Evaluating atmospheric CO₂ effects on gross primary productivity and net ecosystem exchanges of terrestrial ecosystems in the conterminous United States using the AmeriFlux data and an artificial neural network approach

Highlights

• We upscaled AmeriFlux tower data to the conterminous United Stateswith and without considering the atmospheric CO₂.

• GPP/NEE difference between two models exhibits a great spatial and seasonal variability and an annual difference of 200 g C m⁻² yr⁻¹.

• Air temperature played an important role in determining the atmospheric CO₂ effects on carbon fluxes.

• The simulation without considering CO₂ effects failed to detect ecosystem responses to droughts in part of the US in 2006.

Abstract

Quantitative understanding of regional gross primary productivity (GPP) and net ecosystem exchanges (NEE) and their responses to environmental changes are critical to quantifying the feedbacks of ecosystems to the global climate system. Numerous studies have used the eddy flux data to upscale the eddy covariance derived carbon fluxes from stand scales to regional and global scales. However, few studies incorporated atmospheric carbon dioxide (CO₂) concentrations into those extrapolations. Here, we consider the effect of atmospheric CO₂ using an artificial neural network (ANN) approach to upscale the AmeriFlux tower of NEE and the derived GPP to the conterminous United States. Two ANN models incorporating remote sensing variables at an 8-day time step were developed. One included CO₂ as an explanatory variable and the other did not. The models were first trained, validated using eddy flux data, and then extrapolated to the region at a 0.05^o × 0.05^o (latitude × longitude) resolution from 2001 to 2006. We found that both models performed well in simulating site-level carbon fluxes. The spatially-averaged annual GPP with and without considering the atmospheric CO₂ were 789 and 788 g C m⁻² yr⁻¹, respectively (for NEE, the values were −112 and −109 g C m⁻² yr⁻¹, respectively). Model predictions were comparable with previous published results and MODIS GPP products. However, the difference in GPP between the two models exhibited a great spatial and seasonal variability, with an annual difference of 200 g C m⁻² yr⁻¹. Further analysis suggested that air temperature played an important role in determining the atmospheric CO₂ effects on carbon fluxes. In addition, the simulation that did not consider atmospheric CO₂ failed to detect ecosystem responses to droughts in part of the US in 2006. The study suggests that the spatially and temporally varied atmospheric CO₂ concentrations should be factored into carbon quantification when scaling eddy flux data to a region.

Introduction

Net ecosystem carbon exchange (NEE) and gross primary productivity (GPP) are two major fluxes involved in the ecosystem biogeochemical carbon process. Quantification of GPP and NEE and their responses to environmental changes would improve our understanding of the ecosystem carbon cycling and its feedbacks to the global climate system. At present, GPP and NEE at regional scales are often quantified using atmosphere inverse models (e.g. Prince and Goward, 1995), process-based biogeochemical models (e.g. White et al., 2000), and satellite remote sensing approaches (e.g. Running et al., 2000). The estimation of NEE based on the inverse models is limited by the sparseness of the carbon dioxide (CO₂) observation network (Tans et al., 1990); and this approach cannot differentiate the carbon source/sink contributions of each ecosystem (Janssens et al., 2003). Process-based models generally first derive the GPP and ecosystem respiration, which are modeled as a function of physical and biological variables and thus can be applied at regional and global scales. However, large uncertainty still exists in current quantification due to complex model structure, uncertain parameters, and model input (Chen et al., 2011). Satellite remote sensing approaches have advantages in GPP quantification that is based on a near real-time dataset, instead of broadly parameterized ecosystem models. However, they are highly dependent on the accuracy of the model algorithms and the quality of satellite imageries (Baldocchi et al., 2001).

The eddy covariance technique provides direct measurement of net carbon and water fluxes between vegetation and the atmosphere (Baldocchi et al., 1988, Foken and Wichura, 1996, Aubinet et al., 1999). At present, over 500 flux tower sites have been operated to measure the exchanges of carbon fluxes continuously over a broad range of climate and biome types (FLUXNET http://daac.ornl.gov/FLUXNET/fluxnet.shtml). These towers also provide calibrated, validated NEE data and the derived GPP product. To date, numerous studies have been conducted using those flux data to explore the GPP and NEE temporal or spatial variation and their controlling factors in terrestrial ecosystems (Valentini et al., 2000, Law et al., 2002, Baldocchi, 2008, Hirata et al., 2008, Kato and Tang, 2008, Lund et al., 2010, Bracho et al., 2012). Eddy flux data have also been widely used for model calibration (Baldocchi, 1997, Hanan et al., 2002, Reichstein et al., 2002, Hanson et al., 2004) and to upscale from stands to regional levels. The upscaling exercises generally employ the machine learning algorithm that uses meteorological data, vegetation properties, and remote sensing products, which have been conducted for Asia (Zhu et al., 2014, Fu et al., 2014), Europe (Jung et al., 2008, Vetter et al., 2008), the US (Yang et al., 2007, Xiao et al., 2008, Xiao et al., 2010) and even at global scales (Beer et al., 2010, Yuan et al., 2010).

However, few studies have incorporated the atmospheric CO₂ concentrations into regional GPP and NEE extrapolations when using eddy flux tower data. Increasing CO₂, which is mainly due to the burning of fossil fuels and, to a much lesser extent, land-cover change (Keeling et al., 1995; Hartmann et al., 2013), will have substantial direct and indirect effects on the carbon budget (Canadell et al., 2007). Numerous studies have been conducted to understand how plants and ecosystems will respond to elevated CO₂ (Ainsworth and Long, 2005). The large-scale free-air CO₂ enrichment experiment (FACE) showed that increasing CO₂ would reduce the stomatal conductance (Curtis and Wang, 1998, Medlyn et al., 2001, Ainsworth and Rogers, 2007) over a short time period, resulting in decreased transpiration, enhanced water-use efficiency (Conley et al., 2001, Wall et al., 2001) and increased soil moisture (Bunce, 2004). In addition, stimulated photosynthesis (e.g. Li et al., 2014) would increase the above and below-ground biomass production (Piao et al., 2006, Los, 2013, Wan et al., 2007, Deng et al., 2010) and hence accelerate the CO₂ loss from heterotrophic respiration (Luo et al., 1996). Generally, long-lived plant biomass (trees) is more responsive to increasing CO₂ than herbaceous species (Ainsworth and Long, 2005). However, increased plant production with high carbon (C) to nitrogen (N) ratio under elevated CO₂ results in low quality litter input and slows the soil N mineralization, thus triggering the negative feedback to plant biomass over time (Oren et al., 2001, Gill et al., 2002, Luo et al., 2004) if there is no extra nitrogen input. This conceptual framework about progressive N limitation is more obvious in long-lived plants (trees) (Luo et al., 2004) and has been supported by few experiments (Oren et al., 2001, Norby et al., 2010). In addition, the climate drivers and their interaction with soil resources (Reich et al., 2014) would also constrain the ecosystem response to CO₂ fertilization. For example, increasing N mineralization due to warming temperature would further promote plant growth (Peltola et al., 2002, Dijkstra et al., 2010) whereas the drought stress may diminish this positive effect under elevated CO₂ condition (Dermody et al., 2007, Larsen et al., 2011). Moreover, altered water, energy balance (Drake et al., 1997, Keenan et al., 2013), and vegetation physiology due to increasing CO₂ would in turn affect the carbon cycling. Therefore, incorporating atmospheric CO₂ concentrations into the regional GPP and NEE estimation should be a research priority. In addition, previous upscaling flux studies that exclude the atmospheric CO₂ concentration assume that CO₂ variations were spatially and temporally uniform in a region. Although atmospheric CO₂ is generally well-mixed globally since it is chemically inert (Eby et al., 2009), it actually exhibits a large seasonal and spatial variability at the regional scale (Miles et al., 2012). The seasonal and spatial characteristics have previously been reported at site and regional levels (Davis et al., 2003, Haszpra et al., 2008, Miles et al., 2012). For example, the summer measurement of the atmospheric boundary layer CO₂ mole fraction from a nine-tower regional network deployed during the North American Carbon Program's Mid-Continent Intensive (MCI) during 2007 to 2009 shows that the seasonal CO₂ amplitude is five times larger than the tropospheric background and the spatial gradient across the region is four times the inter-hemispheric gradient (Miles et al., 2012).

In this study, we used an artificial neural network (ANN) approach to upscale the AmeriFlux tower derived GPP and NEE to the conterminous US from 2001 to 2006 by considering the atmospheric CO₂ concentrations as an independent factor. We developed two sets of ANN models linking GPP and NEE and remote sensing variables for each vegetation type: the first considered the atmospheric CO₂ concentration effects (CO₂ incorporated simulation, S0) and the other did not (non-CO₂ incorporated simulation, S1). After the training and validation procedure, we then used the two models to extrapolate GPP and NEE to the conterminous US. We hypothesized that: (1) both models would capture the GPP and NEE variation at the calibration stage; (2) the two simulations would exhibit spatiotemporal differences which associate with climate drivers; and (3) ecosystems with high productivity (e.g., forest, cropland) would show greater differences between the two simulations.