A global moderate resolution dataset of gross primary production of vegetation for 2000–2016

Accurate estimation of the gross primary production (GPP) of terrestrial vegetation is vital for understanding the global carbon cycle and predicting future climate change. Multiple GPP products are currently available based on different methods, but their performances vary substantially when validated against GPP estimates from eddy covariance data. This paper provides a new GPP dataset at moderate spatial (500 m) and temporal (8-day) resolutions over the entire globe for 2000–2016. This GPP dataset is based on an improved light use efficiency theory and is driven by satellite data from MODIS and climate data from NCEP Reanalysis II. It also employs a state-of-the-art vegetation index (VI) gap-filling and smoothing algorithm and a separate treatment for C3/C4 photosynthesis pathways. All these improvements aim to solve several critical problems existing in current GPP products. With a satisfactory performance when validated against in situ GPP estimates, this dataset offers an alternative GPP estimate for regional to global carbon cycle studies.


Background & Summary
Vegetation photosynthesis at the ecosystem scale, also known as the gross primary productivity (GPP), is the first step of CO 2 entering the biosphere from the atmosphere. Over the past century, with the increasing carbon release from landcover change and fossil fuel burning, CO 2 accumulation rate in land, ocean and atmosphere has continuously increased 1 . The increase of CO 2 in Earth's atmosphere is the major cause of global climate change 2 . Compared to the ocean sink, the global land sink exhibits much higher inter-annual variability and is also greatly affected by global climate change 1,3,4 . A major contribution of this high variability comes from GPP, as the photosynthesis process is vulnerable to droughts, heatwaves, floods, frost and other types of disturbances [5][6][7][8] . An accurate estimation of GPP will not only provide information about the ecosystem response to these extreme events, but also help to predict the future carbon cycle dynamics.
Multiple approaches to estimating GPP at a global scale have been developed over the past decades. Based on the fundamental theories used, they can be divided into four categories: the enzyme kinetic (process-based) models 9,10 , the light use efficiency (LUE) or production efficiency models 11,12 , machine learning techniques based on eddy covariance (EC) measurements 13,14 , and statistical models using solarinduced chlorophyll fluorescence (SIF) 15 . Amongst all these types of models, LUE models have been widely used because of its simple form and relatively long period of data availability. More importantly, it offers a balance between the temporal (usually 8-days) and spatial (usually up to 500 m or 250 m) resolution, which is suitable for regional to global scale studies. Most LUE models share a common form as follows 16 : where the APAR refers to the absorbed photosynthetically active radiation, and the LUE is an empirical factor that converts this absorbed energy to carbon fixed through photosynthesis process 17 . The most widely used LUE GPP product is the MODIS GPP product 12 (hereafter MOD17, GPP MOD17 ). However, site level validation studies have shown its moderate to low performance in capturing the seasonal and inter-annual variations 18 . Cross-site synthesis studies also reported large scale underestimates in GPP MOD17 , especially for croplands 19,20 . Although improvements have been made in recent updates of MOD17 (C55 and C6), problems still exist, most related to the parameterization of the biome-specific lookup table and the climate drivers 21,22 . MOD17 assumes that the LUE remains constant for individual biome types, and the APAR is the energy absorbed by the entire canopy 12 . However, recent studies have shown that the photosynthetic capacity is not only affected by leaf quantity (amount of leaf area or leaf area index, LAI), but also by leaf quality (the photosynthetic rate of each individual leaf) 23,24 . The leaf quality is mostly related to the leaf chlorophyll content and the leaf nitrogen content 25,26 . There is not only a spatial variation of leaf quality within each biome due to the environment and nutrient availability 24 , but also a seasonal variation caused by the leaf's phenological regulation 27 . MOD17 does not take leaf quality into consideration, and therefore exhibits a relatively low spatial and temporal representation of GPP when compared with in situ GPP estimations. Many recent studies suggest that the fraction of PAR absorbed by chlorophyll (fPAR chl ) can better capture the seasonal variation of vegetation photosynthetic capacity and greatly improve the seasonal representation of GPP 28,29 . The vegetation photosynthesis model (VPM), developed based on this theory, showed superior performance with regards to site-level validations across a wide range of biome types [30][31][32][33] . A recent study also showed satisfactory performance (overall R 2 = 0.82) of regional VPM GPP simulation in North America 34,35 and the potential to apply this model to the entire globe.
The objective of this study is to develop a new global GPP dataset that overcomes the limitation of MOD17. This dataset, dubbed VPM GPP V20, is based on an improved LUE theory which uses the energy absorbed by chlorophyll, and implemented to the entire globe during the MODIS era. The VPM algorithm utilized remotely sensed datasets together with reanalysis climate dataset and land cover classification. The schematic workflow of the VPM algorithm is shown in Fig. 1 and is described in detail in the Methods section. This dataset can serve as an alternative to MOD17 GPP product for regional and global carbon cycle studies.

Data processing
We used multiple datasets as the input of VPM. All datasets being used are summarized in Table 1, and detailed processing methods are described below.

Vegetation indices and gap-filling
Enhanced Vegetation Index (EVI) 36 and Land Surface Water Index (LSWI) 37 are two major inputs for the VPM model, both of which are calculated from the MOD09A1 C6 500 m 8-day land surface reflectance dataset. To generate a continuous GPP product, there are two types of data gaps we need to consider: (1) missing observations: for some tiles at some DOY, e.g., year 2001, DOY 169, 177, because of the sensor calibration, no observations were acquired during those period; (2) bad-quality data: the data quality of this remote sensing reflectance measurement is often affected by the atmospheric conditions (e.g., cloud, aerosols), a rigorous quality check and gap-filling algorithm is required to generate a reliable time series of VI data for model simulation. Since EVI values are often lower if clouds or aerosols exist, while LSWI may be slightly higher, we used different strategies to gap-fill these two vegetation indices.
The raw EVI calculated from the reflectance was first masked for all cloud, cloud shadow and aerosol affected regions based on the quality assurance (QA) layer in the MOD09A1 dataset (MOD35 cloud = 'clear' (bit 0-1 equals '00'); aerosol quantity = 'low' or 'average' (bit 6-7 equals '01' or '10')) 8 and then filled with an advanced gap-filling algorithm (Fig. 2). The median values of valid data for each day of year (DOY) across 17 years (2000-2016) were chosen as the reference for seasonal pattern. Under extreme conditions, some bad quality data may not be identified through the QA check, these data usually have a lower value and can be detected using the Best Index Slope Extraction algorithm 38 . After this process, gaps were filled with a linear interpolation and smoothed using a Savitzky-Golay filter. The reference will serve as a reliable baseline of the seasonal variation of EVI for each pixel. The novelty of this method is that, instead of linearly interpolate the EVI to fill the gaps within each year, we filled the gaps in seasonal anomalies (seasonal EVI−reference). This allows us to give a robust estimation when there is long period of missing/bad data during which EVI changes dramatically ( Supplementary  Information, Supplementary Fig. S1). To reconstruct the gap-filled EVI, we calculated the differences between EVI for one specific year and the corresponding DOY reference. The differences with missing values (generated by the quality check) were then gap filled and smoothed in the temporal domain. These gap-filled differences were added back to the reference to reconstruct the gap-filled and smoothed EVI. Using this method, we not only filled bad observations caused by atmospheric conditions, but also reconstructed the EVI for periods when no observations were made by the Terra satellite, e.g., in the year 2000 before DOY 49. Figure 3 shows the average percentage of the good EVI data for the entire period. Gap-filled EVI data mostly locates in boreal regions, which is caused by high frequency of snow, clouds and aerosols; and tropical regions, which is caused by clouds and aerosols. No significant trend in data quality was found for most areas This approach may introduce uncertainties due to the year to year variation of vegetation water condition and the biased estimation when cloud or snow exist. However, we believe it has limited impact on the GPP estimates since LSWI was not a factor for GPP but was used to indicate water stress. When cloud and snow exist, it usually corresponds to a higher LSWI value and the water stress is limited. We did not use the same algorithm for LSWI gap-filling because (1) LSWI is sensitive to surface water (snow, ice,  flooding, etc.), and will exhibit a higher value which cannot be removed by using the BISE algorithm. (2) LSWI has higher variance which contains information of water stress and using this gap-filling algorithm may smooth this information. (3) atmospheric corruptions have limited effects on LSWI than EVI 8,30 . Future studies may consider using vegetation optical depth (VOD) as an indicator of water stress and upscale to higher spatial resolution 39  Our previous VPM GPP product for North America did take C3 and C4 crop into consideration 34 . However, because of the data availability and spatial and temporal coverage, we did not used the Cropland Data Layer but made a simple assumption that the C3/C4 species are half-half for all cropland pixels. This is not accurate since, for example, the Mid-West corn belt grows much more corn (C4) than the Mississippi river basin. Since we aim to develop a global GPP product in this study, we use two static C4 vegetation maps (one for cropland, one for natural vegetation) for the entire globe. Although these maps are low in spatial resolution (Table 1) and lack inter-annual variation, they provide the best estimates of C4 vegetation distribution based on our current knowledge 43,44 . C4 crop and natural vegetation are considered separately. For C4 crop, we used the Earth Stat global major crop type distribution for the year 2000 44 . Among 19 major crop types, 4 are C4 crops (corn, millet, sorghum, sugarcane). This dataset provides the plant area percentage of each crop type for each 0.083 degree gridcell, through which we calculated the C4/C3 crop ratio: where the Total crop percentage for the 0.083 gridcell was calculated from the MODIS landcover dataset (MCD12C1 C55). Within each 0.083 degree gridcell, all 500-m cropland pixels were considered to have a fixed C3/C4 crop ratio which allowed us to calculate the LUE using the area-weighted average of C3 and C4 LUE. However, this assumption introduces uncertainty when comparing simulated GPP with GPP estimates from cropland flux towers, since most sites are surrounded by either pure C3 or C4 crop. When averaged over a larger scale however, this mismatch may be negligible. We used the ISLSCP II C4 vegetation map for natural C4 vegetation distribution 43 and assumed that all grassland, savannas, woody savannas and wetland within each 1°× 1°gridcell share the same C3/C4 ratio. The corresponding LUE for these C3/C4 mixed pixels can be calculated based on their area-weighted averages.

Climate data
For the global GPP simulation, we used the NCEP reanalysis II dataset. The daily downward shortwave radiation flux at surface (dswrf in W m − 2 ) was averaged over 8 days to match the model simulation interval. We used the average of daily maximum temperature (tmax.2m) and the daily mean temperature (air.2m) as daytime temperature and it was then averaged over each 8-day period. The aggregated datasets were in Gaussian grid (192 × 96) and further downscaled to the 500-m resolution using the same method that was described in previous publications 21,34 .

Vegetation photosynthesis model (VPM) description
The VPM model follows the light use efficiency scheme and estimates GPP as the product of light absorption by chlorophyll of the vegetation (APAR chl ) and the efficiency (ε g ) that converts the absorbed energy to carbon fixed by plants through photosynthesis: where APAR chl is calculated as a product of photosynthetically active radiation (PAR) and the fraction of PAR absorbed by chlorophyll (fPAR chl ): The fPAR chl is calculated as a linear function of EVI, which is modified from previous model framework 32 : The coefficients 0.1 and 1.25 are used to adjust for sparsely vegetated or barren land and have been validated using the solar-induced chlorophyll fluorescence data (data not shown). ε g in equation (3) is down-regulated by temperature limitation (T scalar ) and water stress (W scalar ) from its maximum value (ε 0 ) which only differs by C3/C4 photosynthesis pathways: Both T scalar and W scalar range from 0 to 1 and can be calculated as follows: Where the T, T max , T min and T opt refer to the daytime mean temperature, maximum, minimum, and optimum temperature for photosynthesis, respectively. The last three parameters are biome-based and can be obtained from a look-up table ( Table 2). The land cover product from MODIS (MOD12Q1) is used to provide biome information since this is the only annual land cover product with high spatial resolution and global coverage. The uncertainty of the land cover classification is not assessed but is supposed to have limited effect on the final GPP estimation since it only directly affects the temperature scalar and indirectly affects ε 0 . LSWI max is the maximum LSWI during the snow-free period for each pixel each year. To eliminate potential bias, a temporal smoothing using nearby four years (two years before, two years after) is applied and calculates the second largest LSWI max within this five-year period 34 .

Code availability
The code for EVI gap-filling and VPM algorithms are available at https://github.com/zhangyaonju/ Global_GPP_VPM_NCEP_C3C4.

Data Records
The original VPM GPP V20 dataset is available at 500 m spatial resolution and 8-day temporal resolution. The entire earth land surface is divided into 290 subregions (tiles) under the sinusoidal projection. There are 46 GeoTiff files for each year each tile, each of which represents GPP for an 8-day average. The units are all in g C m − 2 day − 1 with a scalar factor of 0.001. Each file is also associated with a IGBP class ε 0 (g C/mol APAR) T min (°C)  data quality layer: a Boolean value indicates whether the EVI is from the raw data (0) or gap-filled (1). We also provide two coarser spatial resolutions at 0.05°× 0.05°and 0.5°× 0.5°with a longitude-latitude projection under WGS84 datum. Three temporal resolutions, i.e., 8-day, monthly and annual, are available for 0.05°× 0.05°and 0.5°× 0.5°spatial resolution products. The units are g C m − 2 day − 1 , g C m − 2 month − 1 , and g C m − 2 year − 1 , respectively. GPP products at 0.05°× 0.05°and 0.5°× 0.5°spatial resolutions can be accessed at (Data Citation 1); the raw 500 m 8-day GPP product together with the data quality layer can be accessed at (Data Citation 2). Because of the state-of-the-art gap-filling and smoothing algorithms applied, we can provide a continuous spatial temporal GPP estimate with no missing tiles. The global GPP estimation for the period 2000 to 2016 ranges from 121.60 to 129.42 Pg C year − 1 with an increasing rate of~0. 39 Pg C year − 2 . The highest annual GPP occurs mostly in tropical regions, especially in the Amazon and Southeast Asia (Fig. 4a). This corresponds to the highest peak in annual GPP around the equator (Fig. 4b). The maximum daily GPP however, shows its highest value in the Midwest region of the United States. Other crop planting regions, e.g., central Europe, Northeast China, Southeast Africa and South America also show relatively high maximum daily GPP. High maximum productivity in these regions creates another peak around 50°N (Fig. 4d). All other continents except South America and Oceania exhibit an increasing trend of annual GPP over the past 17 years (Table 3). The increasing rate is highest for Europe (0.89% year − 1 ), followed by Asia (0.64% year − 1 ), North America (0.61% year − 1 ), and Africa (0.21% year − 1 ). Annual GPP for Oceania did not show much change throughout the study period while South America experienced a significant decrease (−0.19% year − 1 ).

Technical Validation
Comparison against eddy covariance flux tower This GPP dataset was validated against 113 eddy covariance flux towers across the globe. These flux towers data were obtained from the FLUXNET 2015 Tier 1 dataset (2016 November release, http:// fluxnet.fluxdata.org/data/fluxnet2015-dataset/), and only a small portion (23 out of 136) was excluded due to inconsistency between the flux tower footprint and MODIS pixels (Supplementary Information, Supplementary Table S1). GPP from the flux tower was calculated as an average from both daytime and nighttime partition methods after a rigorous data quality check. The validation was carried out at two scales: (1) at spatial and seasonal scale, we used all 8-day GPP estimates for all sites-years (n = 28,378) and (2) at interannual scale, we calculated annual GPP anomalies for long-term sites from both EC estimates and VPM (more than 5 years of observation between 2000 and 2016, n = 479) (Fig. 5).
The overall accuracy of the VPM GPP V20 dataset is relatively high with an R 2 of 0.74 and a low RMSE of 2.08 g C m − 2 day − 1 . A complete list of sites used and GPP comparisons is available in the Supplementary Information (Supplementary Table S1). For individual biome types, the VPM GPP product underestimated evergreen forest (~27% and 30% for ENF and EBF, respectively) and slightly underestimated cropland (~15%). For other biome types, the VPM GPP did not show a systematic bias for the spatial seasonal variation. Except ENF and CSH, most biome types showed relative high R 2 values (> 0.7) which means VPM GPP captures well the spatial and seasonal variations. In terms of interannual variation, VPM GPP did not show much agreement with EC-based GPP estimates. Across all biome types, VPM GPP showed less inter-annual variation compared with that from the EC tower. The R 2 is low ( o0.5) for most biome types. This low consistency in interannual variation may be caused by insensitivity of stress factors of VPM at an interannual scale, ecosystem memory effects which is not taken into consideration in VPM, and uncertainties in both GPP estimates.

Comparison with other GPP datasets
We also compared our GPP dataset with some other data-driven GPP products, e.g., FluxCom GPP, MOD17 (both C55 and C6). VPM showed higher annual GPP estimates in eastern and central US, midlatitude Eurasia, and subtropical region in South America and Africa (Fig. 6a). These regions are mostly dominated by C4 cropland or grassland. VPM GPP is lower in tropical rainforest. When compared with MOD17C6, VPM gave higher estimations for most mid-to low-latitude regions, which is consistent with a recent study suggesting that MOD17C6 largely underestimates GPP for most ecosystems 20 .
The average global annual GPP estimates are similar in value for VPM and FluxCom, both around 125 Pg C year − 1 , though VPM showed a significant increasing trend (0.39 Pg C year − 2 ) while FluxCom is almost stable during the past 14 years (Fig. 6c). Both MOD17 products showed lower GPP estimates, with  a difference of~14 Pg C year − 1 for MOD17C55 and~26 Pg C year − 1 for MOD17C6, in comparison with VPM GPP. The C6 version of MOD17 also showed a significant increasing trend (0.29 Pg C year − 2 ) that did not appear in the C55 version. This difference may be caused by using different climate data  (GMAO/NASA (Global Modeling and Assimilation Office/National Aeronautics and Space Administration) and NCEP Reanalysis II) or the corrected sensor degradation effect in MODIS C6 data 45,46 . We also calculated the spatial pattern of GPP trend using the 0.5°× 0.5°8-day dataset as a further validation (Fig. 7). This spatial pattern is consistent with many previous studies: (1) high latitude regions and Qinghai Tibet Plateau experienced a continuous increasing GPP caused by warming and extended growing season 47-50 ; (2) afforestation in Northern China 51,52 and vegetation greening in southern Sahara Desert 53,54 also lead to an increase in GPP; (3) Tropical forest in southeast Asia 55 and southern part of Amazon 56 where deforestation happened exhibited a decreasing trend of GPP.
Overall, GPP VPM provides a reliable GPP estimation for non-forest natural vegetation and cropland by considering the C3/C4 vegetation ratio 57 . The underestimation for ENF and EBF may be related to the higher light use efficiency for diffused radiation 58 , but requires further testing. The total global GPP estimation is close to previous studies using different approaches 59,60 , and the trend of annual GPP increase are also in line with some other recent reports 4 .

Usage Notes
The GPP for gridcells in coastal regions (in 0.05°× 0.05 and 0.5°× 0.5°spatial resolution products) are averaged over the entire gridcell but land only, therefore, the land area fraction is not needed when calculating the regional sum.