Diverse photosynthetic capacity of global ecosystems mapped by satellite chlorophyll fluorescence measurements

Abstract

Photosynthetic capacity is often quantified by the Rubisco-limited photosynthetic capacity (i.e. maximum carboxylation rate, V_cmax). It is a key plant functional trait that is widely used in Earth System Models for simulation of the global carbon and water cycles. Measuring V_cmax is time-consuming and laborious; therefore, the spatiotemporal distribution of V_cmax is still poorly understood due to limited measurements of V_cmax. In this study, we used a data assimilation approach to map the spatial variation of V_cmax for global terrestrial ecosystems from a 11-year-long satellite-observed solar-induced chlorophyll fluorescence (SIF) record. In this SIF-derived V_cmax map, the mean V_cmax value for each plant function type (PFT) is found to be comparable to a widely used N-derived V_cmax dataset by Kattge et al. (2009). The gradient of V_cmax along PFTs is clearly revealed even without land cover information as an input. Large seasonal and spatial variations of V_cmax are found within each PFT, especially for diverse crop rotation systems. The distribution of major crop belts, characterized with high V_cmax values, is highlighted in this V_cmax map. Legume plants are characterized with high V_cmax values. This V_cmax map also clearly illustrates the emerging soybean revolution in South America where V_cmax is the highest among the world. The gradient of V_cmax in Amazon is found to follow the transition of soil types with different soil N and P contents. This study suggests that satellite-observed SIF is powerful in deriving the important plant functional trait, i.e. V_cmax, for global climate change studies.

1. Introduction

Plant functional traits are fundamental to ecological studies (Diaz et al., 2016; Kattge et al., 2011; Kunstler et al., 2016; Levine, 2016). The Rubisco-limited photosynthetic capacity (i.e. maximum carboxylation rate, V_cmax) is among these traits that have large interspecific and intraspecific variations. V_cmax also varies strongly with time and space within the same plant function type, depending on radiation and nutrient availabilities (Misson et al., 2006). Our understanding of the spatiotemporal distributions of V_cmax and its response to changing climate can help project carbon sequestration by vegetation and future climate change (Madani et al., 2018; A. P. Walker et al., 2017).

The availability of measured V_cmax data is limited because V_cmax is traditionally determined by measuring A-Ci curves, which is rather laborious (Kauwe et al., 2016; Xu and Baldocchi, 2003). Currently, Earth System Models (ESMs) heavily rely on meta-analysis based V_cmax databases which are derived from the correlation between leaf N content and V_cmax (Kattge et al., 2009). V_cmax also has strong seasonal variations (Jin et al., 2012; Wang, 2008; Wang et al., 2004; Wang et al., 2009a, 2009b). The value of V_cmax is often set as a constant for each plant function type (PFT) in terrestrial biosphere models, and the uncertainty in V_cmax has hampered our understanding of global carbon and water cycles (Bonan et al., 2012; Harper et al., 2016; L. M. He et al., 2017; Madani et al., 2018).

Recent advances in remote sensing make it possible to directly derive V_cmax for large areas from remotely sensed satellite observations. Recently, Alton (2018) mapped global V_cmax from satellite observations of chlorophyll content. However, the derived leaf V_cmax values are lower and vary more narrowly than compiled measurements. Carter et al. (1996) showed the promise of solar-induced chlorophyll fluorescence (SIF) in capturing photosynthetic capacity at the leaf level. Recent studies show that spaceborne SIF has a strong correlation with gross primary productivity (GPP) (Joiner et al., 2018; MacBean et al., 2018; Sun et al., 2018; Sun et al., 2017a) and thus it is promising to derive V_cmax from SIF (Koffi et al., 2015), e. g. for cropland (Y. Zhang et al. 2018; Zhang et al., 2014).

Photosynthesis involves two processes. The first one is light-dependent: light energy is harvested and converted into chemical energy in the form of electron carrying molecules like ATP and NADPH in Photosystems I and II (PS I and PS II) in the thylakoid membranes of chloroplasts. When a chlorophyll pigment absorbs visible light, it becomes excited and will return to the ground state by one of several competing processes, including photosynthesis in PS II, non-radiative decay, nonphotochemical quenching, and fluorescence. The yield fluorescence is generally small (1%–5%) but the signal is enough to provide information on the total amount of light absorbed, as the fluorescence is generally proportional to the total amount of absorbed photosynthetically active radiation (APAR) (Frankenberg and Berry, 2018). In the second process (i.e., the Calvin cycle), which is light-independent, carbohydrate molecules are assembled from CO₂ using the chemical energy harvested during the light-dependent reactions. In this process, Ribulose-1, 5-bisphosphate carboxylase/oxygenase (Rubisco) is the enzyme involved in the first major step of carbon fixation. Rubisco catalyzes the carboxylation of ribulose-1,5-bisphosphate (i.e. RuBP). However, Rubisco is able to fix only 3–10 CO₂ molecules each second per molecule of enzyme (Bar-On and Milo, 2019). The reaction catalyzed by Rubisco is, thus, the primary limiting factor of the Calvin cycle at sunlit conditions. Therefore, V_cmax (the maximum rate of carboxylation of Rubisco; μmol CO₂ m⁻² s⁻¹) is largely determined by the amount of Rubisco in the leaf.

Although the Calvin cycle does not use light as a reactant, it requires the products of the light-dependent reactions (energy carrying molecules, e. g., ATP and NADPH) to drive the construction of new carbohydrate molecules. After the energy is transferred, the energy carrying molecules return to the light-dependent reactions to obtain more energized electrons. In addition, several enzymes of the light-independent reactions are activated by light (Bar-On and Milo, 2019). Therefore, these two processes are tightly coupled.

In a leaf, nitrogen plays an important role in regulating the photosynthetic capacity as it is a component of pigments for light harvest, proteins associated with photosystems and Calvin cycle (including Rubisco), ATP and NADP. Plants require an optimized balance of light intensity and carbon assimilation in the process of photosynthesis for survival when environmental conditions are ideal and for acclimation when environmental conditions are severe (Smith et al., 2019). For a leaf acclimated to the environment, the allocation of N to light harvesting pigments (chlorophylls) and carboxylation enzymes (Rubisco) would be optimized in a way that they become highly correlated. Studies have found that chlorophyll content correlates well with V_cmax which is largely determined by the amount of Rubisco (Croft et al., 2017; Luo et al., 2019). Therefore, chlorophyll can be a useful indicator for the carboxylation capacity of a leaf (Murchie and Lawson, 2013). As sun-induced chlorophyll fluorescence (SIF) is not only determined by chlorophyll content but also by radiation and other environmental variables, we cannot relate SIF to V_cmax directly. However, we can use a numerical technique to estimate and remove the effects of other factors in SIF signals and extract the needed V_cmax. The data assimilation technique presented in this paper is developed for this purpose.

Excessive light may damage the photosynthetic machinery. The correlation between SIF and photosynthesis assimilation rate will certainly not be held for a leaf without water supply under sunlight, as photosynthesis needs water (to provide electrons from hydrogen atoms, and for heat dissipation). At such an extremely stressed condition, the absorbed light energy excites the pigment but the native energy or the electron acceptor is missing. The energy can then be used to excite oxygen and produce singlet oxygen molecules, which decompose chlorophylls. In this case, SIF may not be an indicator of the photosynthesis rate.

However, a plant reacts to mitigate the harmful effects of excess light in natural settings (i.e., via photoprotection). For example, some plants can change the leaf inclination angle to reduce the incident light on the leaf surface. Under persistent droughts over days and weeks, plants can develop mechanisms (photo-acclimatization) of up-regulation of stress response proteins or down-regulation of pigment biosynthesis. For example, PS II is damaged by light irrespective of light intensity, and has a repair cycle as the chloroplast acclimation strategy to irradiance stress. The PS II repair cycle occurs in chloroplasts and in cyanobacteria, consisting of degradation and synthesis of the D1 protein of the PSII reaction center, followed by activation of the reaction center. However, environmental stresses, for example, extreme temperatures, salinity, and drought (Zhu et al., 2017), limit the supply of CO₂ for use in carbon fixation, which in turn decreases the rate of repair of PSII. In autumn, many plants stop making chlorophylls and break down chlorophylls into smaller molecules in responding to cold. The decreases of light-harvesting and activity of the Calvin cycle tend to be synchronized; therefore, SIF has been used to track photosynthesis even under drought conditions (Gonsamo et al., 2019; Jiao et al., 2019; Y. Sun et al., 2015).

A spaceborne instrument is only able to observe a fraction of SIF escaped from the canopy, and these fractions vary greatly with the sun-target-satellite observation geometry and complex canopy structure (He et al., 2017b; Z. Zhang et al. 2018). Therefore, although the SIF-GPP correlation is strong, it is also found that the SIF-to-GPP ratio varies with PFT and location (Li et al., 2018a; Wood et al., 2017; Zhang et al., 2016a, 2016b), hindering the derivation of global V_cmax maps using SIF. Norton et al. (2018) averaged the OCO-2 SIF data into 2° × 2° grids, assimilated these SIF data into a combined ecosystem model (BETHY-SCOPE) with prescribed leaf area index (LAI) as input and found that the simulation of global GPP is improved; however, SCOPE (van der Tol et al., 2009) utilizes the SAIL5 model (Verhoef, 1984) which is a 1D radiative transfer model, therefore, they did not make full use of vegetation structure information (such as clumping) in their framework, neither was global V_cmax map produced from their study.

The objective of this study is to map the spatiotemporal distribution of V_cmax for global terrestrial ecosystems by reconciling the variations in the SIF-to-GPP ratio in ecosystems due to changes in observation geometry and canopy structure. In the following, we present the major findings for the top-of-canopy V_cmax that is normalized to 25 °C (V_cmax, 25⁰) unless specified otherwise.

2. Materials and methods

2.1. Forcing data, SIF data, and key model parameters

A complete description of meteorological forcing data (Rienecker et al., 2011), GOME-2 SIF data (Joiner et al., 2013) and the BEPS model (Boreal Ecosystem Productivity Simulator) (He et al., 2018) for Gross Primary Production (GPP) simulation and its key model parameters is provided in the Supplementary materials. A Leaf Area Index (LAI) product (named GLOBMAP) (Liu et al., 2012) is used to drive BEPS. A clumping index map is used to describe the canopy structure (L. He et al., 2012; He et al., 2016).

Ground measurements of SIF from winter wheat were also made using an automated multi-angle chlorophyll fluorescence measurement system (SI). The measurements took place at a flux tower site (Jurong station, 31°9′N, 119°1′E) in Jiangsu province, China. The SIF was observed during 15th April to 30th May in 2016 when the wheat canopy was closed. The SIF measurements, together with LAI, GPP from this site are used to derive the SIF-GPP relationship.

2.2. The “two-leaf” ecosystem model for simulation of vegetation productivity

A detailed description of BEPS is provided in the supplementary materials. In this section, we summarize the key BEPS modules that are related to V_cmax and GPP modeling. BEPS is spun up from 1981 to 2006 using the forcing data. In this study, we modified the BEPS to output GPP from sunlit leave for SIF retrieval. At the leaf level, Farquhar’s model is used to calculate the net CO₂ assimilation rate (Farquhar et al., 1980). In Farquhar’s model, the CO₂ assimilation rate is either determined by Rubisco-limited or light-limited gross photosynthetic rates (see details in Supplementary material, Section 2.2). At light-saturated condition (for sunlit leaves), V_cmax and leaf temperature are two of the most important parameters that control Rubisco-limited gross photosynthetic rates (μmol m⁻² s⁻¹); while at light-limited condition, J_max, radiation and leaf temperature are three of the most important parameters for the calculation of the gross photosynthetic rate. Although Farquhar’s model is included in many ESMs as the core photosynthesis module for almost four decades, these ESMs still rely on tabulated V_cmax values for each ecosystem with no consideration of their spatial variations. BEPS uses a “two-leaf” approach to calculate the canopy level CO₂ assimilation (Chen et al., 1999; Norman, 1982). The key parameter V_cmax is also upscaled weighted by the leaf nitrogen profile in the canopy (Chen et al., 2012). The objective of this paper is to globally map V_cmax,0, the maximum Rubisco capacity at the leaf level at the top of the canopy that is normalized at 25 °C.

2.3. The normalization of satellite SIF observation to the hotspot direction

Exposed at different radiation conditions, sunlit and shaded leaves differ in their SIFs, which can vary as large as an order of magnitude (Pinto et al., 2016). Additionally, there are large variations of SIF signals observed from different solar and satellite view geometries (Liu et al., 2016; Z. Zhang et al., 2018). The escape probability of SIF is also affected by canopy structure. For example, leaves in cropland are distributed most randomly among all vegetation types; therefore, the sunlit leaves of crops are mostly distributed at the top of the canopy and can be largely observed at different satellite view angles; however, the sunlit leaves of forests can be located in the low parts of a canopy due to their complex 3-dimensional structure and thus the portion of observed sunlit leaves, where the most SIF signal originates, depends on view angles. Therefore, the global maximum SIF signal is often observed from croplands but not from forests, although forests are the most productive ecosystems.

To reliably estimate the ratio of observed SIF (especially the SIF from sunlit leaves) to the total emitted SIF, we need a well-documented geometric optical model, such as the 4-Scale model (Chen and Leblanc, 1997) and vegetation structure parameters, including LAI (Liu et al., 2012) and clumping index (L. M. He et al., 2012). Especially, the global clumping index provides unique information regarding to the leaf distributions in different ecosystems, which frees us from the requirement of homogenous canopy in order to calculate the ratio of observed sunlit leaves, which is key to SIF escape probability. For global applications, we have simplified the 4-Scale model to estimate the escape probability of SIF from canopy at different angles (He et al., 2017b). The methodology is detailed in the SI.

Here, we applied the previously developed method to normalize GOME-2 clear-sky SIF measurements to the hotspot direction (Chen and Leblanc, 1997; He et al., 2017b); in other words, the normalization gives an estimate of the maximum SIF signal that can be observed in a canopy from a satellite from the hotspot direction.

2.4. The relationship between “hotspot SIF” and GPP from sunlit leaves

Previously, the canopy level SIF-to-GPP ratio has been directly used to constrain global GPP simulations (Joiner et al., 2018; MacBean et al., 2018). As aforementioned, the SIF-to-GPP ratio at the canopy level can be affected by both satellite view angle and vegetation structure; therefore, previous studies often relied on time-averaged SIF-to-GPP ratios calibrated for different ecosystems.

In this study, we propose to use “hotspot SIF” to constrain “sunlit GPP” simulation in an ecosystem model, such as BEPS. In theory, this “hotspot SIF” to “sunlit GPP” ratio is independent of both view angle and vegetation structure if assuming a uniform leaf inclination angle, and thus ecosystem-based calibration of the ratio is no longer necessary. We can explicitly use a BRDF model and vegetation structural parameters (i. e. LAI and clumping index) to minimize uncertainties in the SIF-to-GPP ratio at the canopy level among ecosystems.

Moreover, we choose to constrain “sunlit GPP” not the total GPP in our simulations because the photosynthesis at the hotspot direction is light-saturated (see Eq. (2)) under clear-sky condition when GOME-2 SIF is collected. Therefore, the modeling of sunlit GPP in BEPS is less prone to uncertainties induced by the radiation input data.

Another reason to constrain “sunlit GPP” not the total GPP is that the calculation of sunlit LAI in BEPS is less prone to the uncertainties of the total LAI (see Eqs. (6) and (7)). As shown in the SI, sunlit LAI does not increase proportionally to the increase of total LAI; the magnitude of sunlit LAI is largely stable when total LAI is larger than 2.

The “hotspot SIF” to “sunlit GPP” ratio is determined in two ways (see more details in SI). In the first approach, we (1) simulate sunlit-GPP in BEPS using the default V_cmax map which is derived according to land cover type and tabulated V_cmax from Kattge et al. (2009), (2) derive the “hotspot SIF” (corrected from GOME-2) to “sunlit GPP” ratio for each pixel, (3) create a histogram of this “hotspot SIF” to “sunlit GPP” ratio from the global simulation, and (4) choose the mode of this histogram (Fig. S4(a) and (b), SI). In the second approach, we estimate this ratio directly from measured SIF and measured GPP at a flux tower site (Fig. S4(c)). In the end, both approaches reach a similar ratio, 4.1 mW m⁻² nm⁻¹ sr⁻¹/(g C m⁻² h⁻¹) and then this unique ratio is applied to all ecosystems without further calibration. We suggest that the value of this ratio should be further calibrated according to the features of a specific sensor (band width, spectral resolution for SIF retrieval) if it is applied to another SIF product.

2.5. The data assimilation framework for V_cmax retrieval

In this study, we do not directly invert V_cmax from SIF; instead, we use an EnKF filter for parameter estimation (He et al., 2014; Mo et al., 2008; Schöniger et al., 2012). EnKF for parameter optimization is optimal for ill-posed inversion problems (Iglesias et al., 2013) because the parameter (V_cmax in this study) is updated in a system that synthesizes prior knowledge and new observations by statistical minimization of estimation errors. This is ideal for V_cmax retrieval as V_cmax evolves smoothly along the leaf nitrogen changes.

The normalization of GOME-2 SIF to the hotspot direction tightly follows the same strategy to separate sunlit and shaded leaves in BEPS. This enables us to directly constrain sunlit GPP in BEPS using hotspot SIF. We optimize the ecosystem model parameter, top-of-canopy maximum carboxylation rate (V_cmax,0, normalized to 25 °C), in a parallel data assimilation system run on a high performance computer (He et al. 2017a; He et al., 2014). This system has been used to assimilate soil moisture data from the SMAP mission into BEPS; for this study, it follows the same theory to adjust the BEPS model status (i.e., V_cmax) by assimilating GPP “observations” from GOME-2. The description of EnKF is summarized in SI.

We start to run 80 BEPS model replicates in the system since 2007 with perturbed V_cmax having an average value equal to the original BEPS V_cmax input (and one standard deviation of 12 μmol m⁻² s⁻¹) (Fig. 1). Considering that there might be inconsistency of cloud conditions between forcing data and GOME-2 product, we only perform the parameter optimization when both forcing data and GOME-2 observations indicate clear-sky conditions. For each grid, whenever a GOME-2 SIF observation is available, it is normalized to hotspot direction and converted into sunlit GPP to constrain BEPS-simulated “sunlit GPP” for V_cmax retrieval. Currently, there is no much information available regarding the uncertainty of the GOME-2 SIF signal. When SIF is quality controlled and gridded to half degree spatial resolution and biweekly time frame, the estimated absolute error is 0.1–0.4 mW m² nm⁻¹ sr⁻¹ (Joiner et al., 2013; Joiner et al., 2016). Since the instantaneous SIF observations are used, we choose 0.5 mW m² nm⁻¹ sr⁻¹ in this study. This generally means that higher SIF signal has less relative uncertainty and has stronger constraints in the system for V_cmax retrieval. The optimized V_cmax is then used in the next step for the simulation of both sunlit and shaded GPP in BEPS until the next V_cmax retrieval is available. An 11-year-long record of V_cmax (2007–2017) is produced in this study based on GOME-2 SIF observations.

Fig. 1.

The parameter optimization framework for retrieving top-of-canopy V_cmax from sun-induced chlorophyll fluorescence (SIF) measurements in a data assimilation system (bold frames).

3. Results

3.1. Spatial distribution of V_cmax

The retrieved V_cmax has strong seasonal and inter-annual variations. The 11-year mean and standard deviation of V_cmax are shown in Fig. 2 and Table 1.

Fig. 2.

Global top-of-canopy V_cmax normalized to 25 °C derived from 11-year (2007–2017) SIF data. (a) Multi-year average of V_cmax map weighted by GPP. (b) The seasonal and inter-annual variation of V_cmax represented by one standard deviation. (c) The difference between SIF-derived V_cmax and BEPS-default V_cmax.

Table 1.

Top-of-canopy V_cmax (μmol m⁻² s⁻¹) normalized to 25 °C for different plant function types (Evergreen Needle-leaf trees, ENF; Evergreen Broadleaf trees, EBF; Deciduous Needle-leaf trees, DNF; Deciduous Broadleaf trees, DBF). “SIF” indicates V_cmax derived from SIF measurements; “BEPS” indicates the initial V_cmax values in BEPS.

	Count	Vcmax (SIF)	Vcmax 1 std. (SIF)	Vcmax (BEPS)	Vcmax 1 std. (BEPS)
ENF	1954	48.2	5.7	62.8	1.2
EBF	8230	53.6	14.7	42.4	2.2
DNF	667	44.6	2.7	41.7	1.9
DBF	1071	72.7	14.7	59.9	1.7
Shrub	10,420	57.4	31.0	59.5	1.8
Grass	8022	88.7	20.9	90.3	2.0
Crops	2432	82.7	15.2	90.5	2.2

The distinct belts of boreal forests in the northern hemisphere have the lowest V_cmax values and variations (44.6 ± 2.7 μmol m⁻² s⁻¹ for deciduous needleleaf forests, DNF, and 48.2 ± 5.7 μmol m⁻² s⁻¹ for evergreen needleleaf forests, ENF, respectively). These values and their ranges are similar to previous findings at flux tower sites (He et al., 2014; Zheng et al., 2017). Note that there are different definitions in LAI and V_cmax in term of “leaf area” for non-flat leaves. We used a definition by Chen and Black (1992b) which defines LAI as the half the total intercepting area per unit ground surface area; while the V_cmax values reported in the previous literature might use “per projected leaf area”. Considering the difference in “leaf area”, the V_cmax values derived under two LAI definitions will differ by a factor of ~π/2 (projecting cylinder to plate).

Evergreen broadleaf forests (EBF) also have relatively low V_cmax (53.6 ± 14.7 μmol m⁻² s⁻¹) around the equator but with large spatial variations. In the retrieved V_cmax map, we find that the V_cmax gradient in Amazon follows the transition of Oxisol soil (lack of P) to Ultisol soil (non-Oxisol) (Fig. 2c). This is consistent with the study by Kattge et al. (2009) who found that V_cmax is strongly correlated to leaf nitrogen (N) content and reported V_cmax values in tropical trees for oxisols soil and non-oxisols soil, respectively, and the sensitivity of V_cmax to leaf N is increased with increasing leaf phosphorus (P) (Walker et al., 2014). This result also echoes the mechanistic model by Ali et al. (2016) that shows a clear gradient of V_cmax in Amazon (their Fig. 4, and SI).

Fig. 4.

Impacts of the SIF-derived global V_cmax map. (a) Annual GPP in 2007–2016 constrained by SIF. (b) Annual GPP in 2007–2016 using default V_cmax values. (c) The GPP difference between (a) and (b). (d) Comparison of SIF-derived V_cmax and BEPS-default V_cmax for each plant function type. (e) Comparison of GPP simulations w/o SIF-constraint. The values in (d) and (e) are listed in Supplementary tables.

The V_cmax values in shrublands have the largest spatial variation (57.4 ± 31.0 μmol m⁻² s⁻¹); the V_cmax values from this study are close to the median V_cmax (60 μmol m⁻² s⁻¹) of Shrub in the Caatinga biome, eastern Brazil, from two field campaigns (Rezende et al., 2016). Deciduous broadleaf forests (DBF), often mixed with croplands, have a moderately high V_cmax and spatial variation among all PFTs (72.7 ± 14.7 μmol m⁻² s⁻¹). The V_cmax values for grassland are the highest (88.7 ± 20.9 μmol m⁻² s⁻¹). With the inclusion of corn, which is a low V_cmax C4 crop, croplands have the second highest V_cmax (82.7 ± 15.2 μmol m⁻² s⁻¹). V_cmax in croplands also has the strongest seasonal variation (Fig. 3b) due to agricultural management, e. g. cropping rotations. This V_cmax map clearly shows the global distribution of crop belts, e. g. the corn-soybean and wheat belts in US, wheat-corn belt in China, the super-high productive wheat-rice cropping system in Punjab, India (Kang et al., 2009; Khajuria, 2016; Singh et al., 2009), and two cropland belts in Australia (Supplementary figures).

Fig. 3.

Top-of-canopy V_cmax normalized to 25 °C for different plant function types at 36 km resolution. (a) EBF at Amazon. (b) DBF at US. (c) ENF at boreal forest, Canada. (d) Cropland in US corn belt. (e) Cropland in Australia. (f) Cropland in France. (g) Cropland (wheat-corn system) in China. (h) Cropland (wheat-rice system) in Punjab, India. (i) Crop/forest mixed pixel in South America. LAT indicates latitude, and LON indicates longitude.

Interestingly, we reveal a very high V_cmax belt in South America, which is associated with the rapid development of soybean cultivation in recent years (Correia, 2017; Fehlenberg et al., 2017; Gasparri et al., 2016; McKay et al., 2016; Mier y Terán Giménez Cacho, 2016), the so-called soybean revolution in South America (Cattelan and Dall’Agnol, 2018; Richards, 2010; Torres et al., 2017). During the last 35 years, the soybean production in South America is ten-fold increased (SI). According to http://www.fao.org/faostat/en/#data/QC/visualize, currently more than 80% of soybean are produced in America, while recently Brazil surpassed USA as the top soybean producer. The functional group of legume is well known for its high V_cmax (Ainsworth and Rogers, 2007; Feng and Dietze, 2013). As shown in Fig. 3(i), the retrieved V_cmax for croplands is as high as 220 μmol m⁻² s⁻¹, comparable to the ranges of soybean V_cmax measurements (101–190 μmol m⁻² s⁻¹) (Bunce, 2016; Morgan et al., 2004; B. J. Walker et al., 2017; Xu et al., 2016). We found that the soybean production (Cattelan and Dall’Agnol, 2018) in each Brazilian state can explain 52% of variance in V_cmax (SI).

The dervied V_cmax around coastal tundra in high Arctic is higher than orginal BEPS V_cmax input by 10–20 μmol m⁻² s⁻¹. This higher V_cmax retrieval echos to recent measurements by Rogers et al. (2017). They attribed the high V_cmax in high Arctic area to two factors: (1) the higher leaf N content; and (2) more N is allocated in Rubisco. They also found that the V_cmax (normalized at 25 °C) was significantly higher than the inputs in current Earth System models by two- to five-folds.

Five distinctive features are found from our V_cmax map, including (1) gradients across major PFTs, (2) gradients within the Amazon forest from north to south, (3) global major crop belts, (4) high V_cmax in crop belt of South America, and (5) moderate to high V_cmax in northern tundra. Potential drivers (temperature, precipitation, radiation, LAI, root-zone dryness) of these V_cmax features are also examined (Figs. S25–S29). The transition of temperature itself from the tropic to the polar region cannot explain the V_cmax distribution. At low to middle latitudes, lower solar radiation is associated with higher cloud coverages (and more precipitation); the V_cmax gradient (low to high) from this study follows the transition of LAI (high values in forest in wet area to low values in grassland in dry area). However, water stress cannot explain the high V_cmax values in tundra areas. Temperature, precipitation and radiation cannot explain the V_cmax gradient within Amazon forests. The crop belts with high V_cmax values, which are dominated by anthropogenic activities, are not associated with extra high or lower LAI values (Fig. S25). Comparing the V_cmax patterns in crop belts of South America to maps in Figs. S19 to S29, we suggest that the soybean belt best explains these high V_cmax values.

3.2. Seasonal variation of V_cmax

The seasonal variations of V_cmax for nine locations are provided in Fig. 3. Due to clouds, the EBF has the least number of retrievals (Fig. 3(a)). EBF and ENF (Fig. 3(c)) have the least seasonal variations. V_cmax for ENF in the spring is overestimated due to the underestimation of LAI in this season (Wang et al., 2016). DBF (Fig. 3(b)) in US shows strong seasonal variation, increasing in the spring, peaking at early summer (day of year (DOY) = 180), and decreasing gradually in the fall.

Due to management practices, the seasonal variations of V_cmax for croplands show diversified patterns (Figs. 3(d)–3(i)). The US corn belt, a single cropping system (Fig. 3(d)), is characterized with a V_cmax maximum in August. The croplands in the southern hemisphere (Fig. 3(e, i)) have peaked V_cmax during the spring. Fig. 3(f) and (g) show that the high V_cmax of winter wheat in the spring is replaced by another crop in the summer with low V_cmax values (e. g. C4 corn) in Shandong, China. With adequate irrigation and fertilization, V_cmax in Punjabi croplands is among the highest in the world (Fig. 3h).

Only limited validation can currently be performed for this V_cmax map that is at 36 km resolution, because it is difficult to find pure pixels at this resolution and V_cmax has seasonal and inter-annual variations (Xu and Baldocchi, 2003). V_cmax also varies with the vertical N profile in the canopy (Domingues et al., 2005) and with environmental conditions (e. g. soil nutrition, water stress, and shading) (Dechant et al., 2017; Sharwood et al., 2014). Both types of information are limited in the literature. Ground observations suggest that the maximum V_cmax ranges from 105 to 150 μmol m⁻² s⁻¹ (Feng et al., 2011; Hu et al., 2014; Silva-Pérez et al., 2017; J.S. Sun et al. 2015) for wheat and from 30.1 ± 3.5 to 40 ± 1 μmol m⁻² s⁻¹ (Massad et al., 2007; Sharwood et al., 2014) for maize (corn). It can be as high as 140 μmol m⁻² s⁻¹ (Borjigidai et al., 2006; Ikawa et al., 2018) for rice. The V_cmax values retrieved in this study are generally close to the upper range of ground values because retrieved values represent the top-of-canopy values which are higher than the average values for the canopy.

3.3. Impacts on global GPP simulation

Similar to other ecosystem models, the V_cmax values in the default BEPS are set according to vegetation types. After using the spatially and temporally explicit V_cmax maps in BEPS (Fig. 4(a–e)), the GPP estimation is reduced by 13% and 6% for ENF and croplands, respectively, due to decreased V_cmax (Table 2). In contrast, the GPP estimation is increased by 2%, 5%, 7%, 10%, and 1% for EBF, DNF, DBF, Shrub, and grassland, respectively. The GPP estimation for EBF at non-oxisol soil is significantly increased due to increased V_cmax. DBF in the US is highlighted with increased V_cmax and GPP. Although there is a remarkable increase of GPP in the croplands of South America, the global average GPP of croplands is decreased by 6% due to the stronger seasonal variation in derived V_cmax and crop rotations (e. g. with low V_cmax of corn).

Table 2.

GPP values (g C m⁻² s⁻¹) simulated using SIF-retrieved V_cmax and default V_cmax in BEPS for different each plant function types.

	Count	GPP (SIF)	GPP 1 std. (SIF)	GPP (BEPS)	GPP 1 std. (BEPS)
ENF	1954	1143.6	284.7	1307.6	263.3
EBF	8230	2817.6	261.8	2774.0	267.7
DNF	667	770.1	93.4	731.0	73.1
DBF	1071	1827.4	358.6	1708.5	342.4
Shrub	10,420	200.3	119.5	182.5	108.9
Grass	8022	415.7	378.9	411.0	382.9
Crops	2432	986.9	379.8	1049.1	380.9

Driving the model with the same LAI and meteorological data, the global GPP simulation with SIF-derived V_cmax is 126.3 ± 1.5 Pg C yr⁻¹ (here the one standard deviation was derived from global GPP in 2007–2017; it is not an indicator of uncertainty). This contrasts to the GPP simulation using the seasonal invariable V_cmax (130.4 ± 2.2 Pg C yr⁻¹) which could lead to overestimate of GPP in the spring and autumn.

4. Discussion

4.1. Why do we choose “hotspot SIF” to constrain “sunlit GPP”?

Like other remote sensing products, the accuracy of this new V_cmax map depends on the accuracies of the ecosystem model used for V_cmax optimization and the input data, such as LAI, clumping index, SIF data, metrological data and the SIF-to-GPP ratio.

As aforementioned, we have chosen to use “hotspot SIF” to constrain “sunlit GPP” rather than to use “total SIF” to constrain “total GPP”. This approach is necessary for the conversion from the canopy-level SIF measurements to the leaf-level information retrieval. It also has the advantage that that both the modeling of “hotspot SIF” and “sunlit GPP” are less prone to uncertainties in input parameters.

The modeling of sunlit GPP depends on the simulations of (1) sunlit LAI and (2) Rubisco-limited gross photosynthetic rates. In Section 2.4, we have shown that the magnitude of sunlit LAI is not sensitive to the multiplication of “total LAI and clumping index” when LAI is relatively large. This is because the uncertainty in total LAI mostly goes to the shaded part in the BEPS two-leaf scheme. Comparing to light-limited GPP modeling at the leaf level which requires three major parameters (J_max, leaf temperature and radiation) as inputs, the Rubisco-limited GPP modeling does not require radiation as input which has much uncertainty. By using hotspot SIF, our framework is not very sensitive to the uncertainty in the meteorological forcing data (except temperature, which usually has less uncertainty).

4.2. Towards to a unique GPP-SIF ratio

Using “hotspot SIF” frees us from using ecosystem-based calibrations. Determining a unique SIF-to-GPP ratio across biomes is at the cutting edge of SIF studies (Duveiller and Cescatti, 2016; Frankenberg et al., 2011; Guan et al., 2016; Joiner et al., 2014; Li et al., 2018b; Liu et al., 2017; Luus et al., 2017; Verma et al., 2017; Verrelst et al., 2016; Wagle et al., 2016; Walther et al., 2016; Wang, 2014; Wood et al., 2017; Yang et al., 2016; Yang et al., 2015; Yoshida et al., 2015; Zhang et al., 2016a, 2016b), while previous studies were hindered by ignoring the variation of the SIF signal due to the variation of the shadow fraction with view angle. For example, it was found that there are small variations in the “total GPP”-to-“average SIF” slopes (4.60–5.55 g C m⁻² day⁻¹/(mWm⁻² nm⁻¹ sr⁻¹)) (Zhang et al., 2016a, 2016b), which are equivalent to “average SIF”-to-“total GPP” ratio of 4.3 to 5.2 mW m⁻² nm⁻¹ sr⁻¹/(g C m⁻² h⁻¹). The variation of this slope should be reduced once the SIF signal is normalized to the hotspot direction, from which shadows are invisible. Combining a unique clumping index derived from multiple-angular MODIS data and a sophisticated geometric optical model advances the way to derive the “hotspot SIF”-to-“sunlit GPP” ratio in this study (Chen and Leblanc, 1997; He et al., 2017b). Field observations at flux tower sites, although limited, also yield the same SIF-to-GPP ratio (4.1 mW m⁻² nm⁻¹ sr⁻¹/ (g C m⁻² h⁻¹)), further supporting the “meta-based” analysis from the global GPP simulation in this study.

Besides the strong arguments on the similarity of the GPP-SIF slope across PFTs (Li et al., 2018b; Xiao et al., 2019; Z. Zhang et al., 2018), Liu et al. (2017) found that the slope is higher for C4 crops than for C3 plants. This is because the C4 plants have less photorespiration and are more efficient in photosynthesis, and therefore C4 plants emit less SIF for the same amount of photosynthesis than C3 plants. Currently, the information on the GPP-SIF slope for C4 plants is still limited. In this study, we used a C3 model and potentially this suggests that the GPP may be underestimated for regions where C4 plants dominate; however, given our current framework, a separate GPP-SIF slope can be taken whenever the slope information becomes more certain and the location and period of C4 plantation become clearer in future.

4.3. Comparison of result to previous dataset

The SIF-derived V_cmax from this study is compared to a widely used N-derived V_cmax dataset (Fig. 5) (Kattge et al., 2009). Although both datasets show comparable magnitudes (R² = 0.61) and variations, they diverge in the magnitudes of V_cmax for ENF, with higher values in N-derived V_cmax. This may be explained by the use of different LAI definitions for non-flat leaves: we use true LAI, defined as the half total intercepting area, while Kattge et al. (2009) may use projected LAI with lower values (Chen and Black, 1992a).

Fig. 5.

Comparison of SIF-derived V_cmax from this study to N-derived* V_cmax (Kattge et al., 2009) for seven plant function types.

4.4. Future studies

We realize that V_cmax measurements from flux towers at a footprint of ~500 m may not match our retrievals at 36 km resolution. Two future efforts will be highly useful in validating our proposed framework of V_cmax mapping: (1) making use of available SIF measurements at high spatial resolutions. These measurements include OCO-2 and OCO-3 data at the spatial resolution of 1.1 km (Sun et al., 2017b) and TRO-POMI data at the spatial resolution of 3.5 × 7 km² (Köhler et al., 2018). Especially OCO-3 has the capability to point to a specific flux tower site for multi-times in a day. This gives the opportunity to compare V_cmax retrievals from satellite imagery to eddy-covariance based retrievals (Zheng et al., 2017). (2) making ground measurement of SIF using ultra-high resolution hyperspectral sensors (Xi Yang et al., 2018). Hotspot SIF measurements from these ground systems can be used to invert V_cmax values that can be compared to direct V_cmax measurements from A/Ci curves (Manter and Kerrigan, 2004).

5. Conclusions

Based on the strong correlation between SIF and GPP, this study uses a data assimilation approach to derive a long-term product (2007–2017) of top-of-canopy V_cmax at the leaf level (normalized to 25 °C) from GOME2 SIF data. A critical step in this approach is to use angularly-normalized instantaneous SIF in the hotspot direction to constrain simulated GPP from sunlit leaves in order to account for different probabilities of observing sunlit leaves from different view directions relative to the sun. In conclusion, we have the following major findings:

1. The range of the SIF-derived V_cmax retrievals are comparable to documented V_cmax values in the literature. The V_cmax map shows distinct belts associated with the distribution of plant function types (PFT). Within each PFT, V_cmax also varies greatly, except for conifers. In the Amazon evergreen broadleaf forests, the retrieved V_cmax is sensitive to soil type and hence possibly P limitation, with a V_cmax gradient along the transition of soil orders.

2. Seasonal variation of V_cmax depends largely on PFTs and locations. Agricultural activities, such as crop rotation and fertilizer use, can strongly impact both temporal and spatial variations of V_cmax in croplands. All well-known crop belts in different continents are clearly illustrated on the derived V_cmax map. Characterized with high V_cmax, the booming soybean revolution in South America is also clearly shown on the map.

A great deal of effort is needed to validate this V_cmax map. Due to the large spatio-temporal variations of V_cmax, low-cost and standardized V_cmax measurements for the large pixels are yet to be developed.