Quantifying and understanding carbon storage and sequestration within the Eastern Arc Mountains of Tanzania, a tropical ...

PAGE 1

RESEARCHOpenAccessQuantifyingandunderstandingcarbonstorage andsequestrationwithintheEasternArc MountainsofTanzania,atropicalbiodiversity hotspotSimonWillcock1,2*,OliverLPhillips1,PhilipJPlatts3,AndrewBalmford4,NeilDBurgess5,6,JonCLovett1, AntjeAhrends7,JulianBayliss4,NikeDoggart8,KathrynDoody9,EibleisFanning10,JonathanMHGreen11, JaclynHall12,KimLHowell13,RobMarchant3,AndrewRMarshall3,14,BonifaceMbilinyi15,PantaleonKTMunishi15, NishaOwen10,16,RuthDSwetnam17,ElmerJTopp-Jorgensen18andSimonLLewis1,19AbstractBackground: Thecarbonstoredinvegetationvariesacrosstropicallandscapesduetoacomplexmixofclimatic andedaphicvariables,aswellasdirecthumaninterventionssuchasdeforestationandforestdegradation.Mapping andmonitoringthisvariationisessentialifpolicydevelopmentssuchasREDD+(ReducingEmissionsfrom DeforestationandForestDegradation)aretobeknowntohavesucceededorfailed. Results: WeproduceamapofcarbonstorageacrossthewatershedoftheTanzanianEasternArcMountains(33.9 millionha)using1,611forestinventoryplots,andcorrelationswithassociatedclimate,soilanddisturbancedata.As expected,tropicalforeststoresmorecarbonperhectare(182MgCha-1)thanwoodysavanna(51MgCha-1).However, woodysavannaisthelargestaggregatecarbonstore,with0.49PgCover9.6millionha.Weestimatethewhole landscapestores1.3PgC,significantlyhigherthanmostpreviousestimatesfortheregion.The95%Confidence Intervalforthismethod(0.9to3.2PgC)islargerthansimplerlook-uptablemethods(1.5to1.6PgC),suggesting simplermethodsmayunderestimateuncertainty.Usingasmallnumberofinventoryplotswithtwocensuses( n =43) toassesschangesincarbonstorage,andapplyingthesamemappingprocedures,wefoundthatcarbonstorageinthe tree-dominatedecosystemshasdecreased,thoughnotsignificantly,atameanrateof1.47MgCha-1yr-1(c.2%ofthe stocksofcarbonperyear). Conclusions: Themostinfluentialvariablesoncarbonstorageintheregionareanthropogenic,particularlyhistorical logging,asnotedbythelargestcoefficientofexplanatoryvariableontheresponsevariable.Ofthenon-anthropogenic factors,anegativecorrelationwithairtemperatureandapositivecorrelationwithwateravailabilitydominate,having smallerp-valuesthanhistoricalloggingbutalsosmallerinfluence.Highcarbonstorageistypicallyfoundfarfromthe commercialcapital,inlocationswithalowmonthlytemperaturerange,withoutastrongdryseason,andinareas thathavenotsufferedfromhistoricallogging.Theresultsimplythatpolicyinterventionscouldretaincarbon storedinvegetationandlikelysuccessf ullysloworreversecarbonemissions. Keywords: EasternArcMountains;Tanzania;IPCCTier3;REDD+;Forest;Disturbance;Degradation;Ecosystemservice *Correspondence: S.P.Willcock@soton.ac.uk1SchoolofGeography,UniversityofLeeds,LeedsLS29JT,UK2SchoolofBiologicalSciences,UniversityofSouthampton,Southampton SO171BJ,UK Fulllistofauthorinformationisavailableattheendofthearticle 2014Willcocketal.;licenseeSpringer.ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommons AttributionLicense(http://creativecommons.org/licenses/by/2.0),whichpermitsunrestricteduse,distribution,andreproduction inanymedium,providedtheoriginalworkisproperlycredited.Willcock etal.CarbonBalanceandManagement 2014, 9 :2 http://www.cbmjournal.com/content/9/1/2

PAGE 2

BackgroundTropicalforestsaregloballysignificantecosystems;accountingfor~50%ofglobalforestarea[1],storing 45%ofallcarboninterrestrialvegetation[2-4],maintaininghighbiodiversity[5],andprovidingecosystem services,suchastimber,non-timberforestproducts[6], andclimatechangemitigation[7,8].However,withinthe lastfewdecades,vastareasoftropicalforestshavebeen convertedtootherland-usesordegraded.Forexample, between1990and1997,4.4-7.2millionhectaresof humidtropicalforestwereconvertedeachyearandan additional1.6-3.0millionhectaresofforestwerevisibly degraded[9].Thisprocessincreasedintheearly2000s, withanestimated5.1-5.7millionhectaresofhumid tropicalforest(and3.5-4.7millionhectaresofdrytropicalforest)deforestedperyearbetween2000and2005 [10].Thegradualandsustainedreductioninforestqualityandquantityhasresultedinsubstantialemissionsof CO2[11].Globally,deforestationandforestdegradation accountedfor6-20%ofanthropogenicGHGemissions inthe1990sandearly2000s[12-14].Tropicalregions makeasubstantialcontributiontothis,emitting0.7-1.5 PgCyr-1between1990and1999[9,15-17]and0.71.5 PgCyr-1between2000and2007[13,16-18].Theseprocessesalsoimpactthefuturepotentialofforeststoremovecarbonfromtheatmosphere[7,19,20]. Recently,attemptstomitigateincreasinganthropogenicCO2emissionsthroughreducingemissionsfrom degradationanddeforestation(REDD+)havebeeninstigated[21].TheREDD+programmeisaimedatcontributingtoareductioningreenhouseemissionswhilst providingeconomicincentivesforbettermanagement andprotectionofforests.Thispolicyhasbeenwidely welcomedandmayprovideafinancialincentivetosignificantlyreducecarbonemissions[22,23],althoughthe equityandjusticeissuessurroundingtheimpactonlocal livelihoodsareactivelydebated[24,25].KeytechnicalissuesforthesuccessfulimplementationofREDD+include(butarenotlimitedto)theaccuracyofmonitoring systems,preventingleakageandestablishingaccurate historicalbaselines.Thus,thesuccessofREDD+,inpart, restsonrobustscientificinformationonthemagnitude andextentofcarbonstorageintropicalregionsandhow itchangesovertime[26]. TheIntergovernmentalPanelonClimateChange(IPCC) provideathree “ Tier ” systemthroughwhichcarbonstocks andemissionscanbereported,eachwithadifferentlevel ofmethodologicalcomplexityandaccuracy.Tier1isthe simplestmethod,usingglobaldefaultvaluesobtainedfrom theIPCCliterature[27,28].TheintermediateTier2level improvesonTier1byusingcountryspecificdata.Tier 3isthemostrigorousapproach,usinglocalforestinventorydata,focusingonthedirectmeasurementof trees,repeatedoveratimeseries[27-29].Herewe developaTier3methodologyfortheEasternArc Mountains(EAM)watershedarea. Theestimatesbecomeprogressivelymorerobustfrom Tier1to3duetochangesintwomainsystematicerrors [29].Thefirst,completeness,referstothenumberof IPCCcarbonpoolsthatareincluded,withstudiesincludingallfivepools(abovegroundlive,litter,coarse wooddebris[CWD],belowgroundandsoilcarbon)consideredcomplete.Thesecond,representativeness,derivesfromthesubstantialnaturalvariabilityinthe carbonstoredacrosslandscapes,evenwithinabiomeor country[30].Theabovegroundbiomassofaforest withinalandscapemaydifferconsiderablyfromglobal default(Tier1)valuesorevenfromcountry-specific (Tier2)values.Forexample,inthePeruvianAmazon, datafromtheLosAmigosConservationConcession[31] wereshownnottoberepresentativeofforestsnationally.Nearbyforestssituatedtothenorthandsouthof thislocalstudyareestimatedtocontain20-35%lesscarbonperunitarea[32],suggestingthatLosAmigosConservationConcessionisanareaoflocallyhighbiomass. SinceTier3methodsaccountforvariationobserved withinbiomesandcountries,therepresentativenessof thecarbonestimatesishigherthanthoseassociatedwith Tier1and2methodologies[32,33]. However,Tier3methodsaremoreexpensive[34,35] andsomenationsmaylackthecapacitytoadoptsuch methods[36].Whilst,insomecases,thecapabilityto applyTier3guidelinesisbeingrapidlydeveloped,multitemporalinventorydataanddataonhistoricalcarbon stockchangescantakeseveraldecadestoaccrue[37,38]. ItisexpectedthatREDD+requirementswillallowdata provisionsfromseveraltiersinasinglereport.Highly variableand/orsubstantialcarbonpoolsshouldbeestimatedusingTier3methodology(e.g.forestaboveground livecarbon[ALC]),whilstTier1orTier2methodology maybesufficientforsmallercarbonpools(e.g.CWD)or carbonpoorlandcovercategories(e.g.bareground). InTier3methods,inordertoextrapolatefromplot data,itisnecessarytodevelopcorrelationswithremotelysenseddatatoscaletothestudyareaorcountrywideestimates.Generally,carbonstorageiseitherestimatedviastatisticalcorrelationwithelectromagnetic properties,ground-truthedbyvolumetricmeasurements, suchasdiameteratbreastheight(DBH),whichareconvertedtobiomassestimatesusingallometricequations. Avarietyofremotelysenseddatasourceshavebeen employedforcarbonmappingandthesecanbeaggregatedintofourgroups:photographicimagery,RADAR, LiDAR,andancillarygeographicinformationsystems (GIS)data(seeAdditionalfile1:SI1foranevaluationof eachmethod).Here,weuseancillaryGISdataassuch datahavethreemainadvantages:1)wideavailability,often freeofcharge;2)asuitableresolution(e.g.90m[39]);andWillcock etal.CarbonBalanceandManagement 2014, 9 :2Page2of17 http://www.cbmjournal.com/content/9/1/2

PAGE 3

3)correlationswiththeseancillaryGISdatamayindicate whichvariablesdirectlyaffectcarbonstorage.Developing anunderstandingofhowthesevariablesinfluencecarbon storageisvitalforaccuratescenariosoffutureemissions. Here,wecorrelatecarbonstorageestimatesfromtree inventoryplots( n =1,611,mediansize=0.1ha)with dataonclimatic(e.g.temperature,precipitation,and solarradiation),edaphic(e.g.soilwaterholdingcapacity andsoilfertility)andproxyvariablesfordirecthuman interventions(e.g.governancetype,distancefromthe maineconomicdemandcentres,populationpressure, andhistoricallogging),andvariablesthatderivefrom climate-humaninteractions(e.g.burntareaindex)forthe TanzanianwatershedoftheEasternArcMountains(hereafter,EAM[40]),whichcovers33.9millionha(Figure1; seeSwetnametal(2011)[41]forfurtherdetails).WedevelopTier3typecorrelationequationstoestimatethe totalALCstoredacrosstheforestedandwoodedland covercategories,anadvancementonpreviousTier2estimatesfortheregionpresentedinWillcocketal(2012) [42].Additionally,weinvestigatethemostinfluentialcorrelatesofspatialdifferencesincarbonstorageandhow theseresultfromchangesineitherspeciescompositionaffectingwooddensity(specificgravity)orthenumberof largetreespresent.Lastly,asmallernumberofinventory plots(n=43,mediansize0.1ha)havetwocensuses,and byapplyingthesamemappingprocedures,weassess changesincarbonstorageovertime,providingafirstorderestimateofsequestrationacrosstheregion.ResultsCarbonstocksUtilising1,611plotsandscalingtothe33.9millionha studyareaweestimatethat1.32(95%confidence Figure1 TheEasternArcMountainsofTanzaniaandKenya[40]. ThestudyareaistheEasternArcwatershedinTanzania[41]. Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page3of17 http://www.cbmjournal.com/content/9/1/2

PAGE 4

interval[CI]rangesfrom0.89to3.16)PgCwasstored intheabovegroundlivevegetationintheyear2000 (Figure2;Table1).Woodlandandbushlandcontributedmosttotheamountofstoredabovegroundlive carbon(ALC)inthestudyregion,withopenwoodland storingthemostALC(0.49[0.47to1.60]PgCover9.6 millionha);followedbybushland(0.29[0.15to0.51] PgCover5.0millionha)andclosedwoodland(0.18 [0.13to0.61]PgCover1.8millionha). Bestestimatevaluesfromourmethodology,perunit area,ineachlandcoverclass,aregiveninTable2.Forest containedthegreatestALCperunitarea,withhighest valuesinsub-montaneforest(189[95to588]Mgha-1), followedbylowland(182[152-to360]Mgha-1),upper montane(166[69to533]Mgha-1),montane(130[62 to702]Mgha-1),andforestmosaic(121[55to485] Mgha-1).WoodlandsheldlessALCthanforests,with closedwoodlandstoring100(70to331)Mgha-1and openwoodlandstoring51(38to165)Mgha-1(Table2), butmorethanthelandscapeaverageof39(26to93) Mgha-1. Oursequestrationmodelsuggeststhatthelandscape maybelosing0.05(-0.07to0.26)PgCyr-1(meannetflux toatmosphereof1.47[-2.13to7.75]MgCha-1yr-1).Of the12.3millionhaoftree-dominatedlandinourstudy area,only1.4%(0.17millionha)showsacarbondecrease overtheentire95%CIrangeandonly0.8%(0.10million ha)adefinitecarbonincrease(Figure3).Thelocations showingnetcarbonuptakeareintheUdzungwamountains,whilethelocationswithnetreductionsincarbon storagearemainlyinthePareandUsambaramountains.LinksbetweencarbonstockandinfluentialvariablesThevariablesthatinfluencecarbonstorageandsequestrationmaybeinferredfromrelationshipswithinthecorrelationmodels.Forwardselectionresultsarepresentedin thefollowingparagraphsasthesebestindicatecausalrelationships[43-45].Ingeneral,backwardmodelswerein Figure2 Abovegroundlivecarbonstorageinthestudyarea(a),withupper(b)andlower(c)pixelbased95%CI. Seetextfordetails onMethods. Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page4of17 http://www.cbmjournal.com/content/9/1/2

PAGE 5

closeagreementwithforwardmodels(Tables3and4; Additionalfile1:TablesS1-S3). Carbonstorage(adjustedR-squared[AdjR-sq]=0.18) iscorrelatedpositivelywiththenaturallogarithmof thepopulationpressurewithdecayconstantof12.5km (p-value<0.001)andincreasedby1Mgha-1forevery 8700kmfromaroad(p-value<0.010),andevery30,000 unitsinthecostdistancetoDaresSalaam(p-value< 0.010).Carbonstoragedecreasedby1Mgha-1forevery 1Cincreaseinmeanannualmonthlytemperaturerange (p-value<0.001),every2.7%riseinthetotalavailable watercapacityofthesoil(p-value<0.001),andevery Table1Abovegroundlivecarbonstoredwithinthestudyareafortheyear2000,estimatedbythisandprevious studiesStudyAboveground livecarbon,Pg (95%CIrange) MethodologyResolution(m2)Disturbanceincluded?Tanzanian on-thegrounddata? Presentstudy* – Tier31.32(0.89-3.16)Correlationequationsderived usingremotelysensed influentialvariables. 100Anthropogenicvariablesrepresent humandisturbance.Natural disturbancevariablesalso included. Yes Willcocketal(2012)* – OriginalTier2[ 42 ] 1.58(1.56-1.60)Landcoverbasedlook-uptable.100Onlywherelandcovercategories areidentifiedasdisturbed (e.g.croplandmosaics). Yes Willcocketal(2012) – HarmonisedTier2[ 42 ] 1.64(1.52-1.76)Landcoverbasedlook-uptable.100Onlywherelandcovercategories areidentifiedasdisturbed (e.g.croplandmosaics). Yes Baccinietal(2012) – Tier1[ 3 ] 2.03DerivedfromMODISandGLAS LiDARdata. 500Partiallyincludesdisturbance throughimpactsoncanopy heights. Yes Saatchietal(2011) – Tier1[ 4 ] 0.83DerivedfromMODIS,SRTM, QSCATandGLASLiDAR. 1000Partiallyincludesdisturbance throughimpactsoncanopy heights. No Hurttetal(2006) HYDE-SAGE – Tier1[ 46 ] 0.63ModelledfromtheMiamiLU ecosystemmodelwithcropland datafromtheCentrefor SustainabilityandtheGlobal Environment. ~110,000Containssimplesubmodelsof naturalplantmortality,disturbance fromfire,andorganicmatter decomposition,aswellaswood harvesting. No Hurttetal(2006)HYDE – Tier1[ 46 ] 0.41ModelledfromtheMiamiLU ecosystemmodel. ~110,000Containssimplesubmodelsof naturalplantmortality,disturbance fromfire,andorganicmatter decomposition,aswellaswood harvesting. No Baccinietal(2008) – Tier1[ 47 ] 0.34DerivedfromMODISandGLAS LiDARdata. 1000Partiallyincludesdisturbance throughimpactsoncanopy heights. No*ThisstudyandWillcocketal(2012)arenotindependentastheyarederivedfromthesameunderlyingdataandutilisethesamelook-uptablevalues. Table2Themean(and95%CI)estimatesofforestcharacteristicsinvestigatedinthisstudy(carbonstorage,carbon sequestration,WSG,theinterceptfromthepowerlawrelationshipandthegradientfromthepowerlawrelationship) separatedbylandcovercategoryLandcovercategory[ 41 ]Carbonstorage (Mgha-1) Carbonsequestration (Mgha-1yr-1) WSG(gcm-3)Theinterceptfromthe powerlawrelationship Thegradientfromthe powerlawrelationship LowlandForest(<1000m) 182(152to360)-0.91(-7.08to4.29)0.60(0.59to0.60)6.01(2.94to5.17)-0.93(-1.04to-0.82) Sub-montaneforest (1000-1500m) 189(95to588)-2.02(-11.06to1.29)0.58(0.57to0.58)5.95(3.68to8.23)-1.31(-1.48to-1.14) MontaneForest (1500-2000m) 130(62to702)-2.03(-11.85to1.07)0.60(0.59to0.60)6.95(3.51to10.39)-1.57(-1.82to-1.32) Upper-montaneforest (>2000m) 166(69to533)-2.08(-10.49to1.23)0.60(0.58to0.60)7.03(4.60to9.45)-1.61(-1.93to-1.26) Forestmosaic 121(55to485)-1.18(-6.69to2.92)0.56(0.56to0.56)9.22(6.98to11.46)-1.90(-1.99to-1.81) ClosedWoodland 100(70to331)-1.24(-7.91to2.63)0.64(06.2to0.65)6.67(4.95to8.60)-1.55(-1.85to-1.30) OpenWoodland 51(38to165)-1.49(-7.53to2.05)0.61(0.59to0.62)6.38(4.88to7.82)-1.45(-1.70to-1.19) Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page5of17 http://www.cbmjournal.com/content/9/1/2

PAGE 6

4.4monthincreaseinthemeannumberofdry monthsannually(p-value<0.050).Carbonstoragewas 2.1Mgha-1lowerinareaswherehistoricallogging waspresent(p-value<0.010),and4.2Mgha-1higherin areasunderthecontroloflocalcommunities/governments (p-value<0.010).Thus,carbonstorageishighinareas farfromthecommercialcapital,withalowmonthly temperaturerange,withoutadryseason,thathavenot sufferedfromhistoricalloggingandareunderlocal community/governmentcontrol(Figure4;Table3). Therateofcarbonsequestrationcorrelatedwith threeprincipalcomponent(PC)axes(presentedin orderofinfluence;AdjR-sq=0.41).Carbonsequestrationwasnegativelycorrelatedwiththesoilfertilityaxis (PC5;p-value<0.050),warmertemperaturesandlonger dryseasons(PC3;p-value<0.050),andwithincreasedanthropogenicdisturbance(PC1;p-value<0.010).Thus,carbonsequestrationwashighestinlessfertileareaswith littleornodroughtandlittleanthropogenicdisturbance (Table4). Woodspecificgravity(WSG;AdjR-sq=0.28;see Additionalfile1:SI2)wasmoststronglyaffectedbythe annualmeanburnedareaprobability(increasingby 1gcm-3forevery0.04increase;p-value<0.001)and thetotalavailablewatercapacityofthesoil(decreasing by1gcm-3forevery82.0%increase;p-value<0.001). Thus,WSGishigherinburntareaswithlittleavailable water(Additionalfile2:FigureS1;Additionalfile3: FigureS2;Additionalfile1:TableS1). Theinterceptofthepowerlawrelationship(anindicationofpotentialstemdensity[seeAdditionalfile1:SI3]; AdjR-sq=0.30)wasmostaffectedbythenaturallogarithmofthepopulationpressurewithdecayconstantof 12.5km(positivecorrelation;p-value<0.001)andthe meanannualmonthlytemperaturerange(increasingby 1.0forevery1.2Cincrease;p-value<0.001).Thus,the Figure3 Abovegroundlivecarbonsequestrationintree-dominatedlandcovercategorieswithinthestudyarea(a),withupper(b)and lower(c)pixelbased95%CI. SeetextfordetailsonMethods. Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page6of17 http://www.cbmjournal.com/content/9/1/2

PAGE 7

densityofsmallerstemsincreasesinareaswithahigh populationpressureandlargetemperaturefluctuations (Additionalfile3:FigureS2;Additionalfile4:FigureS3; Additionalfile1:TableS2). Correlationsidentifiedfo rthegradientofthepower lawrelationship(anindicationoftheproportionof largerstems;seeAdditionalfile1:SI3)werebroadly theinverseofthoseidentifiedfortheintercept.The gradientofthepowerlawrelationshipwasmostaffectedbythenaturallogarithmofthepopulationpressurewithdecayconstantof20.8km(negativecorrelation; p-value<0.001)andthemeanburnedareaprobabilityin thefourthquarter(decreasingby1.0forevery0.2increase;p-value<0.001).Thus,theproportionoflarge stemswasgreaterinareasexperiencingfewdisturbancesfrompeopleorfire(Additionalfile3:FigureS2; Additionalfile5:FigureS4;Additionalfile1:TableS3). Wheninvestigatingthemostinfluentialcorrelatesof spatialdifferencesincarbonstorageandhowtheseresultfromchangesineitherspeciescompositionaffecting wooddensity(specificgravity)orthenumberoflarge treespresent,wefoundthatthefinalTier3carbonstorageestimateswerepositivelycorrelatedwithbothsizefrequencydistributionestimates(bothinterceptandgradient[p-values<0.001]),andnegativelycorrelatedwith WSGestimates(p-value<0.001)andmaximumheight estimates(p-value<0.001;Additionalfile1:seeSI4).All possibleinteractionswereinvestigatedandweresignificant(AdjR-sq=0.35;p-values<0.001),however,the majorityoftheexplanatorypowerlaywithinthesecond orderinteractions(AdjR-sq=0.33;p-values<0.001; Additionalfile1:TableS5).Broadly,WSGandtheproportionoflargerstemshadlargestinfluenceoverthe carbonstorageestimate.Consideringonlysecondorder Table3Thecoefficientsandassociatedp-valuesofthevariablescorrelatedwithabovegroundcarbonstorageusing bothforwardandbackwardselectionproceduresVariable(whereappropriate,unitsaregiveninbrackets)GroupForwardBackward Coefficientp-valueCoefficientp-value (Intercept) n/a-1.21E+033.14E-03-2.80E+007.55E-01 Naturallogarithmofthepopulationpressurewithdecayconstantof12.5km Anthropogenic1.06E+001.06E-051.42E+002.27E-06 Naturallogarithmofthepopulationpressurewithdecayconstantof16.7km Anthropogenicn/an/a1.42E+002.27E-06 Distancetoroads (km)Anthropogenic1.15E-041.09E-031.78E-041.30E-05 Historicallogging – Partiallylogged (nologging/partiallylogged)Anthropogenic-2.10E+001.09E-03-3.83E+004.97E-07 CostdistancetoDaresSalaam Anthropogenic3.41E-052.00E-032.58E+005.46E-03 Naturallogarithmofthecostdistancetomarkettowns Anthropogenic-6.05E-015.24E-02-9.85E-011.89E-02 Governance – local (national/local/joint/unknown)Anthropogenic4.24E+009.29E-03n/an/a Governance – national (national/local/joint/unknown)Anthropogenic-7.95E-039.78E-01n/an/a Governance – unknown (national/local/joint/unknown)Anthropogenic6.26E-017.10E-01n/an/a Meanannualmonthlytemperaturerange (C)Climatic-9.79E-012.00E-16-1.15E+001.98E-13 Meanannualminimummonthlytemperature (C)Climaticn/an/a1.09E+003.07E-16 Meanannualmaximummonthlytemperature (C)Climaticn/an/a-1.15E+001.98E-13 Meannumberofdrymonthsannually Climatic-2.28E-012.57E-02-3.09E-015.58E-03 Totalavailablewatercapacityofthesoil (vol.%,-33to-1500kPAconformingtoUSDAstandards) Edaphic-3.75E-011.16E-05-8.59E-013.05E-05 Totalnitrogencontentofthesoil (gkg-1)Edaphicn/an/a-4.13E-012.50E-03 Totalcarboncontentofthesoil (gkg-1)Edaphicn/an/a6.18E+001.15E-03 pHofthesoil (pH)Edaphicn/an/a1.73E+002.96E-02 Spatialautocorrelationterm5 Spatial6.45E+013.15E-036.60E+001.18E-01 Spatialautocorrelationterm7 Spatial-8.48E-013.57E-03-1.71E-011.45E-01 Spatialautocorrelationterm4 Spatialn/an/a6.60E+001.18E-01 Spatialautocorrelationterm3 Spatialn/an/a-1.71E-011.45E-01 Table4Thecoefficientsandassociatedp-valuesofthe variablescorrelatedwithabovegroundcarbon sequestrationVariableCoefficientp-value (Intercept) 0.0320.890 PC1 -0.1120.006 PC3 -0.2550.010 PC5 -0.4120.012 Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page7of17 http://www.cbmjournal.com/content/9/1/2

PAGE 8

interactions,inareasoflowpotentialstemdensity,carbonstorageispositivelycorrelatedwithmaximumcanopyheight(Additionalfile6:FigureS5).However,the oppositecorrelationisobservedinareasofhigherstem density.Althoughsimilarinteractionsareobservedbetweenbothsize-frequencydistributionestimates(gradientandintercept),theinteractionbetweenWSGand maximumcanopyheightisinverse,withcarbonstorage onlyshowingpositivecorrelationswithmaximumcanopyheightinareasofhighWSG.Bothsize-frequency distributionestimatesalsointeractedsimilarlywith WSG,withbothshowingpositivecorrelationswithcarbonstorageinareasoflowWSG,butnegativecorrelationsinareasofhighWSG(Additionalfile6:FigureS5). Finallycarbonsequestrationcorrelationvalueswerepositivelycorrelatedwithcarbonstorageestimates(p-value< 0.001),indicatingthatareasstoringthemostcarbonare alsothosethatareincreasinginstockatthefastestrate.DiscussionTier3correlation-basedmethodvs.Tier1and2methodsOurestimatesof1.3PgCstoredacrossthe33.9million hectaresislargerthanmostpreviousTier1estimates [46-48],althoughbelowthemostrecentlyproducedestimate[3](Table1).Underestimationoftheamountof carbonstoredintheEAMregioninglobalanalysescan bearesultoftheirpoorresolutionand/orapplicationof datafromotherregionswhichmaydiffersystematically comparedtoEastAfricanforests,woodlandsandsavannas[42].Whenseparatedbylandcovercategory,our locallyderivedcarbonestimatesarecomparabletothose presentedinotherlocal[49-52]andglobalstudies,the latteroftencontaininglittleornodatafromEastAfrica [3,4,46,47,53].Thissuggestsdifferencesbetweenourestimatesandotherstudieshavearisenbecausemany previousstudiesmappedcarbonstorageatlowerresolution[3,4,46,47,53].Whenconsideringhomogenous landscapes,scaleeffectsareunlikelytocauseadramaticdifferenceincarbonestimates.However,inhighly fragmentedandheterogeneouslandscapes,suchasEast Africa,theeffectsofscalearelikelytobesubstantial.Forestfragments,typicallyofhighcarbonstorage,maybe omittedatlowerresolutions,being ‘ replaced ’ bymore dominant,butlowcarbon,landcovercategories(e.g.open woodland),resultinginunderestimationofcarbonstorage. Itmustbenotedthat,thelandscape-scaleconfidenceintervalssurroundingourTier3estimatesareconsiderably widerthanthosearoundpreviousestimates[3,4,42,47,53]. ThisresultisconsistentwithHilletal(2013),whoalso showedincreasingmethodologicalsophisticationdoesnot (f)Total available water capacity Total available water capacity of the soil (%)(e)Historical logging (d)Governance (c) (b)Temperature range (a)Number of dry months Historical logging Governance Mean annual monthly temperature ran g e (oC)Mean number of dry months annuallyNo logging Figure4 Themodelledeffectofmostinfluential,significantanthropogenic(a,b,andc),climatic(dande)andedaphic(f)variablesof abovegroundlivecarbonstorage. Dashedredlinesindicatethemodelled95%CI.Thedataisindicatedbyblacklinesabovethex-axis. Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page8of17 http://www.cbmjournal.com/content/9/1/2

PAGE 9

necessarilyresultinreduceduncertainty,asisoftenassumed[54].Confidenceintervalsderivedfromlook-up tablevaluesmayshowasystematicbias.Therangesprovidedareanartefactofthestudyarea,thenumberofland covercategoriesandtheresolution,aswhensummed acrossalargenumberofpixels,pixelerrorismostlynegatedasunderestimatesinonepartofthelandscapeare counterbalancedbyoverestimatesinotherparts.The95% CIdevelopedfromcorrelationequationsareeffectively basedonnumerouscontinuousvariables,containingthe uncertaintyrelatingtoanthropogenic,climaticandedaphicvariables,thushavemanythousandsofpossible combinations,severelylimitingtheabilityofthe ‘ lawof averages ’ toact.Hence,the95%CIpresentedinthisinvestigationmaybetterreflectthatoftheactuallandscape, containingmorevariablesthatmake-upthecomplex landscapeheterogeneity(i.e.improvedrepresentativeness), althoughthisisonlytrueforthosepixelsestimatedusing thecorrelationequations(86%oftheEAMbutonly52% ofthestudyarea).Therefore,thelook-uptable95%CI presentedinWillcocketal(2012),andusedinthisstudy, mayunderestimateuncertainty[42].Futurestudiesshould expandtheexistingplotnetwork(Figure1),enablingthe correlationequations(andimproved95%CI)tobeappliedtotheentirestudyarea.Thisprocesshasalready begununderanewWWF-REDD+project(whichfocusses onbettersamplingthedata-deficientlandcovercategories identifiedinthisstudy[55])andtheNationalForestMonitoringandAssessment(NAFORMA)project[56,57].LinksbetweencarbonstockandinfluentialvariablesTheresultspresentedhereindicatethatALCstoragein tree-dominatedecosystemsiscorrelatedwithanthropogenic,climaticandedaphicvariables.However,inallour modelsthereisalargeamountofunexplainedvariation (R-squaredvaluesforourcorrelationmodelsvarybetween0.18and0.41).Thisislikelytobeduetothree mainreasons(Additionalfile1:SI6).Firstly,althoughwe usedthehighestresolutiondatasetsthatarefreelyavailable,severaloftheassociatedvariablesareofrelatively poorresolutionacrosstheEAM(including;wind,light andsoilnutrientvariables[Additionalfile1:TableS6]). ThisisparticularlyimportanthereaslowresolutionGIS dataisunlikelytocorrelatewellwiththeresponsevariablesfromourplotnetworkasmanyplots(withhigh variance[58])mayfallwithinasinglecell[59].Thus, ourstudymaybebiasedagainstretaininglowresolution explanatoryvariablesinourmodels.Secondly,contemporaryforestcharacteristicsaretheresultofgrowth,recruitmentandmortalityovermanyyears.Itisdifficult toobtaindataonhistoricalvariablesandyetthesecould havehadasignificantimpactonpresentdaycarbon storageandotherforestcharacteristics[60].Thirdly, presentdayinformationisalsolacking,forexample datasetsdescribingphysicalsoilpropertiesinthestudy areaareunavailable.Thus,futureworkisneededtodevelopadditionalhighresolutionGISdata,particularly forhistorictimeperiods. Ofthevarianceexplainedinourforwardandbackwardmodels,directanthropogenicfactorsarethemost influentialexplanatoryvariables(asnotedbythelargest coefficientofexplanatoryvariablesontheresponsevariable,incontrasttothose[e.g.temperature]withsmaller p-valuesbutalsosmallerinfluence[Table3])andsoare thefocusofourremainingdiscussion(seeAdditional file1:SI5fordiscussionofclimaticandedaphicvariables). Withinourstudyarea,peopleareclusteredaround highcarbonareas(Figure4).Wesuggestthiscouldbe duetotheseareashavingfavourableclimaticconditions withmoremoistureforplant(andthuscrop)growth. Further,theincidenceofmalariaislowerathighelevations[61],makingtheselocationsmorehabitableforhumanpopulations.Thusthereisapeakinpopulation densitynearthebaseofhigh-carbonmontaneforests [40].Ourinterpretationthatitisthelandscapesuitabilitydrivinghumanpopulationdensityisconsistentwith theobservationthatwhenindividuallocalitiesare followedovertime,degradationatthelocallevelcaused bythepopulationisevident[62,63].Thisemphasises thatourresultsarenotproofofcausationandthatthe driversmaybeacorrelateoftheexplanatoryvariables retainedinourmodels(Additionalfile1:SI6).Ourresultsalsoshowadecreaseincarbonstorageinpreviouslyloggedareasandinareasnearerthecommercial capital,DaresSalaam.Thisconfirmspreviousreports thatareasnearthecapitalhavelowerbiomassdueto thelocaldemandoflowgradetimberbythecity,aswell asinternationaldemandforhighgradetimberviathe city ’ sport[62];emphasisingtheconnectionsbetween theruralandurbanlandscape,andhowthesphereof urbaninfluencedriveschangeinruralecosystems.Futureinvestigationsshouldusesimulationmodellingand directexperimentationtoidentifyiftheinfluentialvariableshighlightedherecanbeconfirmedasdriversof carbonstorageandsequestration,providingadeeperunderstandingoftheprocess-basedrelationships. Thedecreaseincarbonstorageasaresultoflogging (51-77%oftheALCisretained)isofsimilarmagnitude tootherreportedestimates[64].However,thehistorical loggingdataweutilisedwasbasedonexpertopinion (Additionalfile1:TableS6)so,givenitsimportance,furtherworkdevelopingandevaluatinghistoricalvariables isneeded(Additionalfile1:TableS7).Weobservea comparabledecreaseduetodifferinggovernance.Land undernationalcontrolholdsbetween40%and65%of theALCstoredinareasunderdecentralisedgovernance. Thisperhapsindicatesthatdecentralisationofmanagement(e.g.participatoryandcommunityledforestry)isWillcock etal.CarbonBalanceandManagement 2014, 9 :2Page9of17 http://www.cbmjournal.com/content/9/1/2

PAGE 10

successfulinourstudyarea[37,65].However,itisnot possibletoprovecausationwithintheframeworkofthis study.Manylocallymanagedforestsarelocatedinthe south-eastofourstudyareawithinanareaofnaturally highcarbonstorage,whereaslandundernationalcontrolcoversmuchlargerareas,includingthedry,carbonpooreast.Hence,ourfindingthatcarbonstorageis higherinareasunderdecentralisedcontrolmaybean artefactofthedifferingareaswherethistypeofland managementoccurs.Furtherstudiesmonitoringchange incarbonstorageovertimeunderthetwodifferentgovernanceregimeswouldenabletheeffectoflandmanagementtobedetermined. Theoveralleffectsoncarbonstoragearearesultof manychangesinforestcharacteristics.BothWSGand theproportionoflargerstemsdecreasewithincreasing anthropogenicdisturbance,however,stemdensity( = 10cmDBH)increases.Anthropogenicdisturbance,for examplelogging,isoftenacommercialactivityandresultsinthepreferentialremovalofthelargest,most valuablestems[62].Themoreopencanopy,following stemremoval,wouldresultinincreasedrecruitment fromyoungforesttrees[66],leadingtothehighnumbersofsmallstemsobserved.However,theopposite wouldbeexpectedinwoodlandsandsavannas,with moreopencanopiesresultinginmoregrass,highfireintensityandsolessrecruitment[67,68].Ourresultshighlighthowinfluentialthenegativeeffectofpeopleon tropicalforestcarbonstoragecanbe.Thisassertionis supportedbydatafromacrossthetropics[69-71].The significantimpactofanthropogenicactivitiesimplies thatREDD+could,atthelocalscale,havesignificant positiveimpactsoncarbonstorage.However,careful policydesignstolimitleakageofdeforestationandencouragetheinvolvementofthelocalpopulationare neededtoensureREDD+schemesachievetheircarbon storageandsequestrationaims[72]. Likecarbonstorageanditscomponents,carbonsequestrationisalsocorrelatedwithanthropogenic,climaticandedaphicvariables.Weestimatethatsome localities(forexampletheUdzungwaMountainsNationalPark;Figure4)provideacarbonsinkofcomparableper-areamagnitudetomodelledestimatesinEast Africa[73]andtothatobservedoverrecentdecadesin structurallyintactAfricanforest[7].However,many areasofforestandwoodlandwithinthestudyareaexperienceahighlevelofdegradationanddisturbance, andsoarenetsources.Here,wehaveshownthatanthropogenicdisturbanceisakeydeterminantofthe trendincarbonstorageovertimeineasternTanzania. ImportantlocationsofhighcarbonlossesarethePare andUsambaramountains(Table5),whichhistorically haveseenthehighestratesofdegradationanddisturbance [74].ThenationalpopulationofTanzaniaisincreasing [75]andthismayincreasethepressureontree-dominated ecosystemswhichcouldresultinthestudyareabecoming asignificantsourceofcarboninthefuture.Furthermore, theeffectofincreaseinanthropogenicpressurescouldbe compoundedbypotentialdecreaseincarbonstorageasa resultofincreasingtemperatures[76,77]andchangesin soilnutrients(seeAdditionalfile1:SI5).However,these futureeffectscouldbecomplicatedbyincreasinglevelsof atmosphericCO2,varyingeffectivenessoflegallyprotectedareasandshiftingconsumptionpatterns.ConclusionsOurresultsshowthattheamountofcarbonstoredin forestsacross33.9millionhaoftheEasternArcMountainsofTanzaniaisconsiderable:1.32(0.89to3.16)Pg. Ourestimateissignificantlyhigherthanmostprevious estimates.However,ourmoresophisticatedmethodalso hashigheruncertainty,implyingthatothermethods maysubstantiallyunderestimatetheuncertaintyinvolved.Withinthetree-dominatedlandcovercategories, historicalloggingisthemostinfluentialdirectanthropogenicfactor,whilethemeannumberofdrymonthsis themostinfluentialenvironmentalfactor,withanorder ofmagnitudelessimpactoncarbonstorage.Weshow thatWSG,size-frequencydistributionvariablesand heightvariablesareallimportantindeterminingcarbon storage.Ourestimatesindicatethat,between2004and 2008,tree-dominatedcommunitiesacrossthestudy areasshowednosignificantchange,howeversomeareas Table5Carbonstoredandsequesteredacrossthe individualmountainblocksoftheEAMrange(thetotalis denotedinbold)EasternArc Mountain Block[ 40 ] Area,km2Abovegroundlive carbonstorage,Tg Meancarbon sequestration, Mgha-1yr-1Tier3Willcocketal (2012)-Original Tier2[ 42 ] NorthPare5101.932.382.60 SouthPare2,3278.969.592.41 West Usambara 2,94513.5215.963.64 East Usambara 1,1455.917.632.79 Nguu1,5629.3412.711.89 Nguru2,56515.1118.861.79 Ukaguru3,24313.3920.631.42 Uluguru3,05715.9213.911.35 Rubeho7,98436.8440.961.06 Malundwe330.290.291.80 Udzungwa22,788101.73104.051.01 Mahenge2,60623.5812.080.19 Total50,765246.53259.061.19 Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page10of17 http://www.cbmjournal.com/content/9/1/2

PAGE 11

wereidentifiedaslargesinks(0.8%ofthestudyarea) andotherslargesources(1.4%ofthestudyarea),showingtheimportanceoftakingalandscapescaleapproach. Thecarbonmapsproducedandstatisticalrelationships documentedcanassistpolicy-makersindesigningpoliciestomaintainandenhancecarbonstorageforclimate mitigationandotherecosystemservices.MethodWecollateddatafrom2,462treeinventoryplotswithin ourstudyarea(seeAdditionalfile1:SI3),thenapplieda qualitycontrolandstandardisationprotocol.Thisconsistsoftwomainsteps:(1)Metadataqualitycontrol;and (2)Measurementbiasdetection. Firstly,allplotslackingarecordedspatiallocationand afixedareawerediscarded(770plots).Plotswhereone ormorediameteratbreastheight(DBH)datawere knowntobemissingwerealsoexcluded(7plots).Furthermore,plotssmallerthan0.025ha(16plots)were deemedtoproduceunreliablecarbonestimatessoalso removedfromthedataset. Secondly,toassesspossiblemeasurementbias,i.e.not measuringoverbuttressesandsooverestimatingbiomass [78],theremainingplotsweregroupedbytheleadfield researcher.Size-frequencydistributions,using10cmsize classes,werecreatedforeachofthesegroups.Forestsizefrequencydistributionsaresuggestedtoconformtothe-2 powerlawbasedonmetabolicscaling[79].Althoughit hasbeenarguedthatthisruleisnotgloballyapplicable [80],manystudiesacceptthisasatheoreticalmaximum valuefortheabundanceoflargestems[81].Thus,researcherswithmanyplotsabovethismaximumvalue likelymeasuredstemsaroundbuttressesandsowereremoved(1researcher,100Plots). Thequalitycontrolandstandardisationprocedureresultedinadatasetof1,611treeinventoryplots(median 0.1ha,mean0.1ha,mode0.1ha[43plotswithmultiplecensuses;median0.1ha,mean0.5ha,mode 1.0ha];Figure1;seeAdditionalfile1:SI3forafurther information)fromwhichwecalculatedplot-levelstand structureindicesandabovegroundcarbonstorageper unitarea(seeAdditionalfile1:SI2forfulldetails).Weobtainedtheexponentandinterceptofthepopulationsizefrequencydistributionusingthepowerlawfitforeach plotusingthelog-logtransformationmethod.Whereby, foreachplot,wecreated10cmbinsize-frequencydistributionsbasedonDBH,andalinearmodelofthelogarithmoffrequencyagainstthelogarithmofthesizeclass wasfitted.Whilstnotasaccurateasthemaximumlikelihoodestimationmethod,oursimplermethodismore stableformanyofourplots,providingboththeintercept andslopeindicatorsofpopulationstructure[82]. WeobtainedWSGdataviathephylogeneticinformationprovidedbyourtreeinventoryplots.Weuseda globalwooddensitydatabasetoextractspeciesaverage WSG[83].Thisprocedurepr ovidedover32,000trees withWSGdata.Whenthiswasnotpossiblewe adoptedahierarchicalappr oach,firstapplyingtheappropriategenusaverageifavailable(~14,000trees)beforeconsideringfamilyaverage(~9,500trees),plot average(~4,500trees)anddatasetaverage(~80trees) inturn[84].IncludingWSGasanadditionalparameter inallometricequationsreducesthebiomassestimation error[49,85,86]. Inaddition,weestimatedplotbiomassusingmoist foresttreeallometry[86]basedonmeasurementsof DBHfromourtreeinventoryplots,WSG(asdescribed above)andheightdata(derivedfromourdatasetusing thebestfitDBH-heightequationform[Equation5.1;see Additionalfile1:SI4],ifnotmeasuredinthetreeinventoryplots).Finally,carbonwasassumedtobe50%of biomass[7]. Forasmallernumberofplots,multiplemeasurements wereavailableovertime(n=43;meanplotsize=0.5ha; meanmeasurementperiod=3.9years).Wecalculated changesincarbonstorageratesbydividingthedifferenceincarbonstorageestimatesbetweencensusesby thenumberofyearsseparatingthem. Forour1,611geo-referencedtreeinventoryplots,we obtainedfurtherinformationonvariablesfallingintofive broadcategories;anthropogenic,climatic,geographic, edaphic,andpyrologic(medianresolution1.0ha,mean resolution22.0ha,moderesolution1.0ha;Additional file1:TableS6).Anthropogenicdata,furtherdivided intosixsubcategories,wereobtained:(1)population pressurevariables(n=14relatedvariables)wereobtainedfromPlatts(2012)[87](seeAdditionalfile1:SI7); (2)DaresSalaamrelatedvariables(n=3;e.g.distanceto DaresSalaam),(3)markettownrelatedvariables(n=3; e.g.distancetomarkettowns),and(4)infrastructurerelatedvariables(n=2;e.g.distancetoroads)werederived fromavailabletopographicmaps;(5)historicallogging (n=1)fromSwetnametal(2011)[88];and(6)governance(n=1)fromtheWorldDatabaseonProtected Areas[89].Climatedataweredividedintothreesubcategories(precipitation[n=2;maximummeancumulative waterdeficitandmeannumberofdrymonthsannually],temperature[n=4;meanannualtemperature, meanannualminimummonthlytemperature,meanannualmonthlymaximumtemp erature,andmeanannual monthlytemperaturerange]andwindspeed[n=1]) andwerederivedfromtheTropicalRainfallMeasuring Mission[90,91],WorldClim[92,93],andUnitedStates NationalAeronauticsandSpaceAdministrationSurface meteorologyandSolarEnergy[94]datasets.Similarly, geographicdatahavetwovariables(aspect[n=1]andincomingsolarradiation[n=1])derivedfromShuttleRadar TopographyMission[93]andNationalRenewableEnergyWillcock etal.CarbonBalanceandManagement 2014, 9 :2Page11of17 http://www.cbmjournal.com/content/9/1/2

PAGE 12

Laboratory[95,96]datasetsrespectively.Lastly,weextractededaphicdata(n=6)fromtheInternationalSoil ReferenceandInformationCentredatabase[97,98]and fire-relatedvariables(n=5)derivedfromMODISimages[99]. Wethencorrelatedthesevariableswithcarbonstorage,andfollowingthis,itscomponents:WSG,theinterceptofthepowerlawrelationship,andthegradientof thepowerlawrelationship,ineachcaseusinggeneral linearmodels(seeAdditionalfile1:SI2-5).Notransformationswererequiredtoensureanormaldistribution whencorrelatingeitherWSG,theinterceptofthepower lawrelationshiporthegradientofthepowerlawrelationshipwiththeindividualvariables.However,carbon storageestimatesrequiredasquareroottransformation toensureanormaldistributionwithinthegenerallinear models(normalitywasconfirmedusingtheShapiroWilktest;p-value>0.05).Inallmodels,plotswere weightedbythesquarerootoftheirareaasconfidence inbiomassestimationincreaseswiththeareasurveyed [100,101].Landscapescalespatialautocorrelationwas accountedforbyincludingspatialterms(latitude,longitudeandtheinteractionsbetweenthem)inthemodel (Additionalfile1:TableS6)[102].Thenumerouspossibleinteractionswereexcludedfromthemodels,as thesewerefoundtoaddverylittleexplanatorypowerto themodels,onlyincreasingR-squaredvaluesby~0.001 withtheadditionofeachinteractionterm.Allanalyses wereperformedusingR2.12.1[103]andmappedin ArcGISv9.3.1[104]. Whenassessingcarbonsequestration(n=43)fewer degreesoffreedomwereavailable,thereforeexplanatoryvariablesneedtobegrouped.Therefore,weconductedaprinciplecomponents(PC)analysis,obtaining fivePCwhichexplained>90%ofthecumulativevarianceoftheindividualinfluentialvariables(Additional file1:TableS4).Then,covariationofPCwithcarbon sequestrationwasassessedinsteadoftheindividualinfluentialvariables.Carbonsequestrationestimatesrequiredacube-roottransfor mationtoensureanormal distributionwithinthegenerallinearmodels(confirmedusingtheShapiro-Wilktest;p-value>0.05). Thisenabledtheeffectofmultiplevariablestobeexaminedevenwiththislimiteddataset.PCanalysisof thevariableswasperformedonthescaleddatausing theprcomppackage[105]withinR2.12.1[103].All otheraspectsofthemodel(weightingandspatialautocorrelation)wereperformedidenticallytothemodels forcarbonstorageanditscomponents. Themostappropriatemodelwaschosenusingforwardandbackwardstepwiseselection.Forwardmodels aremoreusefulforinferringcausalrelationships[43] andsowerepreferentiallyusedtoinfertheinfluential variablesofcarbonstorageandsequestration.However, averagingforward – backwardsandbackward – forwards predictionsoutperformsconventionalselectionprocedures[43]andsobothmethodswereusedwhenestimatingthespatialdistributionswithinthestudyarea. Akaikeinformationcriterion(AIC)wasusedtoreduce/ expandthemodels,withvariableselectionoccurring whenthevariablereducedthemeansquarederror (MSE)underten-foldcrossvalidation[106].Unlike modelselectionusingR-squared,whichneglectsthe principlesofparsimony,AICconsidersbothmodelfit andcomplexity,resultinginbetterpredictionsand allowinginferencestobemadefrommultiplemodels [107].Modelselectioncontinueduntiltheaddition/removaloffurthervariablesabletoreducecrossvalidation MSEnolongerincreasedAIC,therebyproducingthe best-fitmodelwiththelowestpredictionerror[43]. Withineachcategory(anthropogenic,climatic,geographic,edaphic,andpyrologic),somevariableswere highlycorrelated(Additionalfile1:TableS7)andthis mayconfoundthestepwiseprocedureaseachvariable doesnotcarryenoughdistinctinformation[108].For example,alltemperaturerelatedvariables(Additional file1:TableS7)werecorrelated(R-squared>0.6).However,itisunclearwhichcorrelatedbestwiththevariablesofinterest,e.g.carbonstorageandsequestration. Manystudiesincludemeanannualtemperatureinbiomassmodels[77,109],buttheorysuggeststhatitmaybe thetemperaturerangedrivingthisrelationshipasphotosynthesiscorrelateswithmaximumtemperatures,but respirationwithminimumtemperatures[76,110,111]. Wefoundthat,ifweremovedcorrelatedvariablesprior tomodelselection,thefinalmodelswereartefactsofthe variableswehadselected.Forexample,ifweincludedmean annualtemperatureinthemo del,butnottemperature range,thenthesignificantco rrelationsbetweenmeanannualtemperatureandALCstoragewerefound.However, thesecorrelationswereinsignificantiftemperaturerange wasaddedtothemodel,withthenewlyaddedvariable showingasignificanteffectinstead.Inshort,theresultant modelswereautomaticallybiasedtowards apriori expectations.Toavoidthisbias,wedevisedaprocedurebywhich theinfluentialvariablesincludedinmodelselectionwere selectedbytheirabilitytoexplainvariationwithinthedata ofinterest(e.g.carbonstorage).Allvariables(describe above)wereincludedinmodelselection.Oncethishadrun tocompletionthemodelwasassessed.Thesubcategory withthemostcorrelatedvar iablesretainedwithinthe modelwasselectedandallbutthemostinfluential,significantvariablewereremoved.Forexample,ifallfour temperature-relatedvariableswereincludedintheinitialmodelandthiswasthelargestgroupofvariables thenthisgroupwouldbeselected.Then,ifmeanannual temperaturewasthemostinfluentialandsignificant temperature-relatedvariable,allothertemperature-relatedWillcock etal.CarbonBalanceandManagement 2014, 9 :2Page12of17 http://www.cbmjournal.com/content/9/1/2

PAGE 13

variableswouldbeexcludedinthenextroundofmodel selection.Thus,stepwisemodelselectionwasthenrepeatedforallremainingvariables.Thisprocesswasrepeateduntilnohighlycorrelatedvariablesremained withinthemodelproduced. Sinceonlylandscape-scalevariationwasaccountedfor bythespatialtermsalreadyincludedinthemodel(latitude,longitudeandtheinteractionsbetweenthem; Table1;Additionalfile1:TableS6),itwasnecessaryto investigatetheeffectoflocal-scale(<10km2)spatial autocorrelation[102].Todothis,theseparateforward andbackwardmodels,containingnohighlycorrelated variables(producedabove),weremapped.Then,the sumofthemodelestimateswithinthemapswereextractedat1,3,5,7and10km2resolutions,andincluded asadditionalvariables(representinglocalspatialautocorrelationterms)intothestepwisemodelselection process,whichwasre-runafinaltime[112].However, inallcases,localspatialautocorrelationtermswere rejectedastheydidnotreducecrossvalidatedMSE. Sinceitwasnotnecessarytoincludelocalspatialautocorrelationtermsinthemodels,thepreliminarymaps producedabovecouldberegardedasfinalspatialrepresentationsofthetenbestfitmodels,two(forwardand backward)foreachofthefivevariablesofinterest(carbonstorage,carbonsequestration,WSG,theintercept ofthepowerlawrelationshipandthegradientofthe powerlawrelationship).Eachpairofmaps(forwardand backward)werethencombinedintoasingle,final weightedmeanestimate.Theratiooftherelevantcross validatedMSEoftheforwardandbackwardmodelswas usedtocreatetheweightedmean,withthemodelshowinglowesterrorreceivingthehighestweighting[43]. Thus,weultimatelyproducedfivemaps(fromtenbest fitmodels);oneeachforcarbonstorage,carbonsequestration,WSG,theinterceptofthepowerlawrelationship,andthegradientofthepowerlawrelationship.As ourcarbonstorageestimateswerederivedfromdata representingtreeswithaDBHgreaterthanorequalto 10cm,regionallyestimatesofratiosfromWillcocketal (2012)wereusedtoestimatetheunmeasuredcomponentofALCstorage[42],thiswassummedwithour modelledcarbonstorageestimate,providinganestimate oftotalALCstorage. Althoughthefivemapsproducedcoveredtheentire studyarea,wewereconcernedthatextrapolatingpredictionsbeyondtherangeofobservedpredictorvariables fromourdatasetcouldresultinlarge,unquantifiableerrors.Thus,welimitedthemodelstolocalitieswhereall theassociatevariableswerewithintherangeofthat showninourdataset,thusonlyinterpolatingwithinour correlationmodelsfortree-dominatedlandcovercategories.Foranypixelsoutsidethedatarange,look-up tablemethodswereusedinpreferencetothecorrelation modelestimates.Thus,foreverylandcoverinourstudy areacontainingtrees(openwoodland;closedwoodland; forestmosaic;lowlandforest;sub-montaneforest;montaneforest;anduppermontaneforest[41])thatfell withinthelimitsofourdataset,theestimateofcarbon storagederivedfromthecorrelationequationswasused. Forallotherlandcovercategories,andforthoselocalitiesforwhichpredictorvariablesfelloutsidetheranges ofvaluesusedinmodelconstruction,landcoverbased look-uptablevaluesfromWillcocketal(2012)were usedtoestimateALCstorage[42].Intotal,look-uptable valueswereappliedto52%ofthelandscape,although thiswaspredominantlytolowcarbonlandcovercategories,with86%oftheEAM(whichholdthemajority oftheregionstropicalforest[113])estimatedusingthe correlationapproachdescribedabove.EstimatesofWSG andpopulationstructurewereonlymadeforwooded landcovercategories,withestimatesforareaswithin ourdatasetrangebeingderivedfromtherelevantcorrelationequationsandestimatesforotherareascoming fromlandcoverbasedlook-uptablevaluesderivedfrom themedianvalueofourWSGandpopulationstructure data(weightedbythesquarerootofplotsizeandderivedviasamplingwithreplacement10,000times)for eachlandcovercategory(Additionalfile1:TableS8). Forcarbonsequestration,again,estimateswereonly madeforwoodedlandcovercategoriesforthoseareas insidetherangeofourdatasetestimatesderivedfrom thecorrelationequationswereused.However,unlike carbonstorage,WSGandpopulationstructure,forareas outsidetherangeofourdataset,alandcoverbased look-uptablewasnotusedasseverallandcovercategorieswerepoorlyrepresentedduetothesmallsamplesize available(n=43).Instead,forpixelsoutsidetherangeof thecorrelation-derivedcarbonsequestrationmodel(16% ofpixelswithwoodedlandcover),themedianvalueof datafromourrecensusedplots(againweightedbythe squarerootofplotsizeandderivedviasamplingwith replacement10,000times)wasutilised. Forevery1hapixelofeachmapderivedfromcorrelationequations,weproduced95%confidenceintervals (CI).Ifthepixelestimatewasderivedfromthegeneral linearmodels,thenthepixel95%CIwascalculatedby addingandsubtractingthesquarerootofthecrossvalidationMSE.Forlook-uptablepixelsthelookuptable 95%CIwereused.Thepixel95%CIdescribes,forevery pixel,therangewewouldexpecteachofourestimates toliewithin.However,aswearealsointerestedinestimatingcarbonstorageandsequestrationonalandscape scale,indicationsofuncertaintyarealsorequiredat landscape-scale.Simplysummingthepixel95%CIto derive95%CIoftheoveralllandscape-scaleestimates wouldincorrectlytreatrandomerrorasaregion-wide systematicbias.Thus,toderive95%CIforlandscape-Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page13of17 http://www.cbmjournal.com/content/9/1/2

PAGE 14

scaleestimates,werandomlyallocatedeachpixelanestimatewithintherangedictatedbyits95%pixelCI,and summedthesevaluesacrosstheentirelandscape.This processwasperformed10,000timesandthemedian valueand95%CI(the250thand9,750thrankedvalues, whichmaynotbeequallydistributedaroundthemedian)forabovegroundcarbonstorageandsequestration inthestudyareawereobtained. Forthefinalmodelofcarbonstorageestimates,weinvestigatedhowthecomponentsofcarbonstorage(population structure,WSGandtreeheight)interactedtoultimately producetheecosystemserviceofcarbonstorage.Weobtainedestimatesofmaximumcanopyheightfromthebest fitDBH-heightequation[Equation5.1;seeAdditionalfile1: SI4andAdditionalfile7:FigureS6],andcombinedthis spatiallywithourcorrelatio nmodelderivedestimatesof WSG,theinterceptofthepowerlawrelationshipandgradientofthepowerlawrelationship.Wethencorrelatedthese againstourestimatesofcarbons torage,allowingallpossible interactions,andselectedthebest-fitmodel(viaAIC)using bothforwardsandbackward sstepwiseregression. Ethicalapprovalfortheabovestudywasobtained fromtheFacultyofEnvironmentResearchEthicsCommittee,inaccordancewiththeUniversityofLeedsresearchethicspolicy.AdditionalfilesAdditionalfile1: Supportingtext(includingS1-7andTablesS1-12). Additionalfile2:FigureS1. ThespatialvariationofWSGintreedominatedlandcovercategorieswithinthestudyarea(a),withupper (b)andlower(c)pixelbased95%CI.Seetextfordetailsonmethods. Additionalfile3:FigureS2. Themostinfluential,significantinfluential variablesonWSG(aandb),theinterceptofthepowerlawrelationship (candd),andthegradientofthepowerlawrelationship(eandf). Dashedredlinesindicate95%CI. Additionalfile4:FigureS3. Thespatialvariationintheinterceptof thepowerlawrelationship(aproxymeasureforpotentialstemdensity)in treedominatedlandcovercategorieswithinthestudyarea(a),withupper (b)andlower(c)pixelbased95%CI.Seetextfordetailsonmethods. Additionalfile5:FigureS4. Thespatialvariationinthegradientofthe powerlawrelationship(aproxymeasurefortheproportionoflarger stems)intree-dominatedlandcovercategorieswithinthestudyarea(a), withupper(b)andlower(c)pixelbased95%CI.Seetextfordetailson methods. Additionalfile6:FigureS5. The2ndorderinteractionsrelatingmy carbonstoragederivatives(woodspecificgravity,maximumcanopy height,theinterceptofthepowerlawrelationship,andthegradientof thepowerlawrelationship[shownhereasWSG,height,intercept,and gradientrespectively])toabovegroundlivecarbonstorage.Dashedred linesindicate95%CI. Additionalfile7:FigureS6. TheeffectofMATontreeheightfora rangeofDBH.Thedata(points)correspondtoDBHrangeswhereasthe Gompertzmodelfits(solidlines)illustratetherelationshipformid-point ofthisrangeonly.Dottedlinesrepresentthe95CIofthemodelfits. Abbreviations AIC: Akaikeinformationcriteria;ALC:Abovegroundlivecarbon; CI:Confidenceinterval;CV:Crossvalidation;CWD:Coarsewoodydebris; DBH:Diameteratbreastheight;EAM:EasternArcMountains;eCEC:Effective cationexchangecapacity;GIS:Geographicinformationsystems;HYDE:History DatabaseoftheGlobalEnvironment;IPCC:IntergovernmentalPanelonClimate Change;IUCN:InternationalUnionforConservationofNature;KITE:York InstituteforTropicalEcosystems;MAT:Meanannualtemperature;MSE:Mean squarederror;PC:Principalcomponents;REDD+:ReducingEmissionsfrom DeforestationandForestdegradation;SAGE:CentreforSustainabilityand GlobalEnvironment;WSG:Woodspecificgravity. Competinginterests Theauthorsdeclarethattheyhavenocompetinginterests. Authors ’ contributions Themajorityofthisjointly-authoredpublicationwasledbySW.Contributions tothecollaborativedatasetcamefromPJP,AA,ND,KD,EF,JG,JH,KH,ARM, BM,PKTM,NO,EJTJ,AM,SWandRDS.TheanalysiswasperformedbySW, supervisedbyOLPandSLL.ThemanuscriptwaspreparedbySW,with assistancefromOLP,SLL,AB,PP,NDDandRM.Allauthorsreadandapproved thefinalmanuscript. Acknowledgements ThisstudyispartoftheValuingtheArcresearchprogramme(http:// valuingthearc.org/)fundedbytheLeverhulmeTrust(http://www.leverhulme. ac.uk/).Manuscriptpreparationandlateranalysestookplaceunderthe ‘ WhichEcosystemServiceModelsBestCapturetheNeedsoftheRuralPoor? ’ project(WISER;NE/L001322/1),fundedwithsupportfromtheUnited Kingdom ’ sEcosystemServicesforPovertyAlleviationprogram(ESPA;www. espa.ac.uk).ESPAreceivesitsfundingfromtheDepartmentforInternational Development(DFID),theEconomicandSocialResearchCouncil(ESRC)and theNaturalEnvironmentResearchCouncil(NERC).SLLwasfundedbya RoyalSocietyUniversityResearchFellowship;SWadditionallybythe StokenchurchCharity;OLPwassupportedbyanAdvancedGrantfromthe EuropeanResearchCouncilandisaRoyalSociety-WolfsonResearchAward holder.Thefundershadnoroleinstudydesign,datacollectionandanalysis, decisiontopublish,orpreparationofthemanuscript.WethanktheTanzanian CommissionforScienceandTechnology(COSTECH),theTanzanianWildlife Institute(TAWIRI)andtheSokoineUniversityofAgriculturefortheirsupportof thiswork,aswellasallthefieldassistantsinvolved.Furthermore,wewouldlike tothankthetwoanonymousreviewers,whosecommentsandinsightvastly improvedthemanuscript. Authordetails1SchoolofGeography,UniversityofLeeds,LeedsLS29JT,UK.2Schoolof BiologicalSciences,UniversityofSouthampton,SouthamptonSO171BJ,UK.3EnvironmentDepartment,UniversityofYork,YorkYO105DD,UK.4DepartmentofZoology,UniversityofCambridge,CambridgeCB23EJ,UK.5WWFUS,Washington,USA.6UNEPWorldConservationMonitoringCentre, CambridgeCB30DL,UK.7GeneticsandConservation,RoyalBotanticGarden Edinburgh,Edinburgh,UK.8TanzanianForestConservationGroup,Dares Salaam,Tanzania.9FrankfurtZoologicalSociety,FrankfurtD-60316,Germany.10TheSocietyforEnvironmentalExploration,LondonEC2A3QP,UK.11STEP Program,PrincetonUniversity,Princeton08544,USA.12Departmentof Geography,UniversityofFlorida,POBox117315,Gainesville,Florida,FL 32611,USA.13TheUniversityofDaresSalaam,DaresSalaam,Tanzania.14CentrefortheIntegrationofResearch,ConservationandLearning, FlamingoLandLtd,MaltonYO176UX,UK.15SokoineUniversityof Agriculture,POBox3001,Morogoro,Tanzania.16EDGEofExistence, ConservationProgrammes,ZoologicalSocietyofLondon,London,UK.17DepartmentofGeography,StaffordshireUniversity,Stoke-on-TrentST42DF, UK.18DepartmentofBioscience,AarhusUniversity,AarhusCDK-8000, Denmark.19DepartmentofGeography,UniversityCollegeLondon,London WC1E6BT,UK. Received:4November2013Accepted:31March2014 Published:28April2014 References1.MalhiY,GraceJ: Tropicalforestsandatmosphericcarbondioxide. Trends EcolEvol 2000, 15: 332 – 337. 2.IPCC: Landuse,land-usechange,andforestry. Cambridge,UK:Cambridge UniversityPress;2000.Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page14of17 http://www.cbmjournal.com/content/9/1/2

PAGE 15

3.BacciniA,GoetzSJ,WalkerWS,LaporteNT,SunM,Sulla-MenasheD,Hackler J,BeckPSA,DubayahR,FriedlMA,SamantaS,HoughtonRA: Estimated carbondioxideemissionsfromtropicaldeforestationimprovedby carbon-densitymaps. NatClimChange 2012, 2: 182 – 185. 4.SaatchiSS,HarrisNL,BrownS,LefskyM,MitchardETA,SalasW,ZuttaBR, BuermannW,LewisSL,HagenS,PetrovaS,WhiteL,SilmanM,MorelA: Benchmarkmapofforestcarbonstocksintropicalregionsacrossthree continents. ProcNatlAcadSci 2011, 108: 9899 – 9904. 5.MyersN,MittermeierRA,MittermeierCG,daFonsecaGAB,KentJ: Biodiversity hotspotsforconservationpriorities. Nature 2000, 403: 853 – 858. 6.TimkoJA,WaeberPO,KozakRA: Thesocio-economiccontributionof non-timberforestproductstorurallivelihoodsinSub-SaharanAfrica: knowledgegapsandnewdirections. IntForRev 2010, 12: 284 – 294. 7.LewisSL,Lopez-GonzalezG,SonkeB,Affum-BaffoeK,BakerTR,OjoLO, PhillipsOL,ReitsmaJM,WhiteL,ComiskeyJA,DjuikouoM-N,EwangoCEN, FeldpauschTR,HamiltonAC,GloorM,HartT,HladikA,LloydJ,LovettJC, MakanaJ-R,PeacockJ,PehKS-H,SheilD,SunderlandT,SwaineM,TaplinJ, TaylorD,ThomasC,VotereR,WllH: Increasingcarbonstorageinintact Africantropicalforests. Nature 2009, 457: 1003 – 1006. 8.PhillipsOL,MalhiY,HiguchiN,LauranceWF,NunezPV,VasquezRM, LauranceSG,FerreiraLV,SternM,BrownS,GraceJ: Changesinthecarbon balanceoftropicalforests:evidencefromlong-termplots. Science 1998, 282: 439 – 442. 9.AchardF,EvaHD,StibigH-J,MayauxP,GallegoJ,RichardsT,MalingreauJ-P: Determinationofdeforestationratesoftheworld'shumidtropicalforests. Science 2002, 297: 999 – 1002. 10.HansenMC,StehmanSV,PotapovPV: Quantificationofglobalgrossforest coverloss. ProcNatlAcadSci 2010, 107: 8650 – 8655. 11.PutzFE,ZuidemaPA,PinardMA,BootRGA,SayerJA,SheilD,SistP,EliasM, VanclayJK: ImprovedTropicalForestManagementforCarbonRetention. PLoSBiol 2008, 6: e166. 12.IPCC: ClimateChange2007:ThePhysicalScienceBasis. Availableat http://www.ipcc.ch/pdf/assessment-report/ar4/wg1/ar4-wg1-spm.pdf [Accessed05/01/12].Agenda2007,6. 13.vanderWerfGR,MortonDC,DeFriesRS,OlivierJGJ,KasibhatlaPS,Jackson RB,CollatzGJ,RandersonJT: CO2emissionsfromforestloss. NatGeosci 2009, 2: 737 – 738. 14.DixonRK,BorwnS,HoughtonRA,SolomonAM,TrexlerMC,WisniewskiJ: Carbonpoolsandfluxesofglobalforestecosystems. Science 1994,263: 185 – 190. 15.DeFriesRS,HoughtonRA,HansenMC,FieldCB,SkoleD,TownshendJ: Carbonemissionsfromtropicaldeforestationandregrowthbasedon satelliteobservationsforthe1980sand1990s. ProcNatlAcadSciUSA 2002, 99: 14256 – 14261. 16.HoughtonRA: TRENDS:ACompendiumofDataonGlobalChange. Availableat:http://cdiac.ornl.gov/trends/landuse/houghton/houghton. html.Anupdateofestimatedcarbonemissionsfromlandusechange, typicallyupdatedeveryyear. OakRidgeNationalLaboratory,US DepartmentofEnergy;2008. 17.PanY,BirdseyRA,FangJ,HoughtonR,KauppiPE,KurzWA,PhillipsOL, ShvidenkoA,LewisSL,CanadellJG,CiaisP,JacksonRB,PacalaSW,McGuire AD,PiaoS,RautiainenA,SitchS,HayesD: Alargeandpersistentcarbon sinkintheworld ’ sforests. Science 2011, 333: 988 – 993. 18.HansenMC,StehmanSV,PotapovPV,LovelandTR,TownshendJRG,DeFries RS,PittmanKW,ArunarwatiB,StolleF,SteiningerMK,CarrollM,DiMiceliC: Humidtropicalforestclearingfrom2000to2005quantifiedbyusing multitemporalandmultiresolutionremotelysenseddata. ProcNatlAcad Sci 2008, 105: 9439 – 9444. 19.ChaveJ,ConditR,Muller-LandauHC,ThomasSC,AshtonPS, BunyavejchewinS,CoLL,DattarajaHS,DaviesSJ,EsufaliS: Assessing evidenceforapervasivealterationintropicaltreecommunities. PLoSBiol 2008, 6: e45. 20.FieldCB,BehrenfeldMJ,RandersonJT,FalkowskiP: Primaryproductionof thebiosphere:integratingterrestrialandoceaniccomponents. Science 1998, 281: 237 – 240. 21.REDDoverview: REDDoverview. [http://www.un-redd.org/] 22.StrassburgB,TurnerRK,FisherB,SchaefferR,LovettA: Reducingemissions fromdeforestation — The “ combinedincentives ” mechanismand empiricalsimulations. GlobEnvironChang 2009, 19: 265 – 278. 23.KindermannG,ObersteinerM,SohngenB,SathayeJ,AndraskoK, RametsteinerE,SchlamadingerB,WunderS,BeachR: Globalcostestimates ofreducingcarbonemissionsthroughavoideddeforestation. ProcNatl AcadSciUSA 2008, 105: 10302 – 10307. 24.CattaneoA,LubowskiR,BuschJ,CreedA,StrassburgB,BoltzF,AshtonR: Oninternationalequityinreducingemissionsfromdeforestation. EnvironSciPol 2010, 13: 742 – 753. 25.FarrisM: Thesoundoffallingtrees:integratingenvironmentaljustice principlesintotheclimatechangeframeworkforReducingEmissions fromDeforestationandDegradation(REDD). FordhamEnvironLawRev 2010, 20: 515. 26.BirdseyR,PanY,HoughtonR:Sustainablelandscapesinaworldof change:tropicalforests,landuseandimplementationofREDD+:PartII. CarbonManag 2013, 4: 567 – 569. 27.IPCC: GuidelinesforNationalGreenhouseGasInventories .In National GreenhouseGasInventoriesProgramme. EditedbyEgglestonHS,BuendiaL, MiwaK,NgaraT,TanabeK.Kanagawa,Japan:InstituteforGlobal EnvironmentalStrategies;2006. 28.IPCC: Goodpracticeguidanceforlanduse,land-usechangeandforestry In IPCCNationalGreenhouseGasInventoriesProgramme. EditedbyPenman J,GytarskyM,HiraishiT,KrugT,KrugerD,PipattiR,BuendiaL,MiwaK, NgaraT,TanabeK.Kanagawa,Japan:InstituteforGlobalEnvironmental Strategies;2003. 29.GOFC-GOLDA: ASourcebookofMethodsandProceduresforMonitoring andReportingAnthropogenicGreenhouseGasEmissionsandRemovals CausedbyDeforestation,GainsandLossesofCarbonStocksinForests RemainingForests,andForestation. AB,Canada:GOFC-GOLDGOFC-GOLD ProjectOffice,NaturalResourcesCanadaCalgary;2010. 30.AsnerGP,PowellGVN,MascaroJ,KnappDE,ClarkJK,JacobsonJ,KennedyBowdoinT,BalajiA,Paez-AcostaG,VictoriaE,SecadaL,ValquiM,Hughes RF: High-resolutionforestcarbonstocksandemissionsintheAmazon. ProcNatlAcadSci 2010, 107: 16738 – 16742. 31.InternationalW: CarbonStorageintheLosAmigosConservationConcession, MadredeDios,Per. WashingtonDC,USA:WinrockInternational;2006. 32.AsnerGP,MascaroJ,ClarkJK,PowellG: ReplytoStokeetal.:Regarding high-resolutioncarbonstocksandemissionsintheAmazon. ProcNatl AcadSci 2011, 108: E13 – E14. 33.CludiaDiasA,LouroM,ArrojaL,CapelaI: Comparisonofmethodsfor estimatingcarboninharvestedwoodproducts. BiomassBioenergy 2009, 33: 213 – 222. 34.PedroniL,DutschkeM,StreckC,PorraME: Creatingincentivesfor avoidingfurtherdeforestation:thenestedapproach. ClimPol 2009, 9: 207 – 220. 35.HardcastleP,BairdD,HardenV,AbbotPG,O ’ HaraP,PalmerJR,RobyA, HaslerT,AmbiaV,BranthommeA,WilkieM,ArendsE,GonzlezC: CapabilityandCostAssessmentoftheMajorForestNationstoMeasureand MonitortheirForestCarbon. Edinburg,Scotland:LTSInternationalLtd;2008. 36.RomijnE,HeroldM,KooistraL,MurdiyarsoD,VerchotL: Assessing capacitiesofnon-AnnexIcountriesfornationalforestmonitoringinthe contextofREDD+. EnvironSciPol 2012, 19 – 20: 33 – 48. 37.BurgessND,BahaneB,ClairsT,DanielsenF,DalsgaardS,FunderM, HagelbergN,HarrisonP,HauleC,KabalimuK,KilahamaF,KilaweE,Lewis SL,LovettLC,LyatuuG,MarshallAR,MeshackC,MilesL,MilledgeSAH, MunishiPKT,NashandaE,ShirimaD,SwetnamR,WillcockS,WilliamsA, ZahabuE: GettingreadyforREDD+inTanzania:acasestudyofprogress andchallenges. Oryx 2010, 44: 339 – 351. 38.ManiatisD,MalhiY,AndrLS,MolliconeD,BarbierN,SaatchiS,HenryM, TellierL,SchwartzenbergM,WhiteL:EvaluatingthePotentialof CommercialForestInventoryDatatoReportonForestCarbonStockand ForestCarbonStockChangesforREDD+undertheUNFCCC. IntJForest Res 2011, 2011: 1 – 13. 39. Scienceforachangingworld. http://www.usgs.gov/default.asp. 40.PlattsPJ,BurgessND,GereauRE,LovettJC,MarshallAR,MccleanCJ,Pellikka PKE,SwetnamRD,MarchantR: Delimitingtropicalmountainecoregions forconservation. EnvironConserv 2011, 38: 312 – 324. 41.SwetnamRD,FisherB,MbilinyiBP,MunishiPKT,WillcockS,RickettsT, MwakalilaS,BalmfordA,BurgessND,MarshallAR,LewisSL: Mapping socio-economicscenariosoflandcoverchange:aGISmethodtoenable ecosystemservicemodelling. JEnvironManag 2011, 92: 563 – 574. 42.WillcockS,PhillipsOL,PlattsPJ,BalmfordA,BurgessND,LovettJC,Ahrends A,BaylissJ,DoggartN,DoodyK,FanningE,GreenJMH,HallJ,HowellKL, MarchantR,MarchantR,MarshallA,MbilinyiB,MunishiPKT,OwenN, SwetnamRD,Topp-JorgensenEJ,LewisSL: Towardsregional,error-boundedWillcock etal.CarbonBalanceandManagement 2014, 9 :2Page15of17 http://www.cbmjournal.com/content/9/1/2

PAGE 16

landscapecarbonstorageestimatesfordata-deficientareasoftheworld. PLoSONE 2012, 7: e44795. 43.PlattsPJ,McCleanCJ,LovettJC,MarchantR: Predictingtreedistributions inanEastAfricanbiodiversityhotspot:modelselection,databiasand envelopeuncertainty. EcolModel 2008, 218: 121 – 134. 44.StuartEA: Matchingmethodsforcausalinference:Areviewandalook forward. StatSciRevJInstMathStat 2010, 25: 1. 45.WiegandRE: Performanceofusingmultiplestepwisealgorithmsfor variableselection. StatMed 2010, 29: 1647 – 1659. 46.HurttGC,FrolkingS,FearonMG,MooreB,ShevliakovaE,MalyshevS,Pacala SW,HoughtonRA: Theunderpinningsofland-usehistory:threecenturies ofglobalgriddedland-usetransitions,wood-harvestactivity,andresulting secondarylands. GlobChangBiol 2006, 12: 1208 – 1229. 47.BacciniA,LaporteN,GoetzSJ,SunM,DongH: Afirstmapoftropical Africa'sabove-groundbiomassderivedfromsatelliteimagery. Environ ResLett 2008, 3: 045011. 48.SaatchiSS,HoughtonRA,DosSantosAlvalRC,SoaresJV,YuY: DistributionofabovegroundlivebiomassintheAmazonbasin. Glob ChangBiol 2007, 13: 816 – 837. 49.MarshallAR,WillcockS,LovettJC,BalmfordA,BurgessND,LathamJE, MunishiPKT,PlattsPJ,SalterR,ShirimaDD,LewisSL: Measuringand modellingabove-groundcarbonstorageandtreeallometryalongan elevationgradient. BiolConserv 2012, 154: 20 – 33. 50.MunishiPKT,ShearTH: Carbonstorageinafromontanerainforestsofthe EasternArcMountainsofTanzania:theirnetcontributionto atmosphericcarbon. JTropForSci 2004, 16: 78 – 93. 51.ShirimaDD,MunishiPKT,LewisSL,BurgessND,MarshallAR,BalmfordA, SwetnamRD,ZahabuEM: Carbonstorage,structureandcompositionof miombowoodlandsinTanzania ’ sEasternArcMountains. AfrJEcol 2011, 49: 332 – 342. 52.PfeiferM,PlattsPJ,BurgessND,SwetnamRD,WillcockS,LewisSL,Marchant R: LandusechangeandcarbonfluxesinEastAfricaquantifiedusing earthobservationdataandfieldmeasurements. EnvironConserv 2013, 40: 1 – 12.FirstView. 53.RueschA,GibbsHK: NewIPCCTier1GlobalBiomassCarbonMapFortheYear2000. OakRidge,Tennessee:OakRidgeNationalLaboratory;2008. AvailableonlinefromtheCarbonDioxideInformationAnalysisCenter http://cdiac.ornl.gov/[Accessed15/01/12]. 54.HillTC,WilliamsM,BloomAA,MitchardETA,RyanCM: Areinventorybased andremotelysensedabove-groundbiomassestimatesconsistent? PLoS ONE 2013, 8: e74170. 55.BurgessND,MwakalilaS,MunishiP,PfeiferM,WillcockS,ShirimaD,Hamidu S,BulengaGB,RubensJ,MachanoH,MarchantR: REDDherringsorREDD menace:responsetoBeymer-FarrisandBassett. GlobEnvironChang 2013, 23: 1349 – 1354. 56.KessyJF,AndersonK,DalsgaarS: NAFORMAFieldManual:Socioeconomic survey. DaresSalaam,Tanzania:ForestryandBeekeepingDivision,Ministry ofNaturalResourcesandTourism;2010:96. 57.VesaL,MalimbwiRE,TomppoE,ZahabuE,MaliondoS,ChamuyaN,Nssoko E,OtienoJ,MiceliG,KaayaAK,DalsgaarS: NAFORMAFieldManual: Biophysicalsurvey. DaresSalaam,Tanzania:ForestryandBeekeeping Division,MinistryofNaturalResourcesandTourism;2010:96. 58.ChaveJ,ConditR,LaoS,CaspersenJP,FosterRB,HubbellSP: Spatialand temporalvariationofbiomassinatropicalforest:resultsfromalarge censusplotinPanama. JEcol 2003, 91: 240 – 252. 59.BacciniA,FriedlM,WoodcockC,ZhuZ: Scalingfielddatatocalibrateand validatemoderatespatialresolutionremotesensingmodels. PhotogrammEngRemoteSens 2007, 73: 945. 60.RamankuttyN,GibbsHK,AchardF,DefriesR,FoleyJA,HoughtonRA: Challengestoestimatingcarbonemissionsfromtropicaldeforestation. GlobChangBiol 2007, 13: 51 – 66. 61.BallsMJ,BdkerR,ThomasCJ,KisinzaW,MsangeniHA,LindsaySW: Effect oftopographyontheriskofmalariainfectionintheUsambara Mountains,Tanzania. TransRSocTropMedHyg 2004, 98: 400 – 408. 62.AhrendsA,BurgessND,MilledgeSAH,BullingMT,FisherB,SmartJCR, ClarkeGP,MhoroBE,LewisSL: Predictablewavesofsequentialforest degradationandbiodiversitylossspreadingfromanAfricancity. Proc NatlAcadSci 2010, 107: 14556 – 14561. 63.BayonG,DennielouB,EtoubleauJ,PonzeveraE,ToucanneS,BermellS: IntensifyingweatheringandlanduseinironageCentralAfrica. Science 2012, 335: 1219 – 1222. 64.PutzFE,ZuidemaPA,SynnottT,Pea-ClarosM,PinardMA,SheilD,Vanclay JK,SistP,Gourlet-FleuryS,GriscomB,PalmerJ,ZagtR: Sustaining conservationvaluesinselectivelyloggedtropicalforests:theattained andtheattainable. ConservLett 2012, 5: 296 – 303. 65.Topp-JrgensenE,PoulsenMK,LundJF,MassaoJF: Community-based monitoringofnaturalresourceuseandforestqualityinMontane ForestsandMiomboWoodlandsofTanzania. BiodiversConserv 2005, 14: 2653– 2677. 66.SilvaJNM,deCarvalhoJOP,LopesJCA,deAlmeidaBF,CostaDHM,de OliveiraLC,VanclayJK,SkovsgaardJP: Growthandyieldofatropicalrain forestintheBrazilianAmazon13yearsafterlogging. ForestEcolManag 1995, 71: 267 – 274. 67.GovenderN,TrollopeWSW,VanWilgenBW: Theeffectoffireseason,fire frequency,rainfallandmanagementonfireintensityinsavanna vegetationinSouthAfrica. JApplEcol 2006, 43: 748 – 758. 68.HoffmannWA: Theeffectsoffireandcoveronseedlingestablishmentin aNeotropicalSavanna. JEcol 1996, 84: 383 – 393. 69.ChhatreA,AgrawalA: Trade-offsandsynergiesbetweencarbonstorage andlivelihoodbenefitsfromforestcommons. ProcNatlAcadSci 2009, 106: 17667 – 17670. 70.ChhatreA,AgrawalA: Forestcommonsandlocalenforcement. ProcNatl AcadSci 2008, 105: 13286 – 13291. 71.MbwamboL,EidT,MalimbwiRE,ZahabuE,KajembeGC,LuogaE: Impact ofdecentralisedforestmanagementonforestresourceconditionsin Tanzania .In Forests,TreesandLivelihoods ;2012:1 – 17. 72.FisherB,LewisSL,BurgessND,MalimbwiRE,MunishiPK,SwetnamRD, KerryTurnerR,WillcockS,BalmfordA: Implementationandopportunity costsofreducingdeforestationandforestdegradationinTanzania. Nat ClimChang 2011, 1: 161 – 164. 73.DohertyRM,SitchS,SmithB,LewisSL,ThorntonPK: Implicationsoffuture climateandatmosphericCO2contentforregionalbiogeochemistry, biogeographyandecosystemservicesacrossEastAfrica. GlobChangBiol 2009, 16: 617 – 640. 74.WillcockS: Long-termChangesinLandCoverandCarbonStoragein Tanzania,EastAfrica .In UniversityofLeeds,SchoolofGeography ;2012. 75.NBS: TanzaniaCensus2002-AnalyticalReportVolumeX. DaresSalaam: NationalBureauofStatistics,MinistryofPlanning,Economyand Empowerment;2006:193. 76.ClarkDA,PiperSC,KeelingCD,ClarkDB: Tropicalrainforesttreegrowth andatmosphericcarbondynamicslinkedtointerannualtemperature variationduring1984-2000. ProcNatlAcadSciUSA 2003, 100: 5852 – 5857. 77.RaichJW,RussellAE,KitayamaK,PartonWJ,VitousekPM:Temperature influencescarbonaccumulationinmoisttropicalforests. Ecology 2006, 87: 76 – 87. 78.PhillipsOL,MalhiY,VincetiB,BakerT,LewisSL,HiguchiN,LauranceWF, VargasPN,MartinezRV,LauranceS,FerreiraLV,SternM,BrownS,GraceJ: Changesingrowthoftropicalforests:evaluatingpotentialbiases. Ecol Appl 2002, 12: 576 – 587. 79.EnquistBJ,NiklasKJ: Invariantscalingrelationsacrosstree-dominated communities. Nature 2001, 410: 655 – 660. 80.LiH-T,HanX-G,WuJ-G: Lackofevidencefor3/4scalingofmetabolism interrestrialplants. JIntegrPlantBiol 2005, 47: 1173 – 1183. 81.EnquistBJ,WestGB,BrownJH: Extensionsandevaluationsofageneral quantitativetheoryofforeststructureanddynamics. ProcNatlAcadSci 2009, 106: 7046 – 7051. 82.GoldsteinML,MorrisSA,YenGG: Problemswithfittingtothepower-law distribution. EurPhysJBCondensMatterComplexSyst 2004, 41: 255 – 258. 83.ZanneAE,Lopez-GonzalezG,CoomesDA,IlicJ,JansenS,LewisSL,Miller RB,SwensonNG,WiemannMC,ChaveJ: Globalwooddensitydatabase. 2009.http://hdl.handle.net/10255/dryad.235[Accessed5/12/2008]. 84.BakerTR,PhillipsOL,MalhiY,AlmeidaS,ArroyoL,DiFioreA,ErwinT, KilleenTJ,LauranceSG,LauranceWF,LewisSL,LloydJ,MonteagudoA,Neill DA,PatioS,PitmanNCA,SilvaJNM,VsquezMartnezR: Variationin wooddensitydeterminesspatialpatternsinAmazonianforestbiomass. GlobChangBiol 2004, 10: 545 – 562. 85.DjomoAN,IbrahimaA,SaborowskiJ,GravenhorstG: Allometric equationsforbiomassestimationsinCameroonandpanmoisttropical equationsincludingbiomassdatafromAfrica. ForestEcolManag 2010, 260: 1873 – 1885. 86.ChaveJ,AndaloC,BrownS,CairnsMA,ChambersJQ,EamusD,FolsterH, FromardF,HiguchiN,KiraT,LescureJ-P,NelsonBW,OgawaH,PuigH,Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page16of17 http://www.cbmjournal.com/content/9/1/2

PAGE 17

RiraB,YamakuraT: Treeallometryandimprovedestimationofcarbon stocksandbalanceintropicalforests. Oecologia 2005, 145: 87 – 99. 87.PlattsPJ: SpatialModelling,PhytogeographyandConservationinthe EasternArcMountainsofTanzaniaandKenya. In UniversityofYork, EnvironmentDepartment ;2012. 88.SwetnamRD: Historicallogginginprotectedareas,E.Tanzania.V2. Cambridge,UK:ZoologyDepartment,CambridgeUniversity;2011. 89.IUCN,UNEP-WCMC: TheWorldDatabaseonProtectedAreas(WDPA). Cambridge,UK:UNEP-WCMC;2010.Availableat:www.protectedplanet.net [Accessed03/05/2010)]. 90.ZomerRJ,TrabuccoA,BossioDA,VerchotLV: Climatechangemitigation: Aspatialanalysisofgloballandsuitabilityforcleandevelopment mechanismafforestationandreforestation. AgrEcosystEnviron 2008, 126: 67 – 80. 91. TropicalRainfallMeasuringMission. http://trmm.gsfc.nasa.gov/. 92.HijmansRJ,CameronSE,ParraJL,JonesPG,JarvisA: Veryhighresolution interpolatedclimatesurfacesforgloballandareas. IntJClimatol 2005, 25: 1965 – 1978. 93. Hole-filledSRTMfortheglobeVersion4,fromtheCGIAR-CSISRTM90m Database. http://srtm.csi.cgiar.org. 94. WindSpeedAt50mAboveTheSurfaceOfTheEarth. http://eosweb.larc. nasa.gov/sse/. 95.PerezR,IneichenP,MooreK,KmiecikM,ChainC,GeorgeR,VignolaF: A newoperationalmodelforsatellite-derivedirradiances:descriptionand validation. SolEnergy 2002, 73: 307 – 317. 96. LowResolutionSolarData. http://www.nrel.gov/gis/. 97.BatjesNH: SOTER-basedsoilparameterestimatesforSouthernAfrica. Volume4.Wageningen:ISRIC-WorldSoilInformation;2004:27. 98.ISRIC: SOTERandWISE-basedsoilpropertyestimatesforSouthernAfrica. 2010.Availableathttp://www.isric.org/[Accessed17/2/2010]. 99.RoyDP,JinY,LewisPE,JusticeCO: Prototypingaglobalalgorithmfor systematicfire-affectedareamappingusingMODIStimeseriesdata. RemoteSensEnviron 2005, 97: 137 – 162. 100.HoughtonRA,LawrenceKT,HacklerJL,BrownS: Thespatialdistributionof forestbiomassintheBrazilianAmazon:acomparisonofestimates. Glob ChangBiol 2001, 7: 731 – 746. 101.ClarkDB,ClarkDA: Landscape-scalevariationinforeststructureand biomassinatropicalrainforest. ForestEcolManag 2000, 137: 185 – 198. 102.DormannCF,McPhersonJM,ArajoMB,BivandR,BolligerJ,CarlG,Davies RG,HirzelA,JetzW,DanielKisslingW,KhnI,OhlemllerR,Peres-NetoPR,ReinekingB,SchrderB,SchurrFM,WilsonR: Methodstoaccountfor spatialautocorrelationintheanalysisofspeciesdistributionaldata:a review. Ecography 2007, 30: 609 – 628. 103.RDevelopmentCoreTeam: R:ALanguageandEnvironmentforStatistical Computing. In RFoundationforStatisticalComputing. Vienna,Austria:R FoundationforStatisticalComputing;2010. 104.ESRI: ArcMap9.3.1. 1999-2009. 105.VenablesWN,RipleyBD: ModernAppliedStatisticswithS. Oxford,UK: Springer;2002. 106.VarmaS,SimonR: Biasinerrorestimationwhenusingcross-validation formodelselection. BMCBioinforma 2006, 7: 91. 107.JohnsonJB,OmlandKS: Modelselectioninecologyandevolution. Trends EcolEvol 2004, 19: 101 – 108. 108.ChongI-G,JunC-H: Performanceofsomevariableselectionmethods whenmulticollinearityispresent. ChemomIntellLabSyst 2005, 78: 103 – 112. 109.AsnerG,FlintHughesR,VargaT,KnappD,Kennedy-BowdoinT: Environmental andBioticControlsoverAbovegroundBiomassThroughoutaTropicalRain Forest. Ecosystems 2009, 12: 261 – 278. 110.LloydJ,FarquharGD: EffectsofrisingtemperaturesandCO2onthe physiologyoftropicalforesttrees. PhilTransRoySocBBiolSci 2008, 363: 1811 – 1817. 111.GrahamEA,MulkeySS,KitajimaK,PhillipsNG,WrightSJ: Cloudcoverlimits netCO2uptakeandgrowthofarainforesttreeduringtropicalrainy seasons. ProcNatlAcadSci 2003, 100: 572 – 576. 112.MagginiR,LehmannA,ZimmermannNE,GuisanA: Improvinggeneralized regressionanalysisforthespatialpredictionofforestcommunities. JBiogeogr 2006, 33: 1729 – 1749. 113.BurgessND,ButynskiTM,CordeiroNJ,DoggartNH,FjeldsaJ,HowellKM, KilahamaFB,LoaderSP,LovettJC,MbilinyiB,MenegonM,MoyerDC, NashandaE,PerkinA,RoveroF,StanleyWT,StuartSN: Thebiological importanceoftheEasternArcMountainsofTanzaniaandKenya. Biol Conserv 2007, 134: 209 – 231.doi:10.1186/1750-0680-9-2 Citethisarticleas: Willcock etal. : Quantifyingandunderstanding carbonstorageandsequestrationwithintheEasternArcMountainsof Tanzania,atropicalbiodiversityhotspot. CarbonBalanceand Management 2014 9 :2. Submit your manuscript to a journal and bene“ t from:7 Convenient online submission 7 Rigorous peer review 7 Immediate publication on acceptance 7 Open access: articles freely available online 7 High visibility within the “ eld 7 Retaining the copyright to your article Submit your next manuscript at 7 springeropen.com Willcock etal.CarbonBalanceandManagement 2014, 9 :2Page17of17 http://www.cbmjournal.com/content/9/1/2