Notice: Wiley Online Library will be unavailable on Saturday 27th February from 09:00-14:00 GMT / 04:00-09:00 EST / 17:00-22:00 SGT for essential maintenance. Apologies for the inconvenience.
Biofuel made from conventional (e.g., maize (Zea mays L.)) and cellulosic crops (e.g., switchgrass (Panicum virgatum L.) and Miscanthus (Miscanthus × giganteus)) provides alternative energy to fossil fuels and has been considered to mitigate greenhouse gas emissions. To estimate the large-scale carbon and nitrogen dynamics of these biofuel ecosystems, process-based models are needed. Here, we developed an agroecosystem model (AgTEM) based on the Terrestrial Ecosystem Model for these ecosystems. The model was incorporated with biogeochemical and ecophysiological processes including crop phenology, biomass allocation, nitrification, and denitrification, as well as agronomic management of irrigation and fertilization. It was used to estimate crop yield, biomass, net carbon exchange, and nitrous oxide emissions at an ecosystem level. The model was first parameterized for maize, switchgrass, and Miscanthus ecosystems and then validated with field observation data. We found that AgTEM well reproduces the annual net primary production and nitrous oxide fluxes of most sites, with over 85% of total variation explained by the model. Local sensitivity analysis indicated that the model sensitivity varies among different ecosystems. Net primary production of maize is sensitive to temperature, precipitation, cloudiness, fertilizer, and irrigation and less sensitive to atmospheric CO2 concentrations. In contrast, the net primary production of switchgrass and Miscanthus is most sensitive to temperature among all factors. Nitrous oxide fluxes are sensitive to management in maize ecosystems, and sensitive to climate factors in cellulosic ecosystems. The developed model should help advance our understanding of carbon and nitrogen dynamics of these biofuel ecosystems at both site and regional levels.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Bioenergy is becoming increasingly attractive to many countries, but has sparked an intensive debate regarding energy, economy, society, and environment. Biofuels provide alternative energy to conventional fossil fuels. However, producing biofuels requires a large amount of biomass feedstocks, which may lead to land, water, and nutrient competitions between bioenergy crops and grain crops (Tilman et al., 2009; Pimentel et al., 2010), causing problems such as food insecurity (Fargione et al., 2010; Diffenbaugh et al., 2012). In addition, the environmental impact of producing and using biofuel is another concern to our society. In particular, to what degree, biofuel feedstock producing, biofuel conversion, and biofuel use will mitigate the climate change has been a research focus (Farrell et al., 2006; Searchinger et al., 2008; Melillo et al., 2009).
Biofuel crops can assimilate carbon dioxide (CO2) from the atmosphere and accumulate C into biomass and soils. Using fossil fuels, however, releases CO2. From the perspective of C cycling, biofuels deserve more credits for their C sequestration effect than fossil fuels (Tilman et al., 2006; Clifton-Brown et al., 2007). To date, many studies indicated that, substituting biofuels, especially using cellulosic crops, for fossil fuels (e.g., gasoline) would mitigate GHG emissions, and therefore benefit the environment (e.g., Farrell et al., 2006; Bessou et al., 2011). However, looking beyond agroecosystems and considering land availability and indirect land-use change impacts due to bioenergy expansion, the biofuel effects on the environment are not so clear. Besides using existing cropland to grow crops for bioenergy use, natural ecosystems (mainly forest and grassland) might be converted to biofuel crops to produce biomass feedstocks, which will inevitably cause land-use change. Studies have shown that by considering the GHG emissions caused by indirect land-use change, the C savings or C credit through developing biofuel is significantly reduced or even became negative (Searchinger et al., 2008; Melillo et al., 2009). The discrepancies among different studies are due to a number of uncertainty sources, including the definition of the process of interest, system boundaries of the life cycle of biofuel production, understanding of biogeochemical or physiological mechanisms, data assimilation, and methods applied. These uncertainties are unavoidable when complex systems and human behavior are included in the carbon sink and source analysis of biofuel development and use (Fargione et al., 2010). The high degree of uncertainty highlights the necessity of further research on large-scale bioenergy development.
To estimate regional GHG emissions of land ecosystems, biogeochemical models that represent the C and N processes and dynamics under changing environmental conditions were used (McGuire et al., 2001; Surendran Nair et al., 2012). These models are either empirically based or mechanistically based. Using data from field observations, empirical models represent relationships between a dependent variable (e.g., biomass yield, CO2 emission) and independent variables regarding climate, soil, and management (e.g., Heaton et al., 2004; Jager et al., 2010). This approach is relatively simple but also less accurate as it does not include the biogeochemical and physical processes of ecosystems. In contrast, most process-based models used to quantify the C and N budget of bioenergy ecosystems have been derived from models originally developed for natural ecosystems (Kucharik, 2003; Bondeau et al., 2007; Di Vittorio et al., 2010). These models incorporated with agroecosystem processes can simulate biomass accumulation and allocation as well as C and N dynamics of agroecosytems. For example, Agro-IBIS was developed by taking advantage of the mechanistic nature of a well-tested model, the Integrated BIosphere Simulator (IBIS), which simulates interactions among soil, plant, and the atmosphere. The Agro-IBIS has been used to simulate maize yield (Kucharik, 2003) and cellulosic biomass production (Vanloocke et al., 2010). Similarly, Agro-BGC is a modified version of the Biome-BGC ecosystem model, with processes added to simulate C4 perennial grass functionality and agricultural practices (Di Vittorio et al., 2010). Another example is LPJml, a model for managed land. It was developed based on the well-established Lund–Potsdam–Jena–DGVM. The LPJml can simulate crop yield and C balance (Bondeau et al., 2007). Some species-specific models, such as ALMANAC (Kiniry et al., 1992; for switchgrass and Miscanthus), APSIM (Keating et al., 1999; for sugarcane), MISCANMOD and MISCANFOR (Clifton-Brown et al., 2004; Hastings et al., 2009, for Miscanthus) were also developed to simulate crop growth. These models may have diverse structures and use different algorithms to describe the same biogeochemical process, but all of them can be used to simulate crop biomass production and some can also simulate C and N dynamics (e.g., Agro-BGC, LPJml).
The Terrestrial Ecosystem Model (TEM) is a global-scale biogeochemical model, among the most-used ecosystem models for estimating C, N, and water dynamics of terrestrial ecosystems (e.g., McGuire et al., 1992; Zhuang et al., 2003, 2013). Although many efforts were made toward modifying TEM for agricultural ecosystems, the crop physiology and agroecosystem processes have not been explicitly considered to date (McGuire et al., 2001; Felzer et al., 2004; Melillo et al., 2009). Here, we develop an agricultural version of TEM (AgTEM) to explicitly model the C and N dynamics of agroecosystems.
AgTEM mainly incorporated two sets of processes that are related to agricultural ecosystems: one is on C accumulation and allocation, and the other is on N cycling by introducing nitrification and denitrification processes in soils. In TEM, total C sequestered through photosynthesis is allocated into two major pools of vegetation and soil of natural ecosystems. For agricultural ecosystems, photosynthesis, phenological development, and biomass allocation are crucial for determining ecosystem C fluxes and pools. In addition, agricultural management (e.g., fertilization and irrigation) affects crop development and therefore was considered in AgTEM. For agroecosystems, the N input from outside the ecosystem significantly affects crop N uptake, soil N availability, and the whole N cycle in a plant–soil–atmosphere system. Thus, special attention was paid to the N dynamics in crop soils and the interactions between soil and crop plants in AgTEM.
Materials and methods
Based on TEM, this study developed an agricultural ecosystem model (AgTEM) to simulate the C and N dynamics of crop ecosystems. The site-level observational data of C and N fluxes and pools were used to test the model performance in simulating net primary production (NPP) and nitrous oxide (N2O) emissions. The model sensitivity responding to major input variables was also analyzed. In a companion study, we examined potential N2O emissions from bioenergy ecosystems using the model, as presented in Qin et al. (2013). Below, we first introduce the TEM model, and then detail how AgTEM is developed, followed by descriptions on model parameterization, validation, and sensitivity analysis.
Terrestrial Ecosystem Model
TEM estimates C and N fluxes and pool sizes of ecosystems at a monthly time step and a given spatial resolution (e.g., 0.5° latitude by 0.5° longitude) using spatially referenced information on climate, elevation, soil, vegetation, and water availability, as well as soil- and vegetation-specific parameters. TEM was first documented and applied for regional estimates in the early 1990s (Raich et al., 1991; McGuire et al., 1992), and several major improvements have been made during the past two decades as a result of advance of ecosystem understanding and available computing resources (e.g., McGuire et al., 2001; Zhuang et al., 2003; Felzer et al., 2004). Equilibrium, as well as transient types of simulations, was introduced to TEM in the late 1990s to early 2000s, and inherited thereafter in the later versions. New modules, such as splitting N pools, ozone effects, and soil thermal and hydrological models, were incorporated into TEM to better understand terrestrial C and N dynamics under changing environmental conditions (Zhuang et al., 2002, 2003; Felzer et al., 2004, 2009).
Many efforts have been put into improving understanding of natural ecosystem processes. Managed ecosystems (e.g., agricultural cropland), however, were less studied using TEM. To understand the agricultural ecosystem C and N dynamics, some progress has been made toward modeling land-use change and cropping effects (McGuire et al., 2001; Felzer et al., 2004; Melillo et al., 2009). However, a significant compromise in earlier versions of TEM for modeling agricultural ecosystems was that crop ecosystems were parameterized as grassland ecosystem (e.g., Felzer et al., 2004) (Table 1). Nitrogen oxides (NOX) emitted from agroecosystems, particularly in fertilized croplands, were not included or not mechanistically modeled in TEM (Table 1). In ecosystem models, NPP is the difference between gross primary production (GPP) and autotrophic respiration (RA). It represents the biomass produced by plants and is used to estimate agricultural yield of the agroecosystem (Hicke et al., 2004).
Table 1. Agricultural modules used in AgTEM and historical TEM versions
Notes and references
N/A, not available.
RAP approach indicated relative agricultural productivity, where agricultural productivity was simulated as a multiplier of the original natural vegetation.
GDD approach adopted growing degree days to simulate crop phenology development.
GDD approachc, using single set of parameters for CROP
Inherited from TEM4.1
TEM4.3 was initially designed to simulate ozone effects on C fluxes, and practices such as irrigation and fertilization were discussed (Felzer et al., 2004); it's also used to simulate cellulosic biofuels (Melillo et al., 2009)
Based on TEM4.2, similar RAP algorithms were used
Maize, switchgrass and Miscanthus; crop-specific model parameterization was adopted
Inherited from TEM4.2
Inherited from TEM4.1
First attempt to calibrate TEM for crop-specific C dynamics purposes; it was used for testing potential biomass production from bioenergy crops at ecosystem level (Qin et al., 2012)
AgTEM2.0 (current version)
Agricultural version of TEM; agricultural management, such as irrigation, fertilization
Inherited from AgTEM1.0
GDD approach, using crop-specific parameters
Soil N mineralization, assimilation, nitrification, denitrification; N2O simulations available
Daily version of AgTEM was designed to simulate C and N dynamics in agricultural ecosystems, especially applicable in bioenergy crop ecosystems
AgTEM was developed to estimate C and N dynamics of bioenergy crop ecosystems (namely, maize, switchgrass and Miscanthus) at a daily time step and at any given spatial resolution. In AgTEM, the algorithms of modeling C and N fluxes and pool sizes are inherited from TEM. A majority of the algorithms describing ecosystem biogeochemical processes in TEM are still applicable in agroecosystems (Table 1). Similar to TEM, five differential equations were used to govern the dynamics of state variables and fluxes (Raich et al., 1991):
where CV, NV, CS, NS, and Nav are pools of vegetation C, vegetation N, soil C, soil N, and available N, respectively, determined by corresponding C and N fluxes (see acronyms in Table 2). The terms in parentheses of Eqns (1) to (4) refer to biomass harvest (H) and return (R) in agroecosystems, which were not included in earlier version of TEM for natural ecosystems. In these equations, t refers to the time step used for computation. To assure stability in the integration over time, 4–5th order Runge–Kutta integration procedure (Cheney & Kincaid, 1985) or the Euler method (Atkinson, 1989; Butcher, 2008) can be used for different time steps. In this and the companion studies (Qin et al., 2013), Euler method was used because of its lower computational cost. Other major modifications and new algorithms in AgTEM include temperature effects on GPP, crop phenological process and biomass accumulation, agricultural management, as well as soil N nitrification and denitrification (Table 1). Below, we detail the development.
Table 2. Variables used in AgTEM to govern C and N fluxes and pools
C in soil and detritus
g C m−2
C in vegetation
g C m−2
Available N in soil and detritus
g N m−2
Organic N in soil and detritus
g N m−2
N in vegetation
g N m−2
Decomposition of harvested products
g C m−2 day−1
Gross primary production
g C m−2 day−1
C in harvested products
g C m−2 day−1
C in litterfall
g C m−2 day−1
Net carbon exchange
g C m−2 day−1
Net primary production
g C m−2 day−1
g C m−2 day−1
C in returned biomass
g C m−2 day−1
g C m−2 day−1
N in litterfall
g N m−2 day−1
Net rate of soil N mineralization
g N m−2 day−1
N inputs from outside ecosystem
g N m−2 day−1
N losses from ecosystem
g N m−2 day−1
N uptake by vegetation
gN m−2 day−1
N in returned biomass
g N m−2 day−1
Temperature effects on GPP are modeled in TEM as a multiplier on potential GPP utilizing minimum temperature, maximum temperature, and optimum temperature for plant photosynthesis (Raich et al., 1991). For each time step, the temperature multiplier on GPP (TEMP) is modeled as follows:
where Tair, Tmin, Toptmin, Toptmax, and Tmax are parameters of transient, minimum, maximum, minimum optimum, and maximum optimum air temperatures, respectively. These parameters are crop-specific in AgTEM (Table 3).
Table 3. Minimum, maximum, and optimum temperatures for plant photosynthesis
Tmin, Toptmin, Toptmax, and Tmax are minimum, minimum optimum, maximum optimum and maximum temperatures, respectively.
In early TEM (e.g., TEM4.2, 4.3), crops were parameterized under grass vegetation type (McGuire et al., 2001; Felzer et al., 2004)
In AgTEM2.0, crop-specific sets instead of single set parameters were used for different crop type (Bird et al., 1977; Kim & Reddy, 2004; Sage & Kubien, 2007)
In TEM, plant phenology was empirically simulated using the estimated evapotranspiration and photosynthetic capacity to describe relative changes of mature vegetation (Raich et al., 1991). In AgTEM, however, crop phenology describing crop growth stages can either be imported from historical observational data or modeled according to a crop's response to air temperature. Growing degree day (GDD), a measure of heat accumulation, is used to predict plant development rates (Felzer et al., 2004; Deryng et al., 2011). Cumulative GDD is modeled as a function of daily temperature:
where Tbase and Tceil are base and ceiling temperature parameters, defined as lower and upper temperature thresholds for the process of interest, respectively. These parameters vary among species and possibly cultivars (McMaster & Wilhelm, 1997). In AgTEM (Table 4), GDD are used to predict crop emergence and maturity, using crop-specific threshold parameters.
Table 4. Parameters used to determine growing degree days and simulate crop phenology
Used in AgTEM according to models such as ALMANAC (Kiniry et al., 1992)
Used in AgTEM according to models such as MISCANMOD (Clifton-Brown et al., 2004)
During the growth period between crop emergence and maturity, plants use solar energy to capture atmospheric CO2 through photosynthesis. The total net chemical energy captured by plant, or cumulative NPP, forms the total biomass of a given ecosystem. In agroecosystems, crop grain (e.g., maize) or biomass (e.g., switchgrass) can then be harvested and removed from the ecosystems. Part of the biomass leftover such as residues and dead roots will be returned to the soil C and N cycling. In AgTEM, biomass of interest (YLD, e.g., yield of grain or harvestable biomass) is modeled empirically based on total NPP (NPPtot) following Hicke & Lobell (2004) and Monfreda et al. (2008):
where δhi, δc, Dhi are parameters for determining the proportion of NPP being harvested, the C content in the dry matter, and the dry proportion of YLD, respectively (Table 5). For the grain harvest of food crops (e.g., maize), δhi is a function of the harvest index and a ratio of aboveground-to-belowground biomass (Rhi):
where HI refers to the harvest index, measuring the proportion of total aboveground biological yield allocated to the economic yield of the crop (Table 5). Rhi, also known as ‘shoot-to-root ratio,’ indicates the biomass allocation to aboveground and belowground and is assumed to be constant for a specific crop (Hicke & Lobell, 2004). For crops used for biomass harvest purposes, HI needs to be slightly modified such that harvestable biomass instead of grain can be accounted for in Eqn (9).
Table 5. Values of crop-specific parameters used for biomass harvest in AgTEM
Rhi is parameterized as aboveground biomass/belowground biomass here.
HI refers to maize grain harvested (grain) or the proportion of aboveground biomass harvested (biomass); no biomass harvested for maize at site level and no grain available for switchgrass and Miscanthus. Data sources and references: Prince et al., 2001; Hicke & Lobell, 2004; Mosier et al., 2006; Meyer et al., 2010.
The net carbon exchange between the terrestrial biosphere and the atmosphere is described with Eqn (10) in AgTEM:
where the net carbon exchange (NCE) is the remaining C flux from NPP, after heterotrophic respiration (i.e., decomposition) (RH) and decomposition (EP) of products harvested from ecosystems for human use (e.g., harvested for YLD) (McGuire et al., 2001). A positive NCE indicates ecosystem acting as a CO2 sink whereas a negative NCE means that ecosystem is a CO2 source.
Agricultural management practices, such as irrigation, fertilization, rotation, and cultivar selection, affect mass and energy input and output in agroecosystems. However, the original TEM designated for natural ecosystems has not considered these factors (e.g., McGuire et al., 1992). Using the modified TEM to simulate agroecosystem has some difficulties in modeling C-N-management interactions and integrating time-varying spatially explicit data into regional simulations (e.g., Felzer et al., 2004) (Table 1). In contrast, AgTEM includes two major management practices of irrigation and N fertilization. Besides precipitation percolation, irrigation is considered as an additional direct water input into the soils, which is modeled based on Zhuang et al. (2002). N fertilizer, mainly in the form of NH4+-N and NO3−-N, enters soils, as nutrients to support crop biomass accumulation and soil microbial activities. The N fertilization implementation in AgTEM is modeled as N input from sources outside the ecosystem, affecting NPP, N dynamics and C–N interactions, which were described in Raich et al. (1991) and McGuire et al. (1992). N fertilizers also affect nitrification and denitrification processes in AgTEM.
Nitrification and denitrification
Atmospheric nitrogen enters agroecosystems mainly through atmospheric deposition (e.g., lightning and rainfall), synthetic N fertilizer application, manure application, and litter fall. These N inputs are further mineralized into soil available N such as NH4+ and NO3−. The gaseous NOX emissions from soils, mostly in forms of N2, nitric oxide (NO) and N2O, are mainly produced through nitrification and denitrification processes (Fig. 1). Nitrification describes the process of the biological oxidation of ammonia with oxygen into nitrite and nitrate. Denitrification represents a process of nitrate reduction that eventually produces N2 and N2O (Fig. 1).
In AgTEM, NOX emissions are simulated according to the amount of soil inorganic N, determined by the microenvironment depending on temperature, soil pH, soil water content, and soil biological activity (Fig. S1, S2). Algorithms describing nitrification and denitrification processes from other studies (e.g., Bradbury et al., 1993; Henault et al., 2005) and models (EOSSE, Smith et al., 2010; Bell et al., 2012) were adapted. Three major NOX fluxes (namely, N2O, NO, and also N2) are included in AgTEM. NOX (NOX) is the total NO and N2O emissions from nitrification and N2 and N2O emissions from denitrification:
where N2Ontf, NOntf, N2Odtf, and N2dtf indicate fluxes of N2O from nitrification, NO from nitrification, N2O from denitrification and N2 from denitrification, respectively (Table S1). Total N2O fluxes (N2O) account for both N2Ontf and N2Odtf (more details on nitrification and denitrification modeling can be found in Supporting Information).
Model parameterization and site-level validation
There are a number of constant, vegetation-specific, or soil-specific parameters in AgTEM. Most of them have been defined and determined in earlier studies (e.g. Raich et al., 1991; McGuire et al., 1992; Zhuang et al., 2003). Some vegetation-specific parameters, such as those used to estimate C and N dynamics in maize, switchgrass, and Miscanthus ecosystems, were determined via calibration of the model driven with climate data using observed data of C and N fluxes and pool sizes (Qin et al., 2011, 2012). To determine biomass allocation and biomass-yield conversion, crop-specific parameters used in Eqns (8) and (9) were defined according to previous researches (Table 5). Most parameters used in soil N nitrification and denitrification can be found in earlier studies (Table S2).
Validation investigates models' performance to reproduce the observations from a system within its domain of application (Rykiel, 1996). The model simulations are compared with observed data, and certain criteria are used to determine model performance (Smith et al., 1997). In total, 29 field experiment sites, including 82 site-treatment (i.e., N input level) observational data sets, were organized for validating AgTEM across the United States. These sites cover three bioenergy ecosystems including maize, switchgrass, and Miscanthus (Table 6). For maize, only continuous maize cropping systems were included in the validation. Data of biomass yield (e.g., maize grain, cellulosic biomass) and annual N2O fluxes were used for model and data comparison. Site location, agricultural management, soil properties, and daily climate conditions were used for model simulations. Site annual N2O flux estimates were based on observations during the crop growing season, and accumulated through all growth stages. Possible N2O fluxes from the nongrowing season were not estimated. For site-level data collection and processing (e.g., NPP calculation) procedures, information can be found in earlier studies (Qin et al., 2011, 2012). The climate data of air temperature, precipitation, cloudiness were obtained from the ECMWF (European Centre for Medium-Range Weather Forecasts) Data Server (www.ecmwf.int). For each site-treatment, AgTEM was run for multiple years, using forcing data describing site location, elevation, climate, soil, vegetation, and management. NPP, biomass of interest (i.e., maize grain, harvestable biomass), and N2O flux were analyzed. For all three crops, modeled NPP and N2O were then compared with the observed data.
Table 6. Field experiments studying biomass production and N2O emissions of bioenergy crops, used in this study
Available observational data
No data for the first year.
Maize site selected for model sensitivity analysis.
Switchgrass site selected for model sensitivity analysis.
Miscanthus site selected for model sensitivity analysis.
For comparison, the modeled data were plotted against observations, and a linear regression with a zero intercept was computed to estimate the slope and coefficient of determination (R2). The closer the regression slope to 1, the better the model fits to the observed data. R2 (0 ≤ R2 ≤ 1) indicates the pattern of simulated and observed values (Smith et al., 1997; Huang et al., 2009). The root mean square error (RMSE) and model efficiency (EF) (Loague & Green, 1991) were also reported to show the discrepancies between simulations and observations.
We also estimated the N2O fluxes following the Intergovernmental Panel on Climate Change (IPCC) N-input approach (Tier 1) (De Klein et al., 2006). The annual direct soil N2O emissions were empirically calculated as a factor (0.01) of total N input into soils, including N from fertilizer, manure, water, and residue. Water N was not accounted for in our study, partly because of its scarcity compared to other N sources and also due to a lack of data. Model performance was evaluated in a similar manner to AgTEM.
Model sensitivity analysis
A sensitivity analysis studies the response of the model to different sources of variance in input data (e.g., parameters, forcing data) (Loucks et al., 2005). To study AgTEM sensitivity, three sites with the most accessible information, one for each ecosystem type (Table 6), were selected. Six major input variables representing the climate, management, and CO2 conditions were included in the sensitivity analysis. For a simplified general form of AgTEM Eqn (12), an output corresponding to change in input variables can be written as Eqn (13):
where Xi denotes the i-th input variables, and X1 to X6 are daily air temperature (TAIR), daily precipitation (PREC), daily cloudiness (CLDS), daily N fertilizer application (FTLZ), daily irrigation (IRGT), and annual atmospheric CO2 concentrations (KCO2), respectively. Y indicates the model output whose sensitivity to environment will be evaluated, and here j can refer to NPP and N2O fluxes in AgTEM. Yi corresponds to input Xi. As for Eqn (13), is the model simulation under changing variable Xi, while other variables are fixed . Therefore, the change of model output due to a given changing input can be expressed as follows:
where V(Yi) is the change of output Y responding to changing input Xi, relative to a reference scenario where all input variables are fixed (as in Y0). In this study, all input forcing data collected for each site were used for the reference scenario. In particular, the N fertilizer application rate in the reference scenario was set as 134 kg N ha−1 for maize and 56 kg N ha−1 for switchgrass and Miscanthus. A certain perturbation was exerted to the forcing data to represent input changes:
As in Eqn (15), for each variable X, negative (−1) and positive (+1) changes (C) were added on to the reference (0) forcing data to calculate output sensitivity to increases and decreases of inputs, respectively. For each model simulation regarding the changing variable X, NPP and N2O outputs were analyzed, and a decadal average V(Y) was reported to demonstrate the magnitude of sensitivity for a given Y.
Site-level biomass production and nitrous oxide emissions
The field experiment sites (i.e., maize, switchgrass, and Miscanthus) selected for model validation spread across a majority portion of the maize-producing areas in the conterminous United States, covering a variety of climate zones such as semiarid steppe climate, humid continental climate, and humid subtropical climate (Fig. 2a). Of the 82 site-treatment datasets collected from 29 sites, 65 of them contain N2O observational data (maize: 57, switchgrass: 4, Miscanthus: 4), and 62 have NPP data (maize: 45, switchgrass: 10, Miscanthus: 7). These data were used as dependent variables for comparisons between model simulations and observations. N input at the site-level ranges from 0 to 310 kg N ha−1 for maize and 0 to 156 kg N ha−1 for switchgrass and Miscanthus (Table 6), representing a wide diversity of N treatments.
AgTEM simulations of crop NPP are consistent with the observations (Fig. 2b). The observed NPP of maize has an average of 680 g C m−2, with a range from 287 to 1400 g C m−2. Crop productivity tends to increase with increasing N application. Observed NPP of switchgrass and Miscanthus are relatively higher than maize, about 850 and 1400 g C m−2, respectively. However, the biomass production is not necessarily related to the N input level. For all sites (n = 62), the regression between modeled and observed NPP yields an R2 of 0.74 with a slope of 0.95 (P<0.001). However, two observations (Fig. 2b, circled) evidently deviate from the 1 : 1 line, showing an underestimation in AgTEM. These two observations of Miscanthus from central and southern Illinois show an extremely high biomass production (Heaton et al., 2008), with an average annual NPP flux of about 2150 g C m−2, about three times the average NPP of the rest of the 60 observations. The peak biomass production may be because of favorable climate, management, and proper harvest time during the experiment time (Heaton et al., 2008). To better illustrate the model performance at the majority of sites, observations beyond the range of [mean ± 2SD (standard deviation)] were removed for the comparison. For these sites within 2SD, the indices indicate that fitness of simulations is improved. The slope of regression approximates 1, with a R2 of 0.85; the RMSE decreases from 0.20 to 0.14 and EF increases from 0.83 to 0.88 (Fig. 2b).
N2O fluxes from maize, switchgrass, and Miscanthus were modeled using both AgTEM and an IPCC empirical model. Observations from maize ecosystems show that N2O emitted from croplands with high N application rates are mostly larger than those with lower N input levels (Fig. 2c). As for all sites (n = 65), the average N2O flux is 1.8 kg N ha−1 (1 kg N ha−1 = 0.1 g N m−2), with the maximum flux reaching 13.5 kg N ha−1 observed in a continuous maize field in Indiana (Omonode et al., 2011). Normally, N fertilizers are not applied to switchgrass and Miscanthus, and the highest N application rate tested in the field experiments is 156 kg N ha−1. N2O emissions from soils of these cellulosic crops are comparable with those from maize cropland under similar N input levels (Fig. 2c). The model simulations using AgTEM well estimate the N2O change, at least for fluxes within a reasonable range (e.g., less than 5.0 kg N ha−1). The comparison between modeled and observed N2O results in a slope of 0.83 and R2 of 0.78, for all sites. By moving two maize observations outside the 2SD range (Fig. 2c, circled), one from Stockbridge, MI (Hoben et al., 2011) and the other from West Lafayette, IN (Omonode et al., 2011), the regression generates a higher slope of 0.94 with a greater R2 of 0.86. The RMSE declined from 0.37 to 0.25, and EF slightly improved from 0.81 to 0.88. The discrepancies between modeled and observed fluxes are partly explained by high soil organic matter content (Hoben et al., 2011). Possible maize residues and residual mineral N gains from N fixation by the previous crop (Omonode et al., 2011) contributed to N2O emissions, while AgTEM did not capture these changes.
The IPCC approach relates N2O emissions solely to N input, such as N fertilizer and residue, but fails to consider environmental factors that also significantly affect N dynamics (Grassini & Cassman, 2012). In our study, the predictions from the IPCC model capture a proportion of the observations, with more persuasive indices supporting the fitness for sites within 2SD than for all available sites (Fig. 2d). However, high variances still existed; the RMSE and EF were 0.66 and 0.41, respectively, for all sites (n = 65), and 0.53 and 0.46, respectively, for limited sites (n = 63). The emission factor of 0.01 may not fit all ecosystems. Based on the observations collected in this study, the emission factor of N2O for maize is 0.010 (R2 = 0.44, P < 0.001, n = 63) or 0.013 (R2 = 0.33, P < 0.001, n = 65); for switchgrass, it is 0.013 (R2 = 0.62, P = 0.2, n = 4) and for Miscanthus it is 0.016 (R2 = 0.56, P = 0.2, n = 4).
Compared with the IPCC empirical model in most cases, AgTEM is a better tool to estimate N2O fluxes from maize, switchgrass and Miscanthus ecosystems. The IPCC approach is a good substitute when process-based models are not used due to lacking data or when the estimation accuracy requirement is not high. AgTEM will work under more complicated circumstances, especially when N2O accounting has higher accuracy requirement while the environment conditions are complex. For example, regional, national, or even global large-scale estimates require process-based modeling for better accounting for the complex climate–soil–atmosphere interactions (Bondeau et al., 2007; Del Grosso et al., 2010).
Model sensitivity to environmental and management factors
A sensitivity analysis quantifies the impact of changes in input data on model outputs. Usually, only a subset of input variables dominates outputs in process-based models (Loucks et al., 2005). To identify those input variables, AgTEM simulations were conducted by varying six input variables at three separate locations, one site for each type of crop. The sensitivity of NPP and N2O in terms of percentage change relative to the reference simulation is reported separately for maize, switchgrass, and Miscanthus.
In AgTEM, climate, soil and CO2 conditions, and agricultural management including irrigation and fertilization which determine photosynthesis and autotrophic respiration will ultimately affect NPP. The sensitivity analysis shows that the perturbations to input variables affect NPP for all three crops. However, the magnitudes of sensitivity differ among variables and crops (Fig. 3). For all crops, KCO2, TAIR, PREC, FTLZ, and IRGT (except no IRGT available for cellulosic crops) have positive effects on NPP, where a positive change of input results in a positive change of output, while CLDS has a negative effect on NPP. All crops are comparably sensitive to CO2 and air temperature, but cellulosic crops (i.e., switchgrass and Miscanthus) are much less sensitive than maize to precipitation, cloudiness, and fertilizer application (Fig. 3). In maize ecosystems, NPP is most sensitive to air temperature, where about 20% of the NPP increase was due to a 10% temperature increase and a 16% NPP decrease was due to a 10% temperature decrease, and least sensitive to CO2, where only about a 7% NPP change was due to a 10% CO2 input change (Fig. 3a). In switchgrass and Miscanthus ecosystems, air temperature is still the dominant factor affecting NPP, and a 10% input change caused a 20% NPP change. However, NPP responses are much less noticeable in response to changes in precipitation, cloudiness, and fertilization, only a 1-5% change resulted from a 10% input change (Fig. 3b and c).
These responses may be partly explained by the fact that environmental and management factors directly or indirectly affect the plant photosynthesis and respiration. The atmospheric CO2 positively affects GPP production via photosynthesis. Elevated CO2 significantly increases leaf photosynthetic CO2 uptake rate (Leakey et al., 2004; Oliver et al., 2009). Higher temperature means a longer growth period and higher GDD, which may benefit crops, especially those grown in the relatively colder areas. An example is the selected switchgrass site in the central Upper Peninsula of Michigan, USA. (46.55°N, 86.92°W, 266.1 m a.s.l.) (Nikièma et al., 2011). Abundant but not excessive precipitation can protect crops from drought, providing sufficient water for evaporation and transpiration. Lower cloudiness allows more solar radiation to be absorbed by plants, and therefore more energy to be stored in vegetation. Favorable management practices could always benefit crop production, for example, irrigation for water inputs and fertilization for nutrient inputs. However, switchgrass and Miscanthus seemed to benefit less from increased water and nutrient inputs or less harmed due to less input (Fig. 3b and c). This is because that these biofuel crops have a relatively higher efficiency for using solar radiation, water, and nutrients (e.g., N) compared with maize. Studies reported that switchgrass and Miscanthus could intercept large proportions of the photosynthetically active radiation (Heaton et al., 2008), use much less irrigation than food crops (Fargione et al., 2010), and have no or only slight responses to N fertilization (Lewandowski et al., 2003).
Among the six factors, CO2 generally has the least impact on N2O output in AgTEM among all three ecosystems (Fig. 4). N2O output is negatively related to CO2 input; less than a 0.5% N2O flux change was estimated in response to a 10% CO2 change. For maize ecosystems, the model is more sensitive to fertilization and irrigation, and less responsive to climate factors (Fig. 4a). For switchgrass and Miscanthus ecosystems, the model shows a much higher sensitivity to climate factors than management. A 4–9% change in N2O is observed as a result of a 10% change of temperature or precipitation, and a 2–3.5% N2O change has occurred in response to a cloudiness change (Fig. 4b and c). Low N input level (56 kg N ha−1) partly explains the insensitivity of modeling response to fertilization.
Additional tests using ±20% input change confirmed the pattern of local responses of NPP (Fig. S3) and N2O (Fig. S4) to input perturbations. However, the relative output changes vary among different input variables and ecosystems. It should be noted that the local sensitivity analysis here is not for quantifying the regional impacts of input on outputs. The sensitivity results may change due to change of input data and the sites for conducting the analysis. A global sensitivity analysis at regional levels would be needed to allow full exploration of the input space, accounting for high-dimensionality, interactions, and spatial heterogeneity. However, the global sensitivity analysis requires more information to build probability distributions for the input variables and parameters and expects higher computational complexity (Tang & Zhuang, 2009).
Impacts of N input on biomass production and N2O emissions
Nitrogen, an indispensable nutrient for plants, is often the limiting factor for both crop growth and N2O production. Generally, crop yields and NPP depend on N availability; higher productivity normally requires considerable N inputs, especially for soils with poor nutrient contents (Millar et al., 2010). Many earlier recommendations on crop N application were made based on a positive N-yield relationship (e.g., Stanford, 1973). However, later N response trials and observations questioned the poor N-yield relationship because crop yield may not necessarily increase at excessive N input levels (Nafziger et al., 2004; Millar et al., 2010). N input may enhance crop growth at lower N levels, but may reach a crop yield threshold when the N application is sufficient (Nafziger et al., 2004). For example in the three-year trials in Michigan, McSwiney & Robertson (2005) observed that maize grain yields increased in response to N additions from 0 to 101 kg N ha−1, but then leveled off when more N was added.
When N availability exceeds the needs by plant and competing biota, N2O emissions can be substantial and exhibit exponential responses to the magnitude of N inputs. It has been found in this study (Fig. 2c) and others (McSwiney & Robertson, 2005; Hoben et al., 2011) that the relationship between N2O flux and N input is nonlinear, with a lower emission rate at relatively low N application levels, and a much higher rate when N input increases. N2O emissions are often simulated as an exponential function of the N input rate with empirical models (McSwiney & Robertson, 2005; Van Groenigen et al., 2010), instead of simply applying a linear model like the IPCC tier 1 approach (De Klein et al., 2006). That is, with increasing N, the marginal gain of crop yields decreases while the marginal N2O emissions increase. The recommended rate of N application can only be reached at such a point that the marginal benefit from crop production balances marginal loss or cost via resource input (e.g., N fertilization) and environment pollution (e.g., GHG emissions). More attention should be paid to environmentally or ecologically optimum N rates from the perspective of ecosystem services (Millar et al., 2010; Chen et al., 2011a; Davis et al., 2012).
Approximation and simulation in modeling
Agroecosystem models and crop models share expanding common interests, yet they also have their own specialties. Both groups facilitate the application of models in a system approach to quantifying crop ecosystem dynamics. Both provide a framework to integrate knowledge about soil, climate, plant, and management to transfer the understanding from one location to another, from site to region, supporting decision making with less time and resources required for analyzing complex systems (Raich et al., 1991; Jones et al., 2003; Loucks et al., 2005). However, crop models are mostly used in the agriculture sector to help understand the impacts of environment factors and especially management practices on crop growth and therefore crop yield (grain based) or biomass (non-grain or not interested in grain), and to provide recommendations on agricultural management or hazard protection. Model simulations focus on finer resolutions, for instance, at site- or field- scale for a specific crop type (e.g., CERES-Maize for maize, Hodges et al., 1987) or for specific purposes (e.g., AquaCrop for water management, Steduto et al., 2009). In contrast, agroecosystem models have usually been used to understand the impacts of natural (e.g., climate) or anthropogenic activities (e.g., cropping) on ecosystem dynamics (e.g., McGuire et al., 2001; Felzer et al., 2009). Crop yields or biomass production is part of the C cycle. The spatial scale can be region, nation and even globe (Bondeau et al., 2007).
In our study, AgTEM models the C and N dynamics for agroecosystems with vegetation-specific parameters for each species or crop type. The model structure and algorithms used to describe the biogeochemical and physical processes (e.g., photosynthesis, biomass allocation) are similar, with only minor changes for specific crops. For example, maize has an extra C pool (grain) while switchgrass and Miscanthus do not have one. Vegetation-specific parameters calibrated with observational data were used to capture the magnitude of differences among crops. Some of these parameters can be found from either experiment-based models or crop models (e.g., Tables 3 and 4). Management practices such as irrigation and fertilization were considered in AgTEM, and grain and biomass harvest were estimated.
In the validation and sensitivity analyses, we used the annual total value at multiple sites instead of daily fluxes from a single site to evaluate the NPP and N2O fluxes. We also combined estimates of three species, maize, switchgrass, and Miscanthus, instead of making separate calculations. In the agroecosystem model, biomass (e.g., grain) is estimated based on NPP, a large-scale and long-term average quantity considering both natural and anthropogenic effects. In comparison with crop models, crop yields are small-scale and short-term results of G × E × M (gene/species × environment × management) interactions. Therefore, using agroecosystem models to estimate small-scale C and N dynamics of crop ecosystems, by calibrating parameters to capture short-term (e.g., day-by-day) fluxes, might result in high uncertain ecosystem dynamics (Bell et al., 2012). In addition, observational data might not be in agreement between experiments or repeated samples as a result of measurement uncertainty such as ground disturbance, investigator biases, method divergences and laboratory requirement differences (Müller & Höper, 2004; Kessel et al., 2013). In this study, for example, the N2O experiments collected gas samples at different time intervals during various time courses (e.g., McSwiney & Robertson, 2005; Omonode et al., 2011) at weekly (Parkin & Hatfield, 2010), biweekly (Nikièma et al., 2011) or irregular (Hoben et al., 2011) time steps. Frequency, timing and quantity of N fertilization may affect daily N2O fluxes significantly (Mosier, 1994), and the N2O variations could be principally due to the degree of coincidence of fertilizer application and major rainfall events (Dobbie et al., 1999). It is therefore useful to use seasonal or annual total N2O emissions from several years' data from a certain ecosystem in a variable climate to obtain a robust estimate of mean N2O fluxes (Dobbie et al., 1999).
Estimation uncertainties and future needs
The discrepancies between modeled and observed NPP and N2O come from several sources of uncertainties. Imperfect representation of processes (structural uncertainty) and limited knowledge of parameter value (parameter uncertainty) in a model constitute model uncertainty (Loucks et al., 2005). In addition, AgTEM only considers irrigation and fertilization in terms of agricultural management. Tillage, crop rotation, crop straw management that affect the biomass and N2O emissions (Halvorson et al., 2008; Liu et al., 2011), however, were not considered. This is partly because of the difficulty to quantify the spatial variability of human activities due to a lack of consistent evidence (Millar et al., 2010), and no spatially explicit data concerning these management practices are available for regional simulations (Felzer et al., 2004). Input data are another source of uncertainty. First, the observational data could be biased due to experimental uncertainty. Compared with maize, there are less data for switchgrass and Miscanthus for model validation. More observational data will help to parameterize and validate AgTEM at locations under different environmental conditions (e.g., Europe and China). The forcing data for model simulations were collected from various sources, thus may not represent local environmental conditions. For example, the temperature and precipitation data used in AgTEM were obtained from the ECMWF reanalysis database. The data may be suitable for regional estimation, but not accurate for site-level simulations (Dee et al., 2011). Thus, local climate, soil and vegetation data at the site are desirable.
Uncertainty cannot be removed but can be narrowed, and the model can be improved. From the perspective of observation, better estimates can be achieved via dedication to cross-site experimental research that are of considerable long period with appropriate time intervals during sufficient time courses (e.g., N2O), covering various climate and management (Dalal et al., 2003). The ecosystem C budget quantification can be improved using eddy flux data (e.g., Chen et al., 2011b). In this study, however, the NCE data of crop ecosystems are not available. Among the many Ameriflux sites (http://ameriflux.ornl.gov/), only a very limited number of sites cover croplands (IGBP) with ecosystem C balance data (e.g., NEE, net ecosystem exchange). There are only two sites listed (Rosemount G21 Conventional Management Corn Soybean Rotation/US-Ro1, Minnesota; Mead Irrigated Rotation/US-Ne2, Nebraska) covering maize croplands that can be potentially used for AgTEM. However, the observed fluxes at these sites measure the maize–soybean rotation system, which did not well represent continuous maize ecosystems. Thus, Ameriflux data were not used in this study. Continuous efforts in the maize-, switchgrass-, and Miscanthus-based ecosystem flux measurements, together with agronomic observations (e.g., yield, management) (e.g., Suyker et al., 2004) should be made to improve the model performance.
Our understanding about the underlying ecophysiological and biogeochemical processes shapes the way we interpret and model agroecosystems. Improved observational data will help calibrate and validate models. The AgTEM, as well as many other agroecosystem models can be improved using more data. These models can be appropriately extrapolated to regional scales when they are well calibrated and validated (e.g., McGuire et al., 2001; Bondeau et al., 2007). The developed AgTEM can be used to quantify C and N dynamics of maize, switchgrass and Miscanthus ecosystems at regional scales.
The authors thank Dr. Wen Sun and Ms. Jayne Piepenburg for proofreading the manuscript. The authors are also thankful to anonymous reviewers for their valuable and constructive comments which have led to a significant improvement to the manuscript. Computing is supported by Rosen Center for Advanced Computing (RCAC) at Purdue University. This study is supported through projects funded by the NASA Land Use and Land Cover Change Program (NASA-NNX09AI26G), Department of Energy (DE-FG02-08ER64599), the NSF Division of Information & Intelligent Systems (NSF-1028291), and the NSF Carbon and Water in the Earth Program (NSF-0630319).