GlobalChangeBiology(2014)20,979–991,doi:10.1111/gcb.12375 Evidence for strong seasonality in the carbon storage and carbon use efficiency of an Amazonian forest LUCY ROWLAND*, TIMOTHY CHARLES HILL†, CLEMENT STAHL‡, LUKAS SIEBICKE‡, BENOIT BURBAN‡, JOANA ZARAGOZA-CASTELLS*, STEPHANE PONTON§, DAMIEN BONAL§, PATRICK MEIR*¶ andMATHEW WILLIAMS* *SchoolofGeosciences,UniversityofEdinburgh,EdinburghEH93JN,UK,†EarthandEnvironmentalSciences,UniversityofSt Andrews,FifeKY169AL,UK,‡INRA,UMR-ECOFOG,Kourou,France,§INRA,UMR1137EcologieetEcophysiologie Forestières,Champenoux54280,France,¶ResearchSchoolofBiology,DivisionofPlantSciences,AustralianNationalUniversity, Canberra,ACT0200,Australia Abstract Therelativecontributionofgrossprimaryproductionandecosystemrespirationtoseasonalchangesinthenetcarbon flux of tropical forests remains poorly quantified by both modelling and field studies. We use data assimilation to combinenineecologicaltimeseriesfromaneasternAmazonianforest,withmassbalanceconstraintsfromanecosys- temcarboncyclemodel.Theresultinganalysisquantifies,withuncertaintyestimates,theseasonalchangesinthenet carbonfluxofatropicalrainforestwhichexperiencesapronounceddryseason.Weshowthatthecarbonaccumula- tioninthisforestwasfourtimesgreaterinthedryseasonthaninthewetseasonandthatthiswasaccompaniedbya 5%increaseinthecarbonuseefficiency.Thisseasonalresponsewascausedbyadryseasonincreaseingrossprimary productivity, in response to radiation and a similar magnitude decrease in heterotrophic respiration, in response to dryingsoils.Theanalysisalsopredictsincreasedcarbonallocationtoleavesandwoodinthewetseason,andgreater allocationtofinerootsinthedryseason.Thisstudydemonstratesimplementationofseasonalvariationsinparame- tersbetterenablesmodelstosimulateobservedpatternsindata.Inparticular,wehighlightthenecessitytosimulate theseasonalpatternsofheterotrophicrespirationtoaccuratelysimulatethenetcarbonfluxseasonaltropicalforest. Abbreviations A = GPPfractionallocatedtopooln Fn A = Allocationofcarbontopooln n C = Carbonstockforpooln n Cr= Coarseroots CUE= CarbonUseEfficiency CWD= Coarsewoodydebris f= Foliage fr= Fineroots GPP= GrossPrimaryProduction LAI= LeafAreaIndex L = Litterfall f Lit= Litter R = Autotrophicrespiration a R = EcosystemRespiration eco R = Respiredfractionofcarbonpooln Fn R = Heterotrophicrespiration h R = Respirationfromcarbonpooln n SOM= Soilorganicmatter T = Turnoverrateofcarbonfrompooln n w= Wood Keywords: carbonuseefficiency,DALEC,dataassimilation,ecosystemrespiration,FrenchGuiana,seasonalcarbonfluxes,trop- icalforest Received27June2013;andaccepted14August2013 Correspondence:LucyRowland,tel.+44(0)1316517034, fax+44(0)1316620478,e-mail:[email protected] ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd. ThisisanopenaccessarticleunderthetermsoftheCreativeCommonsAttributionLicense, whichpermitsuse,distributionandreproductioninanymedium,providedtheoriginalworkisproperlycited. 979 980 L. ROWLAND etal. area of Amazonia, particularly the north east, with Introduction futureclimatechange(Coxet al.,2008;Juppet al.,2010; The seasonality of the net carbon flux of Amazonian Marengo et al., 2012). The seasonal dry period at our forests remains uncertain. Existing studies in Amazo- studysitehasbeenshowntobecoincidentwithreduc- nian forests have reported both increases (Goulden tions in total R , soil respiration (including root and eco et al., 2004; Hutyra et al., 2007; Bonal et al., 2008) and litter respiration), tree growth, stem respiration and decreases (Malhi et al., 1998; Chambers et al., 2004; CWD respiration at the site (Bonal et al., 2008; Stahl Keller et al., 2004) in the total carbon sequestered in et al.,2011;Wagneret al.,2012;Rowlandet al.,2013). the dry season. Models struggle to adequately simu- Toachievethemostlikelysummaryofexistingdata, late wet-to-dry season changes in the net carbon flux we adapt the Data Assimilation Linked Carbon Model (Saleska et al., 2003; Baker et al., 2008; Verbeeck et al., (DALEC; Fox et al., 2009; Williams et al., 2005) for use 2011).Theimportanceofseasonalchangesingrosspri- atthesiteinFrenchGuiana(Fig. 1;hereafterreferredto mary production (GPP) and ecosystem respiration as DALEC-FG). We use Metropolis-Hastings data (R )onthenetcarbon fluxoftropicalforestsremains assimilation (DA; Knorr & Kattge, 2005) to combine eco unresolved. uncertain data with the process information and mass Recent model development studies have focused on balance described by the DALEC-FG model, to con- improving the simulation of GPP (Fisher et al., 2007; strain the seasonal response of the ecosystem. The DA Baker et al., 2008; Grant et al., 2009; Kim et al., 2012) schemeisusedtoparameterizethemodelforbothwet rather than the fate of organic matter, and emissions and dry season, which are defined using a soil water from R . Reco is comprised of autotrophic (leaf, root content threshold (see Methods). Using separate eco and stem) and heterotrophic (litter, dead wood and parameters for each season the analysis can attribute, soil) components. Various field studies have estimated with estimates of uncertainty, the seasonal changes in the contribution of each component of respiration to thenetcarbonfluxtochangesinthecomponentcarbon totalR (Malhiet al.,2009;Metcalfeet al.,2010;Malhi fluxesofthistropicalforest. eco et al., 2013). However, there is still uncertainty regard- ing the sensitivity of these individual respiration com- ponents to the seasonal drying of soil and how these responses coincide with the seasonality in GPP, to affectseasonalchangesintheecosystemcarbonbudget (Meiret al.,2008). Carbonuseefficiency(CUE)istheproportionofGPP invested into net primary production (NPP), rather than expended as autotrophic respiration (R ), and is a animportantindicatorofhowefficientanecosystemis at investing assimilated carbon for growth (Waring et al.,1998).However,CUEisdifficulttoquantifyaccu- ratelyusingmeasurementsbecauseofuncertaintyasso- ciatedwithscalingmeasurementsofleaf,stemandroot respiration to the ecosystem scale (Chambers et al., 2004). Similarly, estimating CUE remains a challenge formodellingtropicalsystemsbecause ofuncertainties inparameterizingtheseasonalityofR (Foxet al.,2009; a Verbeecket al.,2011). Fig.1 DiagramoftheDALEC-FGcarbonmodel,anadaptation Thisstudyreportstheresponsesofalowlandtropical of the Data assimilation linked Carbon (DALEC) model (Wil- forest to seasonal variations in environmental condi- liams etal., 2005). The boxes represent a carbon pool and the tions,atasiteinFrenchGuiana,forwhichmultipleeco- arrows represent a carbon flux through the model, the dotted greyarrowsrepresentalossfromrespiration,whichissettoa logical time series data sets are available. These time fixed fraction of the carbon allocated to each pool. All of the series include: dry and wet season measurements of acronyms for the pool and fluxes are explained in the model leaf,stem,soilandcoarsewoodydebris(CWD)respira- parameters table (Table1). The fractions respired from auto- tion; net ecosystem exchange (NEE); litterfall; leaf area trophicpools(foliarcarbon;C,carboninwood;C ,carbonin f w index (LAI); woody biomass; and stem growth. The fineroots;C andcarbonincoarserootsC )arecalculatedasa fr cr study site experiences a strong seasonal change in soil fractionofthecarbonallocatedtothepool.Thefractionrespired moisture (Bonal et al.,2008; Wagneret al., 2011);some- fromthelitter,coarsewoodydebrisandsoilcarbonpools(C , lit thing which has been predicted to occur over a wider C ,C )arecalculatedasafractionofthetotalpool. cwd som ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 EVIDENCE FOR STRONG SEASONALITY 981 Materialsandmethods parameters for the allocation, turnover rate and respiration fromthefoliage,stemandrootpools(seebelow). AswiththeoriginalDALECmodel,thedailytime-stepand Site computational simplicity of DALEC-FG makes it well suited ThestudyfocusedonatropicallowlandforestsiteatParacou to DA, where a large number of model runs are required. Research Station in French Guiana (5°16N, 52°16W). Data Gross primary productivity (GPP) in DALEC-FG was deter- were collected over a periodof 8 years from January2004to minedusingtheAggregatedCanopyModel(ACM;Williams December2011ontwoadjacent709 70mterrafirmeperma- etal.(1997);Fig. 1).ACMisanempiricalsimplificationofthe nent forest plots (Bonal etal., 2008; Stahl et al., 2011, 2013; Soil–Plant–Atmosphere model (SPA; Fisher etal., 2006, 2007; Wagneret al.,2012;Rowlandetal.,2013).Theplotsweresitu- Williams, 1996) which predicts GPP according to daily mini- ated on nutrient-poor acrisols and were similar in ecological mum and maximum temperature, precipitation, radiation, characteristics, including species density (103 and 116 spe- atmosphericCO concentration,soilwaterpotential,hydraulic 2 cies ha(cid:1)1),stemdensity(612and725stems ha(cid:1)1)andlitterfall resistance,leafnitrogenandLAIcombinedwith10optimized (7.28(cid:3) 0.3and6.42(cid:3) 0.3Mgha(cid:1)1 yr(cid:1)1).FrenchGuianahas parameters. To ensure ACM was correctly calibrated for the astrongseasonalrainfallpatterncausedbythemovementof study site, 10 parameters in ACM were optimized to repro- theintertropicalconvergencezonetwiceayear,causingalong ducetheGPPpredictedbyasetofrunsperformedforthesite (August–November) and short (March) dry season. Conse- usingtheSPAmodel.SPA,adetailedecophysiologicalmodel, quently, despite the site receiving an average of 3041mm of has previously been validated at Amazonian forest sites rainperyear(Gourlet-Fleuryet al.,2004),duringthelongdry (Fisheret al.,2007).OnceSPAwascalibratedforoursite(see season rainfall is normally <50mm per month (Bonal et al., Supporting Information) it accurately produced previously 2008).Thedryseasonreductioninrainfallislargeenoughto publishedGPPestimatesforthissite(Bonaletal.,2008;Fig. 2). causesasignificantreductioninleafwaterpotential(seeSup- ACM replicated the SPAGPP with a rootmean square error porting Information), and has been shown to have a small of0.05gC m(cid:1)2 d(cid:1)1. effectonGPP(Bonalet al.,2008andseeFig. 2). SoilmoistureresponsefunctioninDALEC-FG Modeldescription Asoilmoistureresponsefunctionforheterotrophicsoilrespi- The DALEC model (Williams et al., 2005) was adapted for rationwascreatedusingR datameasuredatthesite.TheR s s FrenchGuiana(DALEC-FG)andisasimpleboxcarboncycle data included respiration from root, litter and soil organic modelofcarbonpoolsconnectedbyfluxes(Fig. 1).Theorigi- matter.Tomodelthesoilwaterresponseofheterotrophicres- nal DALEC model has been used in a number of previous piration, we first had to remove the effect of root respiration modelling studies (Williams etal., 2005; Fox et al., 2009; Hill from the R data. We estimate root respiration by assuming s etal., 2012). Our adaptations to the original DALEC model thatitisaconstantandthattheseasonalchangesinsoilrespi- (Williams etal., 2005) included: (i) inclusion of a coarse root rationarecausedbyheterotrophicprocesses.Previousstudies poolandacoarsedeadwood(CWD)pool(Fig. 1);(ii)Model- onoursiteandatothersitesintheeasternAmazonhavedem- lingstem,leaf,finerootandcoarserootrespirationseparately onstrated a strong heterotrophic soil respiration response to (Fig. 1);(iii)Inclusionofamoistureresponsefunctiontopre- reductionsinsoilmoisture(Bonaletal.,2008;Metcalfeetal., dict heterotrophic respiration created using mean daily soil 2007; Sotta et al., 2007). In comparison, only small, and both respiration(Rs)measuredatthesite(seeSupportingInforma- positiveandnegativeseasonalchangesinautotrophicsoilres- tion) and (iv) The use of separate wet and dry season pirationhavebeenfound(Metcalfeet al.,2007;DaCostaetal., Fig.2 Comparisonofthegrossprimaryproduction(GPP)fromthesoil–plant–atmospheremodel(SPA)runattheParacousitewith theGPPcalculatedfromtheeddycovariancedatacollectedatthesitefrom2004to2005andpublishedinBonaletal.(2008).Lightgrey crossesindicatedailyGPP(gCm(cid:1)2d(cid:1)1)fromBonaletal.(2008)andlightgreytrianglestheequivalentfromSPA.Thelinesshowthe 6-dayrunningmeanfromSPA(darkgreydottedline)andBonaletal.,2008(lightgreysolidline). ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 982 L. ROWLAND etal. 2013). We assume that root respiration is a constant value of 1XðM(cid:1)OÞ2 1.9(cid:3) 0.3gC m(cid:1)2 d(cid:1)1;thisishalfofthesoilrespirationwhen Mf ¼2 E2 itisaveragedoverthe2 yearsofmeasurements.Rootrespira- tion has been shown to be approximately half of annual soil whereMisthemodelledresult,OistheobservationsandEis respiration, at our study site (Ponton & Bonal, unpublished theSEontheobservations. data)andatothersitesintheeasternAmazon(Metcalfeetal., Prior information about the parameter distributions was 2007, 2010). To model heterotrophic soil respiration our esti- includedusingthesameformoflikelihoodfunction,butcom- mated root respiration value is subtracted from all daily R data (n= 601,2005–2006)andthese data areused tocreate as paring parameter selections with estimated prior parameters (Table 1; Knorr & Kattge, 2005). Model parameters were modelofheterotrophicsoilrespiration. assumedtobereal,positiveandtohavealognormalprobabil- Theseasonaleffectoftemperatureontheheterotrophicres- ity distribution (Knorr & Kattge, 2005). Therefore, all pro- pirationfrom soilwas removed bysubtracting thechange in cessesofparameterselection,andacceptanceandrejectionof heterotrophic respiration caused by temperature using the parametersinrelationtopriorrangeswereperformedinlog- temperatureresponsefunctioninDALEC-FG,whichassumes adoublingofrespirationratewitha10 °Criseintemperature. normalspace(Knorr&Kattge,2005). ThestepsizefortheDAwassettoarandomdrawfroma Theremainingseasonalityintheheterotrophicsoilrespiration normal distribution, with a mean of 0 and a SD of 0.004 in was regressed against the mean measured daily surface soil log-normal space, resulting in an acceptance rate of 40–45%. watercontent(SWC)whichwascollectedevery30mininthe surface5–10cm(seebelow).Alog-normalcurvewasfittedto The length of the Markov chain was determined using Gelman–Ruben convergence statistic (Brooks & Gelman, these data (Fig. S1) and normalized, so the optimum point (2.5gC m(cid:1)2 d(cid:1)1)wasequalto1.DALEC-FGwasforcedwith 1998). The Gelman–Ruben convergence statistic was calcu- lated using six Markov chains and indicated that after thedailymeanofmeasuredSWCdataandusedthisnormal- 1 200000 steps the Markov chain had adequately sampled izedlog-normalfunctiontoadjustpredictedvaluesofcarbon theposteriordistribution,withaconvergencelevelbelowthe loss from the heterotrophic pools based on soil moisture. It 1.2 threshold (Brooks & Gelman, 1998). A burn point – the should be noted that this moisture response function is an numberofinitialacceptedparametercombinationswhichare empiricalrelationshipandthusissitespecific. thrown away – was set at 200000 to ensure the initial portionofthechainwasnotsampled.Thefinalposteriordis- Definingwetanddryseason tributions for each separate Markov chain was therefore made up of 1 000000accepted parameter combinations. The Dry season was defined using the soil water content data, posteriorparametervaluesandrangeswerecalculatedasthe including all days where the mean daily SWC was 50th, 15.9th and 84.1th percentiles of the 1 million accepted <0.12m3m(cid:1)3.Thisthresholdwassetasthelowerquartileof parameter combinations. These percentiles are equivalent to all the SWC data, which had an annual mean and SD of themeanandplusandminusoneSDforalog-normaldistri- 0.17(cid:3) 0.04m3 m(cid:1)3. In total 733 of 2922 study days were bution. For data storage purposes the output from 1000 of definedasdryseason.Thewet-dryseasondivisionwasusedto the 1 million accepted model runs was randomly selected define when the assimilation switched between wet and dry andsaved. seasonmodelparametersfortheallocation,turnovertimeand respirationparametersfortheautotrophiccarbonpools(foliar carbon(C),carboninwood(C )andcarboninfineandcoarse Assimilateddata f w roots (C , C )). This seasonal shift meant that the DA could fr cr Eddy covariance flux data. Eddy covariance data on a half adjustecosystemdynamicsacrossseasons,testingthehypothe- hourly time-step from 2004 to 2011 were available from a sesthatseasonalvariationinparameterswouldbetterenable towerlocated<50mfromourstudysites.Thereisadetailed themodeltoreplicatetheobservedpatternsinthedata. methodology published for the set-up of the tower (Bonal etal., 2008). The NEE data were processed using ALTEDDY Dataassimilationmethodology software (http://www.climatexchange.nl/projects/alteddy/) and standard quality control checks were used to filter the TheDAschemeoptimized36parameters.Theseincludesepa- data(Fokenet al.,2005).Followingallnight-timeNEEdatafor rate parameters for the wet and dry season allocation and which u* values were <0.15ms(cid:1)1 were filtered out (Bonal turnover rate and respiration parameters for the autotrophic etal., 2008). As some spurious spikes were still visible in the pools were included in these 36 parameters (Table 1). halfhourlycarbonflux(FC)andcarbonstoragedata(SFC)all A Metropolis-Hastings scheme was used to estimate the pos- values of SFC and FC greater than 10 SDs were filtered out teriordistributionofmodelparameters(Knorr&Kattge,2005). fromthedata(inbothcases<0.11%ofthedatawerefiltered). Weassumeobservationerrorsondifferentdatastreamstobe To create daily values of NEE and limit the use of gap-filled uncorrelatedandthereforeminimizethefunction: data,onlydayswith≥40halfhoursperdaywereused.Miss- L¼e(cid:1)Mf ing values for these days were replaced with the mean day- time or night-time value for that day, before fluxes were where L is the likelihood of the model parameters given the summed. From 2004 to 2011, 497 daily values of NEE were dataandMfisthemodeldatamissfit.Mfisdeterminedby: available. Errors for the NEE data were derived from ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 EVIDENCE FOR STRONG SEASONALITY 983 P),priorlowerestimate(PL)andpriorupperestimate(PU),thensandsourcesofthepriorsestimatesfortheDALEC-FGmodel.arametervaluesareshownfollowedbythedryseasonposterior SourceofPosUprior 431EstimatedfromLMAdata&LAIdata23186Seemethodssection568Galbraith20104610Galbraith2010etal.474Malhi2009*2649Seemethodssection 451952006 etal.0.42(0.33)Malhi2009*etal.0.26(0.18)Malhi2009*etal.0.33(0.47)Malhi2009* etal.0.10(0.09)Malhi2009* e-3)1.7e-3(2.2e-3)EstimatedfromLMAandlitterfall(seeMethods)etal.e-5)2.6e-5(3.1e-5)Rutishauser2010 e-4)5.7e-3(2.2e-3)2006 Te-5)6.4e-5(4.5e-5)Assumedtobethesameasw etal.2.0e-3Metcalfe2010 1.3e-4CarbonlostfromCWDperyearwascalculatedusingdecayrateequationsfrometal.(Herault,2010).Assuming75%ofetal.decayedCWDisrespired(Chambers,2001)thetimetodecaywholeCWDpoolbasedon18%ofcarbonlosttosoilpoolwasthencalculated.0.79(0.99)DefaultassumptionforfractionofrespiredcarboninACM0.66(0.83) Table1ParameterdescriptionsfortheDALEC-FGmodel,includingtheirsymbols(s),units,priorvalue(posteriormedian(Pos),the15.9th(PosL)and84.1th(PosU)percentilesontheposteriorparameterdistributioForallocation,turnoverrateandrespirationparametersfortheautotrophicpools,thewetseasonposteriorpparametervaluesinbrackets ParameterSUnitsPPLPUPosPosL (cid:1)2CInitialfoliageCstockgCm384299493421411f(cid:1)2CgCm2355318343302432209321015InitialwoodCstockw(cid:1)2CgCm371289476469373InitialfinerootCstockfr(cid:1)2CInitialcoarserootCstockgCm1593966262729701814cr(cid:1)2CgCm300182495358264InitiallitterCstocklit(cid:1)2CgCm17381354223219481550InitialcoarsewooddebrisCcwdstock(cid:1)2CInitialsoilorganicmatterCgCm2900022585372373682030368somstockAFractionofGPP0.430.260.710.40(0.31)0.38(0.30)AllocationfractiontofoliagefAAllocationfractiontowoodFractionofGPP0.260.160.430.24(0.18)0.22(0.17)wAFractionofGPP0.230.140.370.29(0.45)0.25(0.41)AllocationfractiontofinefrrootsAAllocationfractiontocoarseFractionofGPP0.080.050.130.06(0.06)0.04(0.04)crroots†TFractionofpool2.4e-31.8e-33.0e-31.7e-3(2.1e-3)1.6e-3(2.0Turnoverrateoffoliagefperday†TTurnoverrateofwoodFractionofpool2.5e-51.9e-53.2e-52.2e-5(2.4e-5)1.8e-5(1.9wperday†TFractionofpool1.4e-36.5e-42.9e-34.5e-3(1.5e-3)3.5e-3(9.1Turnoverrateoffinerootsfrperday†TTurnoverrateofcoarserootsFractionofpool2.5e-51.5e-54.1e-53.8e-5(2.8e-5)2.1e-5(1.8crperday†DTurnoverrateoflitterFractionofpool1.0e-34.7e-42.1e-31.1e-36.7e-4litperday†DFractionofpool4.4e-52.1e-59.3e-58.6e-54.9e-5TurnoverrateofCWDcwdperday RAAFractionofRespiredfractionof0.500.300.820.78(0.96)0.77(0.93)fFffperdayRA0.500.300.820.61(0.80)0.57(0.77)RespiredfractionofwFw ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 984 L. ROWLAND etal. respired respired respired asasforCWDisheheCWD Sourceofprior DefaultassumptionforfractionofcarboninACMDefaultassumptionforfractionofcarboninACMDefaultassumptionforfractionofcarboninACMSettoACMdefault CarbonlostfromCWDperyearwD.Assuming75%ofdecayedcwdetal.respired(Chambers,2001),tfractionofcarbonrespiredfromtpoolperdaywascalculated.SettoACMdefault 7) 4) 3 8 osU 53(0. 97(0. 7e-3 9e-4 8e-5 P 0. 0. 1. 2. 7. 9) 4) 2 4 osL 36(0. 61(0. 9e-4 7e-4 2e-5 P 0. 0. 4. 1. 5. nly.365). o/ Pos 0.46(0.33) 0.89(0.65) 9.2e-4 2.3e-4 6.4e-5 Manaussitesvertime(yrs) U 82 82 1e-3 2e-4 1e-4 andurno P 0. 0. 2. 4. 2. ~na/(t a1 PL 0.30 0.30 .7e-4 9.4e-5 4.7e-5 eCaxiuerrate( hv 0 0 e-3 e-4 e-4 mtrno 5 5 0 0 0 ou P 0. 0. 1. 2. 1. sfrat Afw Afr Acr pool pool pool eragedelas of of of of of of avmo Units FractionperdayFractionperdayFractionperdayFractionperdayFractionperday Fractionperday ulatedasintothe S RFfr RFcr RFlit RFcwd RFsom arecalcnserted 9)ei 0r 0a (2rs Afr Acr Clit wd om etal.mete of of of Cc Cs hira d) n n n n n alpa e o o o o o M Table1(continu Parameter Respiredfracti Respiredfracti Respiredfracti Respiredfracti Respiredfracti *Valuesfrom†Turnoverrate ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 EVIDENCE FOR STRONG SEASONALITY 985 previouslypublishedmethodologies(Hollinger&Richardson, accumulation,whichwas0.40(cid:3) 0.09(where(cid:3)indicatesSE). 2005;Hillet al.,2012)(seeSupportinginformation). These data were assimilated annually to provide the model withinformationoftheapproximatemagnitudeanddirection Foliardata. Leafrespirationmeasurementswereavailableon oftheseasonalchangeinwoodybiomassallocation. ourstudyplotsfromtwostudies(Stahletal.,2013;Zaragoza- Castells et al., unpublished results). The data included the Heterotrophic respiration data. Respiration from coarse average and SD of leaf respiration in dark conditions from woody debris (R ) was estimated from 429 measurements cwd fully sunlit leaves for 52–70 leaves measured in November made on 33 samples during 13 periods from July 2011 to 2007, July 2008 and November 2008 (Stahl et al., 2013) and November2011(Rowlandetal.,2013).Fulldetailsofmeasure- from 70 leaves for the dry season of 2010 (Joana Zaragoza- mentsandmethodusedtoscaletheR measurementstoa cwd Castells, unpublished data). Leaf respiration data were plotlevelareavailableinRowlandet al.(2013). adjustedtothemeandailytemperatureoverourstudyperiod Automatic soil respiration (R) data at the study site were s (25.6°C).Thesedatapointswereadjustedtoacanopyaverage measuredfromApril2005toDecember2006(Bonalet al.,2008 value by scaling respiration values according to changes leaf and Ponton & Bonal, unpublished data). R was measured s respirationbetweensunlitandshadedleaves(seeSupporting everyhalfhouronthestudysiteusingfourautomatedcham- Information). bers(Bonaletal.,2008).Thechamberswereplacedontopof Mean LAI and SD were estimated from measurements thesurfacelitterandrespirationmeasurementsthereforerep- made with the Li-2000 (Licor, Lincoln, NE, USA) at between resentthecombinedrespirationfromsurfacelitter,rootlitter 37and49randomlyselectedlocationsperplotinMarch2005, androotandsoil.Halfhourlyvalueswerethenaveragedinto November2005,November2008,September2010,March2011 dailyvalues.ErrorwasderivedfromtheSEonthefour-cham- and September 2011. LAI was compared to model output bermeasurements.Datawereonlyusedwhenthreeormoreof using an estimate of leaf mass per area (LMA) of the soil chambers recorded measurements (577days). There 122.07(cid:3) 2.23g Cm(cid:1)2 (where (cid:3) indicates SE), measured at was significant autocorrelation in the R data, this was s the site on 70 leaves (Zaragoza-Castells etal., unpublished removedbyfilteringthedatatoevery30days(GomezDans, results);weassumedhalfofthismasswascarbon. 2004)(n= 19).Tomaintainconsistencywiththeassumptions On our study sites litterfall was measured monthly from madeinthemodelledsoilmoistureresponse,weassimilateR s January2004toDecember2011usingfour1 m2littertrapson datawhichhasbeenseparatedintoautotrophicandheterotro- eachplot.Materialwascollected,driedtoaconstantmassand phiccomponents,describedearlierinthemethods. thenweighed. Soil water content data. Soil water content data were taken Woody stem data. Respiration fromstems wasmeasuredon every 30min from two probes at the study sites. For 2004– our study plots (Stahl etal., 2011); stem respiration measure- 2008, data were available from a frequency domain sensor ments were made over 11 periods, during both wet and dry (CS615; Campbell Scientific Inc., North Logan, UT, USA) at season,betweenSeptember2007andFebruary2009.Themean 0.05m depth 15m from the flux tower. Data were available andSEofthesemeasurementswerescaledtoplotlevelusing fromasecondfrequencydomainsensor(CS616;CampbellSci- surface area of the stems and large branches per unit of entificInc.)insertedat0.10mdepth,10mfromthefluxtower groundarea(stemareaindex,SAI;Chambersetal.,2004;Rob- for2007–2011.Thesedatasetswereaveragedintodailyvalues ertsonetal.,2010).Theerroronstemrespirationwasderived andcorrectedfortheeffectsofdifferentprobedepth(seeSup- from the measurement error, following scaling and therefore portingInformation). we assume that the scaling error was captured by the measurementerror. Steady-state observations, error estimation and model A census of the diameters of all trees ≥10cm diameter at output. Themodelinitsstandardformmakesnoassumption breastheight(DBH,1.3m)wasconductedin2004,2006,2008 ofsteadystate.Theseprimaryforestsarelikelytoberelatively and2010.Thesemeasurementswereusedtoestimatethetotal close to steady state over decadal timescales. Therefore, to aboveground biomass of the plots using a biomass equation ensure that the modelled carbon pools were close to steady for tropical moist forests (Chave et al., 2005), which included state, we assimilated seven additional pseudo-observations treeheight;treeheightwascalculatedfromdiameterusinga which were the change in size of each of the seven carbon countryspecificequation(Feldpauschetal.,2011).Asnoerror pools in the DALEC-FG model. These observations had a estimation existed for biomass, a SE of 10% of the biomass valueof0andaSDof2%ofthesizeofthepool.Thissolution valuewaspassedintotheDA. was necessary because computational limits prevented run- Treediametergrowthdataweremeasured32timesfor114 ning the model until it was in steady state, as part of the trees on a monthly to bimonthly basis from 2007 to 2010 on assimilationprocess. our study plots (Wagner etal., 2012). Growth data were not SEwasusedasanestimateofuncertaintyontheassimilated scaledtoplotlevelbyWagneret al.(2012)whostatedthatthe data (Richardson etal., 2010). When combining errors (e.g. treestheymeasuredwerenotrepresentativeofthesizestruc- multiplyingleafrespirationbyLAI),theerrorswereassumed tureoftheforest.The11dryseasonand21wetseasongrowth to be random and uncorrelated (Hughes & Hase, 2010). The data measurements from 114 trees from Wagner etal. (2012) numberofdatapointsforeachassimilateddatastreamandthe wereusedtocalculatetheratioofdrytowetseasonbiomass averageerrorforeachdatastreamareshowninTable 2. ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 986 L. ROWLAND etal. Table2 Thenumberofdatapointscontributingtoeachdata the default values from the DALEC model (Williams et al., stream used in the DA and the average error on these data 2005).All the prior valueswere assigneda SD of 0.25, 0.5or (SE,gCm(cid:1)2 d(cid:1)1) 0.75inlog-normalspace(Knorr&Kattge,2005);Table 1).SD values were assigned based on an assessment of the uncer- Datastream No. SE taintyofthedatasourceandoncreatingrealisticlimitsonthe meanestimate. Netecosystemexchange 497 2.66 Leafrespiration 4 0.76 Leafareaindex 6 0.44 Results Litterfall 112 0.20 Stemrespiration 11 0.08 TheresultsoftheanalysisshowthatmeanannualGPP Abovegroundbiomass 4 2258.35 is 3756.7 (cid:3) 19.1 gC m(cid:1)2 yr(cid:1)1, 9.1% greater than R eco Soilrespiration 19 0.52 (3415.3 (cid:3) 38.5 gC m(cid:1)2 yr(cid:1)1); demonstrating that this Coarsedeadwoodrespiration 13 0.07 forest stores carbon on an annual basis. However, our analysis demonstrates that the strength of the carbon sink increases by approximately four times from wet Priorinformation (NEE: (cid:1)0.54 (cid:3) 0.12 gC m(cid:1)2 d(cid:1)1) to dry season Wherepossiblepriorsonstatesandparameterswerebasedon (NEE: (cid:1)2.1 (cid:3) 0.15 gC m(cid:1)2 d(cid:1)1; Table 3; Fig. 3). The data from published sources and unpublished data from the increased strength of the sink was caused by a 0.79 (cid:3) studysite.Wheresitedatawerenotavailable,estimatesfrom 0.07 gC m(cid:1)2 d(cid:1)1increaseinGPPinresponsetohigher nearby sites in northern Brazil were used. Where no data dry season radiation and a simultaneous decrease existedtheparametersweresettoabestapproximationorto of 0.78 (cid:3) 0.20 gC m(cid:1)2 d(cid:1)1 in R . The effects of eco Table3 ThemeancarbonpoolsandfluxespredictedbytheDAanalysisforstudysitefrom2004to2011.Dataareshownasmean valuesforwetanddryseasonandasmeanannualsums.Thevaluesarecalculatedfrom1000randomlyselectedDAmodelruns andshownalongsidetheSDacrossthesemodelruns(SD) Wetseason Dryseason Annual Mean SD Mean SD Sum SD Allocation gCm(cid:1)2 d(cid:1)1 gCm(cid:1)2 yr(cid:1)1 A 4.01 0.19 3.42 0.18 1413.1 54.9 f A 2.36 0.12 1.88 0.07 818.5 38.5 w A 3.04 0.22 4.84 0.22 1272.6 61.7 fr A 0.64 0.14 0.71 0.18 252.5 43.7 cr Respiration gCm(cid:1)2 d(cid:1)1 gCm(cid:1)2 yr(cid:1)1 R 3.13 0.18 3.27 0.15 1158.9 54.2 f R 1.48 0.03 1.53 0.03 544.2 8.5 w R 1.42 0.17 1.40 0.15 501.3 53.7 fr R 0.49 0.16 0.64 0.14 210.8 54.9 cr R 0.40 0.09 0.26 0.06 130.8 30.2 lit R 0.41 0.02 0.26 0.01 134.5 6.5 cwd R 2.23 0.16 1.43 0.11 735.0 54.6 som Ecosystemfluxes gCm(cid:1)2 d(cid:1)1 gCm(cid:1)2 yr(cid:1)1 NEE (cid:1)0.54 0.12 (cid:1)2.11 0.15 (cid:1)341.4 36.3 GPP 10.09 0.05 10.87 0.05 3756.7 (cid:3)(cid:3) R 9.55 0.13 8.77 0.15 3415.3 38.5 eco R 6.53 0.17 6.83 0.14 2415.1 49.7 a R 3.02 0.12 1.93 0.08 1000.2 39.1 h CUE 0.35 0.02 0.37 0.01 0.36 0.02 Stocks gCm(cid:1)2 gCm(cid:1)2 C 398 8 397 8 398 8 f C 22376 1225 22362 1217 22373 1223 w C 465 57 520 52 480 56 fr C 2842 717 2841 714 2842 717 cr C 524 63 530 63 525 64 lit C 2181 364 2179 364 2181 364 cwd C 29579 5668 29462 5676 29550 5670 som ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 EVIDENCE FOR STRONG SEASONALITY 987 (a) (b) (c) (d) (e) (f) (g) (h) Fig.3 Comparisonofdata(blackpoints,shownwithstandarderrorbars)withmodeloutputfromtheDA.Medianresults(redline) withthe15.9thandthe84.1thpercentiles(redshadedarea),whichrepresent1SDfornongaussiandistributions,areshownforthe resultsoftheDA.Thegreyshadedareaindicatestheperiodsclassifiedasthedryseason. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) Fig.4 BoxplotsoftheDAposteriorparameterestimatesfortheallocation(a–c),turnover(d–e)andrespiration(f–i)parameterswhich showeddryandwetseasondifferences.Thegreyshadedareashowsthepriorrangesfortheparametervalues(seeTable1).PanelJ showstheeffectoftheseparameterchangesonthemodelledautotrophicrespiration(R,gCm(cid:1)2d(cid:1)1)inthewetanddryseason(left), a relativetotheseasonalchangeintheheterotrophicrespiration(R ,gCm(cid:1)2d(cid:1)1;right). h decreasing respiration and increasing GPP were there- Mean annual R from the analysis was 2415 (cid:3) a fore equally important for the seasonal change in the 50 gC m(cid:1)2 yr(cid:1)1,morethantwicethesizeoftheannual net carbon flux of this ecosystem. The seasonal reduc- R (1000 (cid:3) 39 gC m(cid:1)2 yr(cid:1)1; Table 3). The R : R ratio h h a tioninR wascausedbyareductioninheterotrophic decreased from 0.46 (cid:3) 0.02 in the wet season to eco respiration(R ),whichnotonly causedthe decreasein 0.28 (cid:3) 0.01 (Table 3, Fig. 4). This seasonal change was h R butalsocompensatedforanincreaseinautotroph- causedbythe36%reductionindryseasonR .TotalR eco h a icrespirationof0.30 (cid:3) 0.22 gC m(cid:1)2 d(cid:1)1(R ;Table 3). onlyincreasedby4%fromwettodryseason;however, a The analysis tightly constrained (SDs <10% of the the reduction in dry season R resulted in R compris- h a mean) the GPP, R , R , R and CUE fluxes (Table 3). ing 80% of the dry season R . Mean annual carbon eco a h eco ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991 988 L. ROWLAND etal. use efficiency (CUE) was 0.36 (cid:3) 0.02, but increases sequestered in the wet than the dry season in the sea- fromwettodryseasonby5.38 (cid:3) 0.3%. sonal tropical forest studied, and that there are signifi- On an annual basis similar proportions of GPP were cantseasonalchangesincarbonallocation,andCUE. allocated to foliage (37.7 (cid:3) 1.5%) and fine roots Thefourfoldincreaseinthenetcarbonsequestration (33.9 (cid:3) 1.7%;Table 3).TheremainderofGPPwasallo- (391.1 (cid:3) 91.2%decreaseinNEE;Table 3)indryseason cated to stem wood (21.8 (cid:3) 1.0%) and coarse roots wastheresult ofthe responseofheterotrophicrespira- (6.7 (cid:3) 1.2%). However, the division of carbon alloca- tiontosoilmoistureandanincreaseinGPPinresponse tion among leaves, coarse wood (which includes both toincreasedsolarradiation.TheincreaseinNEEinthe stems and coarse roots) and fine roots varied dry season is larger than has been modelled for other significantly when analysed at a seasonal timescale. tropical humid forest sites in northern Brazil (Baker The results of the DA indicate increased allocation of et al., 2013). Our estimated values of annual R and R a h carbon to coarse wood and foliage in the wet season, weresimilartoestimatesfromempiricalbottom-upnet and greater allocation to fine roots in the dry season carbon flux studies elsewhere in eastern Amazonian (Fig. 3; Tables 1 and 2). These changes were driven by forests (Malhi et al., 2009; Metcalfe et al., 2010). The significant changes to the allocation parameters from reduction inR fromwetto drywas driven bya mod- h the wet to dry season; A and A decreased elled response to reduced soil water availability (see f w 22.5 (cid:3) 3.1% and 25 (cid:3) 4.4%, respectively, from wet to Methods). Without this modelled moisture response, dry season, whereas A increased 35.5 (cid:3) 10% (Fig. 4, R increasedinthedryseasoninresponsetoincreased fr h Table 1). dry season temperature (data not shown) and conse- There were distinct seasonal differences in nine of quentlytheseasonalityofthesoilrespirationwasincor- the12parametersassociatedwiththeautotrophicpools rectlysimulated,resultinginanunderestimationofdry (Fig. 4). Increases in the respired fraction of the foliar season carbon sequestration and an inability to match andwoodpoolsfromwettodryseason(18.75 (cid:3) 1.3%, theseasonalityofNEE. and 23.75 (cid:3) 3.9% respectively) were contrasted by The low wet to dry season variation in average GPP decreases in the fraction respired from the fine and (Table 3) and the stronger variation in R matched eco coarse root pools (28.3 (cid:3) 12.5% and 27.0 (cid:3) 19.9% patterns observed by Bonal et al. (2008) at this site. In respectively). The analysis predicted high uncertainty 2004, our GPP estimate was 2.74% greater and in 2005, (SD ≥ 40%ofthemean)forcertainparameters:theallo- 5.74% greater than previously estimated from eddy cation of carbon to coarse roots, and the turnover of covariance data at the site (Bonal et al., 2008). In con- coarse and fine roots, and coarse dead wood and litter trast, our R estimates were 3.37% lower in 2004 and eco (Fig. 4andTable 1).Theerrorsontheposteriorparam- 1.41% lower in 2005 than estimates from Bonal et al. eterdistributionsandthesimulatedmodeloutputasso- (2008). Considering the errors associated eddy covari- ciated with both the fine and coarse root pools were ance measurements (Bonal et al., 2008; Hutyra et al., consistentlygreaterthanthoseassociatedwiththefoli- 2008) these differences are low. However, such differ- age and stem pools (Table 1 and 2; Fig. 4). However, ences result in our estimates of carbon sequestered by despiteasignificantincreaseintheturnoverrateoffoli- thisecosystembeing2.18timesgreaterin2004and1.58 age and therefore litterfall in the dry season (Fig. 4d), times greater in 2005 than previously estimated by theDAstillremainedunabletosimulatethehighlitter- eddy covariance data (Bonal et al., 2008). However, in fall values which occurred at this site during a 1– thisstudy,weareabletodeterminewithanassessment 2 monthperiodinearlytomiddryseason(Fig. 3e).The of uncertainty, the importance of the seasonality of R , h litterfalldatathereforeremainedthemostpoorlyfitted GPP and components of R for altering carbon a datainthisstudy(Fig. 3e). sequestrationandCUEestimatesoftropicalforests. Carbon use efficiency (0.36 (cid:3) 0.02) was lower than temperate forest values of ca. 0.5 (Waring et al., 1998) Discussion and closer to the CUE values proposed for two undis- ThisisthefirststudywhichusesDAtooptimizesepa- turbed old-growth forests in the eastern Amazon ratewetand dryseason parametersin atropical forest (0.34 (cid:3) 0.10and0.34 (cid:3) 0.07;Malhiet al.,2009).The5% andtoinvestigatehowfluxesfromdifferentforestcom- increase in CUE in the dry season was caused by a ponents contribute to seasonal changes in net ecosys- greater dry season increases in GPP (8%) than in R a tem carbon flux. The implementation of seasonal (4%; Table 3) suggesting that this forest is more effi- variations in parameters provides a mechanism cient at investing carbon in the dry season, when GPP through which the DALEC-FG carbon model is able to iselevatedbecauseofhighersolarincidentradiation. better simulate the observed patterns in flux data. The The relatively even annual distribution of GPP analysis determines that four times more carbon is between foliage, fine root and coarse wood (stems and ©2013TheAuthorsGlobalChangeBiologyPublishedbyJohnWiley&SonsLtd.,GlobalChangeBiology,20,979–991