Ag

HQEa

b

c

d

Se

f

g

h

i

a

ARRAA

KRMEGLL

n

h1

Ecological Indicators 72 (2017) 153–164

Contents lists available at ScienceDirect

Ecological Indicators

j o ur na l ho me page: www.elsev ier .com/ locate /eco l ind

ssessing the ability of MODIS EVI to estimate terrestrial ecosystemross primary production of multiple land cover types

ao Shia, Longhui Lia,∗, Derek Eamusa,b, Alfredo Huetea, James Cleverlya,b, Xin Tianc,iang Yua, Shaoqiang Wangd, Leonardo Montagnanie, Vincenzo Magliulo f,yal Rotenbergg, Marian Pavelkah, Arnaud Carrara i

School of Life Sciences, University of Technology Sydney, Sydney 2000, NSW, AustraliaAustralian Supersite Network, University of Technology Sydney, Sydney 2000, NSW, AustraliaResearch Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, People’s Republic of ChinaKey Laboratory of Ecosystem Network Observation and Modeling, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy ofciences, Beijing 100101, People’s Republic of ChinaFaculty of Science and Technology, Free University of Bolzano, Piazza Università 1, 39100, ItalyCNR Institute for Agricultural and Forest Systems, Via Patacca 85, 80056 Ercolano, Napoli, ItalyDepartment of Environmental Sciences and Energy Research, Weizmann Institute of Science, Rehovot 76100, IsraelLaboratory of Plants Ecological Physiology, Institute of Systems Biology and Ecology AS CR, Porící 3b, Brno 60300, Czech RepublicFundación Centro de Estudios Ambientales del Mediterráneo (CEAM), Charles Darwin 14, Parc Tecnològic, 46980 Paterna, Spain

r t i c l e i n f o

rticle history:eceived 11 August 2015eceived in revised form 11 June 2016ccepted 11 August 2016vailable online 20 August 2016

eywords:emote sensingODIS

nhanced vegetation indexross primary productionand cover typeseaf area index

a b s t r a c t

Terrestrial ecosystem gross primary production (GPP) is the largest component in the global carbon cycle.The enhanced vegetation index (EVI) has been proven to be strongly correlated with annual GPP withinseveral biomes. However, the annual GPP-EVI relationship and associated environmental regulationshave not yet been comprehensively investigated across biomes at the global scale. Here we exploredrelationships between annual integrated EVI (iEVI) and annual GPP observed at 155 flux sites, whereGPP was predicted with a log-log model: ln(GPP) = a × ln(iEVI) + b. iEVI was computed from MODISmonthly EVI products following removal of values affected by snow or cold temperature and withoutcalculating growing season duration. Through categorisation of flux sites into 12 land cover types, theability of iEVI to estimate GPP was considerably improved (R2 from 0.62 to 0.74, RMSE from 454.7 to368.2 g C m−2 yr−1). The biome-specific GPP-iEVI formulae generally showed a consistent performancein comparison to a global benchmarking dataset (R2 = 0.79, RMSE = 387.8 g C m−2 yr−1). Specifically, iEVIperformed better in cropland regions with high productivity but poorer in forests. The ability of iEVI inestimating GPP was better in deciduous biomes (except deciduous broadleaf forest) than in evergreendue to the large seasonal signal in iEVI in deciduous biomes. Likewise, GPP estimated from iEVI was ina closer agreement to global benchmarks at mid and high-latitudes, where deciduous biomes are morecommon and cloud cover has a smaller effect on remote sensing retrievals. Across biomes, a significant andnegative correlation (R2 = 0.37, p < 0.05) was observed between the strength (R2) of GPP-iEVI relationshipsand mean annual maximum leaf area index (LAImax), and the relationship between the strength and

mean annual precipitation followed a similar trend. LAImax also revealed a scaling effect on GPP-iEVIrelationships. Our results suggest that iEVI provides a very simple but robust approach to estimate spatialpatterns of global annual GPP whereas its effect is comparable to various light-use-efficiency and data-driven models. The impact of vegetation structure on accuracy and sensitivity of EVI in estimating spatialGPP provides valuable clues to improve EVI-based models.

© 2016 Elsevier Ltd. All rights reserved.

∗ Corresponding author at: School of Life Sciences, University of Technology, Syd-ey, PO Box 123, Sydney, NSW 2007, Australia.

E-mail addresses: Longhui.Li@uts.edu.au, lilhchn@gmail.com (L. Li).

ttp://dx.doi.org/10.1016/j.ecolind.2016.08.022470-160X/© 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Terrestrial gross primary production (GPP) is the amount ofcarbon captured from the atmosphere through vegetation photo-synthesis (Beer et al., 2010). Vegetation GPP is a key component of

1 Indica

thasZeottm2

ceeaaems2TlawraatnHXvtleaedM(2bOticadnYrom2

wMeotsb2sl

54 H. Shi et al. / Ecological

he terrestrial carbon balance and is of fundamental importance touman society because plants provide food, fiber and wood supplynd also contribute to the production of environmental conditionsuitable for human habitation (Melillo et al., 1993; Xiao et al., 2005;hao et al., 2005). Therefore, continuous monitoring and accuratestimation of GPP is required to ensure the long term securityf terrestrial ecosystem services and to address issues pertainingo the global carbon cycle including determination of the size ofhe terrestrial carbon sink, prediction of vegetation dynamics, and

anagement of forests and grasslands (Ciais et al., 2005; Ma et al.,013; Sims et al., 2006b).

GPP can be calculated as the sum of vegetation assimilatedarbon flux, partitioned from net carbon exchange measured atddy covariance (EC) tower sites (Baldocchi et al., 2001; Reichsteint al., 2007), but such observations are limited, both temporallynd spatially. Remote sensing technique provides a promisingpproach to overcome these limitations. Various diagnostic mod-ls taking advantage of spatially extensive remote sensing andeteorological data have been developed to estimate GPP across

tand-to-global scales for a relatively long period (e.g., Jung et al.,008; Running et al., 2004; Sims et al., 2008; Xiao et al., 2005).hese models can be generally partitioned into three categories:ight-use-efficiency (LUE) models, machine learning algorithmsnd simple empirical models (Verma et al., 2014). The LUE theoryas first proposed by Monteith (1972), in which GPP is generally

epresented as the product of LUE, photosynthetically active radi-tion (PAR), the fraction of PAR absorbed by vegetation (fAPAR),nd environmental scalars. fAPAR is a strong function of vegeta-ion greenness, as measured by vegetation indices (VIs), such as theormalized difference vegetation index (NDVI; e.g., Goward anduemmrich, 1992) and the enhanced vegetation index (EVI; e.g.,iao et al., 2004a,b). However, it is difficult to estimate LUE, whicharies among plant functional types, and can be down-regulated byemperature, soil water content, vapour pressure deficit (VPD), andeaf phenology (Xiao et al., 2005). Another deficiency of LUE mod-ls is the coarse resolution of climate inputs, which are often onlyvailable at a large scale. This may introduce significant errors tostimations of GPP (Heinsch et al., 2006; Zhao et al., 2005) and hin-er the acquisition of fine-resolution GPP estimates at large scales.achine learning algorithms, such as artificial neural networks

Papale and Valentini, 2003), support vector machines (Yang et al.,007), and model tree ensembles (Jung et al., 2009), predict GPPased on the non-functional patterns extracted in training data set.bviously, the accuracy of machine learning algorithms relies on

he abundance and representativeness of input information includ-ng remote sensed vegetation properties, meteorological, and landover data (Jung et al., 2011). Therefore, the use of machine learninglgorithms is also limited by the coarse resolution of meteorologicalata. Moreover, in many cases machine learning algorithms showo better performance than LUE models in specific ecosystems (e.g.,ang et al., 2007). Consequently, simple empirical models utilizingemote sensing proxies of vegetation photosynthesis activity (withr without meteorological data) gain consistent interest in esti-ating both spatial and temporal variations of GPP (e.g., Jung et al.,

008; Rahman et al., 2005; Sims et al., 2006b).The growing season NDVI and EVI show strong relationships

ith vegetation production over one or two week intervals (e.g.,ao et al., 2014; Rahman et al., 2005; Sims et al., 2006a,b; Wylie

t al., 2003). Vegetation indices per se are transformations of twor more spectral bands to enhance the signal derived from vege-ation properties (Huete et al., 2002). Both NDVI and EVI employurface bidirectional reflectances of red and near-infrared spectral

ands that are sensitive to leaf chlorophyll content (Huete et al.,002), which converts light to energy consumed by photosynthe-is. NDVI is limited due to its saturation over dense vegetation andarge sensitivity to canopy background brightness (Huete et al.,

tors 72 (2017) 153–164

2002), whereas EVI can improve performance in regions of highbiomass through a decoupling of the canopy and background sig-nals and a reduction in the influence of atmospheric conditionsusing a blue spectral reflectance (Huete et al., 2002). This makes EVImore responsive to canopy structural variations and thus EVI is bet-ter correlated with GPP than NDVI in evergreen (Xiao et al., 2004a)and deciduous (Xiao et al., 2004b) forests as well as in croplands(Xiao et al., 2005). Compared to LUE models, the growing seasonEVI or EVI-based models (e.g., Temperature-Greenness model; Simset al., 2008) provide a comparable or better estimation of GPP atboth the 16-day (Sims et al., 2008, 2006b) and annual (Verma et al.,2014) time-scales. As well as EVI, cumulative growing season fAPARwith separate functions for herbaceous plants, evergreen forestsand all other vegetation types has been used to predict annual GPPin Europe (Jung et al., 2008). The disadvantage of selecting fAPARagainst EVI is subtle: fAPAR consists of fractional absorbance of PARabsorbed by both chlorophyll and by non-photosynthetic pigments(Zhang et al., 2005), while EVI is much closer to the fraction of PARabsorbed by chlorophyll. Moreover, fAPAR shows no significant cor-relation with GPP in deciduous broadleaf forests (Jung et al., 2008).Therefore, the use of EVI should be favored over fAPAR in correlat-ing to GPP. However, current studies on EVI-GPP relationships orEVI-based models have been focused within only a limited numberof biomes and these EVI-based models generally need to computethe start and end or the length of the growing season period (Junget al., 2008; Sims et al., 2008, 2006b; Verma et al., 2014), whichconstitutes an extra source of uncertainty. Simultaneously, envi-ronmental influences on the ability of EVI to estimate GPP acrossa wide spectrum of biomes have not yet been investigated (Simset al., 2006b; Sjöström et al., 2011).

In this study, we used the annual integral of MODIS EVI (iEVI),which only needs removal of those values that have been affectedby cold temperature or snow and subtracting the soil backgroundsignal, to regress with annual eddy covariance measured GPP across12 land cover types. The developed set of formulae were thenapplied at the global scale and compared with a widely used GPPbenchmark dataset to evaluate the effectiveness and robustness ofiEVI, thereby determining whether iEVI can serve as a referencefor other GPP models over a fine-to-coarse resolution. The impactsof environmental conditions on iEVI in estimating GPP were fur-ther investigated across biomes, to improve our understanding ofthe underlying mechanistic processes that differentiate responsesof vegetation photosynthetic activity to remote sensing spectralmeasurements among biomes.

2. Data and methods

2.1. Eddy covariance and meteorological data

The eddy covariance method is a micrometeorological tech-nique that directly measures net carbon, water and energy fluxesacross a horizontal plane between vegetation canopies and theatmosphere (Aubinet et al., 2000; Baldocchi et al., 2001). In thepresent study a total of 155 sites (Supplementary Table S1) wereselected, consisting of 624 site-years of data and representing aworldwide spectrum of biomes and climate regimes with excellentcoverage in North America, Eurasia and Oceania (Table 1, Fig. 1;Baldocchi, 2008; Baldocchi et al., 2001; Wang and Dickinson, 2012).

The flux data were obtained from three sources: (1) a smallfraction (mainly high-latitude and wetland sites) was collecteddirectly from published studies, which only included annual val-

ues of flux and meteorological forcing; (2) a larger fraction wascontributed directly from participating site researchers; and (3)the majority were from FLUXNET level 2 or level 4 productsthat were downloaded from the database. Of the latter two cat-

H. Shi et al. / Ecological Indicators 72 (2017) 153–164 155

Fig. 1. Geographical distribution of flux towers overlaid onto the 2

Table 1Summary of number of sites and site-years used for each biome. CNM: crop-land/natural vegetation mosaic; CRO: croplands; CSH: closed shrublands; DBF:deciduous broadleaf forest; DNF: deciduous needle-leaf forest; EBF: evergreenbroadleaf forest; ENF: evergreen needle-leaf forest; GRA: grasslands; MF: mixedforest; OSH: open shrublands; SAV: savannas; WET: permanent wetlands; WSA:woody savannas.

Biome CRO CSH DBF DNF EBF ENF GRA MF OSH SAV WET WSA

eNegmcppa(ttfdFoaitdutsbg

2

ctoa

Sites 16 4 18 4 13 40 24 8 7 7 9 5Site-years 61 11 83 6 55 190 76 28 21 24 34 35

gories, only site-years with small gaps (i.e., individual gaps inEE of less than 5% of the entire annual record) were selectedxcept in certain ecosystems of the boreal region where onlyrowing season data were available. Carbon and water fluxes andeteorological variables in all selected site-years were then pro-

essed through gap-filling and flux partitioning routines. If therincipal investigator at each site had already performed theserocesses, the already gap-filled and partitioned GPP dataset wasggregated from a half-hourly time-step to an annual time-scaleGPPEC). Otherwise, half-hourly GPP derived using one of thewo FLUXNET standard methods, either the marginal distribu-ion sampling (MDS, a local method; Reichstein et al., 2005) or aeed-forward artificial neural network (ANN, trained on an annualataset; Papale and Valentini, 2003), were obtained from theLUXNET products to calculate annual GPP. Both partitioning meth-ds show good performance according to previous studies (Papalend Valentini, 2003; Reichstein et al., 2005). For sites with neithernvestigator’s decomposition nor standardized flux partitioning,he publicly available online MDS tool (http://www.bgc-jena.mpg.e/∼MDIwork/eddyproc/index.php; Reichstein et al., 2005) wassed to gap-fill and partition NEE. The derived half-hourly GPP,emperature, precipitation and vapour pressure deficit (VPD) of allite-years were screened for outliers and linearly interpolated inins representing the measurement hour of the day before aggre-ation into the annual scale.

.2. Benchmark dataset

The model tree ensemble (MTE) approach was used to empiri-

ally up-scale FLUXNET measurements of fluxes (hereafter GPPMTE)o the global scale. Explanatory variables for the model consistedf meteorological variables, the biophysical state of the vegetation,nd vegetation types (Jung et al., 2009, 2011). GPPMTE constitutes

001 MODIS IGBP land cover map at a 0.5◦ × 0.5◦ resolution.

a benchmark for global FLUXNET up-scaling that has been usedas a baseline for evaluating land surface models and estimatingglobal CO2 uptake (e.g., Beer et al., 2010; Bonan et al., 2011). How-ever, GPPMTE has its own weakness, specifically for estimates of GPPin high-production croplands (Guanter et al., 2014). Mean annualGPPMTE was calculated at a spatial resolution of 0.5◦ for the years1982–2008. The same grid was also applied in our global GPP esti-mation.

2.3. Satellite data

2.3.1. EVI dataEVI is widely used as a proxy of canopy “greenness” to address

spatial and temporal variations in terrestrial photosynthetic activ-ity (e.g., Huete et al., 2002; Ma et al., 2013). EVI is defined as (Hueteet al., 1997):

EVI = G�NIR − �red

�NIR + C1 × �red − C2 × �blue + L

where �NIR, �red and �blue are atmospherically corrected, either fullyor partially, values of surface near-infrared (NIR, 841–876 nm), red(620–670 nm) and blue (459–479 nm) spectral reflectance, respec-tively; G is the gain factor (set at 2.5); L (set at 1.0) is the canopybackground adjustment; and C1(set at 6) and C2 (set at 7.5) are thecoefficients of the aerosol resistance term, which uses the blue bandto correct for the influence of aerosols in the red band.

MODIS monthly VI products (MOD13A3.005) for February2000–2013 were obtained from the USGS repository (http://e4ftl01.cr.usgs.gov/MOLT/MOD13A3.005/). This dataset is pro-duced globally over land at 1-km resolution and monthlycompositing periods from atmospherically corrected surfacereflectances. The compositing algorithm is based on a constrained-view angle-maximum value composite (CV-MVC) to minimizeatmospheric and bidirectional reflectance distribution function(BRDF) influences (Huete et al., 2002).

It is difficult to precisely co-locate the pixels that directly cor-respond to the footprint of an EC tower (Sims et al., 2006b).Fluctuations in flux tower footprint size and shape, due to theunderlying topography, vegetation, wind speed and etc., mayinduce a footprint mismatch between the tower and MODIS (Jung

et al., 2009; Sims et al., 2006b; Sjöström et al., 2011). Where thelandscape is homogenous, the scale mismatch is not a seriousproblem and the MODIS pixels can adequately represent flux siteconditions. Discrepancies are typically observed at grassland and

1 Indica

c(ibtdle2oee

2

aSoodeo(

i

wa

X

wagpiaa

Lm6nEpomocitfita2pnacdG

2

ir

56 H. Shi et al. / Ecological

ropland sites, likely due to the fragmentation of these landscapesCescatti et al., 2012). However, sub- and inter-pixel heterogene-ty is unavoidable in most cases and thus introduces additionalias. Consequently, a central 3 × 3 km window surrounding the fluxower was used to extract mean EVI time series. The 3 × 3 km win-ow has been found to reduce scale mismatch relative to a centrally

ocated 1 km pixel or window sizes of 5 × 5 or 7 × 7 km (e.g., Mat al., 2013; Rahman et al., 2005; Sims et al., 2006b; Sjöström et al.,011; Xiao et al., 2005). At sites with spatially varying amountsf mixed vegetation types, averaging across the MODIS window isquivalent to averaging across time in flux measurements (Simst al., 2006b).

.3.2. Smoothing method of EVITo reduce noise and uncertainties in the MODIS EVI time series

t each site, the singular spectrum analysis (SSA) was employed.SA is a data adaptive, non-parametric analysis approach basedn embedding a time series

{X (t) : t = 1, N

}in a vector space

f dimension M and it works well in the analysis of non-linearynamics in geophysical datasets (Kondrashov and Ghil, 2006; Mat al., 2013; Wang and Liang, 2008). The SSA technique consistsf two complementary stages: decomposition and reconstructionHassani, 2007). The one-dimensional time series

{X (t) : t = 1, N

}s first embedded into a trajectory matrixX = [X1, ...XK ] =

(xij

)L,K

i,j=1,

here K = t − L + 1. Next, singular value decomposition (SVD) ispplied to X:

=d∑

j=1

√�iUjV

Tj , Vj = X

′/√

�j

here �j is the jth eigenvalue of XX′, Uj is the jth eigenvector of XX

′,

nd d is the rank of X . The reconstruction includes the eigentriplerouping and diagonal averaging (i.e., Hankelization of a matrix), toroduce a length N time series from the matrix X . The SSA method

s more robust to outliers than linear filtering because it conducts global reconstruction (i.e., convolution) of the whole time seriess with Fourier methods (Alexandrov, 2009).

Key parameters in SSA are the decomposition window length and the number of leading components in reconstruction. In theonthly EVI time series, a window length of 37 (i.e. 37 months) and

leading components best captured the periodicity and simulta-eously reduced random noises during reconstruction. The missingVI value in January 2000 was extrapolated, thereby yielding a com-lete set of 14 years of EVI data. Since bare soil yields an EVI valuef 0.08–0.10, around which GPP is zero (Sims et al., 2008, 2006b),onthly values of EVI reconstructed from SSA were corrected by

ffsetting 0.10 to remove the background signal. EVI data are alsoontaminated by snow effects in mid- and high latitudes that resultn a false positive signal (Huete et al., 2002). To minimise the con-amination, the snow/ice flag in MOD13A3 VI quality assuranceeld was first used to remove the snow-covered EVI values; andhen EVI values were further screened for effects of cold temper-ture (daytime land surface temperature below −2 ◦C; Tan et al.,011; Zhang et al., 2004) using the MODIS land surface temperatureroduct (MOD11C3). Finally, positive values of monthly EVI duringon-snow and non-cold temperature periods were summed intonnual, integrated values (iEVI) to regress against annual GPP cal-ulated at each flux site. In our global GPP estimation, the 1 km EVIata were re-gridded into 0.5◦ resolution to compare with globalPPMTE and then processed the same way as the site level analysis.

.3.3. Leaf area index dataThe MOD15A2.005 leaf area index (LAI) product is compos-

ted every 8 days at 1 km resolution and is available at the USGSepository (http://e4ftl01.cr.usgs.gov/MOLT/MOD15A2.005). LAI is

tors 72 (2017) 153–164

retrieved through a three-dimensional radiative transfer modelthat requires land cover classification (Knyazikhin et al., 1998). Byapplying a procedure similar to that used for EVI, a central 3 × 3 kmwindow was used to extract LAI time series from 2000 to 2013, andSSA was applied to smooth the series. At each site, peak LAI val-ues in individual site-years were averaged to represent the meanannual maximum LAI (LAImax) of the site and then the site LAI val-ues within a land cover type were averaged to get the mean LAI foreach of the land cover classifications except wetlands, which haveno observed values in MOD15A2.005. In our analyses we assigned amean LAI value of 6.3 (±2.3) for the 6 wetland sites, obtained froma global synthesis of LAI observations (Asner et al., 2003).

2.3.4. Land cover typesThe MCD12Q1.005 land cover type product provides options of

five global land cover classification systems. We used the IGBP landcover scheme which includes 17 land cover classes: water, ever-green needle-leaf forest (ENF), evergreen broadleaf forest (EBF),deciduous needle-leaf forest (DNF), deciduous broadleaf forest(DBF), mixed forest (MF), closed shrublands (CSH), open shrub-lands (OSH), woody savannas (WSA), savannas (SAV), grasslands(GRA), permanent wetlands (WET), croplands (CRO), urban andbuilt-up, cropland/natural vegetation mosaic (CNM), snow andice, barren or sparsely vegetated. The IGBP land cover map in2001 was obtained from ORNL DAAC (http://webmap.ornl.gov/wcsdown/wcsdown.jsp?dg id=10004 1). The product has a spatialresolution of 500 m. In our global GPP estimation, this map wasresampled into 0.5◦ resolution to match the GPPMTE product. Fourland cover types were excluded (water, urban and built-up, snowand ice, and barren or sparsely vegetated), and CNM was classedwith CRO. The GPPMTE product and spatial EVI data were overlayedusing the IGBP map.

2.4. Statistical analyses

Annual GPP calculated for each site in each year was corre-lated with the corresponding annual iEVI using log-log regressionsfollowing (Campos et al., 2013). Since the goodness-of-fit for intra-annual GPP-EVI correlations may differ across biomes (Rahmanet al., 2005; Sims et al., 2006b; Wu et al., 2010), the GPP-iEVIrelationship was further investigated within each biome. Theperformance of these GPP-iEVI models within each biome was eval-uated based on leave-one-out cross-validations (CV), which can testthe practical accuracy of these models. The GPP-iEVI models basedon biomes were then applied to the whole globe, and the final globalGPP estimation was compared to GPPMTE. Two standard statisticalmeasures were employed to assess the regression relationships:the coefficient of determination (R2) and the root mean squarederror (RMSE).

R2 = 1 −

∑i

(xi − yi)2

∑i

(xi − x)2

RMSE =

√√√√∑

i

(xi − yi)2

n

where xi denotes the observed data, yi the modeled data, n the num-ber of observations. R2 represents the proportion of total variationof observed data explained by the model. RMSE measures the biasbetween modeled and observed data.

H. Shi et al. / Ecological Indicat

Table 2Summary of the raw and leave-one-out cross-validation (CV) performance measures(coefficients of determination, R2; root mean squared error, RMSE) of ln(GPP)-ln(iEVI) models for all data points and each biome, respectively.

Biomes Raw R2 CV R2 Raw RMSE CV RMSE

All 0.67 0.66 454.7 455.8CRO 0.51 0.48 326.7 335.9CSH 0.90 0.81 195.4 239.6DBF 0.39 0.36 251.5 256.0DNF 0.90 0.51 175.1 284.9EBF 0.36 0.32 477.4 492.1ENF 0.49 0.48 448.8 454.2GRA 0.70 0.67 291.2 296.2MF 0.48 0.33 490.3 529.2OSH 0.61 0.52 104.6 115.0SAV 0.48 0.43 375.3 395.2

3

3b

psGvf(uacHsGiadtt

3

biwRcdanblbiGrotebcRd

WET 0.11 0.01 297.2 311.9WSA 0.93 0.93 214.3 222.8

. Results

.1. The overall relationship between GPP and iEVI withoutiomes categorization

The significant logarithmic-logarithmic regression (R2 = 0.67, < 0.001, Fig. 2, inset) between GPPEC and iEVI at the annual scalehows that there was a good general correspondence betweenPPEC and iEVI across all biomes (Fig. 2). The leave-one-out cross-alidation based performance measures (CV R2 = 0.66, Table 2)urther demonstrated the effectiveness of the logarithmic modellog-log) in regressing GPP against iEVI. The estimated GPPiEVIsing the global GPP-iEVI relationship also showed reasonablegreement with GPPEC (R2 = 0. 62, p < 0.001), although GPPiEVI wasonsistently underestimated relative to field (EC) measurements.owever, the point distribution in the relationship was much more

cattered at medium to high production biomes (approximatelyPP > 800 g C m−2 yr−1). The global relationship was unable to sat-

sfactorily estimate GPP accurately for some of the biomes, suchs evergreen broadleaf forest and woody savannas, although theata from these biomes still occurred within the overall distribu-ion (Fig. 2). Consequently, it was necessary to further investigatehe individual GPP-iEVI relationship within each biome.

.2. Biome-specific relationships between GPP and iEVI

There was considerable variation among the 12 individualiome types in the regression relationships between GPPEC and

EVI (Fig. 3). The poorest performance of iEVI in estimating GPPECas observed in wetlands in raw regression (R2 = 0.11, p < 0.05,MSE = 297.2 g C m−2 yr−1) and in cross validations (Table 2). Theorrelation was strongest in woody savannas, closed shrublands,eciduous needle-leaf forests, grasslands, and open shrublands;nd moderate in croplands, evergreen needle-leaf forests, savan-as, mixed forests, deciduous broadleaf forests, and evergreenroadleaf forests (Fig. 3). Meanwhile, we calculated the anoma-

ies (by subtracting the mean value) of GPP and iEVI within eachiome and investigated correlations between GPP anomaly and

EVI anomaly using linear regressions. Except for the wetlands,PP anomaly and iEVI anomaly showed moderate to strong cor-

elations within biomes (Fig. S1). This result was consistent withur non-linear models using GPP and iEVI themselves (Fig. 3). Fur-her, cross-validations revealed that the GPP-iEVI models withinach biome (except wetlands) performed robustly and thus could

e applied to the global scale. It is notable that the strength of theorrelation within deciduous biomes was generally better (higher2 and lower RMSE) than those within evergreen vegetation, excepteciduous broadleaf forest (still with better iEVI performance than

ors 72 (2017) 153–164 157

evergreen broadleaf forest). The relationship for mixed forests hada similar R2 to evergreen needle-leaf forest but was associated withthe largest RMSE (490.3 g C m−2 yr−1) across all biomes. There wasalso a wide range of values for the fitted slopes of the relation-ship between GPPEC and iEVI. Croplands and deciduous needle-leafforests occupied a similar and narrow iEVI spectrum, but GPPECof the former was more sensitive (larger slope) to iEVI. Amongforested biomes, GPPEC of evergreen broadleaf forest was less sen-sitive to iEVI than evergreen needle-leaf forests. Likewise, GPPEC ofdeciduous broadleaf forests was less sensitive to iEVI than decidu-ous needle-leaf forests. GPPEC of grasslands and savannas, both ofwhich are grass-dominated biomes, exhibited similar responses toiEVI. Although the range of GPPEC and iEVI values differed substan-tially among mixed forests and woody savannas, the biome-specificregression slopes were found to be close, whereas woody savannashad a much smaller RMSE (Fig. 3).

Estimated GPP based on the biome-specific GPP-iEVI formulaeat all sites were then compared with observed GPPEC (Fig. 4). Therelationship between GPPiEVI and GPPEC was significantly strength-ened relative to the regression obtained without biome partitioning(Fig. 2) with increased R2 (from 0.62 to 0.74), larger slope (from0.538 to 0.723), decreased RMSE (from 454.7 to 368.2 g C m−2 yr−1)and smaller intercept (from 492 to 295). There was a large disper-sion of points around the linearly fitted function, but these pointswere mainly obtained from high-GPP locations (>approximately2400 g C m−2 yr−1).

3.3. Global application of biome-specific GPP-iEVI relationships

The set of biome-specific GPP-iEVI relationships were applied tothe global data of iEVI and IGBP land cover types. Per-pixel compar-ison between GPPiEVI and GPPMTE demonstrated the consistency ofbiome-specific GPP-iEVI models when up-scaling GPP from the siteto global scales (R2 = 0.79, RMSE = 387.8 g C m−2 yr−1; Fig. 5). Thevery high accuracy of global, multi-year, averaged GPP (R2 = 0.97)within each biome was unexpected given that the number of fluxtowers was restricted and their distribution was not geographicallyuniform. Individual biomes for which annual GPP was larger than10 Pg C yr−1 were scattered farther from the 1:1 line, resulting inunderestimation of GPP by iEVI in EBF and overestimation in WSA,CRO and CNM in comparison with the benchmark dataset.

The mean spatial pattern of GPP was accurately reproducedby iEVI (Fig. 6). However, GPP was primarily underestimatedby iEVI in the tropics, western Russia and equatorial Africa andwas overestimated in Europe, eastern North America, the high-latitude tropics of Africa, southeastern South America, southeasternAustralia, southeastern Asia, and parts of India and north China. Inthese regions, central Africa was dominated by tropical EBF (areaaround equator) and its north and south edges (area around 5◦ Nand 5◦ S) were dominated by woody savannas; Europe, easternNorth America, southeastern South America and Australia, Indiaand north China were widely covered by cropland/natural vegeta-tion mosaics or croplands (Fig. 1), which were both parameterisedas croplands when calculating GPPiEVI. Latitudinal GPP derived fromthe iEVI showed positive biases from the benchmark in the regions30◦–38◦ S, 8◦–15◦ N, 20◦–28◦ N and 30◦–55◦ N, and negative biasesin the region 10◦ S–5◦ N (Fig. 7).

3.4. Canopy structure effect on biome-specific GPP-iEVIrelationships

Among vegetation and climatic factors (mean annual maximal

LAI, temperature, precipitation and VPD), only LAI and precipitationinfluenced the regression between GPP and iEVI. The strength of thebiome-specific correlations between GPP and iEVI decreased withincreasing mean annual LAI (R2 = 0.37, p < 0.05) (Fig. 8). The strength

158 H. Shi et al. / Ecological Indicators 72 (2017) 153–164

F rianceb ars. Tht

ocmTtmL

4

4

4fl

mwSt(tdfigtmistettu

ig. 2. Relationship between derived annual GPP from iEVI (GPPiEVI) and eddy covaiomes. The inset shows ln-transformed annual GPPEC and annual iEVI for all site-yehe 1:1 line.

f biome-specific GPP-iEVI relationships with mean annual pre-ipitation followed the same negative trend, but the R2 was onlyarginally statistically significant (R2 = 0.33, p = 0.051; Fig. 8 inset).

he slopes of relationships between GPPEC and iEVI (i.e., the sensi-ivity of GPP to EVI) across 12 biomes increased at small values of

aximal LAI and then decreased across larger values of maximalAI (LAI breakpoint was estimated to be 1.98 m2 m−2, Fig. 8).

. Discussion

.1. Uncertainty analysis

.1.1. Gap-filling and partitioning of eddy covariance carbonuxes

Two major concerns in up-scaling of eddy covariance measure-ents of fluxes are (1) the associated propagation of uncertaintyithin the source datasets, and (2) the up-scaling method itself.

ystematic and random errors in measurement, gap-filling and par-itioning procedures can result in uncertainty for estimates of GPPPapale et al., 2006). Flux measurements can be subject to substan-ial random errors, which can be modelled as a double exponentialistribution (Hollinger and Richardson, 2005). To minimise gap-lling errors in this study, we included only site-years without largeaps (less than 5% missing data). A short-term empirical tempera-ure function was used to model ecosystem respiration in the MDS

ethod and the robustness of this function depends on the nois-ness of the flux data and the range of temperatures during thehort period (Reichstein et al., 2005). Therefore, at sites with stableemperatures and noisy eddy covariance data, it can be difficult to

stablish a reliable relationship between ecosystem respiration andemperature (Reichstein et al., 2005). Consequently, datasets fromhe FLUXNET ANN product were preferred above the MDS prod-ct. The total annual error in eddy fluxes has been conservatively

tower measured GPP (GPPEC) since 2000 based on all sites and across 12 differente solid line represents the least squares regression line. The dashed line represents

estimated to be below 200 g C m−2 yr−1 (Reichstein et al., 2007) andthe products from these standard methodologies are widely used inup-scaling and benchmarking models (e.g., Beer et al., 2010; Bonanet al., 2011; Jung et al., 2009; Rahman et al., 2005). However, it isnoteworthy that neither the ANN nor the MDS method may be thebest option in all flux sites.

4.1.2. Ecosystem heterogeneityA further source of error is introduced when scaling EVI to

global GPP. To match the spatial resolution of GPPMTE, MODIS EVI(1 km resolution) and IGBP classification maps (500 m resolution)were resampled into 0.5◦, thereby simplifying prediction of GPPat the global scale. However, resampling unavoidably introducederror in areas with mixed land cover types. The loss of informa-tion concerning landscape heterogeneity within larger pixels cancause misuse of the biome-specific GPP-iEVI formulae at the sub-pixel level. The incongruence between GPPiEVI and GPPMTE in WSA,CNM and CRO could be due to varying proportions of vegetationcomponents within a grid cell. For WSA, the eddy covariance mea-sured flux data were mostly from Australia, where the woody andherbaceous components of woody savannas are substantially het-erogeneous (Hirota et al., 2011). Even within a single continent (e.g.,Africa or Australia), woody savannas display significant variation instructural and phenological patterns (Kutsch et al., 2008; Sjöströmet al., 2011, 2013). However, EVI only has a moderate capacityto predict ecosystem structural and functional attributes such asbasal cover of vegetated patches, perennial plants species richnessand retention of nutrients (Gaitán et al., 2013). Similar situations(e.g., different species, cultivars and fragments of croplands) can be

encountered in CNM and CRO, besides the fact that the CRO specificGPP-iEVI formula was applied in CNM. In addition, crops are gen-erally intensely managed (e.g., irrigation, fertilisation, sowing andharvest), which constrains the reflectance-based greenness indices

H. Shi et al. / Ecological Indicators 72 (2017) 153–164 159

F nd iEVf odnesr re sta

tqGGoTcsLoedtusi

4

g

ig. 3. Biome-specific relationships between tower-estimated annual GPP (GPPEC) aormulas within each biome. Coefficient of determination (R2) represents the fit goepresents the bias between GPP estimated using iEVI and GPPEC. All relationships a

o accurately estimate GPP of crops (Guanter et al., 2014). Conse-uently, the benchmark dataset GPPMTE underestimates croplandPP in large agricultural regions such as the US Corn Belt, the Indo-angetic Plain and the North China Plain but tends to moderatelyverestimate cropland GPP in South America (Guanter et al., 2014).hus, the overestimation of GPPiEVI in comparison with GPPMTE inroplands of North America, north India and north China (Fig. 6)eems reasonable but is still biased in other agricultural areas.imitations arising from global scaling can be overcome using theriginal relatively high-spatial resolution satellite data (Sjöströmt al., 2011; Zhao et al., 2005). However, using these fine resolutionata will inevitably increase modeling complexity. Furthermore,he 500 m resolution MODIS IGBP map has its own weakness andncertainty (Friedl et al., 2010). Errors due to global scaling wereimilar in GPPiEVI and GPPMTE and consequently were comparablen this study (Fig. 5).

.2. The relationships between EVI and GPP

Vegetation greenness indices (VIs) associated GPP models areenerally based on one of the following two hypothetical rela-

I. Solid lines represent GPP-iEVI relationships derived from the ln(GPPEC) ∼ ln(iEVI)s of the ln(GPPEC) ∼ ln(iEVI) relationship, and the root mean squared error (RMSE)tistically significant.

tionships between either LUE or GPP and VIs. The first holds thatVIs provide proxy information for parameterizing LUE or fAPAR(Gitelson et al., 2006; Inoue et al., 2008; Sims et al., 2006a,b; Wuet al., 2012) in photosynthetic (as opposed to non-photosynthetic)tissues (Xiao et al., 2004a,b, 2005). Following the logic of classi-cal LUE theory, various LUE models have been developed basedon eddy covariance observation and satellite data (Gitelson et al.,2006; Peng et al., 2011; Running et al., 2004; Sims et al., 2008; Yuanet al., 2010). Each of these models includes a combination of equa-tions that are scaled by environmental regulation of GPP (Beer et al.,2010). The second is that VIs can estimate GPP alone. Values of EVIfollow changes in the greenness and structure of vegetation regard-less of the cause of those variations (Huete et al., 2002), resulting ina stronger correlation between tower-estimated GPP and EVI thanthe correlations between tower-estimated GPP and MODIS GPP orbetween tower LUE and EVI during the photosynthetic period (Simset al., 2006b). The assumption that EVI can be taken as a proxy ofLUE results in curvilinear relationships between GPP and EVI (Sims

et al., 2006a). This strongly supports our results (Fig. 3), suggestingthe two hypotheses are essentially consistent, and it is thereforereasonable to assume a strong correlation between GPP and EVI.

160 H. Shi et al. / Ecological Indicators 72 (2017) 153–164

Fig. 4. Comparison of modelled annual GPP (GPPiEVI) using biome-specific GPP-iEVI relationships with eddy covariance tower measured GPP (GPPEC). The solid line representsthe linear regression line. The dashed line represents the 1:1 line.

Fig. 5. Comparison between modelled average annual GPP using iEVI (GPPiEVI) (2000–2013) and the benchmark GPP (GPPMTE) (1982–2008) at a grid level (left) and a biomel nd thi TE (198l

HtStfl

evel (right) across the globe. The red solid line represents the linear regression line andicate standard deviations of mean annual biome GPPiEVI (2000–2013) and GPPM

egend, the reader is referred to the web version of this article.)

owever, EVI does not perform satisfactorily across all vegetation

ypes, particularly at evergreen forest sites (Rahman et al., 2005;ims et al., 2006b; Wu et al., 2010). Furthermore, EVI is not ableo capture GPP variations at short time-scales because short-termuctuations in photosynthetic capacity are not reflected by varia-

e black dashed line represents the 1:1 line. Horizontal and vertical error bars (right)2–2008), respectively. (For interpretation of the references to colour in this figure

tions in canopy greenness over physiological timescales (Sims et al.,2006a). For example, low temperature can significantly and rapidlyreduce GPP whilst having little effect on canopy greenness (Wu

et al., 2010).

H. Shi et al. / Ecological Indicators 72 (2017) 153–164 161

Fig. 6. Spatial comparison of (A) the mean annual GPP from iEVI (GPPiEVI, g C m−2 yr−1) (2000–2013) with (B) the benchmark GPP (GPPMTE, g C m−2 yr−1) (1982–2008) andthe distribution of (C) the residual (g C m−2 yr−1) between GPPiEVI and GPPMTE within the 5–95% quantile.

Fig. 7. Latitudinal patterns (0.5◦ bands) of mean annual GPP by iEVI (GPPiEVI, 2000–2013) and the benchmark (GPPMTE, 1982–2008), respectively. The red shaded arearepresents the standard deviation of all GPPiEVI values of cells along the latitude. The grey shaded area represents the standard deviation of all GPPMTE values of cells alongthe latitude. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

162 H. Shi et al. / Ecological Indicators 72 (2017) 153–164

F each oa gressr

4G

4

evbSlsrssdabaiVsdnapGepenidaVEtoEatv

ig. 8. Strength (R2) and slopes of relationships between ln(GPPEC) and ln(iEVI) for

nnual maximum LAI in each biome. Black lines represent the segmented linear reelationship. The shaded area represents 95% confidence band.

.3. Environmental constraints on the ability of EVI to estimatePP

.3.1. LAI affects covariation (R2) of GPP with iEVIThe performance of EVI in estimating GPP is constrained by

nvironmental conditions, including features of both climate andegetation structure. The covariation between GPP and EVI is oftenetter in deciduous sites than evergreen sites (Rahman et al., 2005;ims et al., 2006b; Wu et al., 2010). Deciduous sites experience aarge range between maximal and minimal EVI (as a result of largeeasonal variation) and among sites the range is significantly cor-elated with mean summer rainfall (positive correlation) or meanummer VPD (negative correlation) (Sims et al., 2006b). Our resultshowed that the strength of the correlation between GPP and iEVI ineciduous biomes was generally better than in evergreen biomes,lthough the strength of the correlation in DBF was only slightlyetter than in EBF. Evergreen biomes show smaller seasonal vari-tion in EVI than deciduous biomes. Observable seasonal variationn vegetation greenness may be a prerequisite for successful use ofIs to estimate vegetation production. Deciduous biomes demon-trate distinct seasonal dynamics of leaf greenness, thus satelliteata can accurately capture these large seasonal changes in green-ess (Ma et al., 2013; Verma et al., 2014). In contrast, it is difficult tochieve the same level of accuracy within evergreen biomes. Thisresumably explains the poor performance of iEVI in estimatingPP of EBF, either in wet tropical areas (Figs. 5–7) or in semi-aridvergreen forests, where photosynthetic capacity can vary inde-endently of EVI and LAI in response to dry conditions (Maseykt al., 2008). However, either the seasonal change of EVI or LAI can-ot effectively explain the weak correlation between iEVI and GPP

n DBF. Nagai et al. (2010) found EVI to increase earlier than GPPuring the leaf-expansion period in DBF, and this caused system-tic variability in the GPP-EVI relationship (Richardson et al., 2012;erma et al., 2014). To address the asynchronicity between GPP andVI in DBF, a phenological scalar may be needed in GPP-iEVI equa-ions, as has been applied in the vegetation photosynthesis modelf Xiao et al. (2004b). This suggests that large seasonal variance of

VI does not necessarily imply a good correlation of EVI and GPPnd thus EVI variance is not appropriate to explain the covaria-ion of iEVI-GPP across biomes. Our result also showed that iEVIariance across biomes can be greatly divergent while R2 of iEVI-

f all biomes as a function of either mean annual precipitation (Prcp, inset) or meanion line. The dashed vertical line indicates the breakpoint of the segmented linear

GPP correlations can be close. For example, iEVI standard deviationsfor CRO and ENF are 0.44 and 0.94, whereas R2 are 0.51 and 0.49,respectively (Fig. 3). In contrast, peak LAI can be as a metric of thecomplexity of canopy structures of a biome and thus is appropriateto indicate the covariation strength of iEVI-GPP relationships. Aspeak LAI increases, the iEVI-GPP relationship is weakened (Fig. 8)by the increased structure complexity due to either small seasonalEVI variations in a biome such as EBF or the asynchronicity betweenGPP and EVI in a biome such as DBF.

Another possible factor contributing to the poor correlationbetween GPP and iEVI in vegetation with high LAI that are mostlocated in wet regions may be the extensive cloudy conditions thatreduce the quality of EVI retrievals (Nagai et al., 2010). In aridand semi-arid areas where cloud cover is minimal, precipitationis a controlling factor of vegetation phenology and productivity(Bradley et al., 2011; Cleverly et al., 2013; Huxman et al., 2004;Jolly and Running, 2004; Ma et al., 2013; Schwinning and Sala,2004). Moreover, peak LAI is typically limited by water availabil-ity in arid and semi-arid regions (Eamus and Prior, 2001; Sjöströmet al., 2011). Consequently, the correlation between the strength ofGPP-iEVI relationships with mean annual precipitation showed thesame trend as that for LAImax (Fig. 8 inset). The correlation betweenthe strength of GPP-iEVI relationships and LAImax may help identi-fying the regions where iEVI is most likely to be a good predictorof GPP. Globally, underestimation of GPP in some locations wascompensated by overestimation in other locations within the samebiome type (Figs. 5 and 6), with a consequential minimisation ofbiases on the estimation of GPP due to global patterns of LAI.

4.3.2. LAI scales the sensitivity (fitted slopes) of GPP to iEVILindroth et al. (2008) proposed that LAI is the principal scaling

parameter for GPP in northern deciduous and coniferous forests. Inbiomes with a relatively small peak LAI (e.g., less than 2.5 m2 m−2),the sensitivity of GPP to EVI increases with LAI (Sjöström et al.,2011), although there were too few arid vegetation classes in thestudy to identify a statistically significant trend in sensitivity acrosssmall values of LAI (less than 1.98 m2 m−2, Fig. 8). Conversely,

as vegetation become less water limited, the sensitivity of GPPto EVI tended to decrease across large values of LAI (larger than1.98 m2 m−2, Fig. 8), which to our knowledge has not yet been foundat a global scale. Degradation of the GPP-iEVI relationship at large

Indicat

Li(t

5

meoosIGsiswstodtiaoretE

A

cDRtbbiaFCFmTeFiLdCeB

A

i0

H. Shi et al. / Ecological

AI is due to (1) decreased sensitivity of variation in EVI to changesn canopy structure, including LAI and canopy type, of dense forestsGao et al., 2000) and (2) biased or decreased seasonality of varia-ions in EVI.

. Conclusions

We comprehensively evaluated the ability of MODIS EVI to esti-ate annual GPP across 12 land cover types based on GPP from

ddy covariance. iEVI does not require calculation of the durationf the growing season, which significantly simplifies the estimationf annual GPP by EVI at the global scale. Cross validations demon-trated the robustness of biome-specific ln(GPP)–ln(iEVI) models.n comparison to a global benchmarking dataset of mean annualPP, we showed that the performance of iEVI was consistent fromite to global scales. Compared to GPPMTE, GPPiEVI performed bettern croplands of high productivity but poorer mainly in forests. Thetrength of the GPP-iEVI relationships across biomes was correlatedith peak LAI, by which the slope was also scaled. These findings

uggest that vegetation structure is an important factor regulatinghe accuracy and sensitivity of EVI in estimating spatial patternsf annual GPP across multiple biomes. While LUE models, data-riven models and terrestrial biosphere models are usually difficulto parameterize or are limited by coarse resolution meteorolog-cal inputs, our study provides a promising and very convenientpproach to estimate global spatial patterns of GPP at either a finer coarse resolution. Nevertheless, the use of EVI in estimating GPPequires further study, especially in deciduous broadleaf forest andvergreen biomes. Our findings on impacts of vegetation struc-ure provide valuable information for such efforts in improvingVI-based models of GPP.

cknowledgements

This research was supported by an Australian Research Coun-il Discovery Early Career Research Award (project numberE120103022). X.T. was supported by one of National Basicesearch Program of China (grant number 2013CB733404) andhe National Natural Science Foundation of China (grant num-er 41101379). This work used eddy covariance data acquiredy the FLUXNET community and in particular by the follow-

ng networks: AmeriFlux (U.S. Department of Energy, Biologicalnd Environmental Research, Terrestrial Carbon Program (DE-G02-04ER63917 and DE-FG02-04ER63911)), AfriFlux, AsiaFlux,arboAfrica, CarboEuropeIP, CarboItaly, CarboMont, ChinaFlux,LUXNET Canada (supported by CFCAS, NSERC, BIOCAP, Environ-ent Canada, and NRCan), GreenGrass, KoFlux, LBA, NECC, OzFlux,

COS-Siberia, USCCC. We acknowledge the financial support to theddy covariance data harmonization provided by CarboEuropeIP,AO-GTOS-TCO, iLEAPS, Max Planck Institute for Biogeochem-stry, National Science Foundation, University of Tuscia, Universitéaval, Environment Canada and U.S. Department of Energy and theatabase development and technical support from Berkeley Waterenter, Lawrence Berkeley National Laboratory, Microsoft ResearchScience, Oak Ridge National Laboratory, University of Californiaerkeley, University of Virginia.

ppendix A. Supplementary data

Supplementary data associated with this article can be found,n the online version, at http://dx.doi.org/10.1016/j.ecolind.2016.8.022.

ors 72 (2017) 153–164 163

References

Alexandrov, T., 2009. A method of trend extraction using singular spectrumanalysis. RevStat 7, 1–22.

Asner, G.P., Scurlock, J.M., Hicke, J.A., 2003. Global synthesis of leaf area indexobservations: implications for ecological and remote sensing studies. GlobalEcol. Biogeogr. 12 (3), 191–205.

Aubinet, M., et al., 2000. Estimates of the annual net carbon and water exchange offorests: the EUROFLUX methodology. Adv. Ecol. Res. 30, 113–175.

Baldocchi, D., et al., 2001. FLUXNET: a new tool to study the temporal and spatialvariability of ecosystem-scale carbon dioxide, water vapor, and energy fluxdensities. Bull. Am. Meteorol. Soc. 82 (11), 2415–2434.

Baldocchi, D., 2008. ‘Breathing’of the terrestrial biosphere: lessons learned from aglobal network of carbon dioxide flux measurement systems. Aust. J. Bot. 56(1), 1–26.

Beer, C., et al., 2010. Terrestrial gross carbon dioxide uptake: global distributionand covariation with climate. Science 329 (5993), 834–838.

Bonan, G.B., et al., 2011. Improving canopy processes in the Community LandModel version 4 (CLM4) using global flux fields empirically inferred fromFLUXNET data. J. Geophys. Res.: Biogeosci. 116 (G2), 2005–2012.

Bradley, A.V., et al., 2011. Relationships between phenology, radiation andprecipitation in the Amazon region. Global Change Biol. 17 (6), 2245–2260.

Campos, G.E.P., et al., 2013. Ecosystem resilience despite large-scale alteredhydroclimatic conditions. Nature 494 (7437), 349–352.

Cescatti, A., et al., 2012. Intercomparison of MODIS albedo retrievals and in situmeasurements across the global FLUXNET network. Remote Sens. Environ.121, 323–334.

Ciais, P., et al., 2005. Europe-wide reduction in primary productivity caused by theheat and drought in 2003. Nature 437 (7058), 529–533.

Cleverly, J., et al., 2013. Dynamics of component carbon fluxes in a semi-arid Acaciawoodland, central Australia. J. Geophys. Res.: Biogeosci. 118, 1168–1185.

Eamus, D., Prior, L., 2001. Ecophysiology of Trees of Seasonally Dry Tropics:Comparisons Among Phenologies Advances in Ecological Research. AcademicPress, pp. 113–197.

Friedl, M.A., et al., 2010. MODIS collection 5 global land cover: algorithmrefinements and characterization of new datasets. Remote Sens. Environ. 114(1), 168–182.

Gaitán, J.J., et al., 2013. Evaluating the performance of multiple remote sensingindices to predict the spatial variability of ecosystem structure and functioningin Patagonian steppes. Ecol. Indic. 34, 181–191.

Gao, X., Huete, A.R., Ni, W., Miura, T., 2000. Optical–biophysical relationships ofvegetation spectra without background contamination. Remote Sens. Environ.74 (3), 609–620.

Gitelson, A.A., et al., 2006. Relationship between gross primary production andchlorophyll content in crops: implications for the synoptic monitoring ofvegetation productivity. J. Geophys. Res.: Atmos. 111 (D8), 1984–2012.

Goward, S.N., Huemmrich, K.F., 1992. Vegetation canopy PAR absorptance and thenormalized difference vegetation index: an assessment using the SAIL model.Remote Sens. Environ. 39 (2), 119–140.

Guanter, L., et al., 2014. Global and time-resolved monitoring of cropphotosynthesis with chlorophyll fluorescence. Proc. Natl. Acad. Sci. 111 (14),E1327–E1333.

Hassani, H., 2007. Singular spectrum analysis: methodology and comparison.Heinsch, F.A., et al., 2006. Evaluation of remote sensing based terrestrial

productivity from MODIS using regional tower eddy flux networkobservations. IEEE Trans. Geosci. Remote Sens. 44 (7), 1908–1925.

Hirota, M., Holmgren, M., Van Nes, E.H., Scheffer, M., 2011. Global resilience oftropical forest and savanna to critical transitions. Science 334 (6053), 232–235.

Hollinger, D., Richardson, A., 2005. Uncertainty in eddy covariance measurementsand its application to physiological models. Tree Physiol. 25 (7), 873–885.

Huete, A., Liu, H., Batchily, K., Van Leeuwen, W., 1997. A comparison of vegetationindices over a global set of TM images for EOS-MODIS. Remote Sens. Environ.59 (3), 440–451.

Huete, A., et al., 2002. Overview of the radiometric and biophysical performance ofthe MODIS vegetation indices. Remote Sens. Environ. 83 (1), 195–213.

Huxman, T.E., et al., 2004. Precipitation pulses and carbon fluxes in semiarid andarid ecosystems. Oecologia 141 (2), 254–268.

Inoue, Y., Penuelas, J., Miyata, A., Mano, M., 2008. Normalized difference spectralindices for estimating photosynthetic efficiency and capacity at a canopy scalederived from hyperspectral and CO2 flux measurements in rice. Remote Sens.Environ. 112 (1), 156–172.

Jolly, W.M., Running, S.W., 2004. Effects of precipitation and soil water potential ondrought deciduous phenology in the Kalahari. Global Change Biol. 10 (3),303–308.

Jung, M., et al., 2008. Diagnostic assessment of European gross primary production.Global Change Biol. 14 (10), 2349–2364.

Jung, M., Reichstein, M., Bondeau, A., 2009. Towards global empirical upscaling ofFLUXNET eddy covariance observations: validation of a model tree ensembleapproach using a biosphere model. Biogeosciences 6 (10).

Jung, M., et al., 2011. Global patterns of land-atmosphere fluxes of carbon dioxide,latent heat, and sensible heat derived from eddy covariance, satellite, and

meteorological observations. J. Geophys. Res.: Biogeosci. 116 (G3), 2005–2012.

Knyazikhin, Y., Martonchik, J., Myneni, R., Diner, D., Running, S., 1998. Synergisticalgorithm for estimating vegetation canopy leaf area index and fraction ofabsorbed photosynthetically active radiation from MODIS and MISR data. J.Geophys. Res.: Atmos. 103 (D24), 32257–32275, 1984–2012.

1 Indica

K

K

L

M

M

M

M

M

N

P

P

P

R

R

R

R

R

S

S

S

in a deciduous broadleaf forest using MODIS data and a radiative transfer

64 H. Shi et al. / Ecological

ondrashov, D., Ghil, M., 2006. Spatio-temporal filling of missing points ingeophysical data sets. Nonlinear Processes Geophys. 13 (2), 151–159.

utsch, W., et al., 2008. Response of carbon fluxes to water relations in a savannaecosystem in South Africa. Biogeosci. Discuss. 5 (3).

indroth, A., et al., 2008. Leaf area index is the principal scaling parameter for bothgross photosynthesis and ecosystem respiration of northern deciduous andconiferous forests. Tellus B 60 (2), 129–142.

a, X., et al., 2013. Spatial patterns and temporal dynamics in savanna vegetationphenology across the north Australian tropical transect. Remote Sens. Environ.139, 97–115.

ao, D., Wang, Z., Li, L., Ma, W., 2014. Spatiotemporal dynamics of grasslandaboveground net primary productivity and its association with climaticpattern and changes in Northern China. Ecol. Indic. 41, 40–48.

aseyk, K.S., et al., 2008. Physiology–phenology interactions in a productivesemi-arid pine forest. New Phytol. 178 (3), 603–616.

elillo, J.M., et al., 1993. Global climate change and terrestrial net primaryproduction. Nature 363 (6426), 234–240.

onteith, J., 1972. Solar radiation and productivity in tropical ecosystems. J. Appl.Ecol. 9 (3), 747–766.

agai, S., Saigusa, N., Muraoka, H., Nasahara, K.N., 2010. What makes thesatellite-based EVI–GPP relationship unclear in a deciduous broad-leavedforest? Ecol. Res. 25 (2), 359–365.

apale, D., Valentini, R., 2003. A new assessment of European forests carbonexchanges by eddy fluxes and artificial neural network spatialization. GlobalChange Biol. 9 (4), 525–535.

apale, D., et al., 2006. Towards a standardized processing of Net EcosystemExchange measured with eddy covariance technique: algorithms anduncertainty estimation. Biogeosciences 3, 571–583.

eng, Y., Gitelson, A.A., Keydan, G., Rundquist, D.C., Moses, W., 2011. Remoteestimation of gross primary production in maize and support for a newparadigm based on total crop chlorophyll content. Remote Sens. Environ. 115(4), 978–989.

ahman, A., Sims, D., Cordova, V., El-Masri, B., 2005. Potential of MODIS EVI andsurface temperature for directly estimating per-pixel ecosystem C fluxes.Geophys. Res. Lett. 32 (19), L19404.

eichstein, M., et al., 2005. On the separation of net ecosystem exchange intoassimilation and ecosystem respiration: review and improved algorithm.Global Change Biol. 11 (9), 1424–1439.

eichstein, M., et al., 2007. Determinants of terrestrial ecosystem carbon balanceinferred from European eddy covariance flux sites. Geophys. Res. Lett. 34 (1).

ichardson, A.D., et al., 2012. Terrestrial biosphere models need betterrepresentation of vegetation phenology: results from the North Americancarbon program site synthesis. Global Change Biol. 18 (2), 566–584.

unning, S.W., et al., 2004. A continuous satellite-derived measure of globalterrestrial primary production. Bioscience 54 (6), 547–560.

chwinning, S., Sala, O.E., 2004. Hierarchy of responses to resource pulses in aridand semi-arid ecosystems. Oecologia 141 (2), 211–220.

ims, D.A., et al., 2006a. Parallel adjustments in vegetation greenness and

ecosystem CO2 exchange in response to drought in a Southern Californiachaparral ecosystem. Remote Sens. Environ. 103 (3), 289–303.

ims, D.A., et al., 2006b. On the use of MODIS EVI to assess gross primaryproductivity of North American ecosystems. J. Geophys. Res.: Biogeosci. 111(G4), G04015.

tors 72 (2017) 153–164

Sims, D.A., et al., 2008. A new model of gross primary productivity for NorthAmerican ecosystems based solely on the enhanced vegetation index and landsurface temperature from MODIS. Remote Sens. Environ. 112 (4), 1633–1646.

Sjöström, M., et al., 2011. Exploring the potential of MODIS EVI for modeling grossprimary production across African ecosystems. Remote Sens. Environ. 115 (4),1081–1089.

Sjöström, M., et al., 2013. Evaluation of MODIS gross primary productivity forAfrica using eddy covariance data. Remote Sens. Environ. 131, 275–286.

Tan, B., et al., 2011. An enhanced TIMESAT algorithm for estimating vegetationphenology metrics from MODIS data. IEEE J. Sel. Topics Appl. Earth Observ.Remote Sens. 4 (2), 361–371.

Verma, M., et al., 2014. Remote sensing of annual terrestrial gross primaryproductivity from MODIS: an assessment using the FLUXNET La Thuile dataset. Biogeosciences 11 (8), 2185–2200.

Wang, K., Dickinson, R.E., 2012. A review of global terrestrial evapotranspiration:observation, modeling, climatology, and climatic variability. Rev. Geophys. 50(2), RG2005.

Wang, D., Liang, S., 2008. Singular spectrum analysis for filling gaps and reducinguncertainties of MODIS land products. Geoscience and Remote SensingSymposium, 2008. IGARSS 2008. IEEE International. IEEE, 558–561.

Wu, C., Munger, J.W., Niu, Z., Kuang, D., 2010. Comparison of multiple models forestimating gross primary production using MODIS and eddy covariance data inHarvard Forest. Remote Sens. Environ. 114 (12), 2925–2939.

Wu, C., Niu, Z., Gao, S., 2012. The potential of the satellite derived greenchlorophyll index for estimating midday light use efficiency in maize,coniferous forest and grassland. Ecol. Indic. 14 (1), 66–73.

Wylie, B.K., et al., 2003. Calibration of remotely sensed, coarse resolution NDVI toCO2 fluxes in a sagebrush–steppe ecosystem. Remote Sens. Environ. 85 (2),243–255.

Xiao, X., et al., 2004a. Satellite-based modeling of gross primary production in anevergreen needleleaf forest. Remote Sens. Environ. 89 (4), 519–534.

Xiao, X., et al., 2004b. Modeling gross primary production of temperate deciduousbroadleaf forest using satellite images and climate data. Remote Sens. Environ.91 (2), 256–270.

Xiao, X., Zhang, Q., Hollinger, D., Aber, J., Moore III, B., 2005. Modeling grossprimary production of an evergreen needleleaf forest using MODIS and climatedata. Ecol. Appl. 15 (3), 954–969.

Yang, F., et al., 2007. Developing a continental-scale measure of gross primaryproduction by combining MODIS and AmeriFlux data through support vectormachine approach. Remote Sens. Environ. 110 (1), 109–122.

Yuan, W., et al., 2010. Global estimates of evapotranspiration and gross primaryproduction based on MODIS and global meteorology data. Remote Sens.Environ. 114 (7), 1416–1431.

Zhang, X., Friedl, M.A., Schaaf, C.B., Strahler, A.H., 2004. Climate controls onvegetation phenological patterns in northern mid-and high latitudes inferredfrom MODIS data. Global Change Biol. 10 (7), 1133–1145.

Zhang, Q., et al., 2005. Estimating light absorption by chlorophyll, leaf and canopy

model. Remote Sens. Environ. 99 (3), 357–371.Zhao, M., Heinsch, F.A., Nemani, R.R., Running, S.W., 2005. Improvements of the

MODIS terrestrial gross and net primary production global data set. RemoteSens. Environ. 95 (2), 164–176.