Improved modeling of gross primary production from a better representation of photosynthetic components in vegetation canopy

Agricultural and Forest Meteorology 233 (2017) 222–234

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology journal homepage: www.elsevier.com/locate/agrformet

Improved modeling of gross primary production from a better representation of photosynthetic components in vegetation canopy Zhengjia Liu a,b , Chaoyang Wu b,∗ , Dailiang Peng c,∗ , Sisi Wang b , Alemu Gonsamo d , Bin Fang e , Wenping Yuan f a

State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, 100101, China Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, 100101, China c Key Laboratory of Digital Earth, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, 100101, China d Department of Geography and Program in Planning, University of Toronto, 100 St. George St., Toronto, ON, M5S 3G3, Canada e Department of Earth & Environmental Engineering, Columbia University, 500 W 120th St., New York, NY, 10027, USA f School of Atmospheric Sciences, Sun Yat-Sen University, Guangzhou, 519082, Guangdong, China b

a r t i c l e

i n f o

Article history: Received 8 April 2016 Received in revised form 14 November 2016 Accepted 4 December 2016 Keywords: Gross primary production Light use efficiency Scaled enhanced vegetation index ChinaFLUX

a b s t r a c t Non-photosynthetic components within the canopy (e.g., dry leaves and stem) contribute little to photosynthesis and therefore, remote sensing of gross primary production (GPP) could be improved by the removal of these components. A scaled enhanced vegetation index (EVI), which is usually regarded as a linear function of EVI, was found to have the strongest relationship with chlorophyll level fraction of absorbed photosynthetically active radiation (FPARchl) and can help improve GPP estimation in croplands compared to canopy level FPAR (FPARcanopy). However, the application of the FPARchl theory to other plant functional types (PFTs) and the underlying reasons remain largely unknown. In this study, based on standard MODIS algorithm we comprehensively assessed the performances of FPARcanopy, scaled EVI (FPARchl1), normalized difference vegetation index (NDVI), scaled NDVI (FPARchl2) and EVI as proxies of FPAR for estimating GPP at four forest and six non-forest sites (e.g., grasslands, croplands and wetlands) from ChinaFLUX, representing a wide range of ecosystems with different canopy structures and eco-climatic zones. Our results showed that the scaled EVI (FPARchl1) as FPAR effectively improved the accuracy of estimated GPP for the entire PFTs. FPARchl1 substantially improved forest GPP estimations with higher coefficient of determination (R2 ), lower root mean square error (RMSE) and lower bias. In comparison, for non-forest PFTs, the improvement in R2 between estimated GPP based on FPARchl1 (GPPchl1) and flux tower GPP was less evident than those between flux GPP and GPP estimations from FPARcanopy (GPPcanopy), FPARchl2, NDVI and EVI. The temperature and water attenuation scalars played important roles in reducing the difference of various GPP and indirectly reducing the impact of different FPARs on GPP in non-forest PFTs. Even so, FPARchl1 is an ecologically more meaningful parameter since FPARchl1 and flux tower GPP dropped to zero more synchronously in both forest and non-forest sites. In particular, we found that the improvement of GPPchl1 relative to GPPcanopy was positively correlated with the maximum leaf area index (LAI), implying the importance of site characteristic in regulating the strength of the improvement. This is encouraging for remote sensing of GPP for which vegetation parameter retrieval has often been found to be less successful at high LAI due to saturations in reflective and scattering domains. Our results demonstrate the significance of accurate and ecologically meaningful FPAR parameterization for improving our current capability in GPP modeling. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Accurately accounting of ecosystem level carbon cycle is a key issue in global climate change research (Keenan et al., 2012;

∗ Corresponding authors at: Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, 100101, China. E-mail addresses: [email protected] (C. Wu), [email protected] (D. Peng). http://dx.doi.org/10.1016/j.agrformet.2016.12.001 0168-1923/© 2016 Elsevier B.V. All rights reserved.

Richardson et al., 2012). Gross Primary Production (GPP) is defined as the total amount of carbon dioxide fixed by plants through the process of vegetation photosynthesis which is an important component of the terrestrial carbon cycle (Gitelson et al., 2006; Liu et al., 2014b; Running and Nemani, 1988; Running et al., 2004; Wu et al., 2009). In recent decades, improving ecosystem model parameterizations to reduce uncertainty of GPP estimation has become a critical focus for understanding of vegetation response to global cli-

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

mate change (Cheng et al., 2014; Liu et al., 2015; Yuan et al., 2015, 2007; Zhang et al., 2015). To this regard, the Eddy Covariance (EC) technique from global flux tower networks provides valuable data sources for validating and calibrating the parameters of ecosystem models to improve GPP estimation capability (Schimel et al., 2015; Wu et al., 2011). Satellite remote sensing data provides input data to calculate GPP in time and space for continuous monitoring of dynamic vegetation cover. Running and Nemani (1988) and Running et al. (2004) employed the Normalized Difference Vegetation Index (NDVI) as a proxy of Fraction of absorbed Photosynthetically Active Radiation (FPAR) to compute GPP as NDVI was found to be comparable to FPAR. A scaled NDVI was employed as FPAR in a considering conductance-limited and radiation-limited GPP model (Yebra et al., 2015). Moderate resolution Imaging Spectroradiometer (MODIS) standard GPP algorithm (MOD17) was directly forced by MODIS FPAR (MOD15A2) product derived from a radiative transfer model and a back-up algorithm considering the relationship between NDVI and FPAR when radiative transfer model fails (Knyazikhin et al., 1998; Myneni et al., 2002; Zhao et al., 2005). Huete et al. (1997) first developed the Enhanced Vegetation Index (EVI) using three spectral bands, blue (459–479 nm), red (620–670 nm) and near-infrared (941–876 nm) and demonstrated that EVI was more sensitive to high vegetation compared to NDVI. Therefore, in several light use efficiency (LUE) models, EVI was also employed as a proxy of FPAR, such as Vegetation Photosynthesis Model (VPM) (Xiao et al., 2004a; Xiao et al., 2004b), Greenness and Radiation (GR) model (Gitelson et al., 2006; Peng et al., 2013), and Vegetation Index (VIM) model (Wu et al., 2010). Scaled vegetation index (SVI, e.g. scaled EVI or scaled NDVI) is usually regarded as a linear function of vegetation index (VI), where SVI = a0 × VI + b0 (a0 is the scaling factor, and b0 is the y-intercept). Earlier study employed scaled EVI in Temperature and Greenness (TG) model based entirely on remote sensing data to calculate GPP estimates, and suggested that scaled EVI better captures GPP variations in their empirical model (Sims et al., 2008). A recent study indicated that the empirical linear regression between Absorbed Photosynthetically Active Radiation (APARcanopy) and GPP measured from flux towers did not pass through the zero intercept, which consequently limited the application of vegetation indices for GPP modeling (Zhang et al., 2015). These results consistently suggested that scaled vegetation indices effectively improved the performance of LUE models, and scaled EVI had the strongest capability for GPP estimation because it was more physiologically meaningful due to a better relationship with chlorophyll level FPAR (Zhang et al., 2015). However, such findings mainly focused on croplands that have less canopy heterogeneity and clumping, and therefore its capability is largely unknown for other plant functional types (PFTs, e.g. forests, grasslands and wetlands) with much complicated range of canopy structures and heterogeneity (Zhang et al., 2014a; Zhang et al., 2014b; Zhang et al., 2015). In order to investigate the feasibility of scaled EVI as FPAR for improving GPP estimates in non-crop ecosystems, we employed MOD17 algorithm and different FPAR products to compare simulated GPP accordingly. The MOD17 algorithm follows the Monteith’s equation, in that GPP can be calculated by computing the LUE and APAR, respectively (Monteith, 1972; Zhao et al., 2005). APAR is a product of Photosynthetically Active Radiation (PAR) and FPAR, the latter being approximated using vegetation indices. MOD17 product employs MOD15A2 FPAR which is regarded as a canopy level FPAR (FPARcanopy), comprised of both photosynthetic and non-photosynthetic components together (Cheng et al., 2014; Myneni et al., 2002; Zhang et al., 2014a). Previous study explored the relationship between FPARcanopy and FPAR at chlorophyll level in leaf (FPARchl, that is photosynthetic component of FPARcanopy): FPARcanopy = FPARleaf + FPARstem, and FPAR-

223

leaf = FPARchl + FPARdry matter + FPARbrown pigment (FPARleaf, FPARstem, FPARdry matter, FPARbrown pigment are FPAR of leaf, stem, dry matter in leaf, and brown pigment in leaf, respectively) (Zhang et al., 2012b, 2009, 2005). As only FPARchl contributes to plant photosynthesis, it is more reasonable to use FPARchl for calculating GPP estimates in theory. Therefore, in this study, we separately used five FPARs (FPARcanopy, scaled EVI as FPARchl1, scaled NDVI as FPARchl2, NDVI and EVI) to calculate GPP at ten flux tower sites (four forest sites and six non-forest sites) in ChinaFLUX network. The objectives are (1) to analyze whether using FPARchl (e.g. scaled EVI) can improve GPP modeling in forests, grasslands and wetlands, (2) to investigate the influences of PFTs on these improvements, and (3) to determine the underlying reasons (e.g., site characteristics) for such improvements. 2. Materials and methods 2.1. Sites description We investigated the predicted abilities of LUE-based GPP model with different derived FPARs inputs at 10 flux tower sites in ChinaFLUX network (Fu et al., 2006; Luo et al., 2011; Wang et al., 2006; Wen et al., 2006; Yan et al., 2008; Yu et al., 2006; Zhang et al., 2006). As shown in Fig. 1 and Table 1, four forest sites were comprised of a mixed forest (Changbaishan site (CBS)), an evergreen needleleaf forest (Qianyanzhou site (QYZ)) and two evergreen broadleaf forest sites (Xishangbanna site (XSBN) and Dinghushan site (DHS)); and six non-forest sites included, a wheat-maize rotation cropland (Yucheng site (YC)), four grassland sites (Xilinguolesite (XLGL), Haibei site (HB), Dangxiong site (DX) and Jilichangling site (JLCL)) and a wetland site (Chongmingdongtan1 site (CM1)). The selection of these sites was mainly based on the availability of carbon flux and micrometeorological observations. 2.2. Flux tower data processing The ChinaFLUX website provides carbon-water-energy fluxes and meteorological data at half-hourly scale (http://www.cerndata. ac.cn/). The data including downward shortwave radiation, PAR, air temperature, humidity, ecosystem respiration (Re) and net ecosystem exchange (NEE), were gap-filled and quality-controlled based on guidelines and earlier studies (Yu et al., 2008, 2006, 2013). Daily data were integrated based on the gap-filled half-hourly values. Daily gross primary production (GPP) is derived by subtracting Re from NEE as: GPP = NEE − Re. Then, flux derived GPP, GPP = − GEE, and flux derived GPP (GPPEC ) are presented in the unit of gC/m2 /day (Liu et al., 2015). Finally, 8-day integrated meteorological and carbon flux data from daily values were used for calculating GPP at the site scale. 2.3. MODIS data The MODIS collection 5 products, including 8-day MOD09A1 reflectance and MOD15A2 LAI/FPAR (hereafter, MOD15A2 FPAR is regarded as FPARcanopy and MOD15A2 LAI is regarded as LAI), were obtained from Land Processes Distributed Active Archive Center (LPDAAC, https://lpdaac.usgs.gov/). The retrieval of FPARcanopy is mainly based on a three-dimensional formulation of the radiative transfer process in vegetation canopy, and meanwhile a look-uptable method is used for inversion of three-dimensional radiative transfer problem (Knyazikhin et al., 1998; Myneni et al., 2002). The 8-day temporal resolution MOD09A1 gives seven reflectance spectral bands, red (620–670 nm), NIR1 (841–876 nm), blue (459–479 nm), green (545–565 nm), NIR2 (1230–1250 nm), SWIR1 (1628–1652 nm) and SWIR2 (2015–2155 nm). We com-

224

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

Fig. 1. Spatial distributions of 10 ChinaFLUX network sites overlaid on land cover map. ENF: evergreen needleleaf forest; EBF: evergreen broadleaf forest; MF: mixed forest; GL: grasslands; and CL: croplands.

Table 1 Brief descriptions of 10 flux tower sites in this study. Sites

Land cover

Lat. (◦ N)

Lon. (◦ E)

Alt. (m)

Data period

References

CBS QYZ DHS XSBN YC HB XLGL DX JLCL CM1

MF ENF EBF EBF CL GL GL GL GL WL

42.40 26.75 23.17 21.90 36.95 37.62 43.63 30.83 44.58 31.52

128.10 115.67 112.53 101.27 116.60 101.32 116.70 91.12 123.50 121.96

738 102 300 756 28 3300 1100 4250 160 0

2003–2005 2003–2005 2003–2005 2003–2005 2003–2005 2003–2005 2003–2004 2003–2004 2007–2008 2005

Zhang et al. (2006) Wen et al. (2006) Zhang et al. (2006) Yu et al. (2006) Wang et al. (2006) Fu et al. (2006) Fu et al. (2006) Yu et al. (2006) Luo et al. (2011) Yan et al. (2008)

MF: mixed forest; ENF: evergreen needleleaf forest;EBF: evergreen broadleaf forest; CL: croplands; GL: grasslands; WL: wetlands.

puted 500 m 8-day EVI and NDVI based on MOD09A1 reflectance data (Goward et al., 1985; Huete et al., 1997). The temporal profile of FPARcanopy, LAI, EVI and NDVI were smoothed using the modified Savitzky-Golay (mSG) filter (Chen et al., 2004). The mSG filter is a simple but robust method based on the Savitzky-Golay (SG) filter (Chen et al., 2004; Savitzky and Golay, 1964), which employs vegetation index time-series characterizations to smooth out the noise caused by clouds or poor atmospheric conditions. The maximum LAI (LAImax) was captured via the maximum value composites (Eidenshink and Faundeen, 1994), as follows, LAImax = max (LAIi , · · ·, LAIn )

(1)

LAImax represents the maximum LAI; the subscripts represent No. ith to nth 8-day LAI. For each site, LAImax is the maximum of this site throughout entire data period.

A recent study has indicated that scaled EVI and scaled NDVI possess stronger relationships with retrieved chlorophyll level FPAR (Zhang et al., 2015). But unstable linear relationships (dynamic scaling factor and y-intercept) between FPARchl and VIs (EVI and NDVI) in different land cover types may have limited the application potential of FPARchl for a larger spatial scale. Sims et al. (2008) have shown that GPP drops to 0 when the minimum value of EVI is around 0.1. In this study, the relationships between flux tower GPP and EVI in different land cover types also confirm the above assumption (Fig. 2). Thus, FPARchl1 was calculated as follows,



FPARchl1 =

a × (EVI − b) (EVI > b) 0 (EVI ≤ b)

(2)

where a is set as 1.0 and b is 0.1 as suggested by Sims et al. (2008). Previous study developed a relationship between FPAR and NDVI based on ground measurements from a set of vegetation types

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

225

2.5. MOD17 GPP algorithm The MOD17 algorithm is a LUE type model with the following equation (Zhao and Running, 2010): GPP = εmax × f (Tmin ) × f (VPD) × FPAR × SWrad × 0.45

f (Tmin ) =

⎧ ⎪ ⎪ ⎪ ⎨

(5)

0 (Tmin ≤ TMINmin ) Tmin − TMINmin

(TMINmin < Tmin < TMINmax ) ⎪ TMINmax − TMINmin ⎪ ⎪ ⎩ 1 (Tmin ≥ TMINmax )

. (6)

Fig. 2. The relationships between flux tower GPP (GPPEC ) and EVI for different land cover types.

f (VPD) =

(3)

where a and b are set to 1.24 and −0.168 (Sims et al., 2005), and the ratio of a with b indicates bare soil NDVI when FPAR is zero. 2.4. Regional data for model application FPARcanopy was commonly used for the retrieval of MODIS GPP product. To clarify the enhanced performance of estimated GPP based on the best FPAR compared to GPPcanopy, the regional land cover and meteorological data associated with spatial FPARs (the best FPAR and FPARcanopy) were employed to calculate spatial GPPs, respectively. These were further analyzed in the following part 3.3. The 1 km land cover data for year 2010 derived from Landsat TM/ETM+ and HJ-1 sensors was acquired from Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (http://www.resdc.cn/), which uses the International Geosphere-Biosphere Programme (IGBP) classification scheme. The results of field work and random sampling have shown that the land cover map is accurate to over 80% (Liu et al., 2014a; Zhang et al., 2014c). To show the potential of FPARchl at the regional scale, we selected the most recent year 2014 that has a complete dataset of ancillary data to simulate GPP. Regional remote sensing data, including MOD09A1 reflectance and MOD15A2 FPAR, were collected form LPDAAC similarly. Regional air temperature (including maximum and minimum), relative humidity and incoming shortwave radiation in 2014 were obtained from the National Center for Environmental Prediction (NCEP) Reanalysis II (http://www.esrl. noaa.gov/psd/). Daily NCEP data were bi-linearly interpolated down to the 1-km MODIS grid level (Zhang et al., 2012a). The vapor pressure deficit (VPD) was calculated based on Food and Agriculture Organization of the United Nations (FAO) method as follow:

 VPD = 1000 ×

1.0 −

RH 100

 × 0.6108 ×

exp

 17.27×Tmin  237.3+Tmin

+ exp 2

 17.27×Tmax  237.3+Tmax

(4)

where, the unit of VPD is Pa; RH is relative humidity with unit of %; Tmin and Tmax are daily minimum and maximum air temperature with unit of ◦ C, respectively.

0 (VPD ≥ VPDmax ) VPDmax − VPD

(VPDmin < VPD < VPDmax ) ⎪ VPDmax − VPDmin ⎪ ⎪ ⎩

(7)

1 (VPD ≤ VPDmin )

(Sims et al., 2005). The relationship eliminates the impact of bare soil on NDVI. In our study, the FPARchl2 follows this relationship. FPARchl2 = min(max (a × NDVI + b, 0) , 1)

⎧ ⎪ ⎪ ⎪ ⎨

where εmax presents the maximum LUE (␧); the attenuation scalars f (Tmin ) and f (VPD) are simple linear regression functions of daily Tmin and VPD using prescribed minimum and maximum environmental constrains based on biome properties look-up table; TMINmin is the daily Tmin at which ␧ = 0, TMINmax is the daily Tmin at which ␧ = εmax (for optimal VPD), VPDmax is the daylight average VPD at which ␧ = εmax (for optimal Tmin ), and VPDmin is the daylight average VPD at which ␧ = 0; FPAR is from MODIS productions; and SWrad is incoming solar shortwave radiation. The variables, Tmin, VPD and SWrad, can be obtained or derived from flux measurements in this study. Many studies have suggested that using site-specific or calibrated εmax instead of biome specific look-up table values can effectively enhance in-site GPP estimates (Cheng et al., 2014; Liu et al., 2014b; Zhang et al., 2015). In this study, the εmax values at canopy level and chlorophyll level in leaf were estimated based on Monte Carlo simulations. In brief, based on εmax values recommended by MOD17 look-up table as initial values, we first run 10000 Monte Carlo simulations to select 100 top-performance parameters and computed the minimum and maximum values of these 100 parameters. Then these calculated minimum and maximum values were input into Monte Carlo simulations again. New simulations finally output 10 top-performance parameters, and the mean and 95% confidence interval of the 10 top-performance parameters were regarded as the best-fit εmax value and it varied for each PFT. Based on MOD17 algorithm and εmax (Table 2), FPARcanopy, FPARchl1, FPARchl2, NDVI and EVI were employed as proxies of FPAR to calculate representative GPP (GPPcanopy, GPPchl1, GPPchl2, GPPndvi and GPPevi), respectively. 2.6. Statistical indicators Three metrics were used for evaluating the performance of GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI , including the average difference between simulations and observations (bias), coefficient of determination (R2 ) and root mean square error (RMSE). 1  × (Xsim i − Xobs i ) n n

Bias =

i=1

(8)

226

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

Table 2 Calibrated parameter values for four LUE models based on different FPARs. Parameter εmax (gC/MJ/day)

MF ENF EBF CL GL WL

for FPARcanopy

for FPARchl1

for FPARchl2

for NDVI

for EVI

1.25 ± 0.11 1.07 ± 0.15 0.66 ± 0.16 2.27 ± 0.18 0.89 ± 0.03 2.88 ± 0.36

2.31 ± 0.13 2.69 ± 0.24 2.14 ± 0.23 3.65 ± 0.23 1.51 ± 0.05 4.88 ± 0.56

1.22 ± 0.07 1.20 ± 0.10 0.41 ± 0.16 2.25 ± 0.16 0.81 ± 0.03 2.54 ± 0.41

1.27 ± 0.08 1.24 ± 0.11 0.50 ± 0.17 2.29 ± 0.16 0.81 ± 0.03 2.65 ± 0.43

1.93 ± 0.11 2.21 ± 0.28 1.78 ± 0.21 3.09 ± 0.29 1.20 ± 0.04 4.02 ± 0.47

MF: mixed forest; ENF: evergreen needleleaf forest;EBF: evergreen broadleaf forest; CL: croplands; GL: grasslands; WL: wetlan.



n

1  RMSE =  × (Xsim i − Xobs i )2

3.2. The seasonal dynamics between estimated and measured GPP (9)

n

i=1

⎛ ⎜ ⎝

R2 = ⎜ 

n  i=1

n  i=1

 Xsim i − Xsim

i

2 Xsim i − Xsim

i

×



⎞2

 Xobs i − Xobs

n  i=1

⎟ ⎟ 2 ⎠

i

Xobs i − Xobs

where, n is the total number of sample; Xsim

i

(10)

i

represents the ith

estimated GPPcanopy or GPPchl; Xobs i is the ith GPPEC ; Xsim i and Xobs i represent the average of Xsim i and Xobs i , respectively.

3. Results 3.1. The performances of estimated GPP We first compared the performance of estimated GPP (GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI ) against flux tower GPP (GPPEC ) (Fig. 3). Results showed that GPPchl1 had the best performance in all estimated GPPs for the overall data (R2 = 0.824, P < 0.001), followed by GPP from GPPEVI with R2 of 0.820 (P< 0.001) and GPPcanopy with R2 of 0.617 (P <0.001). However, GPPchl2 gave the lowest performance (R2 = 0.560, P < 0.001). For forest PFTs, R2 were lower than R2 of overall data. GPPchl1 also well matched GPPEC with the best performance (R2 = 0.727, P < 0.001). The R2 values of GPPcanopy, GPPchl2 and GPPNDVI were all below 0.5. For non-forest PFTs, GPP can be well estimated with all methods. For example, GPPchl1 had the best accuracy with R2 of 0.904 (P < 0.001), followed by GPPEVI with R2 of 0.903 (P < 0.001) and GPPchl2 with R2 of 0.889 (P < 0.001). The performance of GPPcanopy was the lowest with R2 of 0.878 (P < 0.001). Table 3 summarizes the statistics (R2 , RMSE and bias) for GPP estimates at each site. For evergreen broadleaf forest sites (DHS and XSBN), all GPP estimates showed low R2 , high RMSE and high bias. By contrast, the best performance was observed in GPPchl1. For evergreen needleleaf forest site (QYZ), GPPchl2 showed the highest R2 and lowest RMSE when compared to other four GPP estimates. GPPcanopy had the highest R2 and lowest RMSE in high LAI grassland sites (HB and JLCL). For other sites, GPPchl1 all showed the best or second-best performances. GPPNDVI showed better performance at evergreen needleleaf forest site (QYZ). GPPEVI had better performance at evergreen broadleaf forest site (DHS and XSBN), cropland site (YC) and wetland (CM1). Across all sites, estimated GPP based on scaled VI generally had lower bias. Overall, using scaled VI (e.g., NDVI, GPPchl2) improved GPP modeling than that based on the original NDVI (GPPNDVI ) at most sites, as expressed in terms of R2 and RMSE, and the improvement for non-forest sites was more evident than for forest sites.

Maximum values of GPP from FPARchl1 (i.e., GPPchl1) and EVI (i.e., GPPEVI ) were larger than other three estimated GPP values (Fig. 4), which matched well with GPPEC . For example, in evergreen broadleaf forest site (XSBN), GPPcanopy, GPPchl2 and GPPNDVI failed to reconstruct the seasonal dynamic of GPPEC , but GPPchl1 and GPPEVI matched well with GPPEC . The differences between estimated GPP (GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI ) and GPPEC in forests were principally caused by those of FPARs because both algorithms use the flux measured meteorological data. For non-forest sites, GPPcanopy, GPPchl2, GPPNDVI and GPPEVI were comparable with GPPchl1, but GPPchl1 still had a slightly better performance in terms of (a) capturing the peak values, and (b) simulating the impact of drought on GPP, such as in 2005 at XLGL site. In addition, we also found that GPPchl1, GPPEVI , FPARchl1 and EVI well captured the start of growing season in mixed forest, croplands and grasslands (Figs. 3 and 5), implying the advantage of EVI or scaled EVI for GPP modeling. By contrast, FPARcanopy, FPARchl2 and NDVI usually showed a longer growing season, higher values in nongrowing season, and early start of growing season (Fig. 5). Besides, scaled EVI (FPARchl1) better followed GPPEC than EVI, suggesting a better ecological meaning of FPARchl1.

3.3. The spatial difference of GPPchl1 and GPPcanopy The statistics suggested that estimated GPP based on the FPARchl1 had the strongest correlation with GPPEC with high R2 , low RMSE and low bias. Spatial difference between GPPchl1 and GPPcanopy was also investigated in the year of 2014 to better show the potential of FPARchl theory. As shown in Fig. 1, we analyzed the difference in estimated GPP in five typical land cover types including mixed forest, evergreen needleleaf forest, evergreen broadleaf forest, croplands and grasslands. Fig. 6 showed the annual and seasonal spatial differences of GPPchl1 and GPPcanopy (GPP = GPPchl1 − GPPcanopy), respectively. At the annual scale (Fig. 6a), the spatial GPP values were negative in mixed forest, evergreen needleleaf forest, grasslands and croplands, and positive in evergreen broadleaf forest. These were in good agreement with site-level results (Table 3). For mixed and evergreen needleleaf forest sites, GPPchl1 resulted in slightly larger bias than GPPcanopy. In cropland sites, as GPPcanopy overestimated GPPEC and GPPchl1 underestimated GPPEC , GPP were largely negative. For evergreen broadleaf forest sites, GPPchl1 had lower bias than GPPcanopy. The spatial variation of GPP was most pronounced in summer (Fig. 6c), mainly due to the higher summer photosynthesis compared to other seasons. The summer GPP of croplands lead to the highest negative difference and GPP of evergreen forests resulted in the highest positive bias and followed by spring, autumn and finally winter. The strongest negative and the strongest positive GPP were in mid-east region (mainly in Northern China Plain) and southern region, respectively. In brief, spatial GPPchl1 showed better performance in forest ecosystems. The spa-

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

227

Fig. 3. Relationships between GPPEC and estimated GPP (GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI ) for forest in blue color (in (a): y = 0.961x + 17.8; in (b): y = 1.015x + 11.4; in (c): y = 0.712x + 24.7; in (d): y = 0.768x + 22.1; and in (e): y = 1.017 x + 8.5) and non-forest in pink color (in (a): y = 0.971 x − 1.5; in (b): y = 1.036x + 0.9; in (c):y = 1.008 x − 0.2; in (d): y = 0.979 x − 1.8; and in (e): y = 1.003 x − 1.2). The dashed line indicates 1:1 line. The solid lines indicate regression lines. The symbol ** represents p < 0.001. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 3 Coefficients of determination (R2 ), root mean square error (RMSE, gC/m2 /8-day) and bias (gC/m2 /8-day) for estimated GPPs (GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI ) compared to flux tower GPP. Bold font indicates the best performance. GPPcanopy 2

CBS QYZ DHS XSBN YC HB XLGL DX JLCL CM1

GPPchl1 2

GPPchl2 2

GPPNDVI

GPPEVI

R

RMSE

Bias

R

RMSE

Bias

R

RMSE

Bias

R

RMSE

Bias

R2

RMSE

Bias

0.801 0.589 0.460 0.000 0.823 0.959 0.718 0.909 0.892 0.857

13.5 11.4 7.2 21.5 16.4 3.3 4.7 1.8 5.5 11.9

−5.4 −14.0 −13.3 −35.2 4.9 2.0 2.8 0.7 −1.1 0.8

0.894 0.782 0.608 0.509 0.874 0.958 0.856 0.910 0.864 0.876

9.8 8.3 6.2 15.1 13.8 3.4 3.4 1.8 6.2 11.1

−7.0 −14.4 −4.9 −21.1 −4.0 0.9 −0.5 0.7 −1.9 −5.8

0.889 0.802 0.322 0.033 0.855 0.958 0.768 0.866 0.861 0.783

10.1 7.9 8.1 21.2 14.8 3.4 4.3 2.1 6.3 14.6

−4.0 −6.8 −17.8 −46.4 0.4 1.4 0.4 0.0 −0.6 −4.1

0.880 0.795 0.380 0.039 0.848 0.940 0.707 0.880 0.876 0.781

10.5 8.0 7.7 21.1 15.1 4.1 4.8 2.0 5.9 14.7

−3.3 −6.0 −15.1 −43.2 4.2 1.9 2.2 1.3 0.5 1.9

0.895 0.777 0.602 0.457 0.864 0.950 0.756 0.912 0.892 0.873

9.8 7.8 6.9 15.9 14.0 3.7 3.7 1.4 6.7 11.0

−5.3 −10.7 −1.5 −18.7 1.4 1.8 1.8 2.0 −2.0 1.0

tial GPP distributions of growing season (i.e., summer, spring and autumn) were overall consistent with the distributions of the PFTs.

4. Discussion 4.1. The importance of optimized maximum light use efficiency The maximum light use efficiency (εmax ) value is one of the most important parameters in MOD17 algorithm. Many studies have reported and suggested that using site-specific or calibrated εmax values instead of biome specific look-up table values can effectively improve in-site GPP estimates (Cheng et al., 2014; Liu et al., 2014b; Zhang et al., 2015). For example, when replacing εmax value (0.68 gC/MJ/day) of cropland with optimized εmax (2.21 gC/MJ/day), Liu et al. (2014b) indicated that RMSE and bias both obviously decreased. In this study, we used optimization method to select the best-fit εmax values for different PFTs and different FPARs. Optimized εmax values for FPARcanopy are comparable with biome specific look-up table values recommended by MOD17 algorithm in mix forest and evergreen needleleaf forest. However, obvious

2

differences are observed for evergreen broadleaf forest, cropland and wetland, suggesting the importance of optimization for these PFTs. Previous study has also evidenced that optimized parameter performed better in these PFTs (Liu et al., 2014b). Besides, our results show that εmax values for different FPARs varied substantially, thus it is necessary to select different parameterizations for different FPARs for model applications.

4.2. The significance of temporal interpolation of vegetation index Most LUE models are derived by continuous remote sensing data which provide spatiotemporally consistent vegetation dynamic information (Dong et al., 2015; Nightingale et al., 2007; Peng et al., 2013; Wagle et al., 2014). Any noise or error existed in remote sensing data are therefore potentially transferred to estimated GPP. Previous studies suggested large uncertainties in FPARcanopy, and these would subsequently propagate higher uncertainty in GPP estimates compared to the input meteorological data (Liu et al., 2015; Sjöström et al., 2013). Therefore, the temporal interpolation methods are widely used to smooth time-series of vegetation index.

228

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

Fig. 4. Seasonal dynamics of estimated and measured gross primary production (GPP) for all flux tower sites, where a–d are forest sites, and e–j are non-forest sites.

Sjöström et al. (2013) indicated that the linear interpolation, which was used for filling unreliable or missing FPARcanopy based on the nearest reliable values prior and after it in MODIS GPP produced constant FPAR, and consequently failed to capture the green up observed in the field FPAR. In this study, the mSG filter was used for smoothing FPARcanopy and EVI time-series, which was regarded as a robust method for removing noise (Chen et al., 2004). Our results also showed that smoothed vegetation indices using mSG filter had much stronger relationships with GPPEC . Compared to previous studies (Liu et al., 2015, 2014b), R2 between FPARcanopy and GPPEC increased by 0.01-0.14, and R2 between EVI and GPPEC increased by 0.03–0.24 for different sites (Fig. 7). 4.3. The impact of different vegetation index on GPP estimates As shown in forest sites of Fig. 7, FPARchl1 had a better relationship with GPPEC than other four FPARs in most sites. For mixed forest (CBS), R2 increased from 0.76 for FPARcanopy to 0.92 for FPARchl1. In the evergreen needleleaf forest (QYZ), FPARchl2 had a higher correlation with GPPEC (R2 = 0.371) than NDVI (R2 = 0.337), which explains the better GPP estimates of GPPchl2. This potentially suggests that scaled NDVI may have the most potential to improve the performance of GPP in evergreen needle leaf forest. Besides, two stress factors of εmax reduce the impact of the differ-

ence of different FPARs on predicted GPP. For evergreen broadleaf forest sites, FPARcanopy and NDVI generally appeared to saturate at higher GPP values (Fig. 7), and therefore they cannot well reflect the seasonal dynamics of GPP. Meanwhile, scaling process could not improve the performance due to the saturate of NDVI. However, FPARchl1 is more sensitive to the seasonal dynamic at relatively high GPP values (Huete et al., 2002; Liu et al., 2015). Increased R2 values (from 0.04 of FPARchl2 to 0.47 of FPARchl1 at DHS and from 0.00 of FPARchl2 and NDVI to 0.59 of FPARchl1 at XSBN) indicate that using FPARchl1 greatly improves GPP modeling, but it cannot completely eliminate the bias. As shown in Fig. 3c and d that GPPchl1 better matched GPPEC than other GPPs, but there were still considerable bias between GPPchl1 and GPPEC . An earlier model evaluation also revealed that seven LUE models used in their study all showed poor performance for evergreen broadleaf forest (Yuan et al., 2014). Thus, further research on removing the effects of bias offsets is still of urgent need. Zhang et al. (2014a) compared the relationships between GPP estimates (based on FPARcanopy and derived FPARchl from canopy radiative transfer model) and GPPEC for maize and soybean, and indicated that the latter has a better performance in terms of GPP phase and amplitude. Their findings are confirmed in our results. The non-forest sites of Fig. 7 showed FPARchl1 presented slightly better correlations with GPPEC than FPARcanopy, FPARchl2, NDVI and EVI in most sites. A recent study indicates that scaled EVI largely

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

229

Fig. 5. Comparison of flux tower GPP (GPPEC ), FPARcanopy (MOD15A2 FPAR), FPARchl1 (scaled EVI), FPARchl2 (scaled NDVI), NDVI and EVI at (a) mixed forest site, (b) cropland site, and (c–f) four grassland sites. The green dashed lines represent the start of growing season at all sites without evergreen forest sites and wetland site. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

improves predict ability of GPP in croplands (Zhang et al., 2015), but the actual difference between GPPchl1 and GPPEVI is small in this study. Different methods used in two studies are responsible for the difference. In this study, two stress factors of εmax in MOD17 GPP algorithm reduce the impact of the difference of different FPARs on predicted GPP. As we knew, the f(Tmin) in MOD17 GPP algorithm plays an important role in non-growing season, and the f(VPD) in growing season, especially in the peak period. When taking Zhang’s study (2015) as an example, we can find that the main impact of the difference of EVI and scaled EVI on predicted GPP focuses on the bottom and top ends. When using LUE method without considering attenuation scalars of εmax (same method being used by Zhang et al. (2015)) to predict GPP, we can find the impact of the difference between FPARchl1 and EVI on predicted GPP (termed as GPPchl1-zh and GPPEVI-zh hereafter). For all data, GPPchl1-zh presented a stronger relationship with GPPEC (R2 = 0.76) than GPPEVI-zh (R2 = 0.7). Similar results can be found in forest (0.63 vs. 0.57), nonforest (0.89 vs. 0.88) ecosystems (Fig. 8) and each site (Fig. 9). These findings suggest that FPARchl1 as FPAR is ecologically more meaning relative to EVI. As all regression lines showed that when GPPEC dropped to zero, FPARchl1 was much closer to zero than other four FPARs in most forest and non-forest sites (Fig. 7). Besides, FPARchl2 overall performed better than FPARcanopy and NDVI. The large x-intercept values of FPARcanopy, NDVI and EVI potentially caused the high bias values in their GPP estimates, although their impacts were indirectly weakened by temperature and water attenuation scalars of MOD17 GPP algorithm when modeling GPP. In addition, our results indicate that R2 (the difference of R2 of FPARchl2-GPPEC and R2 of NDVI-GPPEC ) decreased as LAImax increased (y = −0.0181x + 0.098, R2 = 0.45, P < 0.05), suggesting that scaled NDVI would have a better performance for PFTs of lower LAI.

The spatial distribution of GPP corresponds with the pattern of land cover types. At the annual scale, GPP are positive in most forest regions and slight negative in non-forest regions. These are also in good agreement with site-level results (Figs. 3 and 4) because GPPchl1 usually gives less bias and RMSE than that of GPPcanopy. Furthermore, seasonal GPP also shows evident differences. The annual and seasonal GPP suggest that using FPARchl1 as a proxy of FPAR in LUE models will be helpful to improve the representation of spatial GPP, which is important for monitoring the effects of climate change on regional ecosystem carbon cycling. 4.4. The difference between FPARchl1 and other FPARs In theory, non-photosynthetic components within the canopy contribute little to plant photosynthesis. Only FPARchl contributes to photosynthesis and it should theoretically have the strongest correlation with GPPEC . In turn, the FPAR that has the best correlation with GPPEC should also have the most potential as a proxy of FPARchl. Our results suggest that the scaling process could potentially reduce or eliminate non-photosynthesis components and scaled VI has better performances than the original VIs. This is confirmed in our analysis that the estimated GPP based on scaled VI (especially FPARchl1) commonly matches better with GPPEC than GPPcanopy, GPPNDVI and GPPEVI . From our evaluation, FPARchl1 is the best proxy of FPARchl. A recent study also found that scaled EVI exhibited the best performance in GPP estimation and reconstruction of strong seasonal dynamics of GPP, especially in the spring and fall for agricultural sites and the underlying reason is that scaled EVI showed the strongest relationship with derived FPARchl based on canopy radiative transfer model (Cheng et al., 2014; Zhang et al., 2015). Possible causes for better modeling of GPP include that: (1) soil background

230

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

Fig. 6. The spatial pattern of GPP (the difference of GPPchl1 and GPPcanopy in 2014) at (a) annual, (b) spring, (c) summer, (d) autumn and (e) winter.

information has less impact on EVI or scaled EVI than NDVI or scaled NDVI (Huete et al., 1997); (2) theoretically, only FPARchl can play a role in vegetation photosynthesis. Compared with derived FPARchl from canopy radiative transfer model, scaled EVI is more convenient to simulate spatial and temporal patterns of GPP at large scales. In our results, differences of estimated GPP dynamics are primarily introduced by different FPARs. Theoretically, FPARchl and

non-photosynthetic FPAR together make up FPARcanopy. However, the non-photosynthetic parts, such as FPAR of background soil, dry leaves and stem can also influence the accuracy of GPP estimates when they are included in the model. For example, NDVI and scaled NDVI could potentially include FPAR of soil background and nonphotosynthetic components. Similarly, FPARcanopy could contain the FPAR of dry matter, brown pigment and stem.

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

231

Fig. 7. The relationships between FPARs (FPARcanopy (MOD15A2 FPAR), FPARchl1 (scaled EVI), FPARchl2 (scaled NDVI), NDVI and EVI) and flux tower GPP (GPPEC ) at all flux sites, where a–d are forest sites and e–j are non-forest sites.

Fig. 8. Relationships between GPPEC and predicted GPP (GPPchl1-zh and GPPEVI-zh ) for forest, non-forest and all data.

A further analysis reveals a positive correlation (R2 = 0.70, P < 0.001) between the difference of FPARcanopy-FPARchl1 and FPARcanopy for all data (Fig. 10). For different PFTs, the positive correlations still existed with R2 of 0.55 and 0.69 (both P < 0.001) for forests and non-forests, respectively. The findings suggest that contributions from the non-photosynthetic parts are increasing with canopy heterogeneity. This might also partly accounts for the difference between GPPcanopy and GPPEC in the peak values because the non-photosynthetic parts cannot be used for vegetation photosynthesis. At high FPARcanopy end, the difference of FPARcanopyFPARchl1 in forest is greater than that in non-forest ecosystems (Fig. 10). This is likely caused by the difference of canopy structure, since forest (including leaves, stem and branch) commonly possesses more complex canopy structure than non-forest PFTs (croplands, grasslands and wetlands). As shown in Fig. 11, the

improved model performance R2 (the difference between the R2 of GPPchl1-GPPEC and the R2 of GPPcanopy-GPPEC ) and RMSE (the difference between the RMSE of GPPchl1-GPPEC and the RMSE of GPPcanopy-GPPEC ) are highly correlated with the LAImax (R2 = 0.30, P < 0.1; R2 = 0.49, P < 0.05). It implies that the FPARchl1based model will give better performance for PFTs of high LAI since it removes more non-photosynthetic components within the canopy. This finding is also certified by Fig. 3 that GPPchl1 estimates exhibited much stronger relationships with GPPEC in forest sites. Negative differences of GPP from FPARcanopy-FPARchl1 are possibly caused by the following reasons. On one hand, the complex canopy structures in forests, especially in evergreen broadleaf forest, make FPARcanopy retrievals more challenging. Previous study reported the main algorithm of FPARcanopy based on radiative transfer model was sensitive to cloud contamination, thus it failed more often over 25◦ N-0◦ latitudes (Myneni et al., 2002). On the

232

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

Fig. 9. Relationships between GPPEC and predicted GPP (blue GPPchl1-zh and pink GPPEVI-zh ) at each site. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

other hand, the contaminated data interpolation method can effectively smooth the time-series of remote sensing vegetation index, but it still cannot remove all noises or errors, especially for cloud or atmospheric contaminations.

5. Conclusions

Fig. 10. Relationship between the difference of FPARcanopy-FPARchl1 (y-axis) and FPARcanopy (x-axis). The symbol ** represents P < 0.001. The black solid regression line for overall data. For forest (blue color), the regression line is y = 0.534x + 0.02 (R2 = 0.548, P < 0.001). For non-forest (pink color), the regression line is y = 0.372x + 0.06 (R2 = 0.685, P < 0.001). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Accurate estimates of GPP are extremely useful for quantifying terrestrial ecosystem carbon cycling and for understanding the impacts of intensifying global change on carbon cycling. In this study, we utilized MOD17 GPP algorithm, flux measured data and different FPARs (FPARcanopy, FPARchl1, FPARchl2, NDVI and EVI) to calculate GPP (GPPcanopy, GPPchl1, GPPchl2, GPPNDVI and GPPEVI ) at four forest sites and six non-forest sites, and found that GPPchl1 exhibited much stronger relationships with flux tower GPP (GPPEC ) compared with other four GPP. These are attributed to that FPARchl1 has a stronger relationship with GPPEC than other four FPARs and is a more ecologically meaningful parameter. Two important aspects account for ecological meaning of FPARchl1, (1) GPP changes in a more synchronously way with FPARchl1 than with other four FPARs; and (2) only PAR absorbed by chlorophyll was used for plant photosynthesis (Zhang et al., 2005). Although GPPchl1 only presented slightly better performance compared with GPPcanopy, GPPchl2, GPPNDVI and GPPEVI in non-forest ecosys-

Fig. 11. (a) Relationship between R2 (the difference between the R2 of GPPchl1-GPPEC and the R2 of GPPcanopy-GPPEC ) and the maximum of LAI (LAImax) and (b) Relationship between RMSE (the difference between the RMSE of GPPchl1-GPPEC and the RMSE of GPPcanopy-GPPEC ) and LAImax. The black dot and triangle representnon-forest site and forest site, respectively. The dash line represents the line of y = 0.

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

tems with lower LAI, the improvements were pronounced in forest ecosystems and were interestingly regulated by the site-level maximum LAI. These results suggest that using FPARchl1 based on scaled EVI to remove non-photosynthetic comments within the canopy has a potential to help LUE model in improving GPP estimates, especially in high LAI PFTs due to saturations of traditional FPARcanopy in general. Acknowledgements We are grateful for ChinaFLUX data and MODIS products. Constructive comments from all reviewers are also appreciated. This work was funded by National Natural Science Foundation of China (Grant No. 41371013, 41601582 and 41522109), China Postdoctoral Science Foundation (Grant No. 2016M590149), the Research Fund for International postdoc funding to A. Gonsamo (Grant No. 2015PE030) and Youth Innovation Promotion Association CAS (Dr. Peng D.). References Chen, J., Jönsson, P., Tamura, M., Gu, Z., Matsushita, B., Eklundh, L., 2004. A simple method for reconstructing a high-quality NDVI time-series data set based on the Savitzky–Golay filter. Remote Sens. Environ. 91 (3–4), 332–344. Cheng, Y., Zhang, Q., Lyapustin, A.I., Wang, Y., Middleton, E.M., 2014. Impacts of light use efficiency and fPAR parameterization on gross primary production modeling. Agric. For. Meteorol. 189–190, 187–197. Dong, J., Xiao, X., Wagle, P., Zhang, G., Zhou, Y., Jin, C., Torn, M.S., Meyers, T.P., Suyker, A.E., Wang, J., Yan, H., Biradar, C., Moore Iii, B., 2015. Comparison of four EVI-based models for estimating gross primary production of maize and soybean croplands and tallgrass prairie under severe drought. Remote Sens. Environ. 162, 154–168. Eidenshink, J.C., Faundeen, J.L., 1994. The 1 km AVHRR global land data set: first stages in implementation. Int. J. Remote Sens. 15 (17), 3443–3462. Fu, Y., Yu, G., Sun, X., Li, Y., Wen, X., Zhang, L., Li, Z., Zhao, L., Hao, Y., 2006. Depression of net ecosystem CO2 exchange in semi-arid Leymus chinensis steppe and alpine shrub. Agric. For. Meteorol. 137 (3–4), 234–244. ˜ A., Verma, S.B., Rundquist, D.C., Arkebauer, T.J., Keydan, G., Gitelson, A.A., Vina, Leavitt, B., Ciganda, V., Burba, G.G., Suyker, A.E., 2006. Relationship between gross primary production and chlorophyll content in crops: implications for the synoptic monitoring of vegetation productivity. J. Geophys. Res. 111 (D8), D08S11. Goward, S.N., Tucker, C.J., Dye, D.G., 1985. North American vegetation patterns observed with the NOAA-7 advanced very high resolution radiometer. Vegetatio 64 (1), 3–14. Huete, A.R., Liu, H.Q., Batchily, K., van Leeuwen, W., 1997. A comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote Sens. Environ. 59 (3), 440–451. Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sens. Environ. 83 (1–2), 195–213. Keenan, T.F., Baker, 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., Verbeeck, H., Zhao, M., Richardson, A.D., 2012. Terrestrial biosphere model performance for inter-annual variability of land-atmosphere CO2 exchange. Global Change Biol. 18 (6), 1971–1987. Knyazikhin, Y., Martonchik, J.V., Myneni, R.B., Diner, D.J., Running, S.W., 1998. Synergistic algorithm for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from MODIS and MISR data. J. Geophys. Res. 103 (D24), 32257–32275. Liu, J., Kuang, W., Zhang, Z., Xu, X., Qin, Y., Ning, J., Zhou, W., Zhang, S., Li, R., Yan, C., Wu, S., Shi, X., Jiang, N., Yu, D., Pan, X., Chi, W., 2014a. Spatiotemporal characteristics, patterns, and causes of land-use changes in China since the late 1980. J. Geogr. Sci. 24 (2), 195–210. Liu, Z., Wang, L., Wang, S., 2014b. Comparison of different GPP models in China using MODIS image and ChinaFLUX data. Remote Sens. 6 (10), 10215–10231. Liu, Z., Shao, Q., Liu, J., 2015. The performances of MODIS-GPP and -ET products in China and their sensitivity to input data (FPAR/LAI). Remote Sens. 7 (1), 135–152. Luo, L., Wang, Z., Mao, D., Lou, Y., Ren, C., Song, K., 2011. Remote-sensing estimation and validation of grassland net primary productivity of Western Songnen Plain in Northest China. Chin. J. Grassland 33 (6), 21–29. Monteith, J.L., 1972. Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 9 (3), 747–766. Myneni, R., Hoffman, S., Knyazikhin, Y., Privette, J., Glassy, J., Tian, Y., Wang, Y., Song, X., Zhang, Y., Smith, G., 2002. Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83 (1), 214–231. Nightingale, J.M., Coops, N.C., Waring, R.H., Hargrove, W.W., 2007. Comparison of MODIS gross primary production estimates for forests across the U.S.A. with

233

those generated by a simple process model, 3-PGS. Remote Sens. Environ. 109 (4), 500–509. Peng, Y., Gitelson, A.A., Sakamoto, T., 2013. Remote estimation of gross primary productivity in crops using MODIS 250 m data. Remote Sens. Environ. 128, 186–196. Richardson, A.D., Anderson, R.S., Arain, M.A., Barr, A.G., Bohrer, G., Chen, G., 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, A.K., Schaefer, K., Tian, H., Vargas, R., Verbeeck, H., Xiao, J., Xue, Y., 2012. Terrestrial biosphere models need better representation of vegetation phenology: results from the North American Carbon Program Site Synthesis. Global Change Biol. 18 (2), 566–584. Running, S.W., Nemani, R.R., 1988. Relating seasonal patterns of the AVHRR vegetation index to simulated photosynthesis and transpiration of forests in different climates. Remote Sens. Environ. 24 (2), 347–367. Running, S.W., Nemani, R.R., Heinsch, F.A., Zhao, M., Reeves, M., Hashimoto, H., 2004. A continuous satellite-derived measure of global terrestrial primary production. Bioscience 54 (6), 547–560. Savitzky, A., Golay, M.J., 1964. Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36 (8), 1627–1639. Schimel, D., Pavlick, R., Fisher, J.B., Asner, G.P., Saatchi, S., Townsend, P., Miller, C., Frankenberg, C., Hibbard, K., Cox, P., 2015. Observing terrestrial ecosystems and the carbon cycle from space. Global Change Biol. 21 (5), 1762–1776. Sims, D.A., Rahman, A.F., Cordova, V.D., Baldocchi, D.D., Flanagan, L.B., Goldstein, A.H., Hollinger, D.Y., Misson, L., Monson, R.K., Schmid, H.P., Wofsy, S.C., Xu, L., 2005. Midday values of gross CO2 flux and light use efficiency during satellite overpasses can be used to directly estimate eight-day mean flux. Agric. For. Meteorol. 131 (1–2), 1–12. Sims, D., Rahman, A., Cordova, V., Elmasri, B., Baldocchi, D., Bolstad, P., Flanagan, L., Goldstein, A., Hollinger, D., Misson, L., 2008. A new model of gross primary productivity for North American ecosystems based solely on the enhanced vegetation index and land surface temperature from MODIS. Remote Sens. Environ. 112 (4), 1633–1646. Sjöström, M., Zhao, M., Archibald, S., Arneth, A., Cappelaere, B., Falk, U., de Grandcourt, A., Hanan, N., Kergoat, L., Kutsch, W., Merbold, L., Mougin, E., Nickless, A., Nouvellon, Y., Scholes, R.J., Veenendaal, E.M., Ardö, J., 2013. Evaluation of MODIS gross primary productivity for Africa using eddy covariance data. Remote Sens. Environ. 131, 275–286. Wagle, P., Xiao, X., Torn, M.S., Cook, D.R., Matamala, R., Fischer, M.L., Jin, C., Dong, J., Biradar, C., 2014. Sensitivity of vegetation indices and gross primary production of tallgrass prairie to severe drought. Remote Sens. Environ. 152, 1–14. Wang, J., Yu, Q., Li, J., Li, L., Li, X., Yu, G., Sun, X., 2006. Simulation of diurnal variations of CO2, water and heat fluxes over winter wheat with a model coupled photosynthesis and transpiration. Agric. For. Meteorol. 137 (3–4), 194–219. Wen, X., Yu, G., Sun, X., Li, Q., Liu, Y., Zhang, L., Ren, C., Fu, Y., Li, Z., 2006. Soil moisture effect on the temperature dependence of ecosystem respiration in a subtropical Pinus plantation of southeastern China. Agric. For. Meteorol. 137 (3–4), 166–175. Wu, C., Niu, Z., Tang, Q., Huang, W., Rivard, B., Feng, J., 2009. Remote estimation of gross primary production in wheat using chlorophyll-related vegetation indices. Agric. For. Meteorol. 149 (6–7), 1015–1021. Wu, C., Munger, J.W., Niu, Z., Kuang, D., 2010. Comparison of multiple models for estimating gross primary production using MODIS and eddy covariance data in Harvard Forest. Remote Sens. Environ. 114 (12), 2925–2939. Wu, C., Chen, J.M., Huang, N., 2011. Predicting gross primary production from the enhanced vegetation index and photosynthetically active radiation: evaluation and calibration. Remote Sens. Environ. 115 (12), 3424–3435. Xiao, X., Hollinger, D., Aber, J., Goltz, M., Davidson, E.A., Zhang, Q., Moore Iii, B., 2004a. Satellite-based modeling of gross primary production in an evergreen needleleaf forest. Remote Sens. Environ. 89 (4), 519–534. Xiao, X., Zhang, Q., Braswell, B., Urbanski, S., Boles, S., Wofsy, S., Moore III, B., Ojima, D., 2004b. Modeling gross primary production of temperate deciduous broadleaf forest using satellite images and climate data. Remote Sens. Environ. 91 (2), 256–270. Yan, Y., Zhao, B.I.N., Chen, J., Guo, H., Gu, Y., Wu, Q., Li, B.O., 2008. Closing the carbon budget of estuarine wetlands with tower-based measurements and MODIS time series. Global Change Biol. 14 (10), 2469–2471. Yebra, M., Van Dijk, A.I.J.M., Leuning, R., Guerschman, J.P., 2015. Global vegetation gross primary production estimation using satellite-derived light-use efficiency and canopy conductance. Remote Sens. Environ. 163, 206–216. Yu, G., Wen, X., Sun, X., Tanner, B.D., Lee, X., Chen, J., 2006. Overview of ChinaFLUX and evaluation of its eddy covariance measurement. Agric. For. Meteorol. 137 (3), 125–137. Yu, G.-R., Zhang, L.-M., Sun, X.-M., Fu, Y.-L., Wen, X.-F., Wang, Q.-F., Li, S.-G., Ren, C.-Y., Song, X.I.A., Liu, Y.-F., Han, S.-J., Yan, J.-H., 2008. Environmental controls over carbon exchange of three forest ecosystems in eastern China. Global Change Biol. 14 (11), 2555–2571. Yu, G.R., Zhu, X.J., Fu, Y.L., He, H.L., Wang, Q.F., Wen, X.F., Li, X.R., Zhang, L.M., Zhang, L., Su, W., 2013. Spatial patterns and climate drivers of carbon fluxes in terrestrial ecosystems of China. Global Change Biol. 19 (3), 798–810. Yuan, W., Liu, S., Zhou, G., Zhou, G., Tieszen, L.L., Baldocchi, D., Bernhofer, C., Gholz, H., Goldstein, A.H., Goulden, M.L., Hollinger, D.Y., Hu, Y., Law, B.E., Stoy, P.C., Vesala, T., Wofsy, S.C., 2007. Deriving a light use efficiency model from eddy

234

Z. Liu et al. / Agricultural and Forest Meteorology 233 (2017) 222–234

covariance flux data for predicting daily gross primary production across biomes. Agric. For. Meteorol. 143 (3–4), 189–207. Yuan, W., Cai, W., Xia, J., Chen, J., Liu, S., Dong, W., Merbold, L., Law, B., Arain, A., Beringer, J., Bernhofer, C., Black, A., Blanken, P.D., Cescatti, A., Chen, Y., Francois, L., Gianelle, D., Janssens, I.A., Jung, M., Kato, T., Kiely, G., Liu, D., Marcolla, B., Montagnani, L., Raschi, A., Roupsard, O., Varlagin, A., Wohlfahrt, G., 2014. Global comparison of light use efficiency models for simulating terrestrial vegetation gross primary production based on the LaThuile database. Agric. For. Meteorol. 192–193, 108–120. Yuan, W., Cai, W., Nguy-Robertson, A.L., Fang, H., Suyker, A.E., Chen, Y., Dong, W., Liu, S., Zhang, H., 2015. Uncertainty in simulating gross primary production of cropland ecosystem from satellite-based models. Agric. For. Meteorol. 207, 48–57. Zhang, Q., Xiao, X., Braswell, B., Linder, E., Baret, F., Moore III, B., 2005. Estimating light absorption by chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a radiative transfer model. Remote Sens. Environ. 99 (3), 357–371. Zhang, L., Yu, G., Sun, X., Wen, X., Ren, C., Fu, Y., Li, Q., Li, Z., Liu, Y., Guan, D., Yan, J., 2006. Seasonal variations of ecosystem apparent quantum yield (␣) and maximum photosynthesis rate (Pmax) of different forest ecosystems in China. Agric. For. Meteorol. 137 (3–4), 176–187. Zhang, Q., Middleton, E.M., Margolis, H.A., Drolet, G.G., Barr, A.A., Black, T.A., 2009. Can a satellite-derived estimate of the fraction of PAR absorbed by chlorophyll (FAPARchl) improve predictions of light-use efficiency and ecosystem photosynthesis for a boreal aspen forest? Remote Sens. Environ. 113 (4), 880–888. Zhang, F., Chen, J.M., Chen, J., Gough, C.M., Martin, T.A., Dragoni, D., 2012a. Evaluating spatial and temporal patterns of MODIS GPP over the conterminous U.S. against flux measurements and a process model. Remote Sens. Environ. 124, 717–729.

Zhang, Q., Middleton, E.M., Gao, B., Cheng, Y., 2012b. Using EO-1 hyperion to simulate HyspIRI products for a coniferous forest: the fraction of PAR absorbed by chlorophyll (fAPARchl) and leaf water content (LWC). IEEE Trans. Geosci. Remote 50 (5), 1844–1852. Zhang, Q., Cheng, Y.-B., Lyapustin, A.I., Wang, Y., Gao, F., Suyker, A., Verma, S., Middleton, E.M., 2014a. Estimation of crop gross primary production (GPP): fAPARchl versus MOD15A2 FPAR. Remote Sens. Environ. 153, 1–6. Zhang, Q., Cheng, Y.-B., Lyapustin, A.I., Wang, Y., Xiao, X., Suyker, A., Verma, S., Tan, B., Middleton, E.M., 2014b. Estimation of crop gross primary production (GPP): I. Impact of MODIS observation footprint and impact of vegetation BRDF characteristics. Agric. For. Meteorol. 191, 51–63. Zhang, Z., Wang, X., Zhao, X., Liu, B., Yi, L., Zuo, L., Wen, Q., Liu, F., Xu, J., Hu, S., 2014c. A 2010 update of National Land Use/Cover Database of China at 1:100000 scale using medium spatial resolution satellite images. Remote Sens. Environ. 149, 142–154. Zhang, Q., Cheng, Y.-B., Lyapustin, A.I., Wang, Y., Zhang, X., Suyker, A., Verma, S., Shuai, Y., Middleton, E.M., 2015. Estimation of crop gross primary production (GPP): II. Do scaled MODIS vegetation indices improve performance? Agric. For. Meteorol. 200, 1–8. Zhao, M., Running, S.W., 2010. Drought-induced reduction in global terrestrial net primary production from 2000 through 2009. Science 329 (5994), 940–943. 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. Remote Sens. Environ. 95 (2), 164–176.