Volume 47, Issue 7 e2020GL087474
Research Letter
Free Access

From the Ground to Space: Using Solar-Induced Chlorophyll Fluorescence to Estimate Crop Productivity

Liyin He

Corresponding Author

Liyin He

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Correspondence to: L. He and C. Frankenberg,

[email protected];

[email protected]

Search for more papers by this author
Troy Magney

Troy Magney

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Now at Department of Plant Sciences, University of California, Davis, CA, USA

Search for more papers by this author
Debsunder Dutta

Debsunder Dutta

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Now at Department of Civil Engineering, Indian Institute of Science, Bangalore, India

Search for more papers by this author
Yi Yin

Yi Yin

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Search for more papers by this author
Philipp Köhler

Philipp Köhler

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Search for more papers by this author
Katja Grossmann

Katja Grossmann

Department of Atmospheric Sciences, University of California, Los Angeles, CA, USA

Now at Institute of Environmental Physics, University of Heidelberg, Heidelberg, Germany

Search for more papers by this author
Jochen Stutz

Jochen Stutz

Department of Atmospheric Sciences, University of California, Los Angeles, CA, USA

Search for more papers by this author
Christian Dold

Christian Dold

National Laboratory for Agriculture and the Environment, USDA-ARS, Ames, IA, USA

Search for more papers by this author
Jerry Hatfield

Jerry Hatfield

National Laboratory for Agriculture and the Environment, USDA-ARS, Ames, IA, USA

Search for more papers by this author
Kaiyu Guan

Kaiyu Guan

Department of Natural Resources and Environmental Sciences, University of Illinois at Urbana-Champaign, Urbana, IL, USA

National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign, Urbana, IL, USA

Search for more papers by this author
Bin Peng

Bin Peng

Department of Natural Resources and Environmental Sciences, University of Illinois at Urbana-Champaign, Urbana, IL, USA

National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign, Urbana, IL, USA

Search for more papers by this author
Christian Frankenberg

Corresponding Author

Christian Frankenberg

Division of Geological and Planetary Sciences, California Institute of Technology, Pasadena, CA, USA

Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, USA

Correspondence to: L. He and C. Frankenberg,

[email protected];

[email protected]

Search for more papers by this author
First published: 09 March 2020
Citations: 81
This article was corrected on 20 MAY 2020. See the end of the full text for details.

Abstract

Timely and accurate monitoring of crops is essential for food security. Here, we examine how well solar-induced chlorophyll fluorescence (SIF) can inform crop productivity across the United States. Based on tower-level observations and process-based modeling, we find highly linear gross primary production (GPP):SIF relationships for C4 crops, while C3 crops show some saturation of GPP at high light when SIF continues to increase. C4 crops yield higher GPP:SIF ratios (30–50%) primarily because SIF is most sensitive to the light reactions (does not account for photorespiration). Scaling to the satellite, we compare SIF from the TROPOspheric Monitoring Instrument (TROPOMI) against tower-derived GPP and county-level crop statistics. Temporally, TROPOMI SIF strongly agrees with GPP observations upscaled across a corn and soybean dominated cropland (R2 = 0.89). Spatially, county-level TROPOMI SIF correlates with crop productivity (R2 = 0.72; 0.86 when accounting for planted area and C3/C4 contributions), highlighting the potential of SIF for reliable crop monitoring.

Key Points

  • The photosynthetic pathway (C3, C4) impacts the relationship between CO2 uptake and SIF, which helps to interpret satellite signals
  • TROPOMI SIF agrees well with the seasonality of crop gross primary production (GPP) when accounting for C3/C4 fractionation
  • TROPOMI SIF is highly correlated with USDA reported crop productivity at the county scale

Plain Language Summary

Crop monitoring is essential for ensuring food security, but reliable, instantaneous production estimates at the global scale are lacking. The monitoring of crop production in a changing climate is of paramount importance to sustainable food security. Accurate estimates of crop production are dependent on adequately quantifying crop photosynthesis. Our paper demonstrates that solar-induced chlorophyll fluorescence (SIF), an emission of red to far-red light from chlorophyll is highly correlated with crop photosynthesis. We show that a new high spatial resolution satellite SIF data set is highly correlated with crop productivity in the United States, which is benchmarked by the United States Department of Agriculture county-level crop statistics. These results will improve the understanding of crop production and carbon flux over agricultural lands, as well as provide an accurate, large-scale, and timely monitoring method for global crop production estimates.

1 Introduction

Cropping systems not only provide sustenance for the world's human population and livestock, but they also have a major impact on both local climate (Mueller et al., 2016) and the carbon cycle (Peters et al., 2007). Hence, large-scale crop monitoring and yield forecasting are necessary to support food security and to quantify the overall impact on the climate and carbon cycle.

Crop productivity is reliant upon its ability to convert light energy into sugar via photosynthesis. Therefore, a reliable measure of gross primary production (GPP) is a key step toward crop monitoring. Satellite observations have the potential to provide GPP estimates from regional to global scales (Running et al., 2004; Ryu et al., 2019; Turner et al., 2006; Yuan et al., 2007; Zhao et al., 2005). Most remote sensing-based GPP estimates use spectral information in the visible and near infrared (NIR) regions that are related to greenness (Huete et al., 2002) or thermal (Anderson et al., 2012) and microwave bands, which are sensitive to vegetation/soil water content (Guan et al., 2017; Konings et al., 2017). In most cases, predicting GPP largely depends on empirical estimates of light use efficiency (LUE, efficiency with which absorbed light is used for CO2 fixation), which is highly uncertain.

The emission of red and far-red light from excited chlorophyll-a molecules, denoted solar-induced chlorophyll fluorescence (SIF), offers a physiologically based GPP proxy, potentially avoiding the need for LUE parameterizations. In fact, an empirical linear scaling between SIF and GPP across various vegetation types under different environmental conditions has already been shown in many studies (Frankenberg et al., 2011; Guanter et al., 2014; Guan et al., 2016; Li et al., 2018; Magney, Bowling, et al., 2019; Smith et al., 2018; Song et al., 2018; Sun et al., 2018; Verma et al., 2017; Wood et al., 2017; Zhang et al., 2016; Zuromski et al., 2018), revealing the potential of SIF to monitor GPP across all spatial scales.

The rationale of linking SIF with crop productivity can be shown using the following equations (Guan et al., 2016):
urn:x-wiley:00948276:media:grl60362:grl60362-math-0001(1)
urn:x-wiley:00948276:media:grl60362:grl60362-math-0002(2)
urn:x-wiley:00948276:media:grl60362:grl60362-math-0003(3)
where means the temporal integration of the variable over the growth period. NPP refers to net primary production, calculated as GPP minus the amount of carbon consumed by plants by autotrophic respiration (Ra), which includes maintenance and growth respiration. During the growing period, we can assume that growth respiration dominates the Ra term, which in theory should scale with GPP. CUE is the carbon use efficiency, which varies across species and environmental conditions (Amthor, 1989; DeLucia et al., 2007). fAG is the fraction of the aboveground to total biomass and HI is the harvest index—the mass of harvested grain divided by total aboveground biomass. Both fAG and HI are related to the crop type and environmental conditions but are usually treated as constant parameters for individual crops. These equations exhibit the underlying relationship among SIF, NPP, GPP, and crop productivity and show that even if GPP and SIF are perfectly correlated, variations in CUE, fAG, and HI can still impact the relationship of SIF to crop productivity.

Some efforts have been made to explore the potential of satellite SIF in estimating crop productivity. Guanter et al. (2014) found that the highest SIF values observed from the Global Ozone Monitoring Experiment-2 (GOME-2) satellite are associated with the Corn Belt in the United States Midwest. In addition, SIF captured the photosynthetic activity over highly productive croplands, while traditional vegetation indices (VIs) show saturation effects in dense canopies. In some cropping systems, however, the use of NIRv (Badgley et al., 2017) might track productivity quite well (Dechant et al., 2019) but may be less sensitive to rapid changes in PSII operating efficiency. Thus, SIF has its own unique advantage to track crop photosynthesis over VIs. Guan et al. (2016) used GOME-2 SIF as an approximation of photosynthetic electron transport rate (ETR) to derive GPP and crop yield, which has shown a significant improvement in county-level crop yield estimates. Zhang et al. (2014) proposed the utility of space-based SIF measurements, combined with a process-based model, to estimate the photosynthetic capacity over six crop flux sites in the United States. However, direct comparisons of SIF and crop yields have been restricted by the coarse spatial resolution of available satellite SIF measurements. In the United States, the average area of counties that are predominantly agricultural is ~1,700 km2, with the smallest one ~430 km2. A single GOME-2 footprint (80 km × 40 km) thus covers several counties, making it challenging to compare SIF with the benchmark of county-level annual crop statistics from the National Agricultural Statistics Service (NASS) of the United States Department of Agriculture (USDA). Fine spatial resolution (1.3 km × 2.25 km) SIF data from the Orbiting Carbon Observatory (OCO-2) has been available since September 2014 (Frankenberg et al., 2015; Sun et al., 2017) but does not provide contiguous spatial coverage. To fill this knowledge gap, we leverage SIF inferred from measurements of the TROPOspheric Monitoring Instrument (TROPOMI), with an unprecedented spatial resolution (up to 7 km × 3.5 km) and near-global daily coverage (Köhler et al., 2018), which allows us to achieve robust SIF averages for individual counties.

A thorough understanding of the relationship between SIF and GPP and how it might vary depending on photosynthetic pathway (C3, C4) is important to achieve accurate estimates of crop photosynthesis from SIF. Some studies found stronger relationships between SIF and absorbed light by chlorophyll, than GPP (Li et al., 2020; Yang et al., 2018). As SIF emanates from the light reactions of photosynthesis, it is expected to be more closely related to the ETR in photosystem II than to GPP (Frankenberg et al., 2011; Gu et al., 2019; Porcar-Castell et al., 2014). Thus, the GPP:SIF ratio should vary between C3 and C4 photosynthetic pathways (Gu et al., 2019; Guan et al., 2016; Porcar-Castell et al., 2014), because it is important to know how efficiently electrons from the light reactions are used for carbon fixation. This is fundamentally different for C3 and C4 plants, with the latter inhibiting photorespiration, which generally leads to a higher and less variable efficiency. Typical C3 crops include soybeans, wheat, barley, oats, and rice, whereas typical C4 crops include corn, sugarcane, and sorghum. Several recent studies point to a different GPP:SIF ratio for C3 and C4 plants (Li et al., 2018; Liu et al., 2017; Wood et al., 2017); however, there is still limited understanding of how exactly these mechanisms influence the GPP:SIF relationship across time and space. To fill this knowledge gap, we include both field measurements and a biophysical model to explain the discrepancies of GPP:SIF of C3 and C4 crops and use this to inform our interpretation of the satellite data.

In this paper, the objectives are to (i) compare observed and modeled GPP:SIF relationships for C3 and C4 crops at site level, (ii) examine whether aggregated TROPOMI SIF at local (10 km) scales can represent the seasonality of GPP observations over homogenous croplands, and (iii) evaluate how well aggregated TROPOMI SIF at the county level can be used to estimate crop productivity and NPP in the United States.

2 Data and Methods

2.1 Site-Level SIF and GPP Observations

2.1.1 Eddy-Covariance GPP

In this paper, eddy-covariance (EC) data were collected at long-term and well-characterized USDA agricultural flux towers within C3 (soybeans, Glycine max L. Merr.) and C4 (corn, Zea mays L.) cropping systems in central Iowa. Specific sites are Brooks Field (41.974536°N, −93.693711°W) and Coles Field (42.488414°N, −93.522582°W) for the PhotoSpec comparison in 2017 as well as two nearby towers at Coles Field covering soy (42.488414°N, −93.522582°W) and corn (42.481677°N, −93.523521°W) for the TROPOMI comparison in 2018. These farming systems are typical for those in the Upper Midwest corn belt (more details in Dold et al., 20172019). Data were excluded under unfavorable weather conditions (e.g., rainfall, low wind turbulence, and high humidity) (Baker & Griffis, 2005), screened for outliers (Dold et al., 2017), and gap-filled. Turbulent CO2 fluxes were computed using the EC method (Burba, 2013), and then net ecosystem exchange was partitioned into GPP and ecosystem respiration (Re). Note that we do not have overlapping data sets of TROPOMI, PhotoSpec and EC systems; thus, the EC to PhotoSpec comparison is limited to 2017 and the EC to TROPOMI comparison to 2018.

2.1.2 PhotoSpec SIF

We installed two PhotoSpec instruments in a soybean (Brooks Field, 41.974203°N, −93.695839°W) and corn field (Coles Field, 42.48655°N, −93.52641°W) in central Iowa following planting in late May 2017 and measured SIF until harvest in September 2017 (Magney, Frankenberg, et al., 2019). PhotoSpec consists of a 2D scanning telescope to guide reflected radiances into a set of high-resolution spectrometers (Grossmann et al., 2018) to infer SIF using the solar Fraunhofer line in-filling technique, similar to all current satellite retrievals. The scanning telescope was placed atop a 7-m tower, and we calculate a “canopy average” of all viewing angles at an hourly time step to match the temporal resolution of flux tower data (following Magney, Frankenberg, et al., 2019). Because the scanning telescope revisits every measuring point within an hour, this canopy average is more representative of a hemispherical sensor with a footprint of a few meters around the tower (an average of all viewing directions). Due to the narrow field of view of PhotoSpec, we can isolate between vegetation and nonvegetation signals, where we have determined NDVI at 0.6 to be sufficient for the definition of the peak growing season, when our analysis was conducted (as can be seen in Figure S1 in the supporting information, where the rows can be seen early in the season). Notably, while the escape ratio and angular dependencies are an important consideration for tower-based SIF measurements (Zeng et al., 2019), this makes little difference in our study for the canopy structure is relatively stable during the peaking growing season. More details on PhotoSpec retrievals and instrument specifications can be found in Grossmann et al. (2018) and Magney, Frankenberg, et al. (2019).

2.2 TROPOMI SIF

On 13 October 2017, the TROPOMI instrument onboard the European Sentinel-5 Precursor satellite was launched. TROPOMI is a multiband push-broom imaging grating spectrometer, which also covers the far-red part of the SIF emission spectrum (Köhler et al., 2018). The SIF retrieval window ranges from 743 to 758 nm, a subset of TROPOMI's band 6 (725–775 nm). A daily correction factor is applied to convert instantaneous SIF to a daily average (following Frankenberg et al., 2011).

For the comparison with GPP observations, we extracted TROPOMI SIF within +/−10 km of two nearby flux towers and applied a +/−4-day moving average, sampled every 4 days. The agricultural cover is quite homogeneous within +/−10 km, with ~58% corn and ~30% soybeans derived from the 30-m cropland data layer (CDL; https://nassgeodata.gmu.edu/CropScape/) (Figure S2). Since TROPOMI SIF represents an average of all crops within the sampling area, we assume a ⅔:⅓ weighted average of flux tower GPP for corn and soy, respectively.

For the comparison with county-level crop statistics reported by USDA NASS, we aggregated SIF at the county scale, yielding on average 700 TROPOMI soundings per month for counties with >45% planted areas. The planted ratio of each county is shown in Figure S3.

2.3 Crop Statistics From USDA NASS

We obtained the county-level total crop production and planted area for individual crop types from the USDA NASS Quick Stats Database (quickstats.nass.usda.gov). In most studies, crop yield refers to the amount of grain per unit of harvested land area per crop (Fischer, 2015). Therefore, it is associated with accumulated NPP weighted by aboveground biomass during the crop growth season (Lobell et al., 2002). Thus, NPP acts as a bridge to link crop statistics with satellite observations (Smith et al., 2014).

We used crop production and acreage for individual crop types to calculate county-level crop productivity as urn:x-wiley:00948276:media:grl60362:grl60362-math-0004. From a physical perspective, this definition is more comparable with the county-level aggregated satellite TROPOMI SIF, since satellite SIF records represent the entire footprint, regardless of land type.

Following Lobell et al. (2002), Guan et al. (2016), and Guan et al. (2017), we converted crop productivity to NPP at the county level:
urn:x-wiley:00948276:media:grl60362:grl60362-math-0005(4)
urn:x-wiley:00948276:media:grl60362:grl60362-math-0006(5)
where i represents the crop type (dominated by corn and soybean here), P the reported crop production, MRP the mass per unit of report production, MC the moisture content, and Y the crop productivity based on the definition given before. The list of fAG and HI of common crops can be found in Lobell et al. (2002) and Guan et al. (2016) (see Table S1).

2.4 Site-Level GPP and SIF Modeling

We use the Soil Canopy Observation, Photochemistry and Energy fluxes (SCOPE) model, a 1D integrated radiative transfer and energy balance model, to simulate photosynthesis, fluorescence, the surface energy balance, and reflectance/emission spectra at leaf and canopy scales (van der Tol et al., 2009). This mechanistic model helps to understand the connection between SIF, GPP, quantum yield of photosystem II (PSII) (PSIIyield) and actual ETR (Ja) under different environmental conditions for C3 and C4 plants. We modeled GPP and SIF at nadir viewing using SCOPE at two flux tower sites with concurrent PhotoSpec and GPP observations. We used available meteorological data from the flux tower (i.e., incoming shortwave and longwave radiation, air temperature, ambient atmospheric pressure, vapor pressure, and wind speed) and optimized canopy parameters (leaf area index, VCMAX, and Chl content) via a Bayesian inversion system (Dutta et al., 2019).

3 Results and Discussion

3.1 Observed and Modeled GPP:SIF for C3 and C4 at the Site Level

We compared ground-based PhotoSpec SIF measurements throughout the peak growing season in both C3 (soybean) and C4 (corn) crops with flux tower GPP as well as photosynthetically active radiation (PAR) (Figures 1a1c). The peak growing season is defined as DOY 210-235 for soybean and DOY 175-190 for corn. Here, hourly SIF is linearly correlated with PAR for soybean (R2 = 0.87) and corn (R2 = 0.84), with negligible differences in the SIF:PAR ratio (Figure 1a). This indicates that during the peak growing season in both agricultural sites, SIF is mostly driven by incident radiation and exhibits negligible differences due to the fluorescence yield. To first order, we expect SIF to be tightly related to PAR in systems where there is little stress (i.e., low variations in SIFyield), which can be expected in highly efficient agricultural areas. Little apparent stress is also indicated by the canopy-scale light response curve of GPP vs. PAR (Figure 1b), with a near-linear relationship for corn and some saturation effects for soybean at higher light levels. Consequently, we found that hourly averaged SIF is strongly correlated with GPP in both soybean (R2 = 0.51) and corn (R2 = 0.64), however, with a larger GPP:SIF ratio observed for corn, leading to roughly a 30% higher GPP per unit SIF in corn at higher light levels (Figure 1c). This implies that the light-use efficiency of corn is higher than that of soybean, due to the evolved mechanism to concentrate CO2 at the rubisco site and thereby minimize photorespiration as well as the occurrence of carboxylation limited photosynthetic rates in C4 plants.

Details are in the caption following the image
Canopy-scale relationships in the peak growing season (DOY 175-190 for corn and 210-235 for soybean) based on hourly averaged data from PhotoSpec for soybean (gold) and corn (green): (a) SIF:PAR, (b) GPP:PAR, and (c) GPP:SIF for both soybeans and corn. SCOPE modeled relationships are for the same time-periods and based on hourly averaged data for soybean and corn: (d) SIF:PAR, (e) GPP:PAR, and (f) GPP:SIF. The results of PhotoSpec-measured relative SIF, which is normalized by incoming NIR-reflected radiance to reduce the effects of structural and bidirectional reflectance of the signal (Yang et al., 2018), are shown in Figure S4.

To test whether the GPP:SIF relationship observed in the field is consistent with what is expected from biophysical models, we compared the simulated far-red SIF at 740 nm and GPP against PAR using SCOPE (Figures 1d1f). For both soybean and corn, the PAR:SIF relationship is highly linear yet with slightly different slopes (Figure 1d). In the light response curve for GPP (Figure 1e), soybean GPP saturates at higher light levels, when the rate of photosynthesis is limited by the carboxylation rate. Corn remains highly linear even at high light levels, resulting in a more linear GPP:SIF relationship. Apart from the high SIF values modeled for corn leading to diverging SIF:PAR relationships between corn and soy, these simulations are broadly consistent with our observational evidence. We found that the high SIF to PAR slope for C4 is likely related to how SCOPE computes the PSIIyield for C4 plants, which are in the range of a maximum SIFyield given by the PSIIyield:SIFyield parameterization (Van der Tol et al., 2014). Our canopy spectrometer data supports this, as we observe similar SIF yields for C3 and C4 crops. The overly high modeled SIF values as well as the inconsistency with our field measurements point to an overestimation in the SIFyield for C4 plants in the SCOPE model. Apart from that, the measurement and modeling perspective agree very well.

Both our observations at the site level as well as the SCOPE modeling results showed that the GPP:SIF relationship is more linear for C4 than C3 plants. It is well established that SIF is a better proxy for the actual ETR (Ja) than for GPP, which requires us to separate C3 and C4 photosynthetic mechanisms if we use Ja to estimate GPP. The explicit relationships among the four variables, SIF, Ja, PSIIyield, and SIFyield, can be described as follows (Gu et al., 2019; Porcar-Castell et al., 2014):
urn:x-wiley:00948276:media:grl60362:grl60362-math-0007(6)
urn:x-wiley:00948276:media:grl60362:grl60362-math-0008(7)
where ΦPSII and ΦSIF is the quantum yield of PSII and SIF, respectively; APARgreen is absorbed PAR by green elements (Gitelson & Gamon, 2015); β is the fraction of APARgreen allocated to PSII; ε is the escape probability of fluorescence from the canopy (Zeng et al., 2019). Here, we neglect variations in the canopy structure and thus ε as we focus on the peak growing season, where we observe little structural change (Figure S1).
Combing 6 and 7, we obtain
urn:x-wiley:00948276:media:grl60362:grl60362-math-0009(8)
which demonstrates that Ja:SIF is determined by ΦPSII and ΦSIF, which is dependent on environmental conditions (e.g., light intensity, drought, and heat stress). The less ΦPSII is varying—as is often the case for highly efficient crops—the more constant and linear the SIF to Ja relationship is. Leaf-level measurements have been used to derive the empirical relationship between ΦPSII and ΦSIF (Lee et al., 2013; Van der Tol et al., 2014). For example, Flexas et al. (2002) found that for under stress, ΦPSII and ΦSIF are usually positively correlated if substantial nonphotochemical quenching exists. At the canopy level, ΦPSII and ΦSIF are mostly positively correlated at the seasonal scale (Porcar-Castell et al., 2014; Song et al., 2018; Verma et al., 2017; Yang et al., 2015; Zhang et al., 2016), mainly due to averaging effects of the canopy, which integrates over a variety of ΦPSII and ΦSIF values for all leaves (similar to our PhotoSpec observations).
The linkage between GPP and Ja depends on the photosynthetic pathway (Collatz et al., 1992; Farquhar et al., 1980):
urn:x-wiley:00948276:media:grl60362:grl60362-math-0010(9)
urn:x-wiley:00948276:media:grl60362:grl60362-math-0011(10)
where Cc is chloroplast CO2 partial pressure and Γ* is chloroplastic photorespiratory CO2 compensation point. Comparing 9 and 10, for C3 plants, an additional nonlinearity in GPP:Ja arises, which is attributed to the varying fraction of electrons used in photorespiration. Due to the lack of photorespiration in C4 plants, the relationship between Ja and GPP is strictly linear. For the same reason, the conversion factor from Ja to GPP is mostly higher for C4 than C3 plants, as about ⅓ of the ETR is wasted in photorespiration in C3 plants. Since SIF is mostly proportional to Ja, 9 and 10 directly explain the more linear behavior and steeper GPP:SIF slope for C4 versus C3 plants. It should be noted that at high Cc values, the slope for C3 plants can be steeper (e.g., at low light levels). In general, variations in the GPP:SIF scaling at the leaf scale can thus be attributed to two different effects: (i) variations in the ratio of ΦPSII and ΦSIF in the light reactions as well as (ii) the scaling from Ja to GPP, which depends on the photosynthetic pathway.

3.2 Consistent Seasonal Cycle Seen by TROPOMI SIF and Field GPP Measurements

Here, we evaluate how well TROPOMI SIF represents local GPP patterns in a corn/soybean dominated area in Iowa during the growing season of 2018. The seasonal cycle of footprint-level TROPOMI SIF and flux tower GPP at both soybean and corn sites is shown in Figure 2. Corn GPP increases more rapidly and reaches its maximum of ~25 gC/m2/day around 10 July, while soybean GPP approaches maximum of ~15 gC/m2/day more gradually, around 20 July. As mentioned in Section 2.2, TROPOMI SIF represents an average of all crops in the area, for which we calculate the weighted average of GPP assuming a ⅔:⅓ contribution of corn and soy, respectively. There is strong agreement between TROPOMI SIF and the weighted area-averaged GPP estimate (R2 = 0.89) (Figure 2), motivating our use of TROPOMI SIF in large-scale crop monitoring. In particular, the GPP:SIF ratio at both satellite and flux tower scales is consistent—around 12 (gC/m2/day)/(W/m2/μm/sr). While there is close correspondence between SIF and GPP during senescence, there is a small mismatch early in the growing season, which could be explained by (1) fewer TROPOMI observations due to increased cloud cover during this time of year (see Figure S5), (2) different planting and emergence dates within the satellite footprint, and (3) potentially higher SIF escape probabilities at the beginning of the season as the canopy is more open at that stage. We also did not find any apparent viewing angle dependent behavior in the TROPOMI data sets (see Figure S5). Importantly, it should be noted that this is the first time that satellite-based SIF can be compared with GPP at weekly temporal resolution and ~10 km spatial scales—a significant improvement over previous satellites providing monthly data at 0.5° resolution.

Details are in the caption following the image
The 2018 seasonal cycle of flux tower GPP for soybeans and corn in Iowa and TROPOMI SIF (within +/−0.1° latitude and longitude, shaded area showing 2 sigma uncertainty range). A moving average of +/−4 days and 4-day sampling interval is applied to both data sets. The upscaled GPP (green) is approximated by a weighted average of ⅔ corn and ⅓ soy contributions, determined by computing crop fraction within +/−0.1° (in total, 88% of this area is covered by either corn and soybean).

Tower-level SIF was not included in this comparison as there was no temporal overlap between tower SIF data and TROPOMI in 2018. In recent years, more ground-based SIF measurements have become available (Dechant et al., 2019; Du et al., 2019; Grossmann et al., 2018; Li et al., 2020; Magney, Frankenberg, et al., 2019; Miao et al., 2018; Zhang et al., 2019; Yang et al., 2015; Yang et al., 2018), which will help disentangle various factors (e.g., view angles, canopy structure, and overpass times) that might also impact the SIF to GPP relationship.

3.3 Spatial Correlation Between County-Level TROPOMI SIF and Crop Productivity

Here, we focus on the county level in the United States to evaluate the capability of using TROPOMI SIF to predict crop productivity. We define the growth period for each county as the period in which SIF exceeds 10% of its peak value within the year. The spatial pattern of crop productivity and the average growing-season TROPOMI SIF of 2018 is shown in Figures 3a and 3c. The SIF map closely matches the NASS yield-based NPP pattern in the Corn Belt, the most dominant region of corn/soybean staple crop production in the United States. For counties with >45% planted areas, we found that crop productivity and converted NPP are both highly correlated with SIF, with R2 = 0.72 and R2 = 0.71, respectively (Figures 3b and 3d). As discussed previously, crop productivity: SIF and NPP:SIF relationships are not identical for C3 and C4 crops. At the same level of SIF, higher crop productivity and NPP are observed for counties with more C4, which is expected based on the ground-based measurements and modeling results (Sections 3.1). In contrast to the nonlinear pattern observed for temporal variations at the site level, spatial variations in the county-level analysis are more linear. The main reason is that the nonlinear pattern is only apparent at high light levels. In contrast, the spatial correlation is discussed for seasonal averages, in which the impact of high light levels is smoothed out.

Details are in the caption following the image
Spatial pattern of county-scale (a) crop productivity derived from USDA NASS and (c) average TROPOMI SIF during the growth period of 2018. In counties with planted area >45%, the relationship of (b) crop productivity: SIF and (d) NPP:SIF. The color scheme represents the relative fraction of C3 crops, with blue meaning more C3 and red means more C4. The maps of C3 and C4 crops distributions are shown in Figure S6. The growth period here is defined as the time when SIF exceeds 10% of max (SIF) within a year. The results are robust with different thresholds defining the growing season (see Figure S7) and TROPOMI relative SIF (see Figure S8).

In theory, crop productivity is associated with temporally integrated net photosynthesis over the growing season. Therefore, we also tested the correlation between crop productivity and the corresponding NPP against integrated SIF over the growing period. We found that the crop productivity:integrated SIF and NPP:integrated SIF correlations are still consistent and robust yet weaker than when using the average growing season SIF (see Figure S9).

Overall, the crop productivity (NPP):SIF relationship at the county level is almost linear. This is likely due to the high efficiency of crops (generally high photosynthetic yields) as well as large contributions from C4 crops, where the relationship is more linear. In addition, crop productivity (NPP):SIF ratio is 25–50% higher for C4 than C3, which is consistent with ground-based measurements and modeling outputs in Sections 3.1. Including results from field measurements (PhotoSpec) and a biophysical model (SCOPE) give us more confidence in explaining the discrepancies between C3 and C4 crops and the associated GPP:SIF relationship. A more detailed understanding of the GPP:SIF slope for different crops at different environmental conditions based on field-level measurements would help us to better estimate NPP:SIF for a variety of crops and ultimately improve the prediction of crop productivity at the global scale, where ground-based measurements are lacking.

3.4 Improving the Productivity Prediction Performance Using Additional Information

Although crop productivity can be estimated with SIF observations alone, we can further improve the performance by adding two other explanatory variables, which are easily accessible: planted area fraction per county (Cropfraction) and fraction of C3 plants per county (C3ratio). The SIF signal from crops will be dampened if the planted area fraction is low, and a correction factor should be considered between C4 and C3, since higher productivity is expected for C4 at the same level of SIF. Here, we fit the two multiple linear regression models for crop productivity based on (i) SIF and Cropfraction (ii) SIF, Cropfraction, and C3ratio, (Figure 4). With the added information of Cropfraction, R2 improved from 0.72 to 0.79, and adding the C3ratio, R2 improved to 0.86 (Figures 3 and 4), which also converges into a more normal distribution of the residuals for C3 and C4 dominated counties around the line of best fit. This behavior is consistent with our field-level measurements and modeling, given us mechanistic confidence in such an approach.

Details are in the caption following the image
The relationships between true crop productivity and predicted crop productivity using multiple linear regression based on variables (a) SIF and crop fraction per county and (b) SIF, crop fraction, and relative C3 ratio per county.

It should be noted that we currently use several assumptions and do not use any model information for this upscaling approach. For instance, we assume that respiration scales with GPP, which results in a linear scaling of GPP to NPP. Notably, other growing season environmental conditions (temperature,  vapor pressure deficit, and soil moisture) as well as canopy structure and associated variations in the escape probability of SIF could have an impact on our interpretation of the SIF signal. Theoretically, however, SIF contains the information on environmental conditions and tracks GPP variations better than traditional reflectance-based VIs (e.g., NDVI), though its response to stress is weaker than for GPP (or LUE) (Magney et al., 2017; Van der Tol et al., 2014). Additionally, HI, the process of plant NPP converting to ultimate grain yield, which is also controlled by reproductive processes and environmental conditions during grain fill, is an important consideration. Despite this, it is hard to justify adding more complexity to the analysis given the already strong empirical linear relationship between crop NPP and SIF.

3.5 Toward Finer-Resolution Crop Productivity Estimates

Our SIF-only-based crop productivity proxy approach achieved robust and reliable performance over agriculture dominated counties in the United States. To fully exploit the potential of this approach in other regions of the world, especially for small-scale, diverse, and fragmented agricultural fields, high spatially and temporally resolved remote sensing data sets will be required. This is still challenging for current satellite SIF products (e.g., short record of TROPOMI, sparse spatiotemporal sampling of OCO-2 and large footprint of GOME-2). Recently, some machine learning-based long-term SIF products of high spatiotemporal resolution have been developed (e.g., CSIF, GOSIF, RSIF, and SIFoco2) based on satellite SIF, high-resolution spectral information, and meteorology data (Gentine & Alemohammad, 2018; Yu et al., 2019; Zhang et al., 2018; Li & Xiao, 2019; Turner et al., 2020). Future work across all scales will be needed to test the long-term performance of SIF products on the global scale, particularly as we link these measurements to photosynthesis (Ryu et al., 2019). A combination of high-resolution reflectance-based remote sensing products and crop models (Jin et al., 2017; Lobell et al., 2015) has already shown great potential in fine-scale yield estimates.

4 Conclusions

We investigated the GPP:SIF relationship of crops from ground-based measurements and validated the biophysics with model runs for C3 and C4 crops at two field sites in the Upper Midwest corn belt of the United States during the 2017 growing season. Generally, we find a linear GPP:SIF relationship except for high-light levels when GPP starts to saturate, whereas SIF still increases, especially for C3 crops. This is attributed to two effects: (1) urn:x-wiley:00948276:media:grl60362:grl60362-math-0012 of crops varies less than in other ecosystems, since crops in this region rarely experienced severe stress conditions, which would lead to a strong reduction of urn:x-wiley:00948276:media:grl60362:grl60362-math-0013. Thus, the correspondence between SIF and actual ETR (Ja) is highly linear for crops; (2) Ja:GPP varies between C3 and C4, which is determined by the number of electrons required for carboxylation. The latter consideration also results in a different GPP:SIF relationship between C3 and C4 crops, with a steeper GPP:SIF slope for C4 plants—an important consideration when using SIF as a proxy for crop productivity. A direct comparison of satellite SIF measurements against two flux tower sites in a highly agricultural area in Iowa further supports this finding at larger spatial scales. At the county scale, we found that SIF is highly correlated with crop productivity derived from the USDA NASS database (R2 = 0.72). Using ancillary information on crop-planted fraction and relative C3 crop ratio per county, we can further improve our estimate using a multiple linear regression model to R2 = 0.86. Our model-free SIF-based crop productivity estimation framework appears promising and can provide insights to monitor the crop productivity globally outside the United States, especially in developing agricultural countries. A validation approach against well-documented productivity estimates in the United States represented a necessary proof of concept before applying the method to less well-monitored agricultural areas.

Acknowledgments

Part of this research was funded by the NASA Carbon Cycle Science program (grant NNX17AE14G). L.H. thanks the Resnick Sustainability Institute at Caltech for fellowship support. TROPOMI SIF data generation by P.K. and C.F. is funded by the Earth Science U.S. Participating Investigator program (grant NNX15AH95G). TROPOMI SIF product is available at ftp://fluo.gps.caltech.edu/data/tropomi/. County-level crop statistics is available at the USDA NASS Quick Stats Database (quickstats.nass.usda.gov). Crop-specific land cover data layer CDL can be accessed at https://nassgeodata.gmu.edu/CropScape/.

    Erratum

    The originally published Figure 4 mistakenly included two copies of panel A. The figure has since been corrected, and the present version may be considered the authoritative version of record.