Uf

EMa

b

c

d

e

f

g

h

a

ARR1A

KBCEFNPS

vf

adm(

0h

Agricultural and Forest Meteorology 171– 172 (2013) 46– 56

Contents lists available at SciVerse ScienceDirect

Agricultural and Forest Meteorology

jou rn al h om epa ge: www.elsev ier .com/ locate /agr formet

sing FLUXNET data to improve models of springtime vegetation activity onset inorest ecosystems

li K. Melaasa,∗, Andrew D. Richardsonb, Mark A. Friedla, Danilo Dragonic, Christopher M. Goughd,athias Herbste, Leonardo Montagnani f,g, Eddy Moorsh

Department of Earth and Environment, Boston University, 675 Commonwealth Avenue, Boston, MA 02215, United StatesDepartment of Organismic and Evolutionary Biology, Harvard University, HUH, 22 Divinity Avenue, Cambridge, MA 02138, United StatesDepartment of Geography, Indiana University, 702 North Walnut Grove, Bloomington, IN 47405, United StatesDepartment of Biology, Virginia Commonwealth University, Box 842012, 1000 West Cary St., Richmond, VA 23284-2012, United StatesDepartment of Bioclimatology, Göttingen University, Büsgenweg 2, 37077 Göttingen, GermanyForest Services and Agency for the Environment, Autonomous Province of Bolzano, 39100 Bolzano, ItalyFaculty of Science and Technology, Free University of Bolzano, 39100 Bolzano, ItalyEarth System Science and Climate Change, Alterra Wageningen UR, P.O. Box 47, 6700 AA Wageningen, The Netherlands

r t i c l e i n f o

rticle history:eceived 4 May 2012eceived in revised form9 November 2012ccepted 20 November 2012

eywords:udburstanopy photosynthesisddy covarianceLUXNETet ecosystem exchangehenologyeasonality modeling

a b s t r a c t

Vegetation phenology is sensitive to climate change and variability, and is a first order control on thecarbon budget of forest ecosystems. Robust representation of phenology is therefore needed to supportmodel-based projections of how climate change will affect ecosystem function. A variety of models havebeen developed to predict species or site-specific phenology of trees. However, extension of these modelsto other sites or species has proven difficult. Using meteorological and eddy covariance data for 29 forestsites (encompassing 173 site-years), we evaluated the accuracy with which 11 different models wereable to simulate, as a function of air temperature and photoperiod, spatial and temporal variability in theonset of spring photosynthetic activity. In parallel, we also evaluated the accuracy with which dynamicsin remotely sensed vegetation indices from MODIS captured the timing of spring onset. To do this, weused a subset of sites in the FLUXNET La Thuile database located in evergreen needleleaf and deciduousbroadleaf forests with distinct active and dormant seasons and where temperature is the primary driverof seasonality. As part of this analysis we evaluated predictions from refined versions of the 11 originalmodels that include parameterizations for geographic variation in both thermal and photoperiod con-straints on phenology. Results from cross-validation analysis show that the refined models predict theonset of spring photosynthetic activity with significantly higher accuracy than the original models. Esti-

mates for the timing of spring onset from MODIS were highly correlated with the onset of photosynthesisderived from flux measurements, but were biased late for needleleaf sites. Our results demonstrate thatsimple phenology models can be used to predict the timing of spring photosynthetic onset both acrosssites and across years at individual sites. By extension, these models provide an improved basis for pre-dicting how the phenology and carbon budgets of temperature-limited forest ecosystems may change inthe coming decades.

∗ Corresponding author at: 675 Commonwealth Avenue, Room 132, Boston Uni-ersity, Boston, MA 02215, United States. Tel.: +1 248 343 9278;ax: +1 617 353 8399.

E-mail addresses: emelaas@bu.edu, eli.melaas@gmail.com (E.K. Melaas),richardson@oeb.harvard.edu (A.D. Richardson), friedl@bu.edu (M.A. Friedl),dragoni@indiana.edu (D. Dragoni), cmgough@vcu.edu (C.M. Gough),athias.herbst@forst.uni-goettingen.de (M. Herbst), leonar@inwind.it

L. Montagnani), eddy.moors@wur.nl (E. Moors).

168-1923/$ – see front matter © 2012 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.agrformet.2012.11.018

© 2012 Elsevier B.V. All rights reserved.

1. Introduction

Phenological transitions, including the onset of canopy develop-ment and senescence and associated cycles of vegetation activity(e.g., photosynthesis and growth), exert first-order control onthe seasonality of land–atmosphere exchanges of carbon, water,energy, and other trace gas constituents (Moore et al., 1996; Wilsonet al., 2000; Fitzjarrald et al., 2001). Environmental drivers includ-

ing air temperature, photoperiod, and water availability are keyregulators of vegetation phenology (White et al., 1997; Jolly et al.,2005). Consequently, phenology is recognized to be a robust indi-cator of biological responses to climate change (Cleland et al.,

rest M

2tsa2emofP

aisSuoputca(awa

pca(pFmdatiaacr

iatFsla2in2oi

aimeowito

m2 m−2) collected at 5 DBF sites included in the La Thuile database

E.K. Melaas et al. / Agricultural and Fo

007). Importantly, numerous recent studies have demonstratedhe potential for significant changes to phenology, and by exten-ion, local and regional budgets for carbon, energy, and water, as

result of climate change (Richardson et al., 2010; Dragoni et al.,011). Such changes are expected to be most pronounced in north-rn high latitude ecosystems where climate change is occurringost rapidly, and shifting in a way that could disrupt critical phen-

logy cues such as temperature, thereby resulting in implicationsor large-scale ecosystem–climate interactions and feedbacks (e.g.,iao et al., 2008).

Time series of CO2 and H2O exchanges obtained by eddy covari-nce measurements are an invaluable resource for evaluating andmproving process understanding and model representation of sea-onal vegetation dynamics (Baldocchi et al., 2001). For example,uni et al. (2003) predicted the timing of spring photosyntheticptake across five boreal forest sites using a 5-day running averagef air temperature measurements. Similarly, Baldocchi et al. (2005)redicted the onset of carbon uptake across 12 temperate decid-ous sites based on the date when daily soil temperature equalshe mean annual temperature. Stöckli et al. (2008) used a moreomplex model that combined air temperature, global radiation,nd vapor pressure deficit measurements to predict leaf area indexLAI) dynamics across 22 sites in 7 biome types. Using a differentpproach, Richardson et al. (2009) used a two-parameter springarming model to predict CO2 source–sink transition dates at both

temperate deciduous site and a boreal conifer site.Despite these important efforts, accurate representation of

henology in ecosystem models has proven to be difficult, andurrent models are biased and do not capture interannual vari-tion in the timing of leaf phenology and canopy CO2 fluxesRichardson et al., 2012; Keenan et al., 2012). Here we build onrevious work by testing a suite of phenology models using theLUXNET ‘La Thuile’ database of eddy covariance flux measure-ents (www.fluxdata.org). Specifically, we use sites in the La Thuile

atabase with distinct summer active and winter dormant season-lity, and where air temperature is the primary driver of springransition from a dormant to an active state. To this end, we specif-cally test the hypothesis that existing models do not adequatelyccount for geographic variation in requirements for photoperiodnd thermal forcing. Consequently, they tend to work well for spe-ific species or sites, but they do not generalize well enough to makeobust predictions across sites and functional groups.

Our analysis focuses on the onset of photosynthetic activityn spring. For deciduous broadleaf (DBF) sites, this correspondspproximately to the timing of leaf emergence (Garrity et al., 2011),raditionally recorded by phenologists as the leaf budburst date.or evergreen needleleaf (ENF) sites, the onset of photosynthe-is is largely independent of (and typically precedes) changes ineaf area, but is similarly influenced by climatic controls that initi-te canopy photosynthesis (Ensminger et al., 2004; Monson et al.,005). In both of these forest types, the phenology of CO2 fluxes

n spring significantly influences annual and seasonal integrals ofet ecosystem productivity (Churkina et al., 2005; Richardson et al.,009; Jeong et al., 2012). Thus, being able to successfully model thenset of photosynthesis across a broad geographic area is of criticalmportance.

To test our hypothesis and explore the challenges identifiedbove, our analysis includes four main elements. First, we usedn situ leaf area index measurements to evaluate three different

etrics for the onset of springtime vegetation activity derived fromddy covariance measurements of CO2. Second, we used springnset dates derived from eddy covariance data to optimize 11

idely used phenology models, and used them to evaluate the abil-

ty of each model to simulate spatial and temporal variability in theiming of spring onset across 29 forested FLUXNET sites. As partf this analysis, we tested modified versions of the models that

eteorology 171– 172 (2013) 46– 56 47

included parameterizations for geographic variability in photope-riod and thermal forcing requirements. Third, we compared resultsfrom the 11 models with predictions from three additional mod-els that use different approaches to simulate phenology withouttuning to observations. Fourth, because remote sensing is widelyviewed to be a useful tool for phenology model development andcalibration (e.g., Stöckli et al., 2008; Jolly et al., 2005), we comparedestimates of spring onset obtained from FLUXNET data againstremotely sensed estimates for the timing of spring greenup fromNASA’s Moderate Resolution Imaging Spectroradiometer (MODIS).

2. Data and methods

2.1. Site selection and data processing

The La Thuile dataset provides eddy covariance measurementsof net ecosystem exchange (NEE) and modeled gross ecosystemproductivity (GEP) of CO2 that have been standardized, gap-filledand partitioned into component fluxes using a common set of algo-rithms (Reichstein et al., 2005; Papale et al., 2006; Moffat et al.,2007). Of the 499 sites included in the database, we identified74 sites located in temperate, boreal, and sub-tropical Mediter-ranean ecosystems with DBF or ENF vegetation that are not heavilymanaged and that have not been recently disturbed. While wateravailability plays a major role in determining the seasonality of car-bon fluxes in sub-tropical Mediterranean ecosystems, this is mostlyrelated to summer drought and associated effects of water lim-itations on productivity; we included these sites in our analysisbecause we hypothesized that initiation of photosynthesis duringlate winter in these ecosystems is largely controlled by tempera-ture.

We used daily data for each site and only included sites with atleast 4 years of available data. Because the La Thuile database doesnot include DBF and ENF sites in the Southern Hemisphere, onlysites located in the Northern Hemisphere were used in the analy-sis. In addition, site-years with (1) more than 20 days of missing CO2flux data, (2) more than 20 days with no high-quality data accord-ing to the FLUXNET quality control procedure, or (3) any missingtemperature data between September 21 of the previous year andJune 21 of the current year, were excluded. The resulting datasetconsisted of 11 sites for DBF and 18 sites for ENF, with 66 and 107site-years, respectively (for site details, see Appendix B).

To estimate spring onset dates from the CO2 flux data, we first fitcubic smoothing splines to time series of daily NEE and GEP. Usingthe smoothed daily data, we identified the day of year (DOY) forthree distinct events for each site-year: the DOY when NEE transi-tioned from positive to negative (the CO2 source–sink transition),and the DOY when the ratio of daily GEP to the growing seasonamplitude (GEPratio) reached 5% and 10% (Fig. 1). Because smoothedNEE values can remain negative throughout the winter and earlyspring at some ENF sites, we did not derive source–sink transi-tion dates for ENF sites. At DBF sites, this was not an issue andsource–sink transition dates generally occurred when the GEPratioreached ∼25%.

Our onset metrics are based on CO2 flux measurements, whichprovide spatially integrated data that are diagnostic of ecosystemphenology (cf. Richardson et al., 2009). To assess the relationshipbetween the CO2 flux-derived spring onset metrics and the timingof leaf emergence, we compared values for each flux metric withtime series measurements of relative canopy leaf area index (LAIr,

where LAIr measurements were available. The LAIr estimates werederived from radiometric measurements using gap fraction theoryand agree favorably with coincident measurements from Licor LAI-2000 Plant Canopy Analyzers (see Richardson et al., 2012).

48 E.K. Melaas et al. / Agricultural and Forest M

Fig. 1. Example showing how spring phenology metrics were derived from CO2 timeseries; flux data are indicated with black dots, and smoothing spline by solid red line;relative GEPratio and LAI threshold dates are indicated with vertical bars; minimumand maximum of smoothed GEP are indicated with horizontal bars. Data are from2ov

2

arlbomfioiaicm(m

TG

A

004 for the FLUXNET site DE-Hai, located in Hainich, Germany. (For interpretationf the references to color in this figure legend, the reader is referred to the webersion of the article.)

.2. Spring phenology models

Most models that are used to simulate phenology in temperatend boreal ecosystems assume that leaf development is primarilyegulated by air temperature and can be modeled using cumu-ative thermal units (heating or chilling degree days) above orelow a prescribed reference heating or chilling temperature: Tfr Tc, respectively (Hänninen and Kramer, 2007). The simplestodels only consider heating temperatures accumulated after a

xed date (e.g., January 1), and are designed to predict the datef specific phenophases such as budburst (e.g., the Spring Warm-

ng model; Hunter and Lechowicz, 1992). More complex modelslso consider the effect of chilling temperatures, which some stud-es have suggested control spring phenology in combination with

umulative heating. This family of models includes the Parallelodel (Landsberg, 1974; Hänninen, 1990), the Sequential model

Hänninen, 1990; Cesaraccio et al., 2004), and the Alternatingodel (Murray et al., 1989).

able 1rowing degree-day models and their associated parameters.

Model Parametersa

Thermal base temperature Chilling base temperature Required

SW1.1 Tf F*

SW1.2 Tf = ax + b F*

SW1.3 Tf F* = ax + bSW1.4 Tf = ax + b F* = cx + dSW2.1 F*

SW2.2 F* = ax + b

SEQ1.1 Tf Tc F*

SEQ1.2 Tf = ax + b Tc = cx + d F*

SEQ1.3 Tf Tc F* = ax + bSEQ1.4 Tf Tc F*

SEQ1.5 Tf = ax + b Tc = cx + d F* = ex + f

a The variable ‘x’ is a vector of long-term mean annual temperatures of each site withppendix A.

eteorology 171– 172 (2013) 46– 56

In this paper we test 11 different models that use thermal heat-ing or combined thermal chilling and heating to predict the timingof spring onset of photosynthetic activity. Each model is based onone of three functional forms in which spring onset is predictedto occur when the state of forcing (Sf(t)) reaches a critical sum ofheating units (F*).

In the “Spring Warming 1” model (SW1), the rate of accumulatedheating is linearly related to air temperature:

Sf (t) =tpheno∑

p0

max(Tair − Tf , 0) (1)

where Tair is daily mean air temperature, p0 is the starting photope-riod when heating is prescribed to begin accumulating, and tphenois the date of spring onset when Sf (t) ≥ F∗. In the “Spring Warming2” model (SW2), accumulated heating is related to air temperatureusing a logistic function (Sarvas, 1974):

Sf (t) =tpheno∑

p0

max[

28.41 + exp(−0.185(Tair − 18.4))

, 0]

(2)

Finally, the Sequential model (SEQ1) assumes that heating accu-mulation (using Eq. (1)) does not occur until a critical sum of chillingunits (C*) is reached, and where the state of chilling (Sc(t)) increasesonly after the daily mean air temperature falls below a prescribedtemperature threshold:

Sc(t) =t1∑p0

(1 Tair < Tc

0 Tair ≥ Tc

)(3)

where t1 is the date when chilling requirements are met and heat-ing accumulation begins.

Variants of these three basic model forms have been widelyused to predict leaf phenology and detailed descriptions foreach approach are presented elsewhere (e.g., Chuine et al., 1999;Richardson and O’Keefe, 2009). Here we test 11 different modelsbased on these basic model forms using implementations that areslightly different from previous efforts (Table 1). Specifically, previ-ous efforts initiate accumulation of chilling or heating requirementsbased on a prescribed date (t0). In the models we test here, accu-mulation is instead initiated based on a photoperiod trigger (p0).This is functionally equivalent to allowing t0 to vary with latitude.For the spring warming models (SW1 and SW2), if the minimumphotoperiod at a given site is always greater than p0, accumulation

is prescribed to begin on December 21. Similarly, for the chillingmodels (i.e., SEQ1), if minimum p0 is never reached, we prescribeaccumulation to begin on September 21. These dates were selectedbecause December 21 has the shortest day length in the Northern

thermal forcing Required chilling forcing Minimum photoperiod

p0

p0

p0

p0

p0

p0

C* p0

C* p0

C* p0

C* = ax + b p0

C* = gx + h p0

in each vegetation grouping. Parameter definitions are described in more detail in

rest M

H2iav1ievListi

bouwd1ito0eadomha2

fltCGuaptma(dvsnr

2

tmammlb(

Cir

that understory vegetation is fully developed as much as the 3weeks earlier (Herbst, pers. comm.). Thus, significant increases inGEP associated with understory vegetation may occur well beforethe emergence of overstory leaves. Similarly, bias in the timing

E.K. Melaas et al. / Agricultural and Fo

emisphere and day length is the same everywhere on September1. Finally, previous studies have indicated that the thermal forc-

ng required for spring onset is lower at more northern sites thant southern sites and have suggested parameterizing the criticalalue in thermal forcing (F*) as a function of latitude (White et al.,997; Fisher et al., 2007). Here we adopt a similar approach, but

nstead parameterize both F* and the base temperature (Tf) as lin-ar functions of mean annual temperature, thereby accounting forariability in climatology related to geographic patterns (Kaduk andos, 2011). Table 1 summarizes whether a chilling requirement isncluded in each model and whether parameters are treated as con-tant across sites or vary across sites as a function of mean annualemperature. Complete descriptions for all 11 models are providedn Appendix A.

In addition to model testing and refinement, we estimatedenchmark spring onset dates using three recently published phen-logy prediction schemes: (1) a standard phenology sub-routinesed in several ecosystem models (hereafter, standard GDD), inhich leaf onset occurs when the accumulated heating degree-ays above 0 ◦C starting January 1 exceed 100 (e.g., Aber et al.,996; Levis and Bonan, 2004; Kucharik et al., 2006); (2) the Grow-

ng Season Index (GSI) model (Jolly et al., 2005), in which minimumemperature, photoperiod, and humidity thresholds control phen-logy and spring onset is predicted to occur when the GSI reaches.5; and (3) a modified version of the model described in Baldocchit al. (2005) (TSOIL), which is based on soil and air temperaturesnd where spring onset (identified as the source–sink transitionate) is identified using daily NEE data as described above. Basedn advice from the authors (Baldocchi, pers. comm.), the TSOILodel was only tested for DBF sites. All three of these approaches

ave received substantial attention because of their intuitive appealnd simplicity of implementation (Cleland et al., 2007; Yuan et al.,012).

To assess the potential for regional scaling and monitoring of ourux-derived phenology metrics using remote sensing, we extractedhe observed timing of spring greenup from the MODIS Global Landover Dynamics product for each site-year (Zhang et al., 2003;anguly et al., 2010). The algorithm used to generate this productses logistic functions fit to time series of MODIS vegetation indicest each pixel, including explicit treatment for snow-contaminatedixels (which can bias retrievals). Phenophase transition dates arehen estimated based on the maximum rate of curvature in the esti-

ated logistic functions. Complete details related to this productnd the algorithm used to generate it are provided in Zhang et al.2003) and Ganguly et al. (2010). For this work we used MODIS-erived green-up dates from 2001 to 2005 estimated as the medianalue of 9 × 9 pixel windows (∼4.2 km × 4.2 km) centered on eachite and excluding non-forested pixels. Because this data product isot available before 2001, the number of site-years was somewhateduced for both DBF (n = 43) and ENF (n = 70) sites.

.3. Model calibration and testing

We use phenology metrics derived from CO2 fluxes to identifyhe timing of spring onset. Parameters for each model were esti-

ated using all candidate site-years for DBF (11 sites; 66 site-years)nd ENF (18 sites; 107 site-years) using optimization routines thataximize agreement between modeled and observed phenologicaletrics. This exercise was then repeated for subgroups of DBF sites

ocated in temperate and boreal biomes (9 sites; 56 site-years), fororeal ENF sites (6 sites; 33 site-years), and for temperate ENF sites10 sites; 69 site-years).

To estimate parameter values for each model we used Montearlo techniques (simulated annealing, Metropolis et al., 1953) sim-

lar to Chuine et al. (1998) and Richardson and O’Keefe (2009),epeating the optimization several times using different randomly

eteorology 171– 172 (2013) 46– 56 49

selected initial values to ensure globally optimum solutions. Toselect the best model for each group and subgroup, we usedthe small-sample corrected Akaike Information Criteria (AICc)(Burnham and Anderson, 2002). Finally, we measured the out-of-sample forecast accuracy of each optimized model for eachvegetation group and subgroup using a fourfold cross-validation,wherein three-fourths of the sites were randomly selected fortraining and one-fourth was used for testing. This procedure wasrepeated four times using mutually exclusive subsets of the datauntil all sites were included in the cross-validation test set.

3. Results and discussion

As we described in the introduction, our analysis includesfour main elements: (1) evaluation of FLUXNET-derived springonset metrics using in situ LAIr measurements; (2) assessment ofstandard and refined spring phenology models; (3) comparison ofspring onset dates from FLUXNET with predictions from alternativebenchmark models; and (4) comparison of spring onset dates fromFLUXNET with start-of-spring estimates from MODIS. To provide amore direct comparison of the different modeling strategies that weevaluated, results from all models are discussed together in Section3.2.

3.1. Evaluation of FLUXNET metrics with in situ LAI time seriesmeasurements

Fig. 2 shows that the DOY on which LAIr reaches 10% of itsannual maximum (the detectable stage of leaf development) is sig-nificantly correlated with the timing of increased photosyntheticuptake across sites with available LAIr measurements (R2 = 0.53,p < 0.01). Note, however, that the emergence of understory canopysignificantly precedes over-story development at some sites, whichaffects the phenology of net CO2 exchange. For example, at theHainich site (shown in Fig. 1) over-story leaf emergence gener-ally occurs around DOY 120, while weekly photos clearly show

Fig. 2. The relationship between day of year when relative gross ecosystem produc-tivity ratio (GEPratio) reached 10% of its annual maximum and the day of year whenestimated canopy leaf area index (LAI) reached 10% of its annual maximum acrossfive broadleaf deciduous forests. 4–8 site-years are included for each site.

50 E.K. Melaas et al. / Agricultural and Forest Meteorology 171– 172 (2013) 46– 56

Table 2Model performance across each vegetation group and for each spring onset metric.

Spring onset metric Modified GDD Standard GDD GSI = 0.5 TSOIL

Model RMSE Biasa RMSE Bias RMSE Bias RMSE Bias

All DBF (11/66)b

Source–sink trans. SW2.2 8.6 0.2 68.6 65.4 41.8 40.2 27.3 13.85% GEPratio SEQ1.3 18.8 0.0 44.3 32.5 22.7 7.1 34.1 −19.210% GEPratio SW2.2 17.6 −3.3 53.5 46.9 26.4 21.6 25.4 −4.7

Bor. & Temp. DBF (9/56)Source–sink trans. SW2.2 8.2 −1.2 64.0 61.5 40.5 39.1 25.5 19.35% GEPratio SEQ1.4 14.9 −0.6 41.5 29.6 22.4 7.1 25.9 −12.510% GEPratio SEQ1.3 10.4 −2.1 50.2 43.7 25.9 21.3 18.3 1.5

All ENF (18/107)5% GEPratio SW1.1 15.5 −0.3 23.5 −14.6 39.5 −35.6 n/a n/a10% GEPratio SEQ1.4 15.1 1.3 20.7 −4.5 29.9 −25.3 n/a n/a

Boreal ENF (6/33)5% GEPratio SW2.1 14.8 2.2 25.2 −24.0 30.7 −28.4 n/a n/a10% GEPratio SW2.1 8.9 −0.5 15.8 −15.4 22.2 −19.9 n/a n/a

Temp. ENF (10/69)5% GEPratio SEQ1.1 13.3 −3.0 21.9 −12.4 41.2 −38.1 n/a n/a

1.7

portedte-yea

odeo2

3

mFftoasb

tTaslisctfitado2aptbstro&ya

10% GEPratio SW1.3 13.1 −0.6 2

a Bias is the mean difference between observed and predicted dates; statistics reb Numbers in parentheses next to each group title indicate the number of sites/si

f photosynthetic uptake relative to the start of overstory leafevelopment at Harvard forest is likely caused by the presence ofvergreen needleleaf trees, which account for roughly one-quarterf the basal area within the flux tower footprint (Richardson et al.,009).

.2. Assessment of spring phenology models

Table 2 presents the root mean squared error (RMSE) and bias forodel predictions of spring onset DOY relative to estimates from

LUXNET data. Note that we only include results for the best per-orming modified GDD model in each group. These results showhat the Standard GDD and GSI models consistently predict springnset metrics with a root mean square error greater than 15 days,nd in some cases, as high as 9 weeks. Results from the TSOIL modelhow RMSE values on the order of 3–4 weeks, although with lessias.

Results presented in Table 2 reveal important patterns relatedo plant functional types and bioclimatic controls on phenology.he RMSE and bias associated with the standard GDD modelre significantly lower for ENF sites than for DBF sites becausepring initiation of photosynthetic uptake generally occurs ear-ier in conifers than in hardwoods and the required critical forcings therefore closer to the prescribed threshold (F* = 100) at ENFites. Similarly, including photoperiod in the GSI model signifi-antly decreases both the RMSE and bias across DBF sites relative tohe standard GDD model. These results strongly suggest that using axed start date across sites to initiate accumulation of heating (i.e.,

0) is problematic at DBF sites (Hänninen and Kramer, 2007; Levisnd Bonan, 2004). At ENF sites, on the other hand, GSI model pre-ictions suggest that the minimum photoperiod required for springnset is lower than the 10 h assumed by the model (Jolly et al.,005). This is confirmed by results for the modified GDD modelt ENF sites (SW1.1), which initiates heating accumulation whenhotoperiod reaches roughly 6 h (note: at sites with longer winter-ime photoperiod, this implies that by default heating accumulationegins December 21 as described in Section 2.2). Our results alsouggest that air temperature does not significantly influence theiming of spring onset of photosynthesis in Sub-tropical Mediter-anean forest regions; model performance decreased for all spring

nset metrics when these sites were included (e.g., All DBF vs. Bor.

Temp. DBF in Table 2). However, to fully assess this, more site-ears of CO2 flux data from a greater number of Mediterranean sitesre needed.

−1.6 31.7 −26.8 n/a n/a

for the modified GDD model are based on fourfold cross-validation.rs.

The modified GDD models provide substantial improvement rel-ative to the standard GDD, GSI, and TSOIL models (Fig. 3, Table 2).Root mean square errors produced by the modified GDD modelsrange from 9 to 19 days, with biases of 3 days or less. In gen-eral, model predictions for earlier spring onset metrics (e.g., 5%GEPratio) are not as robust as those for later onset metrics (e.g.,source–sink transition; cf. Richardson et al., 2009). This result isprobably a by-product of high noise levels in GEP data duringthe early spring, which leads to greater uncertainty in the esti-mated DOY corresponding to 5% GEPratio. However, we contendthat correctly modeling earlier onset metrics is more importantthan modeling later onset metrics, because earlier onset metricsmore closely correspond to the initiation of photosynthetic activ-ity and consequently are essential to constraining growing seasonduration. Moreover, the C source–sink transition is a function ofthe shifting balance between GEP and ecosystem respiration, andis therefore is not a direct indicator for the timing of photosyntheticactivity onset.

Across the different plant functional type and climate groups,seven different modified GDD models provided the most accuratepredictions for the timing of spring onset. Indeed, Table 2 showsthat no single model was optimal, even within the subgroups. Thisresult is consistent with previous findings showing that modelswith substantially different assumptions can give equally accu-rate predictions (Hunter and Lechowicz, 1992; Hänninen, 1995).More generally, this result also demonstrates that a “one-size-fits-all” approach introduces errors in predicted phenology, and thatclimate and plant functional type-specific parameterizations arerequired to capture differential sensitivity in spring phenology tophotoperiod and temperature forcing (i.e., DBF vs. ENF or temperatevs. boreal). Further, substantial improvement over standard GDDmodels can be achieved through a modest increase in model com-plexity and calibration of model parameters using data spanning awide biogeographic range of sites (Kaduk and Los, 2011).

The parameters and structure of the best modified GDD modelsreveal significant differences in photoperiod cues, thermal forc-ing requirements, and chilling requirements between the differentsubgroups (Appendix C). DBF sites generally require more cumu-lative thermal forcing than ENF sites to initiate photosynthesis.Moreover, the values shown in Appendix C for F* and C* indicate

that the amount of forcing required across DBF sites varies as afunction of mean annual temperature. Differentiating temperatefrom boreal ENF sites improves model performance, which suggestsfundamental differences in how these subgroups respond to

E.K. Melaas et al. / Agricultural and Forest Meteorology 171– 172 (2013) 46– 56 51

F al andt he dasfi

cpst1fgirmMc

3M

fs

Fati

ig. 3. Predicted versus observed dates corresponding to 10% GEPratio across (a) boreo predictions given by the optimal modified GDD model of the vegetation group. Tgure legend, the reader is referred to the web version of the article.)

hanges in temperature and day length, probably as a result of eco-hysiological differences among genera or species (e.g., boreal Piceapp. vs. temperate Pinus spp.). Boreal ENF sites require relatively lit-le forcing that is accumulated only after photoperiod reaches about1 h. In contrast, a combination of chilling and significant thermalorcing is necessary across temperate ENF sites. These results sug-est that future climate variability and change may have a greatermpact on the timing of photosynthetic onset across temperateegions, while boreal regions may be less affected because they areore tightly controlled by photoperiod (Körner and Basler, 2010;igliavacca et al., 2012). However, more site-years are needed to

onfirm this hypothesis.

.3. Comparison of start-of-season metrics from FLUXNET andODIS Land

Spatial and temporal patterns in spring onset metrics derivedrom CO2 flux measurements were highly correlated with landurface phenology metrics derived from MODIS (Fig. 4). Specifically,

ig. 4. Timing (DOY) of 5% GEP from flux data versus timing of green-up from MODIS Lanll site-years available during 2001–2005. Data points in the plots are distinguished bemperate (blue) and Mediterranean (green); ENF sites include boreal (red), temperate (bnterpretation of the references to color in this figure legend, the reader is referred to the

temperate DBF sites and (b) boreal ENF sites. The blue dots in both plots correspondhed line is the 1:1 relationship. (For interpretation of the references to color in this

the onset of spring green-up from MODIS was strongly correlatedwith 5% GEPratio across DBF sites (r = 0.84; p < 0.01) and ENF sites(r = 0.77; p < 0.01). Note, however, that for ENF sites, green-up fromMODIS was systematically biased late by as much as 50 days rela-tive to the timing of 5% GEPratio. As a result, green-up from MODIScoincided with the timing when the GEPratio was about 50% forENF sites. This result reflects different ecophysiological responsesto climate forcing at ENF sites relative to DBF sites. Specifically, atboreal and cold temperate ENF sites, the MODIS algorithm appearsto be detecting green-up associated with mid-to-late spring recov-ery of pigments in old needles and the emergence of newneedles.

Overall, results from MODIS demonstrate the utility of remotesensing as a tool for extrapolating the onset of spring photosyn-thetic activity from the scale of FLUXNET sites to regional (and

larger) scales. However, modeling applications that use resultsfrom MODIS need to carefully account for bias in the timingof spring onset from MODIS at ENF sites. Further, care shouldbe taken when remote sensing data sources such as MODIS

d Cover Dynamics Product (MCD12Q2) for (a) DBF and (b) ENF sites. Data includey color according to specific vegetation subgroups; DBF sites include boreal andlue), and Mediterranean (green). The dashed lines in the plots are the 1:1 lines. (For

web version of the article.)

5 rest M

atwoieau

4

pe1ptatoiS

mrrtlortroffrtcmtiwMspL

etwattatcilrastsie

2 E.K. Melaas et al. / Agricultural and Fo

re used to prescribe or calibrate spring onset dates in ecosys-em models (e.g., Stöckli et al., 2008; Jolly et al., 2005). Recentork has suggested that alternative variables such as the date

f snowmelt may help to reduce bias in start of spring metricsn high latitude ecosystems (e.g., Böttcher et al., 2011). How-ver, the relationship between the timing of spring phenologynd snowmelt is quite variable, and is therefore difficult tose.

. Conclusions

We used eddy covariance data to explore and refine modelsredicting the start of spring photosynthetic activity in forestedcosystems with distinct seasonalities. To do this, we optimized1 widely used phenology models, and refined them to includearameterizations for geographic variability in photoperiod andhermal forcing requirements. As part of this analysis we alsonalyzed three additional models that are not based on conven-ional heat sum models, and compared estimates of spring onsetbtained from eddy covariance data with estimates for the tim-ng of spring greenup from NASA’s Moderate Resolution Imagingpectroradiometer (MODIS).

We draw two main conclusions from our results. First, theodified GDD models evaluated in this study provide improved

epresentation of vegetation phenology for ENF and DBF biomeselative to commonly used models in boreal and temperate ecosys-ems. This is important because representation of phenology inand surface models is poor, and realistic representation of phen-logy is critical to predictions of how ecosystem function willespond to climate change. However, to fully assess this, addi-ional site-years of flux data that include temperature regimesepresentative of expected warming scenarios are required. Sec-nd, no single model was optimal across the different plantunctional types and climate groupings. This suggests that plantunctional type and climate-specific parameterizations may beequired to realistically predict phenology over large areas. Indeed,he results presented in Table 2 suggest that by estimating uniquealibrations for different climate and plant functional types, theodified GDD models provided significantly improved predic-

ions of spring onset relative to the standard models. Given thatnter- and intra-species spring onset dates can vary by up to 2

eeks within the area of a flux tower footprint (e.g., Fisher andustard, 2006), further model improvement probably requires

ite specific characterization of landscape vegetation composition,ossibly using medium spatial resolution remote sensing (e.g.,andsat).

Realistic models of spring phenology, and specifically mod-ls that predict the timing of phenological transitions duringhe growing season, are important because current and futurearming arising from climate change is generally expected to

ffect growing season dynamics in mid- to high-latitude ecosys-ems. Because biosphere–atmosphere interactions affect climatehrough a variety of pathways, most importantly by influencingtmospheric carbon dioxide and other greenhouse gas concen-rations, characterizing how phenology will respond to climatehange in the coming decades is critically important. Accord-ng to Richardson et al. (2010), a 1-day shift in the timing ofeaf emergence in deciduous broadleaf forests increases GEP byoughly 10 g C m−2. Thus, the nature and magnitude of uncertaintynd biases in existing models have the potential to introduce

ubstantial errors in modeled estimates of ecosystem produc-ivity. The results from this study demonstrate that relativelyimple variants of widely used phenology models can significantlymprove simulation of spring phenology in temperature-sensitivecosystems.

eteorology 171– 172 (2013) 46– 56

Acknowledgements

EM and MF gratefully acknowledge support from NASA throughCooperative Agreement number NNX08AE61A. ADR acknowl-edges support from the National Science Foundation through theMacrosystems Biology program, award EF-1065029, and the North-eastern States Research Cooperative. We are grateful to the sitePIs whose contributions to the FLUXNET database have made thisstudy possible. This work used eddy covariance data acquired by theFLUXNET community and in particular by the following networks:AmeriFlux (U.S. Department of Energy, Biological and Environmen-tal Research, Terrestrial Carbon Program (DE-FG02-04ER63917 andDE-FG02-04ER63911)), AsiaFlux, CarboEuropeIP, Fluxnet-Canada(supported by CFCAS, NSERC, BIOCAP, Environment Canada, andNRCan), GreenGrass, NECC, TCOS-Siberia, USCCC. We acknowledgethe financial support to the eddy covariance data harmonizationprovided by CarboEuropeIP, FAO-GTOS-TCO, iLEAPS, Max PlanckInstitute for Biogeochemistry, National Science Foundation, Uni-versity of Tuscia, Université Laval and Environment Canada andUS Department of Energy and the database development andtechnical support from Berkeley Water Center, Lawrence Berke-ley National Laboratory, Microsoft Research eScience, Oak RidgeNational Laboratory, University of California - Berkeley, Universityof Virginia. Any opinions, findings, and conclusions or recommen-dations expressed in this material are those of the authors and donot necessarily reflect the views of the National Science Foundation.

Appendix A.

The models used here are largely based on those presented inChuine et al. (1999). For each model, the state of chilling (Sc) orforcing (Sf) is the time integral from p0 or t1 of the rate of chilling(Rc) or forcing (Rf), which are functions of daily mean temperaturex(t). Chilling and forcing accumulate relative to base temperature(Tc or Tf) until a critical threshold (C* or F*) is reached. The date thatthe threshold is reached corresponds to the predicted timing of thephenological metric of interest. Chilling and forcing parameters areeither fixed across all sites, or are modeled as a linear function ofthe long-term mean annual temperature (T̄i) of each site, i.

Spring Warming 1.1 (SW1.1)

Rf = x(t) − Tf where x(t) > Tf

Rf = 0 otherwise

Sf (t) =∑

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = F*.

Spring Warming 1.2 (SW1.2)

Rf = x(t) − [aT̄i + b] where x(t) > aT̄ + bRf = 0 otherwise

Sf (t) =∑

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = F*.

Spring Warming 1.3 (SW1.3)

Rf = x(t) − Tf where x(t) > Tf

Rf = 0 otherwise∑

Sf (t) =

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = aT̄i + b.

rest M

S

S

S

S

r

S

Appendix A Reference:

Chuine, I., et al., 1999. Selecting models to predict the timing of flowering oftemperate trees: implications for tree phenology modeling, Plant Cell andEnvironment 22, 1–13.

E.K. Melaas et al. / Agricultural and Fo

pring Warming 1.4 (SW1.4)

Rf = x(t) − [aT̄i + b] where x(t) > aT̄i + bRf = 0 otherwise

Sf (t) =∑

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = cT̄i + d.

pring Warming 2.1 (SW2.1)

Rf = 28.41 + exp(−0.185(x(t) − 18.4))

where x(t) > 0

Rf = 0 otherwise

Sf (t) =∑

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = F*.

pring Warming 2.2 (SW2.2)

Rf = 28.41 + exp(−0.185(x(t) − 18.4))

where x(t) > 0

Rf = 0 otherwise

Sf (t) =∑

p0

Rf (x(t))

Phenology metric is predicted to occur when Sf = aT̄i + b.

equential 1.1 (SEQ1.1)

Rc = 1 where x(t) < Tc

Rc = 0 otherwise

Sc(t) =∑

p0

Rc(x(t))

Rf = x(t) − Tf where x(t) > Tf

Rf = 0 otherwise

Sf (t) =∑

t1

Rf (x(t))

Forcing summation (Sf) begins at t1 when Sc = C*. Phenology met-ic is predicted to occur when Sf = F*.

equential 1.2 (SEQ1.2)

Rc = 1 where x(t) < cT̄i + dRc = 0 otherwise

Sc(t) =∑

p0

Rc(x(t))

Rf = x(t) − [aT̄i + b] where x(t) > aT̄i + b

Rf = 0 otherwise

Sf (t) =∑

t1

Rf (x(t))

eteorology 171– 172 (2013) 46– 56 53

Forcing summation (Sf) begins at t1 when Sc = C*. Phenology met-

ric is predicted to occur when Sf = F*.

Sequential 1.3 (SEQ1.3)

Rc = 1 where x(t) < Tc

Rf = 0 otherwise

Sc(t) =∑

p0

Rc(x(t))

Rf = x(t) − Tf where x(t) > Tf

Rf = 0 otherwise

Sf (t) =∑

t1

Rf (x(t))

Forcing summation (Sf) begins at t1 when Sc = C*. Phenology met-ric is predicted to occur when Sf = aT̄i + b.

Sequential 1.4 (SEQ1.4)

Rc = 1 where x(t) < Tc

Rc = 0 otherwise

Sc(t) =∑

p0

Rc(x(t))

Rf = x(t) − Tf where x(t) > Tf

Rf = 0 otherwise

Sf (t) =∑

t1

Rf (x(t))

Forcing summation (Sf) begins at t1 when Sc = aT̄i + b. Pheno-logy metric is predicted to occur when Sf = F*.

Sequential 1.5 (SEQ1.5)

Rc = 1 where x(t) < cT̄i + dRc = 0 otherwise

Sc(t) =∑

p0

Rc(x(t))

Rf = x(t) − [aT̄i + b] where x(t) > aT̄i + bRf = 0 otherwise

Sf (t) =∑

t1

Rf (x(t))

Forcing summation (Sf) begins at t1 when Sc = gT̄i + h. Pheno-logy metric is predicted to occur when Sf = eT̄i + f .

5

A

A

B

B

B

C

D

D

G

G

G

G

G

G

4 E.K. Melaas et al. / Agricultural and Forest Meteorology 171– 172 (2013) 46– 56

ppendix B. Names, locations and general information of FLUXNET sites used in this analysis

Site ID Site name IGBP class – climate Lat (◦N) Lon (◦E) No. site years Reference

CA-Oas SSA Old Aspen DBF – Bor 53.63 −106.20 8 Black et al. (2000)DE-Hai Hainich DBF – Temp 51.08 10.45 5 Knohl et al. (2003)DK-Sor Soroe DBF – Temp 55.49 11.65 9 Pilegaard et al. (2003)FR-Hes Hesse DBF – Temp 48.67 7.06 8 Granier et al. (2000)IT-Ro1 Roccarespampani 1 DBF – Med 42.41 11.93 6 Tedeschi et al. (2006)IT-Ro2 Roccarespampani 2 DBF – Med 42.39 11.92 4 Tedeschi et al. (2006)JP-Tak Takayama DBF – Temp 36.15 137.42 4 Ito et al. (2006)US-Ha1 Harvard Forest EMS DBF – Temp 42.54 −72.17 8 Goulden et al. (1996)US-MMS Morgan Monroe DBF – Temp 39.32 −86.41 4 Schmid et al. (2000)US-UMB U of M Biological DBF – Temp 45.56 −84.71 4 Gough et al. (2008)US-WCr Willow Creek DBF – Temp 45.81 −90.08 4 Cook et al. (2004)

CA-Ca1 Campbell River 1 ENF – Temp 49.87 −125.33 8 Humphreys et al. (2006)CA-Ca3 Campbell River 3 ENF – Temp 49.53 124.90 4 Humphreys et al. (2006)CA-Man NSA Old Back Spruce ENF – Bor 55.88 −98.48 4 Dunn et al. (2007)CA-Obs SSA Old Black Spruce ENF – Bor 53.99 −105.12 6 Bergeron et al. (2007)CA-Ojp SSA Old Jack Pine ENF – Bor 53.92 −104.69 6 Howard et al. (2004)CA-Qcu Quebec Boreal Cutover ENF – Bor 49.27 −74.04 5 Giasson et al. (2006)DE-Tha Tharandt ENF – Temp 50.96 13.57 9 Gruenwald et al. (2007)DE-Wet Wetzstein ENF – Temp 50.45 11.46 4 Rebmann et al. (2010)FI-Hyy Hyytiala ENF – Bor 62.18 22.78 8 Suni et al. (2003)FI-Sod Sodankyla ENF – Bor 67.36 26.64 4 Suni et al. (2003)FR-LBr Le Bray ENF – Temp 44.72 −0.78 4 Berbigier et al. (2001)IT-Ren Renon ENF – Temp 46.59 11.43 6 Montagnani et al. (2009)NL-Loo Loobos ENF – Temp 52.17 5.74 8 Dolman et al. (2002)RU-Fyo Fyodorvskoye ENF – Temp 56.46 32.92 7 Kurbatova et al. (2008)US-Blo Blodgett ENF – Med 38.90 −120.63 5 Goldstein et al. (2000)US-Ho1 Howland 1 ENF – Temp 45.20 −68.74 8 Hollinger et al. (2004)US-Ho2 Howland 2 ENF – Temp 45.20 −68.74 6 Hollinger et al. (2004)US-Wrc Wind River Crane ENF – Temp 45.82 -121.95 5 Falk et al. (2008)

ppendix B References:

erbigier, P., et al., 2001. CO2 and water vapour fluxes for 2 years above Eurofluxforest site. Agric. For. Meteorol. 108, 183–197.

ergeron, O., et al., 2007. Comparison of carbon dioxide fluxes over three borealblack spruce forests in Canada. Global Change Biol. 13, 89–107.

lack, T.A., et al., 2000. Increased carbon sequestration by a boreal deciduousforest in years with a warm spring. Geophys. Res. Lett. 27, 1271–1274.

ook, B.D., et al., 2004. Carbon exchange and venting anomalies in an uplanddeciduous forest in northern Wisconsin, USA. Agric. For. Meteorol. 126, 271–295.olman, A.J., et al., 2002. The carbon uptake of a mid latitude pine forest growingon sandy soil. Agric. For. Meteorol. 111, 157–170.unn, A.L., Barford, C.C., Wofsy, S.C., Goulden, M.L., Daube, B.C., 2007. A long-termrecord of carbon exchange in a boreal black spruce forest: means, responses tointerannual variability, and decadal trends. Global Change Biol. 13,577–590.iasson, M.-A., Coursolle, C., Margolis, H.A., 2006. Ecosystem-level CO(2) fluxesfrom a boreal cutover in eastern Canada before and after scarification. Agric. For.Meteorol. 140, 23–40.oldstein, A.H., et al., 2000. Effects of climate variability on the carbon dioxide,water, and sensible heat fluxes above a ponderosa pine plantation in the SierraNevada (CA). Agric. For. Meteorol. 101, 113–129.ough, C.M., Vogel, C.S., Schmid, H.P., Su, H.B., Curtis, P.S., 2008. Multi-yearconvergence of biometric and meteorological estimates of forest carbon storage.Agric. For. Meteorol. 148, 158–170.oulden, M.L., Munger, J.W., Fan, S.M., Daube, B.C., Wofsy, S.C., 1996. Exchange ofcarbon dioxide by a deciduous forest: response to interannual climatevariability. Science 271, 1576–1578.ranier, A. et al., 2000. The carbon balance of a young Beech forest. FunctionalEcology, 14, 312–325.ruenwald, T., Bernhofer, C., 2007. A decade of carbon, water and energy fluxmeasurements of an old spruce forest at the Anchor Station Tharandt. Tellus Ser.B 59, 387–396.

Hollinger, D.Y., et al., 2004. Spatial and temporal variability in forest–atmosphereCO(2) exchange. Global Change Biol. 10, 1689–1706.

Howard, E.A., Gower, S.T., Foley, J.A., Kucharik, C.J., 2004. Effects of logging oncarbon dynamics of a jack pine forest in Saskatchewan, Canada. Global ChangeBiol. 10, 1267–1284.

Humphreys, E.R., et al., 2006. Carbon dioxide fluxes in coastal Douglas-fir stands atdifferent stages of development after clearcut harvesting. Agric. For. Meteorol.140, 6–22.

Ito, A., et al., 2006. Seasonal variation in leaf properties and ecosystem carbonbudget in a cool-temperate deciduous broad-leaved forest: simulation analysisat Takayama site, Japan. Ecol. Res. 21, 137–149.

Knohl, A., Schulze, E.D., Kolle, O., Buchmann, N., 2003. Large carbon uptake by anunmanaged 250-year-old deciduous forest in Central Germany. Agric. For.Meteorol. 118, 151–167.

Kurbatova, J., Li, C., Varlagin, A., Xiao, X., Vygodskaya, N., 2008. Modeling carbondynamics in two adjacent spruce forests with different soil conditions in Russia.Biogeosciences 5, 969–980.

Montagnani, L., et al., 2009. A new mass conservation approach to the study ofCO(2) advection in an alpine forest. J. Geophys. Res. D: Atmos. 114.

Pilegaard, K., et al., 2003. Field measurements of atmosphere–biosphereinteractions in a Danish beech forest. Boreal Environ. Res. 8, 315–333.

Rebmann, C., et al., Treatment and assessment of the CO(2)-exchange at a complexforest site in Thuringia, Germany. Agric. For. Meteorol. 150, 684–691.

Schmid, H.P., et al., 2000. Measurements of CO2 and energy fluxes over a mixedhardwood forest in the mid-western United States. Agric. For. Meteorol. 103,357–374.

Suni, T., et al., 2003. Long-term measurements of surface fluxes above a Scots pineforest in Hyytiala, southern Finland, 1996–2001. Boreal Environ. Res. 8, 287–301.

Suni, T., et al., 2003. Air temperature triggers the recovery of evergreen borealforest photosynthesis in spring. Global Change Biol. 9, 1410–1426.

Tedeschi, V., et al., 2006. Soil respiration in a Mediterranean oak forest at differentdevelopmental stages after coppicing. Global Change Biol. 12, 110–121.

A

up

R

A

B

B

B

B

C

C

C

C

C

D

E

F

F

F

G

G

H

E.K. Melaas et al. / Agricultural and Forest Meteorology 171– 172 (2013) 46– 56 55

ppendix C.

Estimated parameter values for each optimal modified growing degree-day model. The values are produced from model calibrationssing all candidate site-years of the corresponding vegetation group or subgroup. If a single value j is listed in the table under a givenarameter P, then P = j in the corresponding model. If two values j and k are listed, then P = T̄ j + k, such that T̄ is mean annual temperature.Spring onset metric Model Tf (◦C) Tc (◦C) F* C* Photomin

All DBF.Source–sink trans SW2.2 −14.0 224 13.35% GEPratio SEQ1.3 −8.6 8.5 6.8 859 61 8.510% GEPratio SW2.2 −8.9 146 12.6

Boreal and temp. DBFSource–sink trans. SW2.2 −13.6 197 13.65% GEPratio SEQ1.4 −11.6 7.3 938 4.1 56 10.510% GEPratio SEQ1.3 −9.9 7.5 −0.9 944 87 9.6

All ENF5% GEPratio SW1.1 −13.8 688 6.210% GEPratio SEQ1.4 −5.7 13.3 118 −6.2 184 13.3

Boreal ENF5% GEPratio SW2.1 11 11.210% GEPratio SW2.1 20 11.4

Temperate ENF5% GEPratio SEQ1.1 −8.9 12.8 340 75 10.610% GEPratio SW1.3 −13.3 17.2 694 6.6

eferences

ber, J.D., Reich, P.B., Goulden, M.L., 1996. Extrapolating leaf CO2 exchange to thecanopy: a generalized model of forest photosynthesis compared with measure-ments by eddy correlation. Oecologia 106, 257–265.

aldocchi, D., Falge, E., Gu, L.H., 2001. FLUXNET: a new tool to study the temporal andspatial variability of ecosystem-scale carbon dioxide, water vapor, and energyflux densities. Bull. Am. Meteorol. Soc. 82, 2415–2434.

aldocchi, D.D., Black, T.A., Curtis, P.S., Falge, E., Fuentes, J.D., Granier, A., Gu, L.,Knohl, A., Pilegaard, K., Schmid, H.P., Valentini, R., Wilson, K., Wofsy, S., Xu, L.,Yamamoto, S., 2005. Predicting the onset of net carbon uptake by deciduousforests with soil temperature and climate data: a synthesis of FLUXNET data.Int. J. Biometeorol. 49, 377–387.

öttcher, K., Kervinen, M., Aurela, M., Mattila, O.-P., Markkanen, T., Pullianinen, J.,2011. Monitoring spring phenology of boreal coniferous forest in Finland usingMODIS time-series. In: Proceedings of the 31st EARSeL Symposium of RemoteSensing and Geoinformation not only for Scientific Cooperation, 30 May–2 June2011, Prague, Czech Republic, ISBN 978-80-01-04868-9, pp. 125–134.

urnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference: APractical Information – Theoretic Approach, 2nd ed. Springer, New York.

esaraccio, C., Spano, D., Snyder, R.L., Duce, P., 2004. Chilling and forcing model topredict bud-burst of crop and forest species. Agric. For. Meteorol. 126, 1–13.

huine, I., Cour, P., Rousseau, D.D., 1998. Fitting models predicting dates of flower-ing of temperate-zone trees using simulated annealing. Plant Cell Environ. 21,455–466.

huine, I., Cour, P., Rousseau, D.D., 1999. Selecting models to predict the timing offlowering of temperate trees: implications for tree phenology modelling. PlantCell Environ. 22, 1–13.

hurkina, G., Schimel, D., Braswell, B.H., Xiao, X.M., 2005. Spatial analysis of growingseason length control over net ecosystem exchange. Global Change Biol. 11,1777–1787.

leland, E.E., Chuine, I., Menzel, A., Mooney, H.A., Schwartz, M.D., 2007. Shifting plantphenology in response to global change. Trends Ecol. Evol. 22, 357–365.

ragoni, D., Schmid, H.P., Wayson, C.A., Potter, H., Grimmond, C.S.B., Randolph, J.C.,2011. Evidence of increased net ecosystem productivity associated with a longervegetated season in a deciduous forest in south-central Indiana, USA. GlobalChange Biol. 17, 886–897.

nsminger, I., Sveshnikov, D., Campbell, D.A., Funk, C., Jansson, S., Lloyd, J., Shibistova,O., Oquist, G., 2004. Intermittent low temperatures constrain spring recovery ofphotosynthesis in boreal Scots pine forests. Global Change Biol. 10, 995–1008.

isher, J.I., Mustard, J.F., 2006. Cross-scalar satellite phenology from ground, Landsat,and MODIS data. Remote Sens. Environ. 109, 261–273.

isher, J.I., Richardson, A.D., Mustard, J.F., 2007. Phenology model from surface mete-orology does not capture satellite-based greenup estimations. Global ChangeBiol. 13, 707–721.

itzjarrald, D.R., Acevedo, O.C., Moore, K.E., 2001. Climatic consequences of leafpresence in the eastern United States. J. Clim. 14, 598–614.

anguly, S., Friedl, M.A., Tan, B., Zhang, X.Y., Verma, M., 2010. Land surface phenologyfrom MODIS: characterization of the Collection 5 global land cover dynamicsproduct. Remote Sens. Environ. 114, 1805–1816.

arrity, S.R., Bohrer, G., Maurer, K.D., Mueller, K.L., Vogel, C.S., Curtis, P.S., 2011.

Hänninen, H., 1995. Effects of climatic change on trees from cool and temperateregions: an ecophysiological approach to modelling of bud burst phenology.Can. J. Bot.: Rev. Can. Bot. 73, 183–199.

Hänninen, H., Kramer, K., 2007. A framework for modelling the annual cycle of treesin boreal and temperate regions. Silva Fenn. 41, 167–205.

Hunter, A.F., Lechowicz, M.J., 1992. Predicting the timing of budburst in temperatetrees. J. Appl. Ecol. 29, 597–604.

Jeong, S.J., Medvigy, D., Shevliakova, E., Malyshev, S., 2012. Uncertainties in terres-trial carbon budgets related to spring phenology. J. Geophys. Res. 117, G01030,http://dx.doi.org/10.1029/2011JG001868.

Jolly, W.M., Nemani, R., Running, S.W., 2005. A generalized, bioclimatic index topredict foliar phenology in response to climate. Global Change Biol. 11, 619–632.

Kaduk, J.D., Los, S.O., 2011. Predicting the time of green up in temperate and borealbiomes. Clim. Change 107, 277–304.

Keenan, T.F., Banker, I., Barr, A., Ciais, P., Davis, K., Dietze, M., Dragoni, D., Gough,C.M., Grant, R., Hollinger, D., Hufkens, K., Poulter, B., Mccaughey, H., Raczka,B., Ryu, Y., Schaefer, K., Tian, H.Q., Verbeeck, H., Zhao, M.S., Richardson, A.D.,2012. Terrestrial biosphere model performance for inter-annual variability ofland–atmosphere CO2 exchange. Global Change Biol. 18, 1971–1987.

Körner, C., Basler, D., 2010. Phenology under global warming. Science 327,1461–1462.

Kucharik, C.J., Barford, C.C., El Maayar, M., Wofsy, S.C., Monson, R.K., Baldocchi, D.D.,2006. A multiyear evaluation of a dynamic global vegetation model at threeAmeriFlux forest sites: vegetation structure, phenology, soil temperature, andCO2 and H2O vapor exchange. Ecol. Modell. 196, 1–31.

Landsberg, J.J., 1974. Apple fruit bud development and growth – analysis and anempirical model. Ann. Bot. 38, 1013–1023.

Levis, S., Bonan, G.B., 2004. Simulating springtime temperature patterns in thecommunity atmosphere model coupled to the community land model usingprognostic leaf area. J. Clim. 17, 4531–4540.

Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953.Equations of state calculations by fast computing machines. J. Chem. Phys. 21,1087–1092.

Migliavacca, M., Sonnentag, O., Keenan, T.F., Cescatti, A., O’Keefe, J., Richardson,A.D., 2012. On the uncertainty of phenological responses to climate change, andimplications for a terrestrial biosphere model. Biogeosciences 9, 2063–2083.

Moffat, A.M., Papale, D., Reichstein, M., Hollinger, D.Y., Richardson, A.D., Barr,A.G., Beckstein, C., Braswell, B.H., Churkina, G., Desai, A.R., Falge, E., Gove, J.H.,Heimann, M., Hui, D.F., Jarvis, A.J., Kattge, J., Noormets, A., Stauch, V.J., 2007.Comprehensive comparison of gap-filling techniques for eddy covariance netcarbon fluxes. Agric. For. Meteorol. 147, 209–232.

Monson, R.K., Sparks, J.P., Rosenstiel, T.N., Scott-Denton, L.E., Huxman, T.E., Harley,P.C., Turnipseed, A.A., Burns, S.P., Backlund, B., Hu, J., 2005. Climatic influenceson net ecosystem CO2 exchange during the transition from wintertime carbonsource to springtime carbon sink in a high-elevation, subalpine forest. Oecologia146, 130–147.

Moore, K.E., Fitzjarrald, D.R., Sakai, R.K., Goulden, M.L., Munger, J.W., Wofsy, S.C.,1996. Seasonal variation in radiative and turbulent exchange at a deciduousforest in central Massachusetts. J. Appl. Meteorol. 35, 122–134.

Murray, M.B., Cannell, M.G.R., Smith, R.I., 1989. Date of budburst of 15 tree speciesin Britain following climatic warming. J. Appl. Ecol. 26, 693–700.

A comparison of multiple phenology data sources for estimating seasonaltransitions in deciduous forest carbon exchange. Agric. For. Meteorol. 151,1741–1752.

änninen, H., 1990. Modeling dormancy release in trees from cool and temperateregions. Acta For. Fenn. 213, 1–47.

Papale, D., Reichstein, M., Aubinet, M., Canfora, E., Bernhofer, C., Kutsch, W.,Longdoz, B., Rambal, S., Valentini, R., Vesala, T., Yakir, D., 2006. Towards astandardized processing of Net Ecosystem Exchange measured with eddycovariance technique: algorithms and uncertainty estimation. Biogeosciences 3,571–583.

5 rest M

P

R

R

R

R

R

6 E.K. Melaas et al. / Agricultural and Fo

iao, S.L., Ciais, P., Friedlingstein, P., Peylin, P., Reichstein, M., Luyssaert, S., Margolis,H., Fang, J.Y., Barr, A., Chen, A.P., Grelle, A., Hollinger, D.Y., Laurila, T., Lindroth, A.,Richardson, A.D., Vesala, T., 2008. Net carbon dioxide losses of northern ecosys-tems in response to autumn warming. Nature 451, 49-U3.

eichstein, M., Falge, E., Baldocchi, D., Papale, D., Aubinet, M., Berbigier, P., Berhofer,C., Buchmann, N., Gilmanov, T., Granier, A., Gruwald, T., Havrankova, K., Ilves-niemi, H., Janous, D., Knohl, A., Laurila, T., Lohila, A., Loustau, D., Matteucci, G.,Meyers, T., Miglietta, F., Ourcival, J.M., Pumpanen, J., Rambal, S., Rotenberg, E.,Sanz, M., Tenhunen, J., Seufert, G., Vaccari, F., Vesala, T., Yakir, D., Valentini,R., 2005. On the separation of net ecosystem exchange into assimilation andecosystem respiration: review and improved algorithm. Global Change Biol. 11,1424–1439.

ichardson, A.D., O’Keefe, J., 2009. Phenological differences between understory andoverstory: a case study using the long term Harvard Forest records. In: Noormets,A. (Ed.), Phenology of Ecosystem Processes. Springer, New York, pp. 87–117.

ichardson, A.D., Hollinger, D.Y., Dail, D.B., Lee, J.T., Munger, J.W., O’Keefe, J., 2009.Influence of spring phenology on seasonal and annual carbon balance in twocontrasting New England forests. Tree Physiol. 29, 321–331.

ichardson, A.D., Black, T.A., Ciais, P., Delbart, N., Friedl, M.A., Gobron, N., Hollinger,D.Y., Kutsch, W.L., Longdoz, B., Luyssaert, S., Migliavacca, M., Montagnani, L.,Munger, J.W., Moors, E., Piao, S.L., Rebmann, C., Reichstein, M., Saigusa, N., Tomel-leri, E., Vargas, R., Varlagin, A., 2010. Influence of spring and autumn phenologicaltransitions on forest ecosystem productivity. Philos. Trans. R. Soc. Lond., Ser. B

365, 3227–3246.

ichardson, A.D., Anderson, R.S., Arain, M.A., Barr, A.G., Bohrer, G., Chen, G.S., Chen,J.M., Ciais, P., Davis, K.J., Desai, A.R., Dietze, M.C., Dragoni, D., Garrity, S.R., Gough,C.M., Grant, R., Hollinger, D.Y., Margolis, H.A., McCaughey, H., Migliavacca, M.,Monson, R.K., Munger, J.W., Poulter, B., Raczka, B.M., Ricciuto, D.M., Sahoo,

eteorology 171– 172 (2013) 46– 56

A.K., Schaefer, K., Tian, H.Q., Vargas, R., Verbeeck, H., Xiao, J.F., Xue, Y.K., 2012.Land surface models need better representation of vegetation phenology: resultsfrom the North American Carbon Program Site Synthesis. Global Change Biol. 18,566–584.

Sarvas, R., 1974. Investigations on the annual cycle of development of forest trees.Autumn dormancy and winter dormancy. Commun. Inst. For. Fenn. 84, 101.

Stoöckli, R., Rutishauser, T., Dragoni, D., O’Keefe, J., Thornton, P.E., Jolly,M., Lu, L., Denning, A.S., 2008. Remote sensing data assimilation for aprognostic phenology model. J. Geophys. Res. Biogeosci. 113, G04021,http://dx.doi.org/10.1029/2008JG000781.

Suni, T., Berninger, F., Vesala, T., Markkanen, T., Hari, P., Makela, A., Ilvesniemi, H.,Hänninen, H., Nikinmaa, E., Huttula, T., Laurila, T., Aurela, M., Grelle, A., Lindroth,A., Arneth, A., Shibistova, O., Lloyd, J., 2003. Air temperature triggers the recov-ery of evergreen boreal forest photosynthesis in spring. Global Change Biol. 9,1410–1426.

White, M.A., Thornton, P.E., Running, S.W., 1997. A continental phenology modelfor monitoring vegetation responses to interannual climatic variability. GlobalBiogeochem. Cycles 11, 217–234.

Wilson, K.B., Hanson, P.J., Baldocchi, D.D., 2000. Factors controlling evaporation andenergy partitioning beneath a deciduous forest over an annual cycle. Agric. For.Meteorol. 102, 83–103.

Yuan, W.P., Liang, S.L., Liu, S.G., Weng, E.S., Luo, Y.Q., Hollinger, D., Zhang, H.C.,2012. Improving model parameter estimation using coupling relationships

between vegetation production and ecosystem respiration. Ecol. Modell. 240,29–40.

Zhang, X.Y., Friedl, M.A., Schaaf, C.B., Strahler, A.H., Hodges, J.C.F., Gao, F., Reed, B.C.,Huete, A., 2003. Monitoring vegetation phenology using MODIS. Remote Sens.Environ. 84, 471–475.