Ecology and Evolution. 2017;7:2585–2594.  | 2585www.ecolevol.org

Received:17August2016  |  Revised:20November2016  |  Accepted:24November2016DOI: 10.1002/ece3.2696

O R I G I N A L R E S E A R C H

Species distribution models predict temporal but not spatial variation in forest growth

Ernst van der Maaten1,2 | Andreas Hamann3 | Marieke van der Maaten-Theunissen1 |  Aldo Bergsma4 | Geerten Hengeveld5 | Ron van Lammeren4 | Frits Mohren2 |  Gert-Jan Nabuurs5 | Renske Terhürne4 | Frank Sterck2

ThisisanopenaccessarticleunderthetermsoftheCreativeCommonsAttributionLicense,whichpermitsuse,distributionandreproductioninanymedium,providedtheoriginalworkisproperlycited.© 2017 The Authors. Ecology and EvolutionpublishedbyJohnWiley&SonsLtd.

1InstituteofBotanyandLandscapeEcology,UniversityofGreifswald,Greifswald,Germany2ForestEcologyandForestManagementGroup,CentreforEcosystemStudies,WageningenUniversity,Wageningen,TheNetherlands3DepartmentofRenewableResources,UniversityofAlberta,Edmonton,AB,Canada4LaboratoryofGeo-InformationScienceandRemoteSensing,WageningenUniversity,Wageningen,TheNetherlands5Alterra,WageningenUniversityandResearchCentre,Wageningen,TheNetherlands

CorrespondenceErnstvanderMaaten,InstituteofBotanyandLandscapeEcology,UniversityofGreifswald,Greifswald,Germany.E-mail:ernst.vandermaaten@uni-greifswald.de

Funding informationNaturalSciencesandEngineeringResearchCouncilofCanada,Grant/AwardNumber:330527;SeventhFrameworkProgramme,Grant/AwardNumber:311970andMOTIVE;SixthFrameworkProgramme,Grant/AwardNumber:EFORWOOD;EuropeanCooperationinScienceandTechnology,Grant/AwardNumber:ECHOES;AlexandervonHumboldt-Stiftung;DutchMinistryofEconomicAffairs,AgricultureandInnovation,Grant/AwardNumber:226544.

AbstractBioclimateenvelopemodelshavebeenwidelyusedtoillustratethediscrepancybe-tweencurrentspeciesdistributionsandtheirpotentialhabitatunderclimatechange.However,therealismandcorrectinterpretationofsuchprojectionshasbeenthesub-jectofconsiderablediscussion.Here,weinvestigatewhetherclimatesuitabilitypre-dictionscorrelatetotreegrowth,measuredinpermanentinventoryplotsandinferredfromtree-ringrecords.WeusetheensembleclassifierRandomForestandspeciesoc-currencedatafrom~200,000inventoryplotstobuildspeciesdistributionmodelsforfour important European forestry species: Norway spruce, Scots pine, Europeanbeech,andpedunculateoak.Wethencorrelateclimate-basedhabitatsuitabilitywithvolumemeasurements from~50-year-oldstands,available from~11,000 inventoryplots.Secondly,habitatprojectionsbasedonannualhistoricalclimatearecomparedwithringwidthfrom~300tree-ringchronologies.Ourworkinghypothesisisthathab-itatsuitabilityprojectionsfromspeciesdistributionmodelsshouldtosomedegreebeassociatedwithtemporalorspatialvariationinthesegrowthrecords.Wefindthatthehabitatprojectionsareuncorrelatedwithspatialgrowthrecords(inventoryplotdata),buttheydopredictinterannualvariationintree-ringwidth,withanaveragecorrela-tionof.22.Correlationcoefficientsforindividualchronologiesrangefromvaluesashighas.82oraslowas−.31.Weconcludethattreeresponsestoprojectedclimatechangearehighlysite-specificandthatlocalsuitabilityofaspeciesforreforestationisdifficulttopredict.Thatsaid,projectedincreaseordecreaseinclimaticsuitabilitymaybeinterpretedasanaverageexpectationofincreasedorreducedgrowthoverlargergeographicscales.

K E Y W O R D S

climatechange,dendrochronology,Europeanbeech(Fagus sylvatica),Norwayspruce(Picea abies),pedunculateoak(Quercus robur),Scotspine(Pinus sylvestris),speciesdistributionmodels,tree growth

2586  |     van der MaaTen eT al.

1  | INTRODUCTION

Treespeciesdistributionsandforestproductivityarestronglylinkedto climatic factors through direct effects of climate conditions ontree physiological processes (Running etal., 2004). As a conse-quence, climate change is expected to affect forest growth andhealth in the short term (Boisvenue & Running, 2006), aswell asgeographicdistributionofspecies in the longtermthroughdemo-graphicprocesses (Davis&Shaw,2001;Parmesan&Yohe,2003).Fortemperature-limitedborealforests,thegeneralexpectationisaprofoundnorthwardshiftofsuitabletreespecieshabitat,whilethesituationintemperateregionstendstobemorecomplexandvariesbetweenMediterranean,continental,andmaritimeclimates(Bonan,2008).Directand indirectclimate impacts,suchaswatershortageandincreasedriskofforestfires,appeartothreatenMediterraneanforests(Allenetal.,2010;Schröteretal.,2005),whileforestgrowthmight benefit in continental andAtlantic forests, but only at siteswhere an increased evaporative demand can be compensated bysufficientwater availability (Lindner etal., 2010; Spathelf, van derMaaten, van der Maaten-Theunissen, Campioli, & Dobrowolska,2014).

Although predictions of climate change (impacts) still carry un-certainties regarding the magnitude of responses, there is grow-ing awareness among forestry professionalsworldwide that climatechangepotentiallyposesalargethreattotheeconomicandecologicalvalueof forest lands (Lindneretal., 2014).ForEurope,Hanewinkel,Cullmann, Schelhaas, Nabuurs, and Zimmermann (2013) estimatedthateconomic lossesmaytotal severalhundredbillioneurosby theendofthecenturywithoutappropriateadaptationstrategiesfortheforestry sector. To guide species choice in reforestation and forestmanagementprescriptionstoaddressclimatechange,speciesdistri-butionmodels(SDMs)arepotentiallyausefultool.Thesecorrelativemodelsemployavarietyofstatisticalormachinelearningtechniquestorelatespeciespresence/absencedatatorelevantpredictorvariablessuchasclimatedata,topo-edaphicvariables,orotherhabitatfactors(e.g., Guisan & Zimmermann, 2000).Although there are exceptions(e.g.,O’Neill,Hamann,&Wang,2008),SDMsnormallypredictthere-alizednichespaceofspecies.

Themodelshavebeenwidelyusedto illustratethediscrepancybetweencurrenttreespeciesdistributionsandtheirpredictedpoten-tialhabitatunderclimatechange(e.g.,Iverson&Prasad,1998;Loarieetal.,2008;Thomasetal.,2004;Thuiller, Lavorel,Araújo,Sykes,&Prentice,2005).Therealismandcorrect interpretationofsuchpro-jections, however, has been the subject of considerable discussion(e.g., Botkin etal., 2007; García-Valdés, Zavala, Araújo, & Purves,2013;Hampe,2004;Thuilleretal.,2008),andthereisconsensusinthescientificcommunitythatSDMsareconceptuallyinadequatetoaccurately predict demographic processesof species under climatechange.Modelingtherealizedclimaticnichedoesnotrevealthefullrangeofthespecies’ecologicaltolerancesand limitations,andpre-dictedgainorlossofsuitableclimatehabitatdoesnotimplyanimme-diatethreattoaspeciespopulationorrapidexpansionofaspecies’range.

DespitethelimitationsofSDMsforclimatechangeimpactassess-ments on complex ecological systems, it has beenpointedout thatspecies distributionmodels are conceptuallywell suited for simplerpractical tasks: guiding climate change adaptation strategies thatinvolve habitat restoration or choosing suitable tree species for re-forestation (Gray&Hamann,2011,2013;Hamann&Aitken, 2013;Schelhaas etal., 2015). For suchmanagement applications, the pri-marytaskistomatchsourceandtargetenvironments.Nevertheless,itisuncertainwhethersubsequentlong-termforestgrowthandforesthealtharewelldescribedbyspeciesdistributionmodelsthatmaybeused to guide initial decisions on species choice for a general geo-graphicregion.

Here, we contribute a retrospective analysis how SDM-derivedhabitat suitability projections for four major European tree species(Norwayspruce—Picea abies(L.)Karst.,Scotspine—Pinus sylvestrisL.,European beech—Fagus sylvaticaL., and pedunculate oak—Quercus roburL.)correlatewithforestgrowthdatafrom long-term inventoryplots and tree-ring chronologies. Our hypothesis is that periods ofmarginalgrowthobservedinthetree-ringrecord(forexample,duringcoldordryepisodes)maybepredictedbyannualhindcastsofhabi-tatsuitabilityfromspeciesdistributionmodels.Suchcorrelationsbe-tweengrowthandmodeledhabitatsuitabilitymayvaryfromsitetosite.Forexample,projectedhabitatlossduringdroughtperiodsmaycorrelatewellwithtree-ringrecordsonwater-limitedsitesbutnotonwetsites,whichwouldpoint to important interactionsbetweencli-maticandnonclimaticabioticfactors.Notwithstandinglargevariationfromnonclimaticsitefactors,projectionsofclimatichabitatsuitabil-ityshouldtosomedegreecorrelatepositivelywithlong-termgrowthrecords.Strongspatial(inventoryplots)ortemporal(tree-ring-based)associationsof growthwith climatemay increaseour confidence inusingspeciesdistributionmodelstoguideclimatechangeadaptationstrategiesinforestryandecosystemmanagement.

2  | METHODS

2.1 | Forest inventory and tree- ring data

Weuseaforestinventorydatabasefor17Europeancountries,includ-ingBelgium(9,075plots),Croatia(39),Estonia(1,598),Finland(1,690),France(1,741),Germany(54,087),Italy(13,972),Lithuania(744),theNetherlands(1,442),Norway(8,629),Romania(196),Slovakia(1,410),Slovenia(38),Sweden(2,784),Ukraine(126),andtheUnitedKingdom(19,166).ThedatabasewasoriginallycompiledbyBrusetal. (2012)andNabuurs(2009).Inaddition,weaddedanewlyavailableSpanishforest inventory database (82,527) that is publicly available fromtheSpanishMinistryofAgriculture, Food andEnvironment (http://www.magrama.gob.es/es/desarrollo-rural/temas/politica-forestal/inventario-cartografia/inventario-forestal-nacional).

Fora subsetof12countries that still representabroad rangeofclimate conditions throughoutEurope, informationon standageandspecies-specificstandingstockvolumeswasavailable(excludingCroatia,theNetherlands,Romania,Ukraine,andUnitedKingdom).Ratherthanattemptingage-basedadjustments,weonlyusevolumedatafor40-to

     |  2587van der MaaTen eT al.

60-year-oldstands(hereafterreferredtoas~50-year-oldplot/volumedata)resultingin11,539plotsfromatotalof199,264inventoryplotswith species-specific presence/absence data. The presence/absencedatawere used for building species distributionmodels that predicthabitatsuitability,whilethesubsetof~50-year-oldplotdatawasusedtoanalyzeassociationsbetweenhabitatsuitabilityandforestgrowth.

Tovalidatepredictionsofinterannualvariationinhabitatsuitability,weobtainedtree-ringdatafrom295sitesbothfromtheInternationalTree-RingDataBank(ITRDB)(NOAA,2016),aswellasfrompreviousstudies byvan derMaaten (2012) andvan derMaaten-Theunissen,Kahle,&vanderMaaten(2013).Wedetrendedallindividualtree-ringseriesbyfittingacubicsmoothingsplinewitha50%frequencycutoffat30yearsinordertoremovenonclimaticgrowthresponses(seealsoCook&Peters,1981),suchasbiologicalagetrendsoreffectsoffor-estmanagement.Indiceswerecalculateddividingtheobservedbythepredictedvalues,resultinginstandardizedseriesthataredimension-lessandhaveameanofone.Finally,thestandardizedchronologiesforindividualtreeswereaveragedpersiteinso-calledsitechronologiesusingabi-weightrobustmean.

2.2 | Climate data

ClimatedataweregeneratedusingthesoftwarepackageClimateEU(Hamann, Wang, Spittlehouse, & Murdock, 2013; Wang, Hamann,Spittlehouse,&Murdock,2012),availableforanonymousdownloadat http://tinyurl.com/ClimateEU. The ClimateEU package is a soft-ware front-end for interpolated climate databases generated withtheParameter-elevationRegressionson IndependentSlopesModel(PRISM)(Dalyetal.,2008).Thesoftwareallowstoquerymonthlyhis-toricalclimatedatafrom1901to2013,aswellastogenerategriddedclimatesurfacesforEuropeforhabitatsuitabilitymodeling.Thesoft-wareimplementsdownscalingalgorithmsthatuseempiricallyderivedlocallapseratesforindividualclimatevariablestoadjustforanydis-crepanciesbetweentheelevationofsamplelocations(tree-ringchro-nologiesandinventoryplots)andgriddedclimatedatabasesthattheClimateEUsoftwarequeries(Hamannetal.,2013;Wangetal.,2012).

Weusethe30-yearclimatenormalperiodfrom1961to1990asaclimatereferenceperiod,anda15-yearclimateaveragefrom1995to2009torepresentrecentobservedclimatechange(inventoryplotandtree-ringdatawerenotavailableformorerecentyears).Weuse10biologicallyrelevantclimatevariablesthataccountformostofthevariance in climate data while avoiding multicollinearity: mean an-nual temperature, themeantemperaturesof thewarmestandcold-est month, the difference between July and January temperatureas an indicator of continentality,mean annual precipitation,May toSeptember(growingseason)precipitation,growingdegreedaysabove5°C,frost-freedays,andtwodrynessindicesafterHogg(1997):anan-nualclimatemoistureindexandaJune–Augustsummerclimatemois-tureindex.ThevariablesareexplainedindetailbyWangetal.(2012).

Forspatialhabitatmodeling,weuse1-kmresolutionclimategridsinAlbersEqualAreaprojection.Speciesdistributionmodelswerebuiltbasedonawidelyusedreferenceperiodthatlargelypredatesasignif-icantanthropogenicwarmingsignal(the1961–1990climatenormal).

Projectionsweremadeforthisreferencenormalperiod,amorerecentaverage(1995–2009),aswellasforthe2020s(2011–2040)and2050s(2041–2070)usinganensembleaveragefromtheCMIP3multimodeldatasetcorrespondingtothefourthIPCCassessmentreport (Meehletal.,2007).SimilartoFordham,Wigley,&Brook(2011),weexcludedpoorlyvalidatedAOGCMs(MIROC3.2,MRI-CGCM2.3.2,MIROC3.2,IPSL-CM4,FGOALS-g1.0,GISS-ER,GISS-EH,andGISS-AOM)andre-tainedtheremainingCMIP3models.Thestudywas initiatedbeforetheCMIP5datasetcorrespondingtothefifthIPCCassessmentreportbecameavailable.However,wenote that the twoAOGCMgenera-tionsyield remarkably similar projections inmagnitude, uncertainty,andspatial resolution (Knutti&Sedláček,2013).Accountingfordif-ferentapproachestodescribeemissionscenarios(SRESversusRCP),wefindthattheprojectionsatthelevelofmultimodelensembleaver-agesremainlargelythesameforbothtemperatureandprecipitationvariables.

2.3 | Species distribution modeling

We built species distributions using the RandomForest ensembleclassifier (Breiman, 2001) implemented by the randomForest pack-age (Liaw & Wiener, 2002) for the R programing environment (RDevelopmentCoreTeam,2016).Thisensembleclassifiergrowsmul-tipleclassificationtrees(here,n=500)frombootstrappedsamplesofthetrainingdataanddeterminesitspredictionbymajorityvoteoveralldevelopedclassificationtrees(Cutleretal.,2007).Importanceval-uesforthepredictorvariableswerecalculatedasthefrequencythata particular climate variable contributed to a correct classification.Habitatprojectionsweremade(1)forclimatedataforthe1961–1990period to analyzeassociationsbetweenprojectedhabitat suitabilityandgrowthobservedonforestinventoryplots,(2)forclimatedataforindividualyearsfrom1901to2009toanalyzeassociationsbetweenprojectedhabitatsuitabilityandgrowthtree-ringwidth,and(3)forarecentclimateaverage (1995–2009)and futureperiods (2020sand2050s)toinferfuturetrendsingrowthpatternsacrossEurope.

TheRandomForestalgorithmwaspreferredoverotheralgorithms,likemaxent, becausewehada ratherunproblematicdatasetwith ahigh number of census records. To account for nonlinearity in thespeciesresponseacrossclimategradientsandforinteractionsamongclimate variables, RandomForest is generally considered the mostpowerfulimplementationofregressiontreetechniques.

Wealsomadeanattempttousetheregressiontree(ratherthanclassification tree) functionality of RandomForest to model a con-tinuous response variable (i.e., ~50-year-old plot volume data) as afunctionofclimate.Forthevolumemodels,webuilt,foreachofthestudiedtreespecies,atrainingdatasetofapproximately9,000sam-ples,whichwereclimaticallycharacterizedandcomprisedequalnum-bersofplotswithvolumedata,absenceplots,andrandomabsences.Randomabsenceswereonlyselectedforcountrieswithoutplot-dataavailabilitybyusinganoverlaywithtreespeciesdistributionsfromtheEuropeanForestGeneticResourcesProgramme(EUFORGEN,2012).We removed pseudo-absences using a p >.5 threshold for speciespresenceusingRandomForesthabitatprojectionsbasedonpresence/

2588  |     van der MaaTen eT al.

absencevariantsof the trainingdatasets.Thisanalysisdidnotyieldacceptablevalidationstatistics,andwebrieflyreportonthisnegativeresultfortheinventorydata-climatemodelingattempt.

Model performancewas evaluatedusing the area under the re-ceiver operating characteristic curve (AUCof ROC) to evaluate thestatistical accuracy of the species distributionmodels for individualtreespecies.TheAUCstatistic isacommonmeasureof theperfor-manceofclassificationrules;itbalancestheabilityofamodeltode-tectaspecieswhenitispresent(sensitivity)againstitsabilitytonotpredict a specieswhen it is absent (specificity) (e.g., Fawcett,2006;Fielding&Bell,1997).Wefurtherreportmodelsensitivity(calculatedasTP/(TP+FN)withTPastruepositivesandFNasfalsenegatives)andmodelspecificity(TN/(TN+FP)withTNastruenegativesandFPasfalsepositives).AllROCandAUCcalculationswere implementedwith the ROCR package (Sing, Sander, Beerenwinkel, & Lengauer,2005)fortheRprogrammingenvironment.

2.4 | Statistical analysis

To assess whether habitat projections of our species distributionmodelsareassociatedwithobservedforestgrowth,weconductedaPearsoncorrelationanalysisbetweentheprobabilityofpresencees-timatesfromtheRandomForest-basedhabitatsuitabilityprojectionsand growthmeasurements from two data sources: inventory plots(to represent spatial variation)or growth increments from tree-ringdata (to represent temporalvariation).All correlationswerevisuallycheckedforlinearity,andtransformationswerejudgedasunlikelytohaveanynotableorconsistenteffectontheresults.Habitatprojec-tionsforcorrelationswith~50-year-oldplotvolumedatawerebasedon one long-term projection for the 1961–1990 normal period (toanalyzespatialgrowth−climateassociations).Habitatprojectionsfor

comparisonwithtree-ringincrementswerebasedonhabitatprojec-tionsfor109individualyears,from1901to2009(toanalyzetemporalgrowth−climateassociations).

Becausetree-ringchronologiesareknowntodisplaytemporalau-tocorrelations(Fritts,1976),weallowedforamaximumlagof3yearsfor the correlation analysis between annual habitat projections andsite chronologies. Further, because forest ecosystems can be wellbuffered against short-term climate fluctuations,we also evaluatedcorrelationsbasedon3-,5-,7-,and9-yearmovingaveragesofbothhabitatprojectionsandsitechronologies.Generally,5-yearmovingav-eragesgeneratedthestrongestcorrelationcoefficients(bothnegativeandpositive),andwasthereforeappliedtoallchronologies.Todeter-minewhethertheaveragecorrelationcoefficientwaslargerthanzero,a single-sample one-tailed t-test across all correlation coefficientsfromeachspecieswascarriedout,testingtoouroverallscientificnullhypothesis that therewas no positive association between growthandclimatichabitatsuitability.

3  | RESULTS

3.1 | Variable importance and model statistics

RandomForest importancevalues indicatethatclimatepredictorsofspeciesrangesarequitespecies-specific,withanotablecontrastbe-tweenconiferousanddeciduoustrees(Table1).Forspruceandpine,themeanwarmestmonth temperature contributes strongly to cor-rectlypredictingpresenceandabsenceofthesespeciesininventoryplotdata.Further, the summerandannual climatemoisture indiceswere important predictor variables for spruce. In contrast, variableimportancevaluesforsummerheatandmoisturevariableswereverylowforthebroadleavedtreespeciesoakandbeech.Forthesespecies,

Climate variable

RF importance

Norway spruce Scots pine European beech Pedunc. oak

Meanannualtemperature(°C)

2.8 2.0 1.8 1.7

Meanwarmestmonthtemperature(°C)

7.8 8.9 3.2 3.0

Meancoldestmonthtemperature(°C)

5.6 2.0 4.9 2.3

Continentality(°C) 4.9 4.1 8.0 4.6

Meanannualprecipita-tionsum(mm)

3.1 2.5 3.2 1.2

Growingseasonprecipitationsum(mm)

4.3 4.3 2.2 1.9

Growingdegreedays>5°C(days)

5.8 6.0 2.4 2.8

Frost-freeperiod(days) 4.6 2.0 2.6 3.0

Annualclimatemoistureindex(cm)

6.7 5.1 2.0 2.6

June–Augustclimatemoistureindex(mm)

8.5 4.9 2.1 2.8

TABLE  1  ImportancevaluesofRandomForestclimatepredictorvariables,calculatedasthenumberoftimesthataparticularvariablecontributedtoacorrectclassification.Reportedvaluesaredividedby100forreadability

     |  2589van der MaaTen eT al.

continentalityandothercold-relatedvariablesstandoutasthebestpredictorsofspeciesranges.Continentalityisthemostimportantpre-dictor for bothbroadleaves,withmean coldestmonth temperaturethesecondmostimportantpredictorforbeechoccurrencesandsev-eralvariables,includingfrost-freeperiod,beingsecondarypredictorsforoak.

Accuracy statistics for the presence/absence predictions areshown inTable2.Totalerror ratesof falsepositivesandfalsenega-tivesrangebetween0.13and0.32,withthewidespreadconiferousspecieshavingthehighesterrorrates.AUCvaluesarefairlyhighrang-ingfrom0.72to0.90,withbeechandoakhavingthebestpredictiveaccuracy.Forallspecies,thenumberoffalse-negativeerrorsishigherthanthenumberoffalse-positiveerrors,whichindicatesthatmodelpredictionerrors aremainly drivenby falsely predicting species ab-sence.Similarly,modelsensitivityislowandmodelspecificityishigh,showingthattruespeciesabsenceswerewellmodeled.

TheresultsofthedistributionmodelforNorwayspruceareshowninFigure1,withcorrespondingmapsforScotspine,Europeanbeech,

andpedunculateoakprovidedasSupportinginformation(FigsS1–S3).Acomparisonofboththepresencedataandthepredictedspeciesdis-tributionwiththeapproximatenaturaldistribution(afterEUFORGEN,2012;seeinsetFigure1)revealssubstantialdifferences.Today’sdis-tributionofspruceacrossEuropehasadistinctivelywiderrangethanitsnaturaldistributionsuggests,duetoNorwaysprucebeingwidelyplantedas ahighlyvalued forestry species fortimberproduction inEurope.

3.2 | Habitat projections versus growth

The analysis of temporal growth–climate associations based onhabitat suitabilitypredictions andannual growth increments in295tree-ringchronologiesrevealsthatthesecorrelationsarehighlyvari-ableinmagnitudeandevendirection.Overall,wefoundanaveragecorrelationacrossallfourspeciesof .22butwithcorrelationcoeffi-cientsfromindividualchronologiesashighas.82,oftenapproachingzero, and regularlybeingnegativewith valuesup to−.31 (Table3).Examplesfortimeserieswithpositive,neutral,andnegativeassocia-tions are shown in Figure2, and amapof the resulting correlationcoefficientsforallindividualchronologiesofNorwayspruceisshowninFigure3.Nospatialpatternsareapparent,andelevationorclimatenormalvariablesfor theoriginofsample locationswerenotassoci-atedwiththestrengthordirectionsofcorrelationcoefficients (datanotshown).

Even though there are no apparent associations of correlationcoefficientswithgeographicorclimaticvariablesofthesampleloca-tions,theaverageassociationbetweenclimatesuitabilityandgrowthinferredfromringwidthispositiveandoverallhighlysignificant.The

TABLE  2 Predictiveaccuracystatisticsfortheprojecteddistributionareasofthefourstudyspecies

Species Error rate Specificity Sensitivity AUC

Norwayspruce 0.25 0.71 0.65 0.81

Scotspine 0.32 0.63 0.60 0.72

Europeanbeech 0.16 0.85 0.62 0.90

Pedunculateoak 0.13 0.79 0.70 0.90

Error rate=(False Positive+False Negative)/(Total Positive+TotalNegative).

F IGURE  1 SampleplotdatafortheNorwaysprucepresence(●)andinventoryplotsthatalsocontainedheight,diameter,andvolumedata(Δ).Themodeledspeciesdistributionisbasedonprobabilityofpresenceestimateabove.4( ),wherefalse-positiveandfalse-negativepresence/absencepredictionsareminimized(seeTable2forstatistics).Notethatabsencedatawereomittedfromthefigure.TheinsetshowstheapproximatenaturaldistributionofthespeciesaccordingtoEUFORGEN(2012)

Picea abies (Norway spruce)

2590  |     van der MaaTen eT al.

probabilityof the trueaveragecorrelationhavingavalueofzeroorsmaller (one-sided t-test) is negligible, except for European beech,wherewelackthesamplesizetorejectthenullhypothesiswithcon-fidence(Table3).

Habitatprojectionsforinventoryplotlocationsbasedonthe30-yearclimatenormalperiod (1961–1990) foreach locationwerenotsignificantlycorrelatedwithstandingvolumeatage~50.Infact,visualexaminations suggest no associations at all with correlation coeffi-cientsapproachingzeroforallspecies(datanotshown).Similarly,cor-relationsbetweendirectRandomForestmodelpredictionsofstandingvolume trained with ~50-year-old plot volume data and validatedagainstawithhelddatasetofplotmeasurementsapproachzero(datanotshown).

3.3 | Future habitat suitability projections

Theaboveanalysis shows thatmodeloutputsmustbe interpretedwith caution. Our validations against growth data from tree ringsandinventoryplotssuggestthathabitatprojectionsarenotinforma-tiveataforeststandlevelandonlyrepresentaverageexpectationsoverlargergeographicareas.Weremindthereaderthatstandlevelvariation in effects of climate changeon tree growthmaybe verylarge (and occasionally reversed) when compared to the generalexpectationofgrowthtrendsunderclimatechange.Withtheseca-veatsclearlystated,Figure4 illustrateschangesinmodeledhabitatsuitability and by inference also expected changes in tree growthforNorwayspruceunderrecentlyobservedclimatechange(1995–2009)andfutureclimateperiods(2020sand2050s).FiguresS4–S6showcorrespondingprojectionsforScotspine,Europeanbeech,andpedunculateoak.

Withinthecurrentextentofthespeciesrange,habitatsuitabilityofsprucegenerallydecreasesinsouthernandcentralEurope,whereasitincreases(orremainsstable)inthenorthorathigherelevation.Similartrendsareobservedforpine,beech,andoak(FigsS4–S6).Further,ourmodelprojectionsindicatethatclimatechangeishappeningasfastasorfasterthanprojectedbygeneralcirculationmodels.Namely,hind-castprojectionsfor1995–2009periodalreadyappearapproximatelyequalto2020sprojections.

4  | DISCUSSION

4.1 | Species distribution model evaluation

Our species distribution models appear to be reasonably accuratewhen validated against withheld presence/absence data. The twobroadleaved tree species (oak andbeech) had very good validationstatisticswithAUCsof .9,while the twoconifers (pineandspruce)hadmoderatepredictiveaccuracieswithAUCsbetween.7and.8.Theimportancevaluesofpredictorvariablesappearlargelysensible:Theconiferousspecieshadnorangelimitationstowardcoldnortherncli-matesandtheirrangeswithinthestudyareawerethereforebestpre-dictedbywarmtemperaturesanddryness,delineatingthesouthernandlow-elevationrangelimits(Table1,Figures1andS1).Incontrast,thespeciesrangeofthetwobroadleavedtreeswereboundtowardnorthernandhigh-elevationrangelimitswithinthestudyareaaswell.

Species N

Distribution of correlation coefficients (r)

p (r ≤ 0)Minimum Mean Maximum

Norwayspruce 126 −.31 .25 .82 <.0001

Scotspine 128 −.27 .18 .61 <.0001

Europeanbeech 4 −.10 .31 .49 .1083

Pedunculateoak 37 −.22 .19 .72 <.0001

Correlationcoefficients(r)werecalculatedbetween5-yearmovingaveragesofthehabitathindcasts(1901–2009),andcorresponding5-yearmovingaveragesofsitechronologies.Thenumberofchro-nologies(N),theminimum,mean,andmaximumr,aswellastheprobabilitythatthemeancorrelationcoefficientforeachspeciesissmallerorequaltozero,arereported.

TABLE  3 ResultsofPearsoncorrelationanalysesbetweensitechronologiesoftree-ringdataandannualhabitatsuitabilityhindcasts

F IGURE  2 Timeseriesoftree-ringindicesandpredictionsofhabitatsuitabilityforthreesamplechronologiesofNorwaysprucewithahigh(a),alow(b),andanegative(c)correlationcoefficient.ForamapofNorwaysprucecorrelationcoefficientsseeFigure3.Fordistributionstatisticsofallcorrelationcoefficientsforallspecies,seeTable3

0.6

1

1.4

0.4

0.6

0.8

0.7

1

1.3

0.3

0.6

0.9

1900 1920 1940 1960 1980 2000

0.8

1

1.2

0.3

0.6

0.9

(a)

(b)

(c)

Tree

-rin

g ch

rono

logy

Probability of presence

Year

Tree-ring chronology

Probability of presence r = 0.80

r = 0.33

r = –.02

     |  2591van der MaaTen eT al.

Therefore,abroaderrangeofclimatevariableshadpredictivevalue(Table1,FigsS2andS3).

Notably,when comparing the insets of Figure1 and the corre-spondingFigsS1–S3,whichreflecttheoriginalnativespeciesranges(EUFORGEN2016)withthemodeledspeciesrangesandtheoccur-rencerecords(showninthemainmapofthesamefigures),itappearsthatthespeciesdistributionmodelgenerallyoverpredicts.However,treespeciesinEuropehavebeenplantedoutsideoftheirnaturalrangeforcenturies(Lindneretal.,2014),andthespeciesdistributionmodeltherefore likely reflects the larger fundamentalnicheof thespecies.Whilethefundamentalniche (i.e., theabsoluteenvironmental toler-ances) cannot be comprehensively inferred fromplot data (becausewedonotknowwhichenvironmentswerenevertested),ourmodeling

effort might approach the species’ fundamental niche (because ofplanting efforts outside the natural range). Habitat loss projectedby themodels in this study (putatively approaching the fundamen-talniche)should thereforebeofhigherconcern thanprojectionsofamore restricted realized nichemodel (based on a just the naturalspecies’ range).Habitat loss in thisstudycould imply thatprojectedclimatesareoutsidethespeciestolerancesratherthan just favoringcompetitorsinthelongterm.

4.2 | SDM projections versus growth records

Incorrelatingmodeloutputsofhabitatsuitabilitywithgrowthrecordsfromtreerings,allspeciesshowedaverysimilarrangeofcorrelation

F IGURE  3 CorrelationcoefficientsbetweenpredictionsoftheNorwaysprucespeciesdistributionmodelfor5-yearmovingaveragesofhabitatsuitabilityhindcasts(1901–2009),andcorresponding5-yearmovingaveragesofsitechronologies.Thelocationofsitechronologiesisindicatedbycircles,thesizeofcirclesrepresentsthestrengthofthecorrelation,andchronologiesfromsiteshigherthan1,500maremarkedbythickborders

Picea abies (Norway spruce)

Correlation coefficients

< 0.2

0.2 – 0.4

0.4 – 0.6

0.6 – 0.8

2592  |     van der MaaTen eT al.

coefficients (Table3),which reflectsasimilarlybalancedsetofpro-jectionsofchanges (compareFigures4andS4–S6).Thespeciesdonotdifferdramaticallyintheirsensitivitytoclimatechangebasedoncontinental-scaleprojections,nordotheydifferdramaticallyintheirresponseto interannualclimatevariationatthespecificsiteswheretheyoccurbasedontree-ringchronologies.Further,therangeofcor-relationcoefficientsissurprisinglylargeandincludesnegativeaswellaspositiveassociationswithclimatesuitabilityprojections.Thisresultmaynotbeunexpectediftheanalysiswascarriedoutforindividualclimate variables. For example, a high-temperature anomaly would

beexpectedtoyieldgoodgrowthat thenorthernedgeof thespe-ciesdistributionbutlowerthanaverageperformanceatthesouthernrangelimit.Thisis,however,notaplausibleexplanationfortheresultat hand. RandomForest is a regression tree classifier and thereforewell capable of modeling nonlinear growth responses to predictorvariables,aswellasinteractionsamongthosevariables.ThenegativeresultfurtherconformstoFigure3whichdoesnotshowanyspatialpatternsthatareassociatedwiththestrengthofthecorrelations.

We conclude that tree responses to projected climate changeare highly site-specific and that local suitability of a species for

FIGURE  4 PredictedclimatichabitatsuitabilityforNorwaysprucebasedontheclimateperiodofthetrainingdataset(1961–1990),andchangesinhabitatsuitabilityforarecent15-yearclimateperiod(1995–2009),andensembleprojectionsforthe2020sand2050softheCMIP3multimodeldatasetfortheemissionscenarioA2.Notethatpredictionsforallclimateperiodshavebeenlimitedtothecurrentextentofthespeciesrange.Iftheabsoluteprobabilityofpresencewaspredictedtobebelow.4,thehabitatismarkedaslostrelativetothe1961–1990baselineprojection

Picea abies (Norway spruce)

PoP differencefor 1995-2009climate

Probability ofpresence (PoP) for 1961-1990climate

PoP differencefor 2050sclimate

PoP differencefor 2020sclimate

Probability ofPresence (PoP) <0.4 0.4 1.0

Difference in PoP –1.0 0 +1.0 Loss

     |  2593van der MaaTen eT al.

reforestationisdifficulttopredict.Thisisfurtherconfirmedbyourun-successfulattempttopredictstandingvolumeatanageof~50yearsdirectly with a RandomForest classifier. RandomForest projectionsexplainedlessthan5%oftheoverallvarianceinthiscontinuouspre-dictor variable, indicating that local site factors overwhelm climatevariablesaspredictors.Itshouldbenoted,however,thatthisappliestositeswhere thespeciesoccursandhasbeengrowing~50years.Thus, the inference that climate has no influence on tree growthshouldnotbemadefromthisobservation.Rather,weconcludethatforwell-establishedtreesat localsites, thecumulativevolumeoveraperiodof50yearsisnotafunctionofclimatebutpredominantlyafunctionofothersitefactors.

4.3 | Regional- scale growth projections

Despite thewide range of correlation coefficients between habitatsuitabilityprojectionsandtree-ringrecords(Table3),wehavestrongindicationsthattheydonotrepresentrandomnoise,butrealclimatehabitat–growthassociations.Thecorrelationcoefficientsaregener-allypositivewithanaverageof.22anddifferhighlysignificantlyfromzero.Theprobabilitytoobtainanequalorlargeraveragecorrelationcoefficientbyrandomchanceisextremelysmallforallspecies,exceptforEuropeanbeech,wherewewereonlyabletoevaluatefourtree-ring chronologies. Thus,while local site factors such as soils, topo-graphicexposure,andgroundwateravailabilitymayaccountforhowclimatechangemayaffectforestgrowthatindividualsites,significantpositivecorrelationsacrossmultiplesites(Table2)suggestthathabi-tatprojectionsarecapableofcorrectlypredictinggrowthtrendsonaverage. Therefore, we conclude that SDM projections should beinterpretedasaverageexpectationsofincreasedorreducedgrowthoverlargergeographicscales.

Modelpredictionsforthe1995–2009climateperiodcomparedtothe1961–1990baseline,representinga28-yearwarmingtrend(i.e.,midpoint2002minusmidpoint1975), reveal that habitat suitabilityofNorway sprucedeclined in themore southern anddrierpartsofits currentdistribution,whereas it increased in thenorth (Figure4).Interestingly,projectionsdrivenbyalreadyobservedclimatechangeverycloselyresemblethe2020sprojection.Thisimpliesthatclimatechangeappears tomaterialize faster thanprojectedevenunder thepessimisticA2SRESscenario,whichisthebasisfortherevisedRCP8.5scenario,thatis,anemissionstorylinethatassumeshighpopula-tiongrowthand lower incomes indevelopingcountries (vanVuurenetal.,2011).Whilethisoutlookshouldraiseconcern,weshowinthispaperthattheclimateimpactsonforestgrowthwillbehighlyvariableatlocalscales.Asaconsequence,climatechangeadaptationstrategiesfortheforestrysectorthataimatmovingspeciestoappropriatelocalsiteconditionsappearatleastaspromisingaslarge-scalegeographicassistedmigrationefforts.

ACKNOWLEDGMENTS

WegratefullyacknowledgefundingfromEuropeanUnionFP6projectEFORWOODtoGHandGNandtheEuropeanUnionFP7FORMIT

project (grant agreement#311970) toFM.Additional fundingwasprovided from the European Union FP7 project MOTIVE and theDutchMinistryofEconomicAffairs,AgricultureandInnovationgrant#226544toEM,GH,andGN,aswellasfromtheCOSTActionFP0703ECHOES(short-termscientificmissiontoEM).Further,wewouldliketorecognizethecollaborationwithAHarisingfromanAlexandervonHumboldtFellowshipandNSERCDiscoverygrant#330527.Theau-thors are grateful to all data correspondents that contributed theiraggregatedinventorydatatotheEU-NFIplotdataset(withinthepro-jects:EFORWOOD-IP,CarboEurope-IP,ADAM-IP,andCOSTActionE43ENFIN),andtoallcontributorswhoprovidedtheirdendrochro-nology data to the International Tree-Ring Data Bank. This studywouldnothavebeenpossiblewithouttheircontributions.Wethankoneanonymousreviewerforconstructivecommentsthathelpedtoimproveanearlierversionofthismanuscript.

CONFLICT OF INTEREST

Nonedeclared.

REFERENCES

Allen,C.D.,Macalady,A.K.,Chenchouni,H.,Bachelet,D.,McDowell,N.,Vennetier,M.,…Cobb,N. (2010).Aglobaloverviewofdrought andheat-inducedtreemortalityrevealsemergingclimatechangerisksforforests.Forest Ecology and Management,259,660–684.

Boisvenue,C.,&Running,S.W.(2006).Impactsofclimatechangeonnatu-ralforestproductivity–Evidencesincethemiddleofthe20thcentury.Global Change Biology,12,862–882.

Bonan,G.B.(2008).Forestsandclimatechange:Forcings,feedbacks,andtheclimatebenefitsofforests.Science,320,1444–1449.

Botkin, D. B., Saxe, H., Araújo, M. B., Betts, R., Bradshaw, R. H. W.,Cedhagen,T.,…Stockwell,D.R.B.(2007).Forecastingtheeffectsofglobalwarmingonbiodiversity.BioScience,57,227–236.

Breiman,L.(2001).Randomforests.Machine Learning,45,5–32.Brus,D.,Hengeveld,G.,Walvoort,D.,Goedhart,P.,Heidema,A.,Nabuurs,

G.,&Gunia,K.(2012).StatisticalmappingoftreespeciesoverEurope.European Journal of Forest Research,131,145–157.

Cook,E.R.,&Peters,K.(1981).Thesmoothingspline:Anewapproachtostandardizing forest interior tree-ringwidthseries fordendroclimaticstudies. Tree- Ring Bulletin,41,45–53.

Cutler,D.R.,Edwards,T.C.,Beard,K.H.,Cutler,A.,Hess,K.T.,Gibson,J.,&Lawler,J.J. (2007).Randomforestsforclassificationinecology.Ecology,88,2783–2792.

Daly,C.,Halbleib,M.,Smith,J.I.,Gibson,W.P.,Doggett,M.K.,Taylor,G.H.,… Pasteris,P.P. (2008).Physiographicallysensitivemappingofclimato-logicaltemperatureandprecipitationacrosstheconterminousUnitedStates.International Journal of Climatology,28,2031–2064.

Davis,M.B.,&Shaw,R.G.(2001).Rangeshiftsandadaptiveresponsestoquaternaryclimatechange.Science,292,673–679.

EUFORGEN(2012)Distributionmaps–Piceaabies,Pinussylvestris,FagussylvaticaandQuercusrobur.Retrievedfromwww.EUFORGEN.org

Fawcett,T. (2006).An introduction to ROC analysis.Pattern Recognition Letters,27,861–874.

Fielding,A.H.,&Bell, J. F. (1997).A reviewofmethods for the assess-mentofpredictionerrors in conservationpresence/absencemodels.Environmental Conservation,24,38–49.

Fordham,D.A.,Wigley,T.M.L.,&Brook,B.W.(2011).Multi-modelclimateprojections for biodiversity risk assessments. Ecological Applications,21,3317–3331.

2594  |     van der MaaTen eT al.

Fritts,H.C.(1976).Tree rings and climate.London:AcademicPress.García-Valdés, R., Zavala,M.A., Araújo,M. B., & Purves, D.W. (2013).

Chasing a moving target: Projecting climate change-induced shiftsin non-equilibrial tree species distributions. Journal of Ecology, 101,441–453.

Gray,L.,&Hamann,A.(2011).Strategiesforreforestationunderuncertainfutureclimates:GuidelinesforAlberta,Canada.PLoS One,6,e22977.

Gray,L.,&Hamann,A.(2013).TrackingsuitablehabitatfortreepopulationsunderclimatechangeinwesternNorthAmerica.Climatic Change,117,289–303.

Guisan,A., & Zimmermann, N. E. (2000). Predictive habitat distributionmodelsinecology.Ecological Modelling,135,147–186.

Hamann, A., & Aitken, S. N. (2013). Conservation planning under cli-mate change: Accounting for adaptive potential and migration ca-pacity in species distributionmodels.Diversity and Distributions, 19, 268–280.

Hamann,A,Wang,TL, Spittlehouse,DL,&Murdock,TQ (2013).A com-prehensive, high-resolution database of historical and projected cli-mate surfaces for western North America. Bulletin of the American Meteorological Society,1307–1309.

Hampe,A.(2004).Bioclimateenvelopemodels:Whattheydetectandwhattheyhide.Global Ecology and Biogeography,13,469–471.

Hanewinkel, M., Cullmann, D. A., Schelhaas, M.-J., Nabuurs, G.-J., &Zimmermann,N.E. (2013).Climatechangemaycausesevere loss inthe economicvalue of European forest land.Nature Climate Change,3,203–207.

Hogg,E.H.(1997).Temporalscalingofmoistureandtheforest-grasslandboundary inwesternCanada.Agricultural and Forest Meteorology,84,115–122.

Iverson,L.R.,&Prasad,A.M.(1998).Predictingabundanceof80treespe-ciesfollowingclimatechange intheeasternUnitedStates.Ecological Monographs,68,465–485.

Knutti, R., & Sedláček, J. (2013). Robustness and uncertainties in thenew CMIP5 climate model projections. Nature Climate Change, 3,369–373.

Liaw,A.,&Wiener,M.(2002).ClassificationandRegressionbyrandomFor-est. R News,2,18–22.

Lindner,M.,Fitzgerald,J.B.,Zimmermann,N.E.,Reyer,C.,Delzon,S.,vanderMaaten,E.,…Hanewinkel,M.(2014).ClimatechangeandEuropeanforests:Whatdoweknow,whataretheuncertainties,andwhatarethe implications for forest management? Journal of Environmental Management,146,69–83.

Lindner,M.,Maroschek,M.,Netherer,S.,Kremer,A.,Barbati,A.,Garcia-Gonzalo, J., …Marchetti,M. (2010). Climate change impacts, adap-tivecapacity,andvulnerabilityofEuropeanforestecosystems.Forest Ecology and Management,259,698–709.

Loarie,S.R.,Carter,B.E.,Hayhoe,K.,Mcmahon,S.,Moe,R.,Knight,C.A.,&Ackerly,D.D.(2008).ClimatechangeandthefutureofCalifornia’sendemicflora.PLoS One,3,e2502.

vanderMaaten,E.(2012).ClimatesensitivityofradialgrowthinEuropeanbeech(Fagus sylvaticaL.)atdifferentaspectsinsouthwesternGermany.Trees – Structure and Function,26,777–788.

vanderMaaten-Theunissen,M.,Kahle,H.-P.,&vanderMaaten,E.(2013).Drought sensitivityofNorwayspruce ishigher than thatof silverfiralonganaltitudinalgradientinsouthwesternGermany.Annals of Forest Science,70,185–193.

Meehl,G.A., Stocker,T. F., Collins,W.D., Friedlingstein, P.,Gaye,A.T.,Gregory, J. M., … Zhao, Z.-C. (2007) Global Climate Projections. InS. Solomon,D.Qin,M.Manning,Z.Chen,M.Marquis,K.B.Averyt,M.Tignor&H.L.Miller (Eds.),Climate change 2007: The physical sci-ence basis. Contribution of working group I to the fourth assessment re-port of the intergovernmental panel on climate change (pp. 747–845).Cambridge,UKandNewYork,NY:CambridgeUniversityPress.

Nabuurs, G.-J. (2009).NFI plot level database gathered from 18 European countries – digital database. Alterra, Wageningen: European ForestInstitute,Joensuu.

NOAA(2016)Contributors of the international tree-ring data bank.Boulder,CO: IGBP PAGES/World Data Center for Paleoclimatology, NOAA/NCDC Paleoclimatology Program. www.ncdc.noaa.gov/data-access/paleoclimatology-data/datasets/tree-ring

O’Neill,G.A.,Hamann,A.,&Wang,T. (2008).Accountingforpopulationvariationimprovesestimatesoftheimpactofclimatechangeonspe-cies’growthanddistribution.Journal of Applied Ecology,45,1040–1049.

Parmesan,C.,&Yohe,G.(2003).Agloballycoherentfingerprintofclimatechangeimpactsacrossnaturalsystems.Nature,421,37–42.

RDevelopmentCoreTeam(2016).R: A language and environment for statisti-cal computing.Vienna:Austria,RFoundationforStatisticalComputing.

Running,S.W.,Nemani,R.R.,Heinsch,F.A.,Zhao,M.S.,Reeves,M.,&Hashimoto,H.(2004).Acontinuoussatellite-derivedmeasureofglobalterrestrialprimaryproduction.BioScience,54,547–560.

Schelhaas,M.-J.,Nabuurs,G.-J.,Hengeveld,G.,Reyer,C.,Hanewinkel,M.,Zimmermann, N. E., & Cullmann, D. (2015).Alternative forestman-agementstrategiestoaccountforclimatechange-inducedproductiv-ity and species suitability changes in Europe.Regional Environmental Change,15,1581–1594.

Schröter,D.,Cramer,W.,Leemans,R.,Prentice,I.C.,Araujo,M.B.,Arnell,N.W.,…Zierl,B.(2005).Ecology:Ecosystemservicesupplyandvulner-abilitytoglobalchangeinEurope.Science,310,1333–1337.

Sing, T., Sander, O., Beerenwinkel, N., & Lengauer, T. (2005). ROCR:VisualizingclassifierperformanceinR.Bioinformatics,21,3940–3941.

Spathelf,P.,vanderMaaten,E.,vanderMaaten-Theunissen,M.,Campioli,M.,&Dobrowolska,D. (2014). Climate change impacts in Europeanforests:Theexpertviewsof localobservers.Annals of Forest Science,71,131–137.

Thomas,C.D.,Cameron,A.,Green,R.E.,Bakkenes,M.,Beaumont,L.J.,Collingham,Y.C.,…Williams,S.E.(2004).Extinctionriskfromclimatechange. Nature,427,145–148.

Thuiller,W.,Albert,C.,Araújo,M.B.,Berry,P.M.,Cabeza,M.,Zimmermann,N.E.(2008).Predictingglobalchangeimpactsonplantspecies’distri-butions:Futurechallenges.Perspectives in Plant Ecology, Evolution and Systematics,9,137–152.

Thuiller,W.,Lavorel,S.,Araújo,M.B.,Sykes,M.T.,&Prentice,I.C.(2005).Climate change threats to plant diversity in Europe. Proceedings of the National Academy of Sciences of the United States of America,102,8245–8250.

van Vuuren, D. P., Edmonds, J., Kainuma, M., Riahi, K., Thomson, A., Hibbard, K., … Rose, S. K. (2011). The representative concentrationpathways:Anoverview.Climatic Change,109,5–31.

Wang, T. L., Hamann, A., Spittlehouse, D. L., & Murdock, T. Q. (2012).ClimateWNA:High-resolutionspatialclimatedata forwesternNorthAmerica.Journal of Applied Meteorology and Climatology,51,16–29.

SUPPORTING INFORMATION

AdditionalSupportingInformationmaybefoundonlineinthesupport-inginformationtabforthisarticle.

How to cite this article:vanderMaatenE,HamannA,vanderMaaten-TheunissenM,BergsmaA,HengeveldG,vanLammerenR,…SterckF.Speciesdistributionmodelspredicttemporalbutnotspatialvariationinforestgrowth.Ecol Evol. 2017;7:2585–2594.https://doi.org/10.1002/ece3.2696