Satellite-based reflectances capture large fraction of variability in global gross primary production (GPP) at weekly time scales

Highlights

• Seven MODIS bands and potential PAR capture a large fraction (77%) of GPP variability.

• Meteorological data has little impact on upscaled GPP with high quality satellite data.

• Machine learning can be used to estimate uncertainties in upscaled GPP.

• Our global mean GPP (142.5+/-7.7 PgC/y) implicitly includes radiation effects on LUE.

• Vegetation indices may reduce information content of reflectances in compression.

Abstract

A perception has emerged, based on several studies, that satellite-based reflectances are limited in terms of their ability to predict gross primary production (GPP) globally at weekly temporal scales. The basis for this inference is in part that reflectances, particularly expressed in the form of vegetation indices (VIs), convey information about potential rather than actual photosynthesis, and they are sensitive to non-green substances (e.g., soil, woody branches, and snow) as well as to chlorophyll. Previous works have suggested that processing and quality control of satellite-based reflectance data play an important role in their interpretation. In this study, we use high quality reflectance data from the MODerate-resolution Imaging Spectroradiometer (MODIS) data to train neural networks that are used to upscale GPP estimated from eddy covariance flux tower measurements globally. We quantify the ability of the machine learning approaches to capture GPP variability at daily to interannual time scales. Our results show that MODIS reflectances, when paired only with potential short-wave radiation, are able to capture a large fraction of GPP variability (approximately 77%) at daily to weekly time scales. Additional meteorological information (temperature, water vapor deficit, soil water content, ET, and incident radiation) captures only a few more percent of the GPP variability. The meteorological information is used most effectively when information about plant functional type and climate classification is included. We show that machine learning can be a useful tool for estimating GPP uncertainties as well as GPP itself from upscaling methods. Our estimated global annual mean GPP for 2007 is 142.5 ± 7.7 Pg C y${}^{-1},$ which is higher than some other satellite-based estimates but within the range of other reported observation-, model-, and hybrid-based values.

Introduction

Gross primary production (GPP), the amount of carbon dioxide (CO₂) that is assimilated by plants through photosynthesis, is one of the most variable and uncertain components of the global carbon cycle (e.g., Anav et al., 2015). Global GPP has been estimated with a number of different process-based models, data-driven, and hybrid approaches (e.g., Anav et al., 2015; Anav, Friedlingstein, Kidston, Bopp, Ciais, Cox, Jones, Jung, Myneni, Zhu, 2013, Beer, Reichstein, Tomelleri, Ciais, Jung, Carvalhais, Rödenbeck, Arain, Baldocchi, Bonan, Bondeau, Cescatti, Lasslop, Lindroth, Lomas, Luyssaert, Margolis, Oleson, Roupsard, Veenendaal, Viovy, Williams, Woodward, Papale, 2010, Koffi, Rayner, Scholze, Beer, 2012, Parazoo, Bowman, Fisher, Frankenberg, Jones, Cescatti, Pérez-Priego, Wohlfahrt, Montagnani, 2014; Schaefer et al., 2012). In particular, dynamic global vegetation models, driven by observed environmental changes, are used for global carbon budget assessments (Friedlingstein et al., 2019). Benchmarking these models globally with data-driven GPP estimates is critical for understanding the land sink (e.g. Baldocchi, 2018, Peters, Le Quéré, Andrew, Canadell, Friedlingstein, Ilyina, Jackson, Joos, Korsbakken, McKinley, Sitch, Tans, 2017).

Data-driven GPP estimates have become a primary tool for GPP studies as well as model evaluation. One such approach to estimate GPP involves upscaling relatively sparse flux tower eddy covariance GPP estimates with machine learning (ML) methods that use satellite and other meteorological data as inputs (e.g., Jung, Reichstein, Bondeau, 2009, Jung, Reichstein, Margolis, Cescatti, Richardson, Arain, Arneth, Bernhofer, Bonal, Chen, Gianelle, Gobron, Kiely, Kutsch, Lasslop, Law, Lindroth, Merbold, Montagnani, Moors, Papale, Sottocornola, Vaccari, Williams, 2011, Stocker, Zscheischler, Keenan, Prentice, Peñuelas, Seneviratne, 2018, Tramontana, Ichii, Camps-Valls, Tomelleri, Papale, 2015, Tramontana, Jung, Schwalm, Ichii, Camps-Valls, Ráduly, Reichstein, Arain, Cescatti, Kiely, Merbold, Serrano-Ortiz, Sickert, Wolf, Papale, 2016, Wei, Yi, Fang, Hendrey, 2017). It has been shown that ML models that use only remotely sensed data perform as well as models that in addition use meteorological data (Tramontana, Ichii, Camps-Valls, Tomelleri, Papale, 2015, Tramontana, Jung, Schwalm, Ichii, Camps-Valls, Ráduly, Reichstein, Arain, Cescatti, Kiely, Merbold, Serrano-Ortiz, Sickert, Wolf, Papale, 2016).

Some ML methods have been implemented using the framework of the light-use efficiency (LUE) model (Monteith, 1972, Monteith, 1977) expressed by

$$\text{GPP}={\text{APAR}}_{\text{canopy}}\times \text{LUE},$$

where

$${\text{APAR}}_{\text{canopy}}={\text{FAPAR}}_{\text{canopy}}\times {\text{PAR}}_{\text{inc}},$$

PAR_inc is photosynthetically active radiation incident at the top of canopy, and FAPAR_canopy is the fraction of PAR_inc absorbed by the canopy. While the LUE model is attractive in its simplicity and the ease in which satellite data can be incorporated, it should be noted that its parameters only approximate processes inherent in the real-world complex system and furthermore depend on the operational definitions of the component terms in Eq. 3 (Gitelson and Gamon, 2015). For example, as a canopy is composed of components containing chlorophyll, other pigments, and non-photosynthetic vegetation, the LUE model can be revised as

$$\text{GPP}={\text{APAR}}_{\text{green}}\times {\text{LUE}}_{\text{green}},$$

(e.g., Gitelson and Gamon, 2015) where

$${\text{APAR}}_{\text{green}}={\text{FAPAR}}_{\text{chl}}\times {\text{PAR}}_{\text{inc}},$$

and FAPAR_chl is the fraction of PAR_inc absorbed by chlorophyll.

Another class of data-driven GPP models relies more explicitly on the LUE model without machine learning. In these approaches FAPAR_chl is estimated using satellite-based reflectances or vegetation indices along with explicit parameterizations of LUE (hereafter, LUE will refer to LUE_green) with dependences on meteorological inputs, such as surface temperature and vapor pressure deficit (VPD), that reduce its value from the potential maximum under limiting conditions such as temperature and water stress (e.g., Cheng et al., 2014; Running, Nemani, Heinsch, Zhao, Reeves, Hashimoto, 2004, Xiao, Zhang, Braswell, Urbanski, Boles, Wofsy, Moore, Ojima, 2004, Xiao, Zhang, Saleska, Hutyra, Camargo, Wofsy, Frolking, Boles, Keller, Moore, 2005, Yuan, Cai, Xia, Chen, Liu, Dong, Merbold, Law, Arain, Beringer, Bernhofer, Black, Blanken, Cescatti, Chen, Francois, Gianelle, Janssens, Jung, Kato, Kiely, Liu, Marcolla, Montagnani, Raschi, Roupsard, Varlagin, Wohlfahrt, 2014, Yuan, Liu, Zhou, Zhou, Tieszen, Baldocchi, Bernhofer, Gholz, Goldstein, Goulden, Hollinger, Hu, Law, Stoy, Vesala, Wofsy, 2007, Zhang, Cheng, Lyapustin, Wang, Gao, Suyker, Verma, Middleton, 2014b, Zhang, Middleton, Margolis, Drolet, Barr, Black, 2009; 2017). Different satellite-based estimates of FAPAR_chl have been used and compared within this context (e.g., Joiner, Yoshida, Zhang, Duveiller, Jung, Lyapustin, Wang, Tucker, 2018, Peng, Gitelson, Keydan, Rundquist, Moses, 2011, Zhang, Xiao, Wu, Zhou, Zhang, Qin, Dong, 2017).

Following Gamon and Berry (2012) and Gitelson et al. (2018), LUE can be interpreted as being comprised of two components: a “facultative LUE” that is defined as changing on daily time scales and a “constitutive LUE” that changes over seasonal time scales. The facultative LUE is primarily influenced by irradiance and is characterized by decreasing LUE with increasing irradiance (e.g., Gamon, Berry, 2012, Gitelson, Arkebauer, Suyker, 2018). Some LUE-based implementations of Eq. 3 have considered and/or included the dependence of LUE on PAR_inc (Gitelson, Peng, Masek, Rundquist, Verma, Suyker, Baker, Hatfield, Meyers, 2012, Joiner, Yoshida, Zhang, Duveiller, Jung, Lyapustin, Wang, Tucker, 2018, Peng, Gitelson, Sakamoto, 2013, Turner, Ritts, Styles, Yang, Cohen, Law, Thornton, 2006c, Zhang, Song, Sun, Band, McNulty, Noormets, Zhang, Zhang, 2016) as this was found to be an important factor (Yuan, Cai, Xia, Chen, Liu, Dong, Merbold, Law, Arain, Beringer, Bernhofer, Black, Blanken, Cescatti, Chen, Francois, Gianelle, Janssens, Jung, Kato, Kiely, Liu, Marcolla, Montagnani, Raschi, Roupsard, Varlagin, Wohlfahrt, 2014, Yuan, Liu, Yu, Bonnefond, Chen, Davis, Desai, Goldstein, Gianelle, Rossi, Suyker, Verma, 2010). For example, LUE is generally higher in cloudy conditions with a larger fraction of diffuse light as compared with clear-sky conditions (e.g., Gu et al., 2002). To implicitly account for the facultative LUE dependence on PAR_inc, Eq. 3 can be simplified as

$$\text{GPP}=k\times {\text{FAPAR}}_{\text{chl}}\times {\text{PAR}}_{\text{TOA}},$$

where PAR_TOA is the top-of-atmosphere PAR or potential PAR at canopy level and k is a proportionalty constant with units corresponding to those of GPP/PAR. Equation 5 is functionally equivalent to Eqs. 3-4 with the assumption LUE $=k$ PAR_TOA/PAR_inc. In other words LUE increases with decreasing PAR_inc and vice-versa in such a way as to keep GPP constant (for fixed values of FAPAR_chl). This formulation had superior performance for GPP estimation as compared with implementations of Eq. 3 regionally (Gitelson, Peng, Masek, Rundquist, Verma, Suyker, Baker, Hatfield, Meyers, 2012, Peng, Gitelson, Sakamoto, 2013).

GPP was estimated globally using a slightly modified form of Eq. 5 and calibrated with eddy covariance flux tower measurements, considered a gold standard for GPP estimates, by Joiner et al. (2018), i.e.,

$$\text{GPP}=k\times {\text{FAPAR}}_{\text{chl}}\times {\text{PAR}}_{\text{TOA}}\times {\text{LUE}}^{\prime },$$

where LUE′ was a function of satellite reflectances and could be considered as a constitutive LUE component. In that work, FAPAR_chl was expressed as a linear combination of seven reflectance bands from the MODerate-resolution Imaging Spectroradiometer (MODIS). Vegetation was separated into two classes using satellite measurements of far-red solar-induced fluorescence (SIF): 1) highly productive (in crop regions such as C4 maize with high relative annual maximum SIF values) and 2) all others. Joiner et al. (2018) also showed that satellite data quality plays an important role in determining GPP and that a relatively simple model, relying primarily on high quality satellite reflectances and potential PAR (PAR_TOA), captures a large fraction of GPP variability including both spatial (site to site) and temporal (seasonal to inter-annual).

Here, we build on the work of Joiner et al. (2018) by using the same high quality MODIS reflectances but within a machine learning LUE framework to upscale GPP globally with eddy covariance flux tower measurements. We quantify the impact of additional data sets, such as meteorological variables, in addition to visible, near infrared (NIR), and shortwave infrared (SWIR) reflectances for estimation of global GPP. We also revisit the use of different remote sensing proxies for FAPAR_chl. The primary questions we seek to answer are how much GPP variability can be captured by high quality satellite reflectances on weekly to annual time scales and at a global spatial scale, and how much improvement is obtained by using additional data sets. Finally, we explore the use of machine learning for estimation of GPP uncertainties as well as magnitudes.