• Open Access

Modeling spatial and dynamic variation in growth, yield, and yield stability of the bioenergy crops Miscanthus × giganteus and Panicum virgatum across the conterminous United States

Authors


Correspondence: Fernando E. Miguez, tel. + 515 294 5980, fax + 515 294 3163, e-mail: femiguez@iastate.edu

Abstract

C4 perennial grasses are being considered as environmentally and economically sustainable high yielding bioenergy feedstocks. Temporal and spatial variation in yield across the conterminious United States is uncertain due to the limited number of field trials. Here, we use a semi-mechanistic dynamic crop growth and production model to explore the potential of Miscanthus × giganteus (Greef et. Deu.) and Panicum virgatum L. across the conterminous United States. By running the model for 32 years (1979–2010), we were able to estimate dry biomass production and stability. The maximum rainfed simulated end-of-growth-season harvestable biomass for M. × giganteus was ca. 40 Mg ha−1 and ca. 20 Mg ha−1 for P. virgatum. In addition, regions of the southeastern United States were identified as promising due to their high potential production and stability and their relative advantage when compared with county-level maize biomass production. Regional and temporal variation was most strongly influenced by precipitation and soil water holding capacity. Miscanthus × giganteus was on average 2.2 times more productive than P. virgatum for locations where yields were ≥10 Mg ha−1. The predictive ability of the model for P. virgatum was tested with 30 previously published studies covering the eastern half of the United States and resulted in an index of agreement of 0.71 and a mean bias of only −0.62 Mg ha−1 showing that, on average, the model tended to only slightly overestimate productivity. This study provides with potential production and variability which can be used for regional assessment of the suitability of dedicated bioenergy crops.

Introduction

The success of a bioenergy industry based on dedicated crops depends on both high productivity and yield stability of feedstock supply. For this industry to have a major impact on the bioenergy demands of the United States, large-scale deployment of dedicated biomass crops will be required. However, limited land resources results in an inevitable conflict for its use when optimizing economic, social and environmental objectives. Since dedicated biomass crops are in their infancy, it is of great importance to be able to forecast their productivity and stability under different environments, with particular emphasis on marginal land. Empirical yield data, in contrast to the major US grain crops and commercial forestry, is sparse. As a result, predictions from empirically based yield models will have considerable uncertainties both spatially and temporally until a much larger and broader database of field studies is available. This necessitates the development of mechanistically rich models that may project yield from first principles, to avoid dependence on the limited empirical yield data. Such a model is also important for predicting fine scale productivity, e.g. at the scale of fields around a proposed biorefinery. Equally, it allows projection of how alteration of specific crop characteristics in a breeding program, such as increased leaf angle or altered shoot : root partitioning, would affect yield and yield stability. Here, we use such a detailed biochemical and biophysical crop model to develop spatial forecasts for the United States of growth, yield, and yield stability of Panicum virgatum (switchgrass) and Miscanthus × giganteus (Miscanthus).

A recent quantitative analysis of the P. virgatum literature showed that measured shoot dry matter yields ranged from 1 to almost 40 Mg ha−1 with most results in the 10–14 Mg ha−1 range (Wullschleger et al., 2010). Miscanthus × giganteus has been studied mostly in Europe where biomass productivity ranged from 10 to 40 Mg ha−1 (Miguez et al., 2008) with peak shoot dry biomass levels of 50 Mg ha−1 (Heaton et al., 2004, 2010; Clifton-Brown et al., 2004). No published US studies of this crop report yields beyond a third year, which is generally considered the first year in which the yield potential of this perennial is realized. Although the limited data for these crops provide bounds for the expected productivity, there is a critical need for spatially explicit forecasts that also consider the year-to-year variability. An estimate of this temporal variability will be critical information for biorefineries planning to use dedicated crops as feedstocks for cellulosic biofuel which will need to take account of potential fluctuations in supply. Typically it takes around 3 years after planting for these crops to reach their maximum annual yield, reflecting the time needed for the below-ground perennating organs to develop fully, but depending on the location yields might be higher in year four or later. For planning new plantings and feedstock supply, the model must therefore also be able to predict the dynamics of yield increase over this period and its spatial variability.

Evaluating yield stability, or the ability of a genotype to maintain relative performance across a range of environments (Tollenaar & Lee, 2002), depends on field testing of biomass crops across a range of environments involving geographic and temporal extents. However, long-term studies of P. virgatum and M. × giganteus that evaluate yield stability across a wide range of geographies are lacking. Substantial commitment to the production of biofuels from crop biomass feedstocks in the future will require evaluating not only their biomass productivity potential but also yield stability of biomass production both in response to inter-annual variability in weather and in response to anticipated changing climate. The latter is significant, given that a biorefinery established in the United States today may have a life expectancy of 30–50 years. Siting of a bioenergy facility, whether it is combusting the biomass to provide heat and power, or converting it to a liquid fuel, will require a stable predictable supply, so that the processing plant can be used year round and year after year. A supply that varies markedly between years would result in periods of closure and would be a suboptimal use of investment. The ability to predict biomass production of different bioenergy crop alternatives is crucial for evaluating their economic, agronomic, and environmental feasibility (Sommerville et al., 2010). Since land suitable for cultivating biomass crops is limited, there is a need to determine strategic regions where biomass production and profitability are maximized while negative environmental impacts are minimized (Zhang et al., 2010).

Previous Miscanthus × giganteus and Panicum virgatum biomass production simulation models

Biophysical models of varying complexity have been developed with the objective of simulating spatial and temporal variation in yield of M. × giganteus and P. virgatum. For example, Grassini et al. (2009) developed a model specifically for P. virgatum based on empirical relationships and the concept of radiation use efficiency (RUE) that was able to capture variation in date of anthesis and above-ground biomass at three locations in the United States. Wullschleger et al. (2010) performed a comprehensive analysis of P. virgatum potential in the United States and derived a regression-based model to predict biomass productivity in different regions.

For M. × giganteus, Clifton-brown et al. (2000), using a simple model based on RUE, developed potential productivity for Ireland with yields ranging from 16 Mg ha−1 in northern Ireland to 26 Mg ha−1 in southern Ireland, where total annual solar radiation and the length of the growing season due to higher temperatures are longer. These predictions, however, were only based on radiation and temperatures and ignored limitations due to water stress. Price et al. (2004) using a similar approach, but including effects due to water stress, produced yields in the range 7–24 Mg ha−1 for England and Wales. In addition, Price et al. (2004) estimated an inter-annual coefficient of variation in yield of 10–25% (i.e. standard deviation divided by the mean, multiplied by 100). Interannual variability is typically poorly estimated from short-duration field trials, but as noted above, it is a crucial component in planning for feedstock availability for a biorefinery. In a later study, incorporation of site-specific information about soil water availability significantly improved predictions (Clifton-Brown et al., 2004), showing that the rainfed potential for M. × giganteus biomass production in Europe ranged from 17 Mg ha−1 in Sweden to 41 Mg ha−1 in Portugal. In the absence of water limitation, i.e. irrigation as needed, it was estimated that the peak yield was of 60 Mg ha−1, which reflects the maximum potential of M. × giganteus and it is close to the highest values measured in Italy and Greece (50 and 54 Mg ha−1, respectively) and in central Illinois reaching 60 Mg ha−1 (Heaton et al., 2008).

Nutrient recycling, and therefore agronomic sustainability, in these crops is achieved by translocation of nitrogen to the underground perennating organ in the fall. This requires that the shoot is allowed to complete senescence and dry-down before harvest. After this point, it is inevitable that mass will be lost with time until harvest. Clifton-Brown et al. (2004) estimated this to be 0.36% loss per day and therefore 27.5% if harvest is delayed by 90 days. The model MISCANMOD, developed by Clifton-Brown et al. (2004), has been further refined (now MISCANFOR; Hastings et al., 2009) and incorporated improved descriptions of the relationship between potential and actual evapotranspiration, which impacts calculation of water stress; variable RUE which depends on temperature, nutrient, and water stress; and additional modifications that reflect recent findings in M. × giganteus physiology such as photoperiod sensitivity (Hastings et al., 2009). Their results suggest that although M. × giganteus can be highly productive in southern Europe, a ± 20% variability should be expected due to year-to-year fluctuations in weather patterns.

More detailed predictions particularly of growth, establishment, and inter-annual variability of M. × giganteus and P. virgatum require representation of detailed biophysical and physiological processes underlying carbon, water, and nutrient dynamics in a coupled and layered soil, plant and microclimate continuum as presented, for example, in wimovac (Windows Intuitive Model of Vegetation response to Atmospheric and Climate Change) (Humphries, 2003). This is arguably also a more suitable framework to guide future experiments and crop improvement programs (Humphries & Long, 1995). Such detailed models allow exploration of the value of potential genetic and agronomic modifications, as well as impacts of fine scale spatial and temporal variation in weather and soil. For example, using wimovac it was shown that, theoretically, M. × giganteus can increase its productivity by 4 Mg ha−1 if the threshold temperature for growth could be lowered by 2 °C and the degree days requirements were increased so that flowering occurred uniformly (Clifton-Brown et al., 2001). Miguez et al. (2009) showed that in addition to peak productivity, wimovac was able to accurately simulate CO2 uptake, leaf area index (LAI), and partitioning between leaf, stem, root, and rhizome for M. × giganteus.

Here, we develop a new model, BioCro, which is a generic vegetation model based on the previously published wimovac (Humphries & Long, 1995) as adapted for M. × giganteus in Miguez et al. (2009). Based on wimovac's objectives of ease of use, modularity, and interactivity, we developed a new model and implemented it in r (R Development Core Team, 2006) and further parameterized it to simulate P. virgatum, in addition to M. × giganteus. Although previous semi-mechanistic models have been developed for one or other of these feedstocks, few models accommodate both within the same model structure and assumptions about process. Combining both within the same framework and assumptions avoids the danger of confounding differences in model structure with biological differences when comparing the two feedstocks. The new model was necessary for conducting simulations at a regional level (conterminous United States) since it is computationally more efficient, it is cross-platform and therefore more easily integrated with other software. The main algorithms in wimovac were implemented in the c programming language (Kernighan & Ritchie, 1988), and the interface was written in r (R Development Core Team, 2006). r was chosen because it is cross-platform, and it allows for access to optimization methods (parameter estimation, Monte Carlo, and Bayesian) as well as a range of graphics engines. Thus, BioCro has incorporated the biochemical, physiological, and environmental biophysics mechanism implemented in wimovac, plus parameter estimation capabilities and graphical procedures used to evaluate the agreement between the observed and simulated data. BioCro made it possible to run the model efficiently millions of times to allow optimization routines and hourly timesteps, across multiple years with a high degree of spatial resolution (details below).

Our objectives were the following:

  1. Test and parameterize simulations of M. × giganteus and P. virgatum against the only ‘long-term’ (>5 years) replicated side-by-side yield trials of the two species, which have currently been confined in the United States to Illinois.
  2. Map rainfed M. × giganteus and P. virgatum establishment, mature yields, and yield stability in relation to soil and weather variability over the past 32 years across the conterminous United States taking into account soils and climate.
  3. Compare M. × giganteus biomass productivity to representative maize total biomass production.
  4. Test predictions of P. virgatum biomass productivity against observed data from 30 published field trials.

Methods

Model description

The biophysical model has been detailed previously (Humphries & Long, 1995; Miguez et al., 2009), and important changes between BioCro and wimovac are listed in the supplemental information. Briefly, the model simulates hourly leaf-level photosynthesis and stomatal conductance based on Collatz et al. (1992) and scales up to the crop level using a multilayer canopy architecture (10 layers in this study), which discriminates between sunlit and shaded leaves for each layer. Light, wind, and relative humidity are allowed to vary at different heights within the canopy, based on established micrometeorological principles (Humphries & Long, 1995; Miguez et al., 2009). The total canopy-level carbon fixed, plus the leaf/rhizome remobilization, is allocated every hour according to a phenological schedule and carbon allocation coefficients into plant structural pools (i.e. leaf, stem, root, and rhizome) (Table S1). The new leaf area, stem, or root length was simulated based on allocated carbon resources for each tissue and the specific leaf area, specific stem length, and specific root length, respectively. New leaf growth was assumed to occur uniformly with respect to height in the canopy. Additionally, new stem growth was associated with an increase in canopy height and new root growth with an increase in root density at a specific soil depth. Canopy transpiration is modeled after Penman–Monteith as described in Miguez et al. (2009). Canopy transpiration drives soil water depletion, together with soil evaporation and deep percolation. The available water is constrained by field capacity, wilting point, and maximum rooting depth. For the simulation of water stress leaf expansion rate and stomatal conductance are reduced as the water supply from the soil is reduced, via declining plant water potential. Leaf expansion rate is reduced exponentially as water stress increases (Boyer, 1970), and stomatal conductance is reduced linearly with decrease soil water content (Boyer, 1970; Hsiao et al., 1976) and leaf water potential (see supplemental information for details).

The version of the BioCro r package used here was 0.254-9, and it can be obtained from the first author (F. E. M.). The photosynthetic parameters for P. virgatum were obtained by optimizing model simulations against observed data from Illinois (Dohleman et al., 2009) (Supporting Information, Table S2 and Fig. S1). For simulation of LAI, the model simulations were compared with observations from Heaton et al. (2008) resulting in acceptable agreement (Fig. S2). For this reason, we considered there was no need to adjust additional parameters for P. virgatum beyond those previously described in Table S2. We used data from Illinois (Heaton et al., 2008; Dohleman et al., 2009) for detailed testing of the model simulations of key physiological variables of M. × giganteus.

For a given simulation of a growing season, the initial rhizome biomass for the first growing season was 0.6 Mg ha−1, and we also assumed that the crop was planted about a month later after the first frost in the spring. The rhizome biomass was allowed to grow and accumulate biomass through the third growing season and further with values stabilizing as the growing seasons continued. There is some evidence that M. × giganteus rhizome biomass can be as high as 20 Mg ha−1 from a long-term experiment in Ireland (Clifton-Brown et al., 2007), but it was slightly less than 10 Mg ha−1 in a mature M. × giganteus in England (Beale & Long, 1997). Since the initial growth rate at the start of the growing season depends on the amount of rhizome biomass, this is an important parameter for accurately simulating the above- and below-ground biomass increase during the first and subsequent years.

A comprehensive comparison of predictions and observed biomass production was possible for P. virgatum, with over 30 peer reviewed publications of trials studies of P. virgatum within the 48 states (Table S3). The locations used for comparison provide a representative coverage of P. virgatum production areas (Fig. S3). The development and analysis of this data set is described in Maughan (2011). From this database, we only used studies which did not apply irrigation and considered yields of P. virgatum stands that were at least in their second growing season, since P. virgatum generally, but not always, achieves full production by this time (Grassini et al., 2009). These data represent a range of agronomic management practices (tillage, fertilizer application, early and late harvests, etc.), genetic materials (e.g. lowland and upland ecotypes) soil types, climates, and geographical locations.

Meteorological and soil data

For the simulations in Illinois, weather data were obtained from the network of weather stations (http://www.isws.illinois.edu/warm/weatherdata.asp) at daily time steps, and the weather generator was used to produce hourly realizations for running the model. For the US scale simulations, meteorological data were obtained from the National Oceanic & Atmospheric Administration (http://www.noaa.gov). These data are available at 3 h intervals for 32 years (1979–2010) on a 32 by 32 km grid for the continental US National Center for Environmental Prediction (NCEP) Reanalysis data provided by the NOAA/OAR/ESRL PSD, Boulder, CO, USA (http://www.esrl.noaa.gov/psd/). The variables used were air temperature, accumulated total precipitation, relative humidity, wind speed, vegetation and soil moisture and, where appropriate, were interpolated to 1 h intervals for the purpose of running the model. For solar radiation, we used the NASA Land Data Assimilation Systems (NLDAS – http://ldas.gsfc.nasa.gov/nldas/), which applies a correction factor to the reanalysis radiation data to correct its bias (Pinker et al., 2003).

We used the interpolation method developed for the macroclimate modeling in wimovac to simulate hourly conditions (Miguez et al., 2009). The soil moisture variable was used only to derive the initial value of soil moisture at the start of the growing season, and the model then predicted hourly values from the water balance of precipitation and predicted evapotranspiration.

The simulations were carried out over a 32 by 32 km grid for the United States following the resolution of the meteorological data, but the soil data are available in polygons of the relevant area of interest. Therefore, soil properties within the point in the US grid were aggregated for the purpose of the simulations. Soil data were obtained from the US General Soil Map (STATSGO2) database (Soil Survey Staff, 2009). The soil variables used in the model were field capacity, wilting point, and soil depth (as determined by impediments to root growth). Available water content was computed as a weighted average by layer thickness and also weighted by the proportion of the component in the area of interest. Each polygon in this data set is composed of several soils, each representing a component of that polygon. For this reason, the simulations provide a regional average representing the productivity of planting the crop in the whole polygon. Ideally, we would run the model for each soil type and then average the results, but this would result in an increase in computing time by a factor of 10–50. Our approach was then to run the model three times: two for the most prevalent soil types and a third one for the average soil characteristics of the remaining soil types weighted by their proportion in the component.

For the comparisons of M. × giganteus production and maize total above-ground biomass, we obtained grain yield from the USDA/NASS for the period 2006–2010, which are provided as county-level averages. Some of the counties do not distinguish between dry land and irrigated values; therefore, in some cases, we are comparing irrigated maize vs. rainfed M. × giganteus. We assumed a harvest index of 0.5 to convert grain to total above-ground biomass. From the western United States, we included Texas, Oklahoma, Kansas, Nebraska, North Dakota, South Dakota, but excluded states to the west of these states.

Model runs.

Some aspects of the model runs needed to be automated to accommodate different environments. The model estimated the length of the growing season as the interval between the first and last occurrence of air temperatures below 0 °C in the spring and fall (frost-free period). An important parameter in the model, which is not currently determined by environmental variables but rather needs to be determined empirically, is the accumulated thermal time at which sequential leaf senescence starts. Based on observations of M. × giganteus and P. virgatum growth at many locations, we established that significant leaf senescence starts when 60% of the growing season has elapsed (Beale & Long, 1997; Miguez et al., 2008; Grassini et al., 2009; Miguez et al., 2009).

For M. × giganteus and P. virgatum, the thermal periods describing each of the phenological stages (Miguez et al., 2009) were allowed to vary relative to the length of the growing season reflecting some plasticity to local conditions. This is an area where detailed data are lacking and we expect that current research in biomass crops will be incorporated into the model as it becomes available. At the moment, we have not made any assumptions about the susceptibility of rhizomes of Mgiganteus and rootstock of P. virgatum to winter kill because there is insufficient information about the conditions that will prevent these crops from being cultivated at northern locations, although this is known to be an important factor (Heaton et al., 2010; Jain et al., 2010).

To simulate the development of the crop, the model was run for the establishment year as if started from planting rhizomes (M. × giganteus) or seeds (P. virgatum). Second and third growing seasons were simulated using the projected overwintering rhizome biomass following the previous season of growth as input. Mature biomass production was assumed to equal the peak biomass in the fourth growing season, multiplied by 0.67 to take account of losses in senescence, postsenescence, and during harvest (Clifton-Brown et al., 2004; Heaton et al., 2008). When the model was run for multiple years at a location, we used the same data from a specific year, but we did this in parallel for all 32 years. We assumed that nutrients did not limit growth and that biotic stresses (i.e. pathogens, weeds, and arthropods) were controlled.

The results were summarized by computing the mean for each grid location over the 32 years of simulated biomass production. The variability produced by these different meteorological environments was also used to compute the coefficient of variation as a percentage (standard deviation/mean × 100). For the comparison of predicted and observed P. virgatum biomass production, 95% quantile confidence intervals were computed for the predicted data and observed data (Harrell, 2009). To compare the predictions of the model with observations, some additional statistics were computed: index of agreement (Willmott, 1981), the root mean squared error, mean bias (Wallach et al., 2006), and the concordance correlation (Lin et al., 2002).

Results

Validation of modeled leaf photosynthesis, leaf area, and biomass production

Core to the model is prediction of leaf photosynthetic CO2 uptake rate (A) and LAI as two critical inputs to calculating productivity. Based only on hourly weather records, the M. × giganteus A diurnal course was modeled over 9 days of the 2005 growing season on which direct measurement of A was made (Dohleman et al., 2009). Modeled and measured A showed reasonable agreement with no obvious bias or effect of time of day or year with one exception (Fig. 1). In October (day 286) when observed rates were low, the model consistently overestimated measurements. Modeled values tracked the development, peak, and late season decline of LAI, very closely, and in turn, the seasonal course of above-ground biomass was accurately predicted (Fig. 1). Notice that the model stops when the temperature in the fall dropped below 0 °C, and this is why the simulations of above-ground biomass terminate abruptly at day of the year 300. The model was also effective at predicting observed intra- and inter-annual variation in the observed above-ground biomass and productivity of the crop across three sites spanning 5° of latitude (Fig. 2). Importantly, differences between observed and predicted biomass showed no directional bias, except a possible under-estimation of the standing biomass at the end of the year and subsequent to death of the above-ground material in the late autumn. The year-combined concordance correlation for the three sites ranged from 0.66 to 0.88, Wilmot's index of agreement from 0.74 to 0.79, and the root mean square error from 6.76 to 9.13 Mg ha−1 (Table 1). As with M. × giganteus, A diurnal measurement of P. virgatum was used to validate the photosynthesis model, and this resulted in a reasonable agreement with simulated data (Fig. S4). The model, parameterized for P. virgatum, was able to reproduce the growing season dynamics of LAI (Fig. S2).

Figure 1.

Observed and simulated leaf-level CO2 uptake and stomatal conductance, above-ground biomass, and leaf area index of Miscanthus × giganteus in Urbana, IL, in 2005. Data from Heaton et al. (2008) and Dohleman et al. (2009). The two upper panels have nine subpanels with the day of the year (1–365) on the upper label.

Figure 2.

Observed (points) with 95% confidence intervals and simulated (lines) biomass of Miscanthus × giganteus in Urbana (CMI – a, b, c), Dixon Springs (DXS – d, e, f), and Dekalb (DEK – g, h, i), IL, during three growing seasons (2004–2006); data points from Heaton et al. (2008).

Table 1. Agreement between observed and simulated biomass for Miscanthus × giganteus for three growing seasons and three locations in IL. CMI, Champaign, IL; DEK, Dekalb, IL; DXS, Dixon Springs, IL. The indexes presented are root mean squared error (RMSE), Wilmot's index of agreement (WIA), and concordance correlation (CCorr)
LocationYearRMSE (Mg ha−1)WIACCorr
CMI200412.20.70.58
20055.70.90.83
20068.30.580.39
Combined9.130.790.66
DEK20043.550.950.97
20059.480.670.72
20068.75−0.08−0.02
Combined7.50.740.85
DXS20048.020.480.55
20054.660.910.94
20067.520.690.55
Combined6.760.780.88
AllCombined7.950.750.86

Regional predictions of biomass productivity and inter-annual variability

While on a regional scale, M. × giganteus and P. virgatum have a high potential in areas of the Midwest cornbelt, equally high yields are predicted for areas of the south and east, with much lower row crop density and production than the cornbelt (Fig. 3a, b). By running the model for 32 years, we were able to simulate the expected year-to-year variability of M. × giganteus and P. virgatum biomass productivity (Fig. 3c, d). Particularly interesting are the areas where high yields of M. × giganteus are predicted, but with low potential maize total biomass yields such as Tennessee and Mississippi (Fig. 4). Conversely, the upper Midwest seems to be relatively more favorable to maize production. For locations where dry biomass production was at least 10 Mg ha−1, M. × giganteus was, on average, 2.2 times more productive than P. virgatum, which is less than the average threefold difference observed by Heaton et al. (2008) and closer to the twofold difference simulated by Jain et al. (2010).

Figure 3.

Dry biomass productivity and coefficient of variation for sites with at least a productivity higher than 10 Mg ha−1. The four panels (a) predicted average (over 32 years) annual biomass productivity for a mature stand (fourth year) of Miscanthus × giganteus for the United States; (b) predicted average (over 32 years) annual biomass productivity for a mature stand (fourth year) of Panicum virgatum for the United States; (c) coefficient of variation of annual biomass productivity for M. × giganteus; and (d) coefficient of variation of annual biomass productivity for P. virgatum.

Figure 4.

Difference in biomass productivity between Miscanthus × giganteus and maize for the US negative values favor maize production and positive values M. × giganteus. States to the west of Texas, Oklahoma, Kansas, Nebraska, South Dakota, and North Dakota were excluded from the comparison.

Our ability to test the model simulations for M. × giganteus is limited by the available data outside Illinois. In a recent study, dry biomass production of M. × giganteus averaged 3.3 and 12.8 Mg ha−1 for first and second growing seasons, respectively, in two locations in northestern Kansas (Propheter et al., 2010). The model simulations compare very favorably as it simulated 3.4 and 12.1 Mg ha−1 potential average biomass production for this location, for first and second growing seasons, respectively.

A correlation analysis among relevant variables in the simulation of M. × giganteus biomass productivity (Table 2; Fig. S5) showed that the variable which most strongly correlated with productivity was precipitation (0.81), followed by soil water content at the start of the growing season (0.43) and available water capacity (0.27). Conversely, radiation was negatively correlated with productivity (−0.43) most likely because radiation is itself negatively correlated with the previously mentioned variables which are indicators of water availability (Table 2).

Table 2. Correlation matrix of variables in the simulation of Miscanthus × giganteus biomass production
 DBTempPrecipRadAWCiWatMdep
  1. DB, dry biomass; Temp, temperature; Precip, precipitation; Rad, radiation; AWC, available water content; iWat, soil water content at the start of the growing season; Mdep, maximum rooting depth.

DB10.240.81−0.430.270.43−0.1
Temp 10.200.63−0.01−0.20−0.09
Precip  1−0.340.160.560.08
Rad   1−0.24−0.36−0.05
AWC    1−0.03−0.09
iWat     10.16
Mdep      1

A more extensive testing was possible for P. virgatum by using a database of published studies. The predictions for a particular location and the observed P. virgatum showed reasonable agreement (Fig. 5). Considering that the model was parameterized for P. virgatum (cv. Cave-in-Rock) using data from a single location (Urbana, IL) and that the observed data represent a range of agronomic managements (e.g. early or late harvest), genetic materials (e.g. upland and lowland P. virgatum ecotypes) and are subjected to weed and pest pressures, the agreement between observed and predicted productivity is acceptable for the purpose of regional evaluation of bioenergy crops. The index of agreement (Willmott, 1981) was 0.71, the mean bias was −0.62 Mg ha−1, the root mean squared error was 4.2 Mg ha−1, and the concordance correlation was 0.52.

Figure 5.

Comparison of observed and predicted productivity of Panicum virgatum in locations in the United States. Predicted (open circle, predicted median; bars, 95% quantile intervals) and observed (obs, square median; 5% lower limit, cross; and 95% upper limit, plus). The numbers in the top axis are the number of observations for each location.

For perennial crops, it is important to be able to estimate the increase in biomass production during the first 3–5 years. Our analysis simulated this progression and showed an increase in yields up to year sixth for above- and below-ground biomass (Figs S6 and S7). Without any constraints, the model stabilized after year fifth or sixth, and it attained, on average, a constant above ground to belowground ratio (Fig. S8). The simulations also predict that at lower latitudes which receive ample precipitation, the maximum productivity is attained at least by year 3 (Fig. S9).

Discussion

We have thoroughly tested the performance of the model against crop physiological data, mostly obtained from studies conducted in Illinois (Figs 1 and 2; Figs S1, S2, and S4) for M. × giganteus and P. virgatum. The strength of this approach is that a model which accurately describes the fundamental physiological responses to its environment should be able to capture the processes occurring elsewhere when the model is used for prediction outside of the range of conditions where it has been tested. This description is also important when simulating the year-to-year variability which can only be accurately estimated from long-term field experiments.

At the moment biomass productivity data in the United States for M. × giganteus is scarce and testing of the model has been limited to Illinois (Heaton et al., 2008; Dohleman et al., 2009; Dohleman & Long, 2009). The model proved to be suitable for modeling several physiological variables in this environment (Fig. 1). Establishing the robustness of the model in other regions will be possible after a network of trials is established, and this will require a time frame of several years before at least 3–4 years of data are collected on mature stands of M. × giganteus. Although we are extrapolating yields to regions where there is no published data on M. × giganteus productivity, we believe our projected yields are achievable given breeding efforts that will produce materials adequately adapted to southern latitudes and drier climates. For example, since M. × giganteus can be interbred with sugarcane (called miscane), crops similar to M. × giganteus but with a phenology better adapted to southern latitudes should be able to realize the biomass yield potentials obtained in this study.

Our regional predictions of biomass production for M. × giganteus and P. virgatum should be considered in the context in which they were developed. The weather and soil databases used are appropriate for large-scale assessments of the potential of these biomass crops and for the purpose of evaluating the limitations to their potential productivity. Our simulations are appropriate for regional forecasts, however, for a given site average productivity is expected to be higher if better soils or landscape positions are used or, conversely, lower if marginal soils are chosen. This approach can be used as well to assess the feasibility of siting a biorefinery in a particular location, but in this case, it would be recommended that specific weather information and soil maps be used and not extrapolate directly from the maps produced in this study. Also, although M. × giganteus was shown to be 2.2 times more productive than P. virgatum, the relative advantage of the feedstocks has to be considered on a local context, since one feedstock may still be preferred due to agronomics or other considerations. However, our simulations of P. virgatum were tested against data from specific field trials, and they showed a remarkable agreement both for the average and for the variability (Fig. 5).

It is expected that current breeding efforts will develop varieties which will be better adapted to different environments. As the physiological bases of this adaptations are elucidated, they can be incorporated into a physiological-based model such as the one used in this manuscript and new predictions can be developed, taking this breeding improvements into consideration (Boote et al., 1996). At the same time, this information can be valuable for focusing breeding efforts in regions where the potential of these biomass crops is more promising when other important agronomic, social, environmental, and economic variables are considered.

There are some caveats to the comparison between M. × giganteus and maize dry biomass production comparison: first, the prediction is based on maize cultivars selected for high grain yield rather than total biomass. Second, not all of the total above-ground maize biomass can be harvested for biofuel production, and it seems an appropriate practice to leave in the field 30–50% of the stover as a soil conservation practice (Karlen, 2010). Third, grain energy content is higher than that of stem and leaf and a more appropriate comparison would be based on energy terms and ethanol production (or equivalent). Nevertheless, Fig. 4 provides crucial information for the strategic selection of regions more favorable for bioenergy crops.

The predictions for P. virgatum biomass production are in close agreement with previous reports analyzing the range of biomass production potential in the United States with yields ranging from 1 to 40 Mg ha−1, but mostly concentrated in the 5–23 Mg ha−1 range with an average of 11.6−1Mg ha−1 (Wullschleger et al., 2010). The calibration in our model for P. virgatum was not as extensive as for M. × giganteus (Miguez et al., 2009); however, in this study, we have tested the simulated harvestable biomass of P. virgatum with observations from over 30 studies resulting in remarkable agreement in many contrasting environments (Fig. 5).

The correlation analysis shows that generally water availability strongly influences biomass production across regions (Table 2), but the relationships are far from simple and linear (Fig. S5). However, Richter et al. (2008) found that soil available water and the relative average potential soil water deficit explained 70% of yield variability from 14 trials in the United Kingdom which is in close agreement with the values found here. It is important to remark that while the values from Richter et al. (2008) are derived from empirical data, the correlations for M. × giganteus biomass production are derived from simulated biomass production. The relationship between precipitation, available water, and productivity highlights the concern about the impact of widespread cultivation of perennial crops for bioenergy on the hydrological cycle (McIsaac et al., 2010; VanLoocke et al., 2010; Le et al., 2011). Depending on the area, and placement in the landscape, the impact of M. × giganteus and P. virgatum can be beneficial by alleviating nitrate leaching problems (Ng et al., 2010) or runoff, but on the other hand, it can be detrimental by compromising water resources (McIsaac et al., 2010; VanLoocke et al., 2010).

Few models simulate the initial phase of the perennial crop starting from the initial biomass of planted rhizomes. Without any constraints, our simulations showed a stabilization of biomass production after 4–6 years. Although we present averaged data (Figs S6–S8), it is possible to display the simulations in a specific location to evaluate the time taken to reach full productivity. The limitation of our simulations is that we do not simulate a possible decline in productivity in later years because it is estimated that productivity will be sustained for 10 years or more (Lewandowski et al., 2000), but it will likely depend on the location (Clifton-Brown et al., 2007).

In summary, the large-scale deployment of dedicated biomass crops required for a cellulosic-based biofuel industry requires careful agronomic considerations of spatially explicit potential biomass production, time needed until maximum biomass is attained, and year-to-year variability supply as affected by soil and weather variables. Here, we used a carefully parameterized biophysical model to explore the productivity of two potential feedstocks throughout the conterminous United States. Miscanthus × giganteus was, on average, 2.2 times more productive than P. virgatum and also compared favorably to maize when considering above-ground biomass production. This information should be useful to decision making regarding siting of biorefineries as well as for risk assessment of feedstock supply.

Acknowledgements

This research was supported by the Energy Biosciences Institute (EBI) and Iowa State University (ISU). F. E. M. would like to thank Matt Hudson for access to computing resources at the EBI, Daniel Davidson and David Slater for computer support (IGB), Daryl Herzmann for computer resources access at ISU, Brian Gelder with help with NASS/USDA maize data, and Michael Dietze for general discussion. F. E. M. would also like to thank Steven Speidel (STATSGO2), Charlie Zender (NCO), Pavel Michna (RNetCDF), and Chris Garrard (osgeo). The manuscript was greatly improved by the suggestions of three anonymous reviewers.

Ancillary