Evaluating the Performance of Sentinel-3A OLCI Land Products for Gross Primary Productivity Estimation Using AmeriFlux Data

: Accurate and reliable estimation of gross primary productivity (GPP) is of great signiﬁcance in monitoring global carbon cycles. The fraction of absorbed photosynthetically active radiation (FAPAR) and vegetation index products of the Moderate Resolution Imaging Spectroradiometer (MODIS) are currently the most widely used data in evaluating GPP. The launch of the Ocean and Land Colour Instrument (OLCI) onboard the Sentinel-3 satellite provides the FAPAR and the OLCI Terrestrial Chlorophyll Index (OTCI) products with higher temporal resolution and smoother spatial distribution than MODIS, having the potential to monitor terrain GPP. OTCI is one of the red-edge indices and is particularly sensitive to canopy chlorophyll content related to GPP. The purpose of the study is to evaluate the performance of OLCI FAPAR and OTCI for the estimation of GPP across seven biomes in 2017–2018. To this end, OLCI FAPAR and OTCI products in combination with insitu meteorological data were ﬁrst integrated into the MODIS GPP algorithm and in three OTCI-driven models to simulate GPP. The modeled GPP (GPP OLCI-FAPAR and GPP OTCI ) were then compared with ﬂux tower GPP (GPP EC ) for each site. Furthermore, the GPP OLCI-FAPAR and GPP derived from the MODIS FAPAR (GPP MODIS-FAPAR ) were compared. Results showed that the performance of GPP OLCI-FAPAR was varied in di ﬀ erent sites, with the highest R 2 of 0.76 and lowest R 2 of 0.45. The OTCI-driven models that include APAR data exhibited a signiﬁcant relationship with GPP EC for all sites, and models using only OTCI provided the most varied performance, with the relationship between GPP OTCI and GPP EC from strong to nonsigniﬁcant. Moreover, GPP OLCI-FAPAR (R 2 = 0.55) performed better than GPP MODIS-FAPAR (R 2 = 0.44) across all biomes. These results demonstrate the potential of OLCI FAPAR and OTCI products in GPP estimation, and they also provide the basis for their combination with the soon-to-launch Fluorescence Explorer satellite and their integration with the Sentinel-3 land surface temperature product into light use models for GPP monitoring at regional and global scales.


Introduction
Terrestrial gross primary production (GPP), which is the total amount of organic carbon fixed by green plants through photosynthesis at an ecosystem scale, determines the initial amount of energy and material entering the terrestrial ecosystem [1,2]. Terrestrial ecosystems can partly mitigate global warming and offset increasing CO 2 emissions through GPP [3]. Therefore, accurately quantifying GPP is essential for assessing global climate variation and carbon cycles [4]. performed better than the MODIS EVI products for estimating GPP across a diversity of ecosystems and climatic conditions. The MTCI product was discontinued in 2012, but the successful launch of the Ocean and Land Colour Instrument (OLCI) onboard the Sentinel-3 satellite provides the continuity of the MERIS.
As a legacy of the MERIS, the Sentinel-3 OLCI has recognized the pivotal deficiencies of MERIS and has undergone many enhancements [45]. The double satellite system of the OLCI instruments (i.e., Sentinel-3A and 3B) enables a revisit period less than two days, and 21 spectral bands exist with wavelengths ranging from optical to near-infrared in OLCI [46]. There are mainly two OLCI products levels, i.e., level-1B and level-2 (land and water), with a spatial resolution of approximately 300 m or 1.2 km. Specifically, the OLCI level-2 land products include the FAPAR and OTCI products, named the OLCI Global Vegetation Index (OGVI) and the OLCI Terrestrial Chlorophyll Index (OTCI), respectively. OGVI and OTCI are the continuity of the MERIS FAPAR (MERIS Global Vegetation Index (MGVI)) and the MTCI datasets, respectively [47]. The apparent difference of the OLCI products compared with MODIS and MERIS is the data composition algorithm. OGVI and OTCI used the actual value for the day while MOD15 and MTCI were composited using the arithmetic mean and maximum value across eight days, respectively [48]. The fine temporal and spectral resolution of OLCI products opened opportunities for GPP estimation. Nevertheless, to the best of our knowledge, only Zhang et al. [49] focused on GPP estimation using the OLCI FAPAR product so far. The main objective of their study was to investigate the relationships between a series of satellite FAPAR products (including OLCI FAPAR) and solar-induced chlorophyll fluorescence (SIF) products. Only the relationship between APAR and flux-tower based GPP was established, and neither LUE nor VI-driven models were used.
The main objective of this research is to assess the performance of the Sentinel-3 OLCI FAPAR and OTCI products in estimating GPP across different biome types. Thus, this study aims to (1) evaluate the GPP obtained from the OLCI FAPAR and three OTCI-driven models, (2) compare these obtained values of GPP with those derived from EC measurements, and (3) compare the performance between the OLCI FAPAR and MODIS FAPAR products.

Study Sites Description
Seven AmeriFlux eddy covariance flux towers including a variety of vegetation ecosystems types with latitudes ranging from 38 • to 45 • N and longitudes from −122 • to −89 • W were selected for this study ( Figure 1). These sites cover a range of climate conditions and land use classes, including two cropland (CRO) sites (C3: alfalfa and C4: corn, respectively), one grassland (GRA) site, one deciduous broadleaf forest (DBF) site, one closed shrubland (CSH) site, one open shrubland (OSH) site, and one woody savanna (WSA) site ( Table 1). The reason that we selected these sites was mainly in terms of the availability of satellite images, GPP measurements, and meteorological observations data. Detailed site descriptions and other related information can be obtained via the AmeriFluxwebsite (https://ameriflux.lbl.gov/sites/site-search/).

Flux and Meteorological Data
The fluxes and meteorological data at a half-hourly scale are provided by AmeriFlux website. The AmeriFlux network provides continuous observations of CO2, water, and energy fluxes at the ecosystem and landscape levels. It also provides standard datasets of climate and CO2 fluxes to the

Flux and Meteorological Data
The fluxes and meteorological data at a half-hourly scale are provided by AmeriFlux website. The AmeriFlux network provides continuous observations of CO 2 , water, and energy fluxes at the ecosystem and landscape levels. It also provides standard datasets of climate and CO 2 fluxes to the public, after various levels of data processing have been completed. Individualsite data were Remote Sens. 2020, 12, 1927 5 of 22 standardized and rigorously quality-filtered according to established standards within AmeriFlux [54]. The data of the seven flux towers, including GPP (g C m −2 d −1 ), vapor pressure deficit (VPD, hPa), air temperature (T a , • C), sensible heat (H, W m −2 ), downward shortwave radiation (W m −2 ), and latent heat (LE, W m −2 ), were used in our study. The daily GPP and downward shortwave radiation were calculated as the sum of the half-hourly GPP and downward shortwave radiation. The half-hourly VPD was averaged to obtain the daily VPD. The daily minimum T a was derived from the minimum half-hourly T a . Specifically, at the corn site, the GPP and the meteorological data only during the growing season were used to preclude the possible disturbance caused by satellite observations during the nongrowing season [55].

Sentinel-3 OLCI Land Products
There are three processing levels (i.e., level-0, level-1, and level-2) of OLCI with various data products. OLCI level-1 and 2 data products are freely available to the general public. Products are delivered in full resolution (FR, 300 m) and reduced resolution (RR, 1.2 km) for the same coverage area. The two OLCI level-2 land products of FAPAR and OTCI at FR were used in our study.
OLCI FAPAR is essential for studying the plant photosynthetic process. The OLCI FAPAR algorithm consists of two steps. First, thebidirectional reflectance factors are "rectified", that is, angular effects are removed. Then, the information from the blue channel is used to decontaminate the red and near-infrared bands from any atmospheric influence. Such an approach does not require any assumption on the ambient atmospheric properties [56]. Seasonality comparisons against ground-based measurements have shown that OLCI FAPAR effectively describesthe seasonal changes over various types of crops, and more ground-truth is under collection for validation purposes.
OTCI is an indicator of canopy chlorophyll content and is a continuation of the MTCI. The formula of OTCI is as follows: where Band 12 , Band 11 , and Band 10 are the reflectance in the band centered respectively at 753, 709, and 681 nm of the OLCI sensor. OTCI is easily calculated and strongly correlated with the red-edge position (REP). It is more sensitive to the high chlorophyll content compared to REP [41]. Information on canopy chlorophyll content is an important indication of plant photosynthetic capacity. OTCI ranges from 1 to 6.5. Further validation work of OTCI with additional in situ data is currently performed. All of the Sentinel-3 OLCI products in this study were downloaded from the ESA (https: //scihub.copernicus.eu). The data preprocessing operations include reprojection, resampling and subset; these batch processing operations were accomplished through the ENVI Modeler tool in ENVI 5.5.

MODIS FAPAR Product
Different series of MODIS FAPAR products have been widely used in modeling GPP. The latest version is Collection 6 (C6) MOD15A2H, which is composited at an eight-day interval with a spatial resolution of 500 m. The main retrieval method is based on the 3-D radiative transfer (RT) model setting canopy spectral properties for a given biome type [49]. A Look-up- Table (LUT), which is generated by 3-D RT, is used to retrieve FAPAR [57]. A back-up method based on the statistical relationships between NDVI and FAPAR is chosen as an alternative when the main RT retrieval method fails [58].

MODIS GPP Algorithm
In this study, we adopted the MODIS GPP algorithm combined with the Sentinel-3 OLCI FAPAR product to calculate GPP so that the result is comparable with the GPP derived from MODIS FAPAR. The MODIS GPP is a representative LUE model algorithm that calculates GPP by employing the amount of PAR absorbed by vegetation [59]. The algorithm is developed as follows: VPD < VPD min (5) where ε max represents the maximum LUE, TMIN min is the daily minimum temperature (T min ) at which ε = 0, TMIN max is the daily T min at which ε = ε max , VPD max is the daylight average VPD at which ε = ε max , VPD min is the daylight average VPD at which ε = 0, FAPAR is from the OLCI FAPAR product (OGVI), and SW rad is the incident solar shortwave radiation. Here, ε max , TMIN min , TMIN max , VPD max , and VPD min were determined according to the biome properties LUT for each biome types (Table 2), while T min , VPD, and SW rad were obtained from flux tower sites. Additionally, a 1.5 km × 1.5 km area centered on every flux tower site was extracted to represent the flux footprint, FAPAR, and OTCI [5].

Modeling of OTCI-Driven Models
Three types of OTCI-driven models in combination with different input variables were established. Model 1 used the OTCI alone to estimate the GPP directly (GPP = ƒ (OTCI)), according to the fact that GPP is directly related to chlorophyll content, which is also closely associated with OTCI. Therefore, OTCI is regarded as a proxy for LUE and APAR [60]. Model 2 combines OTCI and the incoming shortwave radiation that uses wavelength in the 400-700 nm region to estimate GPP (GPP = ƒ (OTCI × PAR)), according to Yoder and Waring [61], who assumed a direct linear relationship between GPP and the product of photosynthetic efficiency and incoming shortwave radiation over a period of the growing season. Therefore, OTCI related to canopy chlorophyll content represents the photosynthetic efficiency here. Model 3 (GPP = ƒ (OTCI × PAR × FAPAR)) introduces the actual incoming shortwave radiation by considering only absorbed PAR (APAR = PAR × FAPAR) to be functioning for photosynthesis. Therefore, this model theoretically overcomes the limitation of the previous two models and is expected to obtain the highest potential for GPP simulation.

Statistical Analysis
To make a comparisonwith the GPP OLCI-FAPAR , the eight-day period of the MODIS FAPAR product was first averaged to obtain the daily value for the FAPAR. Then, the daily GPP MODIS-FAPAR was calculated based on the site meteorological data and the MODIS GPP algorithm. The GPP MODIS-FAPAR values with dates corresponding to GPP OLCI-FAPAR were selected. Moreover, it is well known that GPP EC encompasses many uncertainties. In this study, random uncertainties in GPP EC were estimated following Richardson and Hollinger [62]. Uncertaintiesare expressed as a 95% confidence interval estimated from the 2.5th and 97.5th percentiles of the daily sum in g C m −2 d −1 . Three statistical metrics were used to evaluate the accuracies ofthe modeled GPP, including the coefficient of determination (R 2 ), root mean square error (RMSE), and bias. A higher R 2 and a lower RMSE and bias mean a better model performance. The RMSE and bias are calculated as follows: where M i and E i are the modeled and EC-measured GPP values, respectively.

Variability of T min , VPD, GPP EC , and Input Model Variables
The flux-tower based GPP varied among the study sites and land use classes. The CRO sites of US-Bi2 and US-Bi1 showed the highest average GPP EC : 12.69 ± 7.04 g C m −2 d −1 at US-Bi2 and 6.11 ± 4.51g C m −2 d −1 at US-Bi1. The GRA, DBF and WSA sites showed a relatively lower average GPP EC : 4.72 ± 1.88g C m −2 d −1 at US-KFS, 3.53 ± 4.80 g C m −2 d −1 at US-WCr, and 2.54 ± 1.88 g C m −2 d −1 at US-Ton. Two types of shrubland (i.e., CSH and OSH) sites, Rls and Rws, had the lowest average GPP EC of 1.68 ± 1.59 g C m −2 d −1 and 1.21 ± 0.99 g C m −2 d −1 , respectively. The temporal dynamics in GPP EC at the seven sites is shown in Figure 2. Apparently, the seasonal change patterns in the GPP EC of alfalfa and corn are different. The corn site (US-Bi2) is a single croppingsystem, whereas the alfalfa site (US-Bi1) is a multiple cropping system. At the US-Bi2 site, GPP EC rose rapidly and peaked in early August 2017 and late July 2018, respectively. At the US-Bi1 site, the harvest cycle of alfalfa is approximately a month. Alfalfa is harvested multiple times per growing season, depending on the level of rainfall and magnitude of production. Moreover, the maximum GPP EC of corn was found to be substantially larger than that of alfalfa. For example, in 2018, the highest GPP EC of corn was 23.75 ± 4.02 g C m −2 d −1 , whereas alfalfa was only 16.33 ± 3.46 g C m −2 d −1 . The OSH and CSH presented a similar temporal variation characteristic, reaching a peak in early May 2017 and the end of May 2018, respectively. The maximum GPP EC of the CSH site (US-Rls) was slightly higher than that of the OSH site (US-Rws) in 2017 and 2018. The DBF site (US-WCr) reached maximum GPP EC in late July 2017 and in late June 2018. The maximum GPP EC in 2017 (11.87 ± 3.02 g C m −2 d −1 ) was lower than in 2018 (18.77 ± 4.78 g C m −2 d −1 ). The GPP EC in the GRA site (US-KFS) exhibited a great variation, ranging from 0.02 ± 0.01g C m −2 d −1 to 21.36 ± 4.80 g C m −2 d −1 . The WSA site (US-Ton) reached maximum GPP EC earlier than US-WCr and US-KFS, and the maximum GPP EC was 11.42 ± 2.73 g C m −2 d −1 in 2017 and 9.13 ± 2.19 g C m −2 d −1 in 2018.The temporal dynamics of T min and VPD for all sites areshown in Figure 3. Overall, T min and VPD presented a similar variation trend. For T min , great variation was found at the two shrublands, i.e., the GRA and DBF sites. Amongst them, maximum T min can reach 20 • C while minimum T min was even below −15 • C. At the two croplands sites, the overall growing season T min was higher in 2017 than 2018. At the WSA site, the T min was approximately 0 • C and the maximum T min was higher in 2017 than 2018. In regard to VPD, the two Remote Sens. 2020, 12, 1927 8 of 22 shrublands and GRA sites were generally higher than other sites. Two croplands sites had a similar range during growing season. The maximum VPD was the lowest at the WSA site compared with other sites.
during the corn-growing season. The magnitude of PAR was similar for all sites, approximately ranging from 400 to 1400 hPa. Several minimal values of PAR at the US-Ton site may due to the changeable weather condition. Two scalar factors, ƒ(Tmin) and ƒ(VPD), showed anopposite temporal change trend at most sites, except for the US-Bi2 and US-Ton sites. At the US-Bi2 site, a great variation of ƒ(Tmin) was observed within days. At the US-Ton site, VPD was not responsible for the LUE restriction since the ƒ(VPD) values were equal to 1. At other sites, ƒ(Tmin) plays an important role in regulating LUE during the nongrowing season and ƒ(VPD) during the growing season. Obviously, the actual Tmin being higher and the actual VPD being lower than the model prescribed threshold lead to the ƒ(Tmin) and ƒ(VPD) values equaling to 1 during the growing and nongrowing season, respectively.   The temporal dynamics of environmental variables corresponding to the temporal resolution of OLCI products used in our study at the seven sites are shown in Figure 4. The FAPAR show similar seasonal patterns that are approximately consistent with the variation in GPP EC at all sites. At the two shrublands sites, the value of FAPAR in peak-growing season waslower than other sites on account of the low vegetation cover of shrublands. Similarly, PAR has strong seasonal variability at most sites, except the US-Bi2 site. The reason may be that the PAR data we collected were only during the corn-growing season. The magnitude of PAR was similar for all sites, approximately ranging from 400 to 1400 hPa. Several minimal values of PAR at the US-Ton site may due to the changeable weather condition. Two scalar factors, ƒ(T min ) and ƒ(VPD), showed anopposite temporal change trend at most sites, except for the US-Bi2 and US-Ton sites. At the US-Bi2 site, a great variation of ƒ(T min ) was observed within days. At the US-Ton site, VPD was not responsible for the LUE restriction since the ƒ(VPD) values were equal to 1. At other sites, ƒ(T min ) plays an important role in regulating LUE during the nongrowing season and ƒ(VPD) during the growing season. Obviously, the actual T min being higher and the actual VPD being lower than the model prescribed threshold lead to the ƒ(T min ) and ƒ(VPD) values equaling to 1 during the growing and nongrowing season, respectively.

Agreement between GPP OLCI-FAPAR and GPP EC
We first compared the performance of GPP OLCI-FAPAR against that of the GPP EC across all sites, including for 2017 and 2018 ( Figure 5). Results showed that the US-WCr site obtained the best performance (R 2 = 0.76), followed by the US-Ton site with R 2 = 0.65, the US-Rls site with R 2 = 0.64, and the US-Bi2 site with R 2 = 0.55. The R 2 values between the GPP OLCI-FAPAR and GPP EC at US-Bi1, US-Rws, and US-KFS were all below 0.5. In terms of the RMSE, the US-Bi2 site produced the maximum error (RMSE = 9.77 g C m −2 d −1 ), which is nearly twice or more than any other sites (Table 3). This situation was mainly caused by thesubstantial underestimation of GPP OLCI-FAPAR during the peak growth period of corn. This cause of underestimation is also applicable to alfalfa and grass sites. US-Rws had the lowest error with an RMSE of 1.23 g C m −2 d −1 ( Table 3). Part of the difference between the US-Rws site and other sites in terms of RMSE was due to the low mean GPP in CSH. Table 3. Coefficient of determination (R 2 ), root mean square error (RMSE, g C m −2 d −1 ) and bias (g C m −2 d −1 ) between GPP OLCI-FAPAR and GPP EC for seven sites.  Figure 6 illustrates the temporal variation of GPP OLCI-FAPAR and GPP EC at all sites. Overall, the GPP OLCI-FAPAR effectively matched the GPP EC and generally captured the seasonal variations consistent with the GPP EC at all sites. However, a remarkable difference remains between the GPP EC and GPP OLCI-FAPAR at several sites. Notably, the GPP EC values in most sites were underestimated by the GPP OLCI-FAPAR , except for US-Ton site. As shown in Table 3, the GPP OLCI-FAPAR tracked the GPP EC well at the US-Ton site, with the second lowest bias of 0.64 g C m −2 d −1 . Similar to RMSE, the US-Bi2 site had the highest bias (−8.44 g C m −2 d −1 ). Interestingly, the US-WCr site with a large RMSE also showed the lowest bias. The reason is that part of the bias is offset by the compensation of GPP underestimation in several months in 2017 with overestimation in 2018.   US-Rws, and US-KFS were all below 0.5. In terms of the RMSE, the US-Bi2 site produced the maximum error (RMSE = 9.77 g C m −2 d −1 ), which is nearly twice or more than any other sites (Table  3). This situation was mainly caused by thesubstantial underestimation of GPPOLCI-FAPAR during the peak growth period of corn. This cause of underestimation is also applicable to alfalfa and grass sites. US-Rws had the lowest error with an RMSE of 1.23 g C m −2 d −1 ( Table 3). Part of the difference between the US-Rws site and other sites in terms of RMSE was due to the low mean GPP in CSH.  Figure 6 illustrates the temporal variation of GPPOLCI-FAPAR and GPPEC at all sites. Overall, the GPPOLCI-FAPAR effectively matched the GPPECand generally captured the seasonal variations consistent with the GPPECat all sites. However, a remarkable difference remains between the GPPEC and GPPOLCI-FAPAR at several sites. Notably, the GPPEC values in most sites were underestimated by the GPPOLCI-FAPAR, except for US-Ton site. As shown in Table 3, the GPPOLCI-FAPAR tracked the GPPEC well at the US-Ton site, with the second lowest bias of 0.64 g C m −2 d −1 . Similar to RMSE, the US-Bi2 site had the highest bias (−8.44 g C m −2 d −1 ). Interestingly, the US-WCr site with a large RMSE also showed the lowest bias. The reason is that part of the bias is offset by the compensation of GPP underestimation in several months in 2017 with overestimation in 2018.

Agreement between GPPOTCI and GPPEC
The time series of GPPEC and GPPOTCIthat was derived from the three OTCI-driven models in 2017-2018 are presented in Figure 7. Overall, the day-to-day variation in the time series of GPP is evident, whether in GPPEC or GPPOTCI. The consistency in temporal dynamics between GPPEC and GPP from models 1, 2, and 3 were high at most sites. For instance, the growth process of alfalfa from growth to harvest to regrowth was tracked well by the GPP estimated from the OTCI-driven models. However, at the US-Rws site, the GPP obtained from model 3 matched well with the GPPEC while the GPP derived from both models 1 and 2 failed to capture the variation trend. This

Agreement between GPP OTCI and GPP EC
The time series of GPP EC and GPP OTCI that was derived from the three OTCI-driven models in 2017-2018 are presented in Figure 7. Overall, the day-to-day variation in the time series of GPP is evident, whether in GPP EC or GPP OTCI . The consistency in temporal dynamics between GPP EC and GPP from models 1, 2, and 3 were high at most sites. For instance, the growth process of alfalfa from growth to harvest to regrowth was tracked well by the GPP estimated from the OTCI-driven models. However, at the US-Rws site, the GPP obtained from model 3 matched well with the GPP EC while the GPP derived from both models 1 and 2 failed to capture the variation trend. This mismatch can be attributed to the fact that when GPP EC started to decline, the modeled GPP from models 1 and 2 continued to increase. The magnitude of the GPP derived from models 1, 2, and 3 varies considerably across sites. Model 2 gained the maximum value compared with models 1 and 3 among all sites in 2017 and 2018. It is worth noting that the GPP produced by model 3 was close to the GPP EC at the two shrubland sites (US-Rls and US-Rws). At the DBF site, the GPP values modeled at models 2 and 3 were highly consistent, whether in magnitude or in variation trend during the growth period. The reason is that the FAPAR inputted in model 3 is nearly equal to 1, which has a minimal effect on GPP estimation.
Remote Sens. 2020, 12, x FOR PEER REVIEW 13 of 24 modeled at models 2 and 3 were highly consistent, whether in magnitude or in variation trend during the growth period. The reason is that the FAPAR inputted in model 3 is nearly equal to 1, which has a minimal effect on GPP estimation. The scatter plots between the GPPOTCI based on three OTCI-driven models and the GPPEC for each site are shown in Figure 8. The performance of the three models varies at different sites, thereby demonstrating that the applicability of the three OTCI-driven models is dependent on the biome types. For instance, insignificant correlations of model 1 were obtained at the US-Rws site, whereas the strongest relationship was exhibited with the GPPEC at the US-WCr site. Even for the same type, their performance also differs due to disparate input parameters. For example, model 3 performed much better than model 1 and 2 at the US-Ton, US-Rls and US-Rws sites. Obviously, the inclusion of FAPAR into the model lead to this result. Additionally, strong relationships were established between GPPEC and all the OTCI-driven models at the US-Bi2, US-WCr, and US-KFS sites. At the Bi1 site, a strong to weak relationship was observed from model 3 to model 1. To sum up, model 3, which includes APAR data, provided a significant relationship with GPPEC at all sites, and model 1 is superior to the other two at several sites. The scatter plots between the GPP OTCI based on three OTCI-driven models and the GPP EC for each site are shown in Figure 8. The performance of the three models varies at different sites, thereby demonstrating that the applicability of the three OTCI-driven models is dependent on the biome types. For instance, insignificant correlations of model 1 were obtained at the US-Rws site, whereas the strongest relationship was exhibited with the GPP EC at the US-WCr site. Even for the same type, their performance also differs due to disparate input parameters. For example, model 3 performed much better than model 1 and 2 at the US-Ton, US-Rls and US-Rws sites. Obviously, the inclusion of FAPAR into the model lead to this result. Additionally, strong relationships were established between GPP EC and all the OTCI-driven models at the US-Bi2, US-WCr, and US-KFS sites. At the Bi1 site, a strong to weak relationship was observed from model 3 to model 1. To sum up, model 3, which includes APAR data, provided a significant relationship with GPP EC at all sites, and model 1 is superior to the other two at several sites.   Figure 9 describes the temporal dynamics of GPP EC , GPP MODIS-FAPAR , and GPP OLCI-FAPAR at the seven sites during 2017 and 2018. Obviously, both GPP MODIS-FAPAR and GPP OLCI-FAPAR followed a similar variation trend observed in GPP EC . Nevertheless, GPP MODIS-FAPAR and GPP OLCI-FAPAR , which overestimated or underestimated GPP EC to varying degrees, were also observed at several sites. For example, a substantial underestimation of GPP at two crop sites during peak growth period and a slight overestimation of GPP at the US-Ton site during growing season were clearly identified. As shown in Figure 10, the relationship between GPP MODIS-FAPAR and GPP EC was established. In addition, Tables 3 and 4 show a comparison of GPP EC , GPP MODIS-FAPAR , and GPP OLCI-FAPAR for each site across two years. As a result, GPP OLCI-FAPAR performed better than GPP MODIS-FAPAR at the US-Bi1, Us-WCr, US-KFS, and US-Ton sites, with a higher R 2 . With respect to RMSE, the GPP MODIS-FAPAR was superior to GPP OLCI-FAPAR for most sites, except for the Us-WCr and US-Ton sites. In terms of bias, GPP MODIS-FAPAR provided the best performance at most sites, except the US-Ton site. Moreover, the relationship of GPP OLCI-FAPAR and GPP MODIS-FAPAR with GPP EC in the combined sites was established ( Figure 11). Overall, the GPP OLCI-FAPAR and GPP MODIS-FAPAR produced acceptable GPP estimates, and if only R 2 is considered, the GPP OLCI-FAPAR (R 2 = 0.55) performed better than GPP MODIS-FAPAR (R 2 = 0.44) across all biomes, demonstrating the applicability of the OLCI FAPAR product for estimating vegetation GPP.

PerformanceAnalysis of GPPOLCI-FAPAR Using MODIS GPP Algorithm
The significant relationship between the GPPOLCI-FAPAR and GPPEC across all sites indicates the potential of Sentinel-3 OLCI FAPAR in estimating GPP. However,the performance of the model varies at different sites. The US-WCr site obtained the best performance across all sites. This finding isconsistent with previous studies [5,63]. The reasons can be summarized as follows. Firstly, the seasonal fractional vegetation cover and vegetation structure of DBF were obvious, which can be easily tracked by FAPAR [12]. Secondly, DBF grows in the middle and high latitudes with less cloud cover, and thus time series images with high quality can be obtained [30]. The performance of the US-Ton site was only worse than the US-WCr site. Many studies have reported that the savanna GPP estimation accuracy was lower than other vegetation types due to the misclassification of the savanna [32,64,65]. Unlike savanna, which has a forest canopy cover between 10%-30%, woody savanna's canopy cover is greater than 30%. Dense vegetation cover can minimize the impact from

Performance Analysis of GPP OLCI-FAPAR Using MODIS GPP Algorithm
The significant relationship between the GPP OLCI-FAPAR and GPP EC across all sites indicates the potential of Sentinel-3 OLCI FAPAR in estimating GPP. However, the performance of the model varies at different sites. The US-WCr site obtained the best performance across all sites. This finding isconsistent with previous studies [5,63]. The reasons can be summarized as follows. Firstly, the seasonal fractional vegetation cover and vegetation structure of DBF were obvious, which can be easily tracked by FAPAR [12]. Secondly, DBF grows in the middle and high latitudes with less cloud cover, and thus time series images with high quality can be obtained [30]. The performance of the US-Ton site was only worse than the US-WCr site. Many studies have reported that the savanna GPP estimation accuracy was lower than other vegetation types due to the misclassification of the savanna [32,64,65]. Unlike savanna, which has a forest canopy cover between 10%-30%, woody savanna's canopy cover is greater than 30%. Dense vegetation cover can minimize the impact from understory vegetation, which may improve the GPP estimation accuracy. The performance of two shrubland sites was relatively inferior than that of the US-Ton site in our study. Previous studies have indicated that the estimation accuracy of the GPP at shrubland sites varied with regions [36,66]. Temperate regions aregenerally superior than tropical regions. The growth of shrubland is mainly controlled by water ability [32]. As shown in Figure 3c,d, ƒ(VPD) plays a dominant role, especially during shrubland growing season. Moreover, due to the sparse vegetation at the shrubland region, the background soil conditions may have a great impact on GPP estimation [67]. The relationship between GPP OLCI-FAPAR and GPP EC was moderate at the US-Bi1 and US-Bi2 sites. Previous studies have demonstrated that GPP estimation accuracy varied with regions and crop types [68][69][70]. Our study also confirmed this conclusion. It is noteworthy that the RMSE was large at the two cropland sites (Table 3). There are mainly two reasons for this situation. Firstly, the rotation period is different for the two crop types. The US-Bi1 site is a single croppingsystem whereas the US-Bi1 site is a multiple cropping system. However, the threshold of the environmental variables in the MODIS GPP algorithm was set the same for different crop types. Secondly, the algorithm assumed one value of ε max for all C3 and C4 crops, which led to the underestimation of GPP. The GPP estimation accuracy was lowest at the US-KFS site. This was not consistent with other studies, which obtained great performance for grassland [71,72]. The seasonal changes of grassland is obvious and can be tracked well by satellite images [73]. Thus, the flux and climate data uncertainty may lead to the discrepancy between GPP OLCI-FAPAR and GPP EC .

Performance of OTCI-Driven Models in GPP Estimation
The R 2 between GPP OTCI and GPP EC at seven sites showed that at least one of the three OTCI-driven models agreed well with the GPP. The US-Wcr site performed best, whether the model used OTCI only or introduced either PAR or FAPAR. This result was also consistent with that in Section 3.2 using the OLCI FAPAR product. Interestingly, the inclusion ofradiation information (PAR and FAPAR) did much to enhance model performance at the savannas and the two shrubland sites (Figure 4c,d,g). The probable reason is that the impact of soil background on the low vegetation coverage at these sites leads to failure of the OTCI in tracking GPP. By contrast, the addition of PAR and FAPAR appears to have little contribution to the GPP estimation at the US-Bi2, US-WCr, and US-KFS sites (Figure 4b,e,f), as OTCI can directly capture the pronounced seasonal variation of GPP at these sites [17,74].

Difference between OLCI FAPAR and MODIS FAPAR Product
As described in Section 3.4, the performance of GPP OLCI-FAPAR and GPP MODIS-FAPAR varies at different sites. The main difference between GPP OLCI-FAPAR and GPP MODIS-FAPAR results from the FAPAR product being incorporated in the model. Therefore, the differences between OLCI FAPAR and MODIS FAPAR, which may lead to inconsistency between GPP OLCI-FAPAR and GPP MODIS-FAPAR , are summarized mainly from two aspects. First, the temporal compositing periods of the two products are different. The MODIS FAPAR product is provided in aneight-day interval while the OLCI FAPAR is at a daily interval. For temporal aggregation purposes, MODIS FAPAR is resampled to daily, but its value does not change throughout the eight days.The maximum FAPAR is chosen as the final output value representing the eight days [35]. Second, the MODIS FAPAR and OLCI FAPAR are retrieved from different algorithms. The OLCI FAPAR retrieval algorithm belongs to the family of the JRC (Joint Research Center), which is based on the 1-D RT model, while MODIS FAPAR is derived from the 3-D RT model. The difference between these algorithms is mainly derived from the various definitions and the leaf/wood spectral values assumptions [49].

Other Sources of Uncertainty, Limitation, and Future Prospects
Additional sources of uncertainty, limitation, and further prospects in simulating GPP are illuminated in the study. First, the quality of the instrument, the randomness of turbulence, the partitioning methods, and gap-filling techniques leading to the random and systematic errors in the flux-derived GPP data is inevitable [75,76]. Second, the mismatches in scale between Sentinel-3 pixels and tower flux footprints lead to further uncertainties. The limitations exist, such as the slow update of the latest EC data and less reliable Sentinel-3 data during the December-April period due to cloudiness contamination. In the future, improvements in the OLCI products through the use of more robust gap-filling methods and in combination with Sentinel-3B data (only Sentinel-3A data wereused in our study to match the flux tower data) may provide even better GPP estimates [28,77].
Although our results showed that GPP OLCI-FAPAR did not always perform better than GPP MODIS-FAPAR , it needs to be emphasized that the objective of our study is not to distinguish which FAPAR product is superior in GPP estimation. Our study demonstrated the potential of OLCI FAPAR and OTCI products for GPP simulation and providedmore complementary solutions to estimate GPP. For example, in terms of data fusion, the OLCI FAPAR product with 10m spatial resolution and nearly daily temporal resolution can be generated by blending Sentinel-2 images and the OLCI FAPAR product using downscaling method and spatiotemporal fusion approaches [45,78]. Then, the same spatial and temporal resolution of the GPP product can be obtained by assimilating FAPAR into LUE models. Moreover, with regard toSIF, the Fluorescence Explorer satellite, which is ESA's eighth Earth Explorer and is proposed as a tandem with Sentinel-3, is expected to be launched by 2022, and it will provide an integrated package of measurements as well as all the necessary auxiliary information to improve GPP assessment together with Sentinel-3 OLCI and the Sea and Land Surface Temperature Radiometer (SLSTR) [79]. Additionally, this study demonstrated that the MODIS GPP algorithm is also suitable for OLCI FAPAR besides MODIS FAPAR. In years to come, attempts in integrating OLCI FAPAR and the SLSTR land surface temperature (LST) product into other LUE models (e.g., CASA and VPM) to estimate GPP at a large scale are also essential.

Conclusions
In this study, we evaluated the performance of two Sentinel-3A OLCI products (i.e., FAPAR and OTCI) in estimating the GPP across seven biomes in 2017-2018. OLCI FAPAR and OTCI products in combination with meteorological data were integrated into the MODIS GPP algorithm and three OTCI-driven models respectively to simulate GPP. GPP OLCI-FAPAR and GPP OTCI were compared with the GPP EC for each site, and a comparison between GPP OLCI-FAPAR and GPP MODIS-FAPAR was established. The main conclusions are summarized as follows: (1) The relationship between GPP OLCI-FAPAR and GPP EC is significant across all sites. The GPP OLCI-FAPAR correlated best with GPP EC at the US-WCr site (R 2 = 0.76) while it performed worst at the US-KFS site (R 2 = 0.45).
(2) The inclusion of APAR data in OTCI-driven models exhibited significant relationship with GPP EC for all sites. Nonetheless, the model using only OTCI provided the most varied performance, with the relationship between GPP OTCI and GPP EC from strong to nonsignificant.
(3) The comparison of GPP OLCI-FAPAR and GPP MODIS-FAPAR suggested that they all produced reasonable GPP estimates. In terms of R 2 , GPP OLCI-FAPAR performed better than the GPP MODIS-FAPAR across all biomes. In the aspect of RMSE and bias, GPP MODIS-FAPAR was superior than GPP OLCI-FAPAR . The main difference between GPP OLCI-FAPAR and GPP MODIS-FAPAR result from the inherent retrieval algorithm and assumptions in the FAPAR product.
In conclusion, the results of this study demonstrate the potential of OLCI FAPAR and OTCI products in GPP estimation, and these results open the possibility of applying them to GPP monitoring at regional and global scales for future studies.