31,835, 1998.

Oecologia,109, 49-58, 1997.

Maisongrande, P.,A. Ruimy,G. Dedieu, andB. Saugier, Monitoring Reid, J. S., P. V. Hobbs, R. J. Ferek, D. R. Blake, J. V. Martins, M. seasonal andinterannual variationsof grossprimaryproductivity,

netprimary productivity, andnetecosystem productivity usinga diagnostic modelandremotelysenseddata,Tellus,47B, 178-190, 1995.

Malhi, Y., A.D. Nobre, J. Grace, B. Kruijt, M. G. P. Pereira, A. Culf, and S. Scott, Carbon dioxide transfer over a central Amazonian rain forest, d. Geophys.Res., 10 3, 31,593-31,612, 1998.

Maimstrom,C. M.,'M. V. Thoropson,G. P. Juday,S. O. Los, J. T. Randerson,and C. B. Field, Interannualvariation in global scale net primary production: Testing model estimates, Global Biogeochem.Cycles,11,367-392, 1997. •,,•.,,•1 version 2 A q,•o•l,• •th cnCt Inn Wash., 1996.

Melillo, J. M., A.D. McGuire, D. W. Kicklighter,B. Moore III, C. J. V6r6smarty, and A. L. Schloss, Global climate change and te•estfial net primaw production,Nature, 363,234-240, 1993. Ministdrio das Minas e Energia (MME), PrQeto RADAMBRASIL report,Rio de Janeiro,Brazil, 1981. Monteith, J. L., Solar radiation and productivity in tropical ecosystems, X Appl. Ecol., 9, 747-766, 1972. Nepstad,D., C. R. de Ca•alho, E. Davidson,P. Jipp,P. Lefebvre,G. H. Negreiros, E. D. da Silva, T. Stone, S. Trumbore, and S. Vieira, The role of deeprootsin the hydrologicandcarboncycles of Amazoniaforestsandpastures, Nature, 372, 666-669, 1994.

R. Dunlap, and C. Liousse, Physical, chemical, and optical propertiesof regionalhazesdominatedby smokein Brazil, J. Geophys.Res., 103, 32,059-32,080, 1998. Richey,J. E., C. Nobre, and C. Deser,AmazonRiver dischargeand climatevariability: 1903 to 1985, Science,246, 101-103, 1989a. Richey, J. E., L. A. K. Mertes, T. Dunne, R. L. Victoria, B. R. Forsberg, A. C. M. S. Tancredi, and E. Oliveira, Sources and routing of the Amazon River flood wave, Global Biogeochem. Cycles,3, 191-204, 1989b. Roberts, J., and O. M. R. Cabral, ABRACOS: A comparison of climate, soil moisture,and physiologicalpropertiesof forestsand pasturesin the Amazon Basin, Common.For. Rev., 72, 310-315, !993.

Rosenburg,N.J., B. L. Blad, S. B. Verma, Microclimate. John Wiley, New York, 495 pp., 1983. Running, S. W., and S. T. Gower, FOREST-BGC, A generalmodel of forest ecosystemprocessesfor regional applications. II. Dynamic carbonallocation and nitrogenbudgets,Tree Physiol., 9, 147-160, 1991.

Saxton, K. E., W. J. Rawls, J. S. Romberger,and R. I. Papendick, Estimating generalized soil-water characteristicsfrom texture, Soil Sci. Soc. Am. d., 50, 1031-1036, 1986. Sellers, P. J., C. J. Tucker, G. J. Collatz, S. O. Los, C. O. Justice, D.

A. Dazlich, and D. A. Randall, A global lxl NDVI data set for climate studies,2. The generation of global fields of terrestrial biophysicalparametersfrom the NDVI, Int. d. RemoteSens., 15,

Nepstad,D.C., et al., Large-scaleimpoverishment of Amazonian 3519-3545, 1994. forestsby loggingandfire, Nature,398, 505-508, 1999. Shuttleworth,W. J., Evaporation from Amazonian forest, Proc. R. New, M., M. Hulme, and P. Jones,Representingtwentieth centu• Soc. London, Set. B, 233, 321-346, 1988. space-timeclimatevariability. II. Developmentof 1901-1996 monthly grids of te•estrial surhce climate, d. Clim., 13, 2217- Silver, W. L., J. Neff, M. McGroddy, E. Veldkamp, M. Keller, and 2238, 2000.

Pa•on, W. J., B. McKeown, V. Kirchner, and D. Ojima, CENTURY Users Manual. Nat. Resour. Ecol. Lab., Colo. State Univ., Fo• Collins, Colo., 1992. Pereira,A. C., Jr., A. W. Setzer, and J. R. dos Santos,Fire estimates in savannas

of central

Brazil

with

thermal

AVHRR/NOAA

calibrated by TM/Landsat, paper presented at the 24th InternationalSymposiumon RemoteSensingof Environment,Rio de Janeiro,Brazil, May 27-31, 1991. Potter, C. S., and S. A. Klooster, Global model estimatesof carbon andnitrogenstoragein litter andsoilpools:Response to changein vegetation quality and biomassallocation, Tellus, 49B, 1-17, 1997.

Potter,C. S., and S. A. Klooster,Interannualvariability in soil trace

R. Cosme, Effects of soil texture on belowground carbon and nutrient storage in a lowland Amazonian forest ecosystem, Ecosystems,3, 193-209, 2000. Skole, D., and C. Tucker, Tropical deforestation and habitat fragmentationin the Amazon: Satellite data from 1978 to 1988, Science,260, 1905-1910, 1993.

Stone,T. A., P. Schlesinger,R. A. Houghton,and G. M. Woodwell, A map of vegetationof SouthAmericabasedon satelliteimagery, Photogramm.Eng. RemoteSens.,60, 541-551, 1994. Stroppiana, D., S. Pinnock, and J.-M. Grfgoire, The global fire product:Daily fire occurrencefrom April 1992 to December1993 derived from NOAA-AVHRR 1279-1288, 2000.

data, Int. d. Remote Sens., 21,

Tans, P. P., I. Y. Fung and T. Takahashi,Observationalconstraints

on the global atmosphericCO2 budget, Science,247, 1431-1438, gas(CO2,N20, NO) fluxesandanalysisof controllerson regional 1990. to globalscales,GlobalBiogeochem. Cycles,12, 621-637, 1998. Potter, C. S., and S. A. Klooster, Dynamic global vegetation Terborgh,J., C. N. Flores, P. Mueller, and L. Davenport,Estimating modeling(DGVM) for predictionof plant hnctional types and the ages of successionalstandsof tropical trees from growth biogenictrace gas fluxes, Global Ecol. Biogeogr.Lett., 8, 473increments,d. Trop. Ecol., 14, 833-856, 1997. 488, 1999a. Thornthwaite,C. W., An approachtoward rational classificationof Potter, C. S., and S. A. Klooster, Detecting a te•estrial biosphere climate,Geogr.Rev., 38, 55-94, 1948. sink for carbondioxide' Interannualecosystemmodeling for the Tian, H., J. M. Melillo, D. W. Kicklighter, A.D. McGuire, J. V. K. mid-1980s,Clim. Change,42, 489-503, 1999b. Helfrich III, B. Moore III, and C. J. Vorosmarty, Effect of Potter, C. S., J. T. Randerson, C. B. Field, P. A. Matson, P.M. interannualclimate variability on carbon storagein Amazonian Vitousek, H. A. Mooney, and S. A. Klooster, Terrestrial ecosystems, Nature, 396, 664-667, 1998. ecosystemproduction:A processmodelbasedon global satellite Tomasella, J., and M. G. Hodnett, Estimating soil water retention characteristicsfrom limited data in Brazilian Amazonia, Soil Sci., andsurhcedata,GlobalBiogeochem. Cycles,7, 811-841, 1993. Potter,C. S., R. H. Riley, and S. A. Klooster,Simulationmodelingof 163, 190-202,1998. nitrogentrace gas emissionsalong an age gradient of tropical Verchot, L. V., E. A. Davidson, J. H. Cattanio, I. L. Ackerman, H. E. forest soils, Ecol. Modell., 97, 179-196, 1997. Erickson,and M. Keller, Land use changeand biogeoch6mical Potter,C. S., E. A. Davidson,S. A. Klooster,D.C. Nepstad,G. H. de controls of nitrogen oxide emissions from soils in eastern Negreiros,and V. Brooks,Regionalapplicationof an ecosystem Amazonia,Global Biogeochem.Cycles,13, 31-46, 1999. productionmodel for studiesof biogeochemist• in Brazilian Vose, R. S., R. L. Schmoyer,P.M. Steurer,T. C. Peterson,R. Heim, Amazonia,Global ChangeBiol., 4, 315-334, 1998. T. R. Karl, and J. Eischeid, The Global Historical Climatology Potter, C. S., E. A. Davidson,D.C. Nepstad,and C. R. Ca•alho, Network:Long-termmonthlytemperature,precipitation,sealevel Ecosystemmodelinganddynamiceffectsof deforestation on trace pressure,and stationpressuredata, Rep. ORNL/CDIAC-53, NDPgas fluxes in Amazon tropicalforests,For. Ecol. Manage., in 041, CarbonDioxide Inf. Anal. Cent., Oak RidgeNatl. Lab., Oak press,2001. Ridge,Tenn., 1992. Priestly,C. H. B., andR. J. Taylor,On the assessment of surhceheat Wassmann,R., U. G. Thein, M. J. Whiticar, H. Rennenberg,W.

10,446

POTTERET AL.: MODELINGAMAZONCARBONCYCLES

Seiler, and W. J. Junk, Methane emissions from the Amazon

floodplain:Characterizationof productionand transport,Global Biogeochem.Cycles,6, 3-13, 1992. Wilson, S. D., and D. Tilman, Componentsof plant competition along an experimentalgradientof nitrogenavailability,Ecology, 72, 1050-1065, 1991.

Zeng, N., Seasonalcycle and interannualvariability in the Amazon hydrologiccycle,J. Geophys.Res., 104, 9097-9106, 1999. Zobler, L., A world soil file for globalclimatemodeling,NASA Tech. Memo., 87802, 32 pp., 1986.

J. Coughlan andC. Potter, NASAAmesResearch Center, Mail Stop 242-4, Moffett Field, CA 94035. (cpotter•gaia.arc.nasa.gov) C. R. deCarvalho, EMBRAPA-Amazonia Ocidental, Laboratorio

deEcofisiologia Vegetal, Be16m, Parfi,Brazil.

J.Dungan, S.Klooster, andA. Torregrosa, EarthSystems Science andPolicy, California State University Monterey Bay,Seaside, CA

93955.

M. Bobo and V. Brooks Genovese, Johnson Controls World

Services,Mail Stop242-4, NASA Ames Operations,Moffett Field, CA

94035.

(Received March16,2000;revised August 14,2000;accepted

August22, 2000.)