Evaluation of Agricultural Production Systems Simulator as yield predictor of Panicum virgatum and Miscanthus x giganteus in several US environments

Simulation models for perennial energy crops such as switchgrass (Panicum virgatum L.) and Miscanthus (Miscanthus x giganteus) can be useful tools to design management strategies for biomass productivity improvement in US environments. The Agricultural Production Systems Simulator (APSIM) is a biophysical model with the potential to simulate the growth of perennial crops. APSIM crop modules do not exist for switchgrass and Miscanthus, however, re‐parameterization of existing APSIM modules could be used to simulate the growth of these perennials. Our aim was to evaluate the ability of APSIM to predict the dry matter (DM) yield of switchgrass and Miscanthus at several US locations. The Lucerne (for switchgrass) and Sugarcane (for Miscanthus) APSIM modules were calibrated using data from four locations in Indiana. A sensitivity analysis informed the relative impact of changes in plant and soil parameters of APSIM Lucerne and APSIM Sugarcane modules. An independent dataset of switchgrass and Miscanthus DM yields from several US environments was used to validate these re‐parameterized APSIM modules. The re‐parameterized modules simulated DM yields of switchgrass [0.95 for CCC (concordance correlation coefficient) and 0 for SB (bias of the simulation from the measurement)] and Miscanthus (0.65 and 0% for CCC and SB, respectively) accurately at most locations with the exception of switchgrass at southern US sites (0.01 for CCC and 2% for SB). Therefore, the APSIM model is a promising tool for simulating DM yields for switchgrass and Miscanthus while accounting for environmental variability. Given our study was strictly based on APSIM calibrations at Indiana locations, additional research using more extensive calibration data may enhance APSIM robustness.


Introduction
Many studies throughout US have reported the extraordinary potential for high biomass production of switchgrass (Panicum virgatum L.) (Vogel et al., 2002;Kiniry et al., 2012;Burks, 2013;Arundale et al., 2014;Trybula et al., 2014) and Miscanthus (Miscanthus x giganteus) (Heaton et al., 2004(Heaton et al., , 2008Khanna et al., 2008;Jain et al., 2010;Kiniry et al., 2012;Mishra et al., 2013;Trybula et al., 2014), both perennial rhizomatous grasses with C4 photosynthesis. Stakeholders involved in developing biomass crops for bioenergy are therefore increasingly interested in estimating potential yields of both species over large geographical domains . Direct measurements of dry matter (DM) yields of these species are scarce relative to corn (Zea mays L.), soybean (Glycine max [L.] Merr.) and other grain crop species, and this lack of data over large geographies and at a fine spatial resolution remains a limitation to informed decision making .
Satisfactory predictions of switchgrass biomass production were achieved with models like ALMANAC in Texas, Arkansas and Louisiana (Kiniry et al., 2005) and SWAT in Indiana (Trybula et al., 2014). Stampfl et al. (2007) achieved satisfactory simulations of Miscanthus biomass production across diverse climate and soil conditions in Europe using the MISCANMOD model developed by Clifton-Brown et al. (2000. Likewise, European studies for renewable energy used a FORTRAN version of MISCANMOD (Hastings et al., 2008) and showed satisfactory simulation of Miscanthus biomass production derived by model improvements in the drought stress function, temperature effect in radiation use efficiency (RUE) and the inclusion of photoperiodism effects (Hastings et al., 2009). Parameterization of WINOWAC was also performed for Miscanthus (Miguez, 2007). Other examples of modelling growth/ adaptation of these species include the use of the STELLA software (Pallipparambil et al., 2015) to identify Ohio, Missouri, Arkansas and Illinois as suitable locations for Miscanthus, as well as to determine sensitive parameters for biomass production. In a recent study (Strullu et al., 2015) the STICS crop-soil model accurately predicted Miscanthus biomass production and environmental impacts in various environments in France and the UK. Despite this important progress in the calibration, development, and modification of several simulation models, the ability to predict DM yield both of switchgrass and Miscanthus by a single model has not yet been achieved.
In this context, a model scaled for a large geographic region and demonstrating adequate performance to predict DM yield is needed. The Agricultural Production Systems Simulator (APSIM) (Keating et al., 2003) is a biophysical model with potential to simulate growth of annual and perennial crops. The APSIM model has been developed in Australia to simulate, on a daily time step, the main biophysical processes of a generic plant in response to management and weather (Keating et al., 2003;Holzworth et al., 2014). However, without preexisting APSIM crop modules to simulate switchgrass and Miscanthus, the re-parameterization of other APSIM crop modules such as the APSIM Lucerne (Robertson et al., 2002) and APSIM Sugarcane modules (Keating et al., 1999) could act as alternatives to simulate growth of both crops. In order to allow the use of APSIM for this purpose, a supervised calibration with a detailed data base and an evaluation of its predictive ability over a broad range of soils and environments is required. Our objectives were to (i) calibrate APSIM Lucerne module for switchgrass and APSIM Sugarcane module for Miscanthus using experimental field data collected in several locations across Indiana and (ii) validate these re-parameterized APSIM modules with independent data from numerous US locations where the accuracy and biases were evaluated.

Materials and methods
The calibration of the APSIM Lucerne and APSIM Sugarcane modules was made using the following steps: (i) data on climate, soil, and management were collected for model inputs; (ii) soil parameterization by location; (iii) adaptation of original plant modules to model switchgrass and Miscanthus growth using actual data from literature or field experiments and (iv) sensitivity analysis to evaluate parameter influence on LAI and the DM yield. Outcomes of the sensitivity analysis by successive iterations directed the compilation of existing data and additional field measurements used to develop model parameters. The model was calibrated through graphical comparison and statistical analyses of observed and modelled leaf area index (LAI) and DM yield data from IN locations with the objective to increase the concordance correlation coefficient (CCC, Tedeschi, 2006) and decrease the bias of the simulation from the measurement (SB, Kobayashi & Us Salam, 2000). These data included not only detailed measurements of LAI and DM yields at the final harvest, but also during crop growth and development. Model validation was made by using graphical comparisons and statistical analyses of observed and modelled DM yield data from 35 locations across the US. Data for switchgrass were grouped by region (southern vs. northern locations) and ecotype (upland vs. lowland). A complete description of datasets used for calibration and validation are provided in the supplementary information (Tables S3 and S4).

Data for model simulations
The data used for model calibration were obtained from field trials across IN (Table 1). For switchgrass model calibration, data from the Water Quality Field Station at Purdue University Agronomy Center for Research and Education (ACRE) near West Lafayette (40°28 0 11.99″N; 87°0 0 36.00″W) and Throckmorton Purdue Agricultural Center (TPAC) five miles south of Lafayette in Tippecanoe County (40°17 0 59.99″N; 86°54 0 0.00″W) ( Table 1). The Miscanthus calibration included two additional IN locations: Northeast Purdue Agricultural Center (NEPAC) in Whitley County between Fort Wayne and Columbia City (41°8 0 24.00″N; 85°29 0 23.99″W), and the Southeast Purdue Agricultural Center (SEPAC) six miles east of North Vernon in Jennings County (39°1 0 48.00″N; 85°31 0 11.99″W) ( Table 1). A complete description of datasets used for calibration and validation of the model is shown in the supplementary information (Tables S3 and S4 for switchgrass and Miscanthus, respectively). Subsequent model validation used data of DM yields gathered across the US, which were collected from published and unpublished studies from 34 dryland locations and one irrigated location (Davis, CA) in 16 states ( Fig. 1; Table 1).

Climate data sources
Daily meteorological data for each location were derived from two data sources. Maximum and minimum air temperatures and rainfall were obtained from National Climatic Data Center (NOAA, http://www.ncdc.noaa.gov), while daily solar radiation was obtained from the NASA Prediction of Worldwide Energy Resource (POWER) -Climatology Resource for Agroclimatology (http://power.larc.nasa.gov). This long-term database also was used as a secondary source of maximum and minimum air temperatures to replace missing values from the NOAA database. Interestingly, recent evaluations of the NASA-POWER solar radiation data indicate very good agreement with measured solar radiation data in areas with flat topography (White et al., 2011;Wart et al., 2013) and with maximum and minimum air temperatures across the US (White et al., 2008). Our evaluations demonstrated a similar fit for daily solar radiation (n = 59031 daily observations) and maximum and minimum air temperatures (n = 69505 daily observations) using measured data from 19 weather stations near the experimental locations used in this study ( Fig. 2; Table S1). The number of air temperature data corrections/filled data was always lower than 2% for all variables. Long-term monthly mean minimum air temperature ranged from À19.7 to 23.3°C and the monthly mean maximum air temperature between À10.7 to 32.5°C. The mean annual rainfall varied from 452 to 1340 mm for Munich, ND and Milan, TN respectively. A summary of climate information by location is reported in Table 1.

Soil parameterization
APSIM requires several soil parameters to adequately reflect the variability among locations (Probert et al., 1998;www.apsim.info). As the soil database of both the 7.3 and 7.5 release versions do not include the soils where these biophysical experiments were conducted, new APSIM soil profiles were created using the following process. First, dominant soil series were identified for each location based on data provided in the literature and in consultation with agronomists and local scientists (Table 1). Second, for each soil series actual soil data (texture, organic carbon [OC] and pH) were obtained from the National Cooperative Soil Survey Soil Characterization Database (NCSS, http://ncsslabdatamart.sc.egov.usda.gov) (see actual data in Table 2). Estimates of the drained upper limit (DUL) and the drained lower limit (LL) were estimated using the HYDRAULIC PROPERTIES CALCULATOR Software developed by Saxton & Rawls (2006) based on soil texture and OC data obtained from NCSS. The estimating equations reported by Saxton & Rawls (2006) were developed by correlation of an extensive data set (1722 samples) provided by the USDA/NRCS National Soil Survey Laboratory. As measured data of soil water parameters were not available for each evaluated location, the accuracy of the HYDRAULIC PROPERTIES CALCULATOR Software (Saxton & Rawls, 2006) to predict soil water parameters was gauged using the observed data of LL (mm mm À1 ) and DUL (mm mm À1 ) from soil series near the locations used in this study. An example of the soil parameterization for Drummer soil series at Water Quality Field Station, West Lafayette IN is presented in Table 2. A complete description of actual and estimated soil parameters used for the calibration/validation of APSIM are provided in the supplementary information for all location evaluated in this study (Table S5).
The APSIM modules were configured for soil N and C (APSIM SoilN), crop residue dynamics (APSIM Surface Organic Matter) and soil water (APSIM SoilWat). Actual OC (%) values were used for initialization (Table S5). To initialize the soil nitrogen pool for the simulations a 10-year simulation of previous management at the experimental locations (corn-soybean rotations), the location-specific climate and soil physical data were used. Crop growth data from these simulations were excluded from subsequent analysis.
For each soil, organic matter (OM, %, OM = OC*1.72; Dalgliesh & Foale, 1998); soil pH 1 : 5 (pH measured for a ratio of 1 part soil and 5 parts water solution according to GlobalSoilMap, 2012; estimated by Libohova et al., 2014); texture class; air dry (AD, mm mm À1 ) corresponding to the moisture limit for dry evaporation of the soil; saturated volumetric water (SAT, mm mm À1 ); bulk density (BD, Mg m À3 ); hydraulic conductivity (ks, mm day À1 ); total porosity (PO, 0-1 calculated as 1-BD/2.6); drainage coefficient (SWCON, day À1 ) were estimated (Table S5). Saturated water content was calculated from BD as described by Dalgliesh & Foale (1998). The parameter AD was estimated as 0.5 9 LL in 0-0.15 m soil layer, 0.9 9 LL in 0.15-0.3 m soil layer and equal to LL at depths >0.3 m (Cresswell et al., 2009). The SWCON, the rate at which water drains, was estimated from DUL and BD (Jones & Kiniry, 1986). For each soil layer within each soil series the water extraction coefficient (KL, mm day À1 ) was set at 0.08 mm day À1 (Robertson et al., 1993a,b;Dardanelli et al., 1997Dardanelli et al., , 2004. The root exploration factor (XF, 0-1) was set to 1 for up to 1 m depth and then decreased exponentially to 0.6 at the maximum soil depth (Monti & Zatta, 2009). The maximum rooting depth was set according to maximum soil depth when there were no impediments to crop rooting. A complete description of actual and estimated soil parameters used for the calibration/validation of APSIM are provided in the supplementary information for each location (Table S5).
Initial soil water values were not available at most locations. Hence, an analysis of soil moisture data at sowing in some locations was performed.  [Diamond et al., 2013;Bell et al., 2013]). The average of soil moisture at 0.2 m depth from these locations, from January to late April, was compared with the DUL of the 36 soils used in this study. With the exception of the TX sites, the initial soil water moisture in spring was always close to DUL. Based on this analysis, 100% of the initial soil water content was used at the onset of all simulations (not shown).

APSIM configuration
Without pre-existing APSIM modules for simulating switchgrass and Miscanthus we re-parameterized the APSIM Lucerne (Robertson et al., 2002) and Sugarcane (Keating et al., 1999) modules to simulate the growth of switchgrass (Table 4) and Miscanthus (Table 5) respectively. Switchgrass and Miscanthus simulations were undertaken using a daily time-step of the APSIM Version 7.5 and 7.3 respectively (Keating et al., 2003;Holzworth et al., 2014). After exhaustive and comparative analysis of plant modules, the re-parameterized APSIM Lucerne module best simulated switchgrass growth in terms of phenological and physiological functions (Table 4). Similarly, the phylogenetic proximity between Saccharum officinarum and Miscanthus, and similarity in physiology, phenology and growth, was the main justification for using the Sugarcane module as a starting point for re-parameterizing APSIM for predicting DM yield of Miscanthus. All the changes in the APSIM Lucerne and Sugarcane modules were implemented through changing parameterization in the initialisation file in extensible mark-up language (XML) format. Different management rules (i.e. sowing, harvesting, fertilization, irrigation, plant density, row spacing, etc.) were created according to practices used in the field and are reported in detail in the supplementary information for switchgrass (Table S3) and Miscanthus (Table S4). The harvesting rules were set to remove the biomass up to 0.03 m (Ojeda et al., 2016). When the dates of management interventions were not available, local average dates for the application of these practices were used. A complete description of management practices used in the simulations is reported in the supplementary information (Tables S3 and S4 for switchgrass and Miscanthus, respectively). The used model output from each simulation was crop DM yield (kg ha À1 ). The simulations in West Lafayette IN included the additional analysis of LAI as another model output. The cultivars used in the field experiments were created as two generic switchgrass genotypes (generic lowland and generic upland) and one Miscanthus genotype (generic) differing in thermal time requirements needed to attain specific phenological stages (Table S6).

Sensitivity analysis
A sensitivity analysis enables users to determine the responses of key model outputs (e.g., harvestable biomass, hereafter DM yield) to variations in selected input parameters. Hence, as part of model calibration, a sensitivity analysis of the APSIM Lucerne and Sugarcane module's to plant and soil parameters (Table 3) was performed using the one-at-a-time method to evaluate parameter influence on LAI and DM yield. Three soil datasets were chosen to represent a range in relevant soil textures (silty, loamy and sandy). We used these soils to analyse the sensitivity of the model parameters through a large range of plant available water capacity (PAWC).
Based on an exhaustive review of the literature, and fieldmeasured data, and in order to adequately predict the growth of switchgrass and Miscanthus with the APSIM Lucerne and Sugarcane modules, the most sensitive plant parameters for switchgrass ( Fig. 3; Fig. 6) and Miscanthus ( Fig. 3; Fig. 9) were identified (Tables 4 and 5 for switchgrass and Miscanthus, respectively). Thereafter, in extensible mark-up language (XML) format, these parameters were modified (Tables 4 and 5 for switchgrass and Miscanthus, respectively). In all cases, the modified parameters were calculated as an average of reported values in the literature or field measurement based on the range of each parameter. For all locations, we used the same values of parameters to simulate the DM yield in the re-parameterized modules. We followed the same parameterization process for both crops, although there were more sensitive parameters for switchgrass than for Miscanthus, which explain the differences in the number of parameters listed in Tables 4 and 5. It should be noted that we only showed the modified parameters, since default (original) values might be easily obtained from the XML file available in the free APSIM version.
In the APSIM Sugarcane module crop growth is divided into two sections, plant and ratoon crop. The parameters in the plant and ratoon crop sections determine the crop growth for the first and second harvests onwards. Hence, the model modifications were made in both sections of XML file (plant and ratoon crop). To assess potential errors in soil datasets, after plant model modifications, a sensitivity analysis was undertaken for PAWC ( Fig. 4a for switchgrass and Fig. 4b for Miscanthus). The maximum variation (%) in the parameters that determine the maximum PAWC in the soil -AD, LL, DUL and SAT -was determined based on the 36 soils used in this study (Table S2). Therefore, for the sensitivity analysis AD, LL, DUL and SAT were modified in AE29%, AE23%, AE10% and AE5%, respectively, in order to provide realistic boundaries. Second, sensitivity of KL, XF and initial OC was evaluated by modifying the range AE50% of initial values (Fig. 4c,e,g for switchgrass and Fig. 4d, f, h for Miscanthus). Using the same approach explained previously, the maximum pH variation (%) was determined (Table S2). Model sensitivity to pH change was evaluated by increasing and decreasing soil pH by 14% of the actual soil values of switchgrass ( Fig. 4i) and Miscanthus (Fig. 4j). The total number of simulations necessary to complete the sensitivity analysis of soil parameters was 958.

Re-parameterization of switchgrass plant module
Crop phenology in APSIM is controlled by the sum of heat units from sowing to maturity. Accordingly, the parameter y_tt (thermal time requirements needed to attain specific phenological stages) was set to the growth habit of switchgrass (Kiniry et al.,  (1 : 1), pH in a 1 : 1 suspension of soil in water; pH (1 : 5), pH in a 1 : 5 suspension of soil in water; OM, organic matter; AD, air dry; LL, lower limit; DUL, drained upper limit or field capacity; SAT, saturated volumetric water content; ks, hydraulic conductivity; BD, bulk density; PO, total porosity; SWCON, drainage coefficient.
2005). Similarly, the stage_stem_reduction_harvest parameter was modified so that, after harvest, switchgrass starts a new regrowth. In order to achieve this initial point of growth, stage_stem_reduction_harvest was reduced from 4 to 3 (Table 4). The LAI and, hence, DM yield in APSIM are defined directly by the RUE and the transpiration efficiency coefficient (Kc) parameters both fixed in each phenological stage. Modifications of the physiological parameters reported elsewhere for the APSIM Lucerne module (Dolling et al., 2005;Brown et al., 2006;Chen et al., 2008) were also used here to simulate switchgrass DM yields. After sensitivity analysis, RUE (coded by y_rue; Madakadze et al., 1998;Kiniry et al., 1999Kiniry et al., , 2012Heaton et al., 2008;Jain et al., 2010;Trybula et al., 2014) and Kc (coded by transp_eff_cf; Byrd & May, 2000) were set based on switchgrass values obtained in the literature (Table 4). For all locations, we used the same RUE and Kc value. In addition, two other parameters (the temperature response of photosynthesis, y_stress_photo, and the extinction coefficient, y_extinct_coef) directly associated with DM yield were modified as follows. The temperature response of photosynthesis was modified based on previous corn studies (Andrade et al., 1993;Louarn et al., 2008) validated for switchgrass (Grassini et al., 2009) (Table 4). Similarly, y_extinct_coef was modified based on the differences in leaf structure between lucerne vs. switchgrass (Kiniry et al., 1999;Trybula et al., 2014).  (Cosentino et al., 2007), UK (Clifton-Brown et al., 2001) and IL (Dohleman & Long, 2009), respectively. Hence, given these discrepancies in RUE values among studies, the parameter y_rue was modified from 1.8 g MJ À1 to 3.0 g MJ À1 for all phenological stages (Table 5). This used value of RUE was calculated from these studies as an average of the ratio between accumulated yield (from emergence to peak biomass) and total annual incident radiation. For all locations, we used the same RUE value (3.0 g MJ À1 ).

Re-parameterization of Miscanthus plant module
The light extinction coefficient (coded by y_ extinct_coef) through the leaf cover of the crop provides a measurement of the absorption of light by leaves (Zub & Brancourt-Hulmel, 2010). Miscanthus achieves y_ extinct_coef values between 0.45 (Trybula et al., 2014) to 0.68 (Clifton-Brown et al., 2000). Based on the insensitivity to the changes of this parameter in the range reported in the literature (Fig. 3d), the y_ extinct_coef default value for Miscanthus (0.38) was unchanged. Miscanthus partition biomass has been parameterized for the WIMOVAC model using data from Beale & Long (1997) and has been validated using data from European studies (Miguez, 2007). Based on the data collected by Burks (2013) and Trybula et al. (2014) in West Lafayette IN, the ratio_root_shoot parameter was modified in APSIM for all phenological stages (Table 5). These authors measured the Miscanthus aboveground and root biomass in different crop growth stages. Therefore, we used these data to re-parameterize the ratio_root_shoot into the Sugarcane  (Table S7).

Evaluation of model performance
Initially, model performance was visually assessed by comparing scatter plots of observed values in the y-axis vs. modelled values in the x-axis (Piñeiro et al., 2008). When multiple data points were available for a particular treatment in an experiment, standard deviations are included as an estimate of error. The evaluation of model performance described in Tedeschi (2006) and Kobayashi & Us Salam (2000) were used to statistically evaluate model performance. The parameters used were: observed and modelled mean and standard deviation of the DM yield, the concordance correlation coefficient (CCC), and mean square error (MSE). The CCC integrates precision through Pearson's correlation coefficient, which represents the proportion of the total variance in the observed data that can be explained by the model, and accuracy by bias which indicates how far the regression line deviates from the concordance (y = x) line. Similarly, the MSE was partitioned into bias (SB, %, the bias of the simulation from the measurement) and mean square variation (MSV, %, the difference between the simulation and the measurement with respect to the deviation from the means), using IRENE software (Fila et al., 2003). Bias and MSV are orthogonal and, consequently, can be analysed independently (Kobayashi & Us Salam, 2000). Model calibration was deemed complete when the CCC and SB were higher than 0.7 and <30%, respectively, for the LAI and DM yield of both crops.
In both crops, the growth period from sowing was simulated including the establishment phase, during which time rhizome biomass, root depth, and DM yield are increasing, and the post-establishment phase, in which perennial organs and root system are fully developed and the DM yield is fairly constant. This is influenced more by variation in weather than changes in plant establishment/underground organ development. However, only the observed DM yield from the post-establishment phase was included in this analysis to evaluate the accuracy of the model to predict DM yield with the established crop. The duration of the establishment phase varied from two to four years, depending on the experimental site (Tables S3  and S4). For switchgrass validation, the data sets were grouped by northern locations (IN, Indiana; IL, Illinois; TN, Tennessee; NE, Nebraska; IA, Iowa; SD, South Dakota; NY, New York; ND, North Dakota) and southern locations (TX, Texas; VA, Virginia; OK, Oklahoma; LA, Louisiana; AR, Arkansas). The same grouping was not applied to Miscanthus, because DM yields in southern US locations are extremely low and difficult to find in the literature. Therefore, the capability of APSIM to simulate the Miscanthus DM yield was not evaluated in southern US locations.

Switchgrass
The most sensitive parameters of plant module were RUE and the light extinction coefficient coded by y_rue and y_ extinct_coef parameters, respectively. The sensitivity of the model to the modification of these parameters (Fig. 3a,c) was high in the selected soils. The largest change in DM yield (29% and 44%) occurred when y_rue was increased from 1.7 to 4 g MJ À1 and 4.9 g MJ À1 in the loamy and silty soils, respectively (Fig. 3a). In contrast, increasing y_rue to 4.9 g MJ À1 reduced switchgrass DM yield by~15% in the sandy soil (Fig. 3a). The trend in DM yield to decreased y_extinct_coef was similar for these soils with declines 5-11% and 25-34% when y_extinct_coef was decreased from 0.8 to 0.5 and 0.8 to 0.2, respectively (Fig. 3c).
The sensitivity analysis carried out to identify possible effects of changing soil parameters in the re-parameterized model (Fig. 4) on switchgrass DM yields also showed soil-specific responses with the greatest responses in DM yield occurring in the sandy soil. For example, when PAWC was decreased, predicted DM yield declined 1%, 7% and 11% for silty, loamy and sandy soils, respectively (Fig. 4a). When PAWC was increased, the DM yield was enhanced 3%, 4% and 5% for silty, loamy and sandy soils, respectively (Fig. 4a).
The highest DM yield response to changes in XF, KL and pH also occurred in the sandy soil (13%, 14% and 21%, respectively; Fig. 4e,c,i). By comparison, the model was less sensitive (<4%) to the changes in the initial OC (Fig. 4g).
The model with the default settings demonstrated a poor ability to simulate LAI and DM accumulation of switchgrass. Summary statistics comparing observed and modelled LAI from original and modified model parameters at the Water Quality Field Station in West Lafayette IN demonstrated the improvement in LAI predictions, as indicated by the increased CCC values (0 to 0.81) and a reduction in the SB (93 to 30%) ( Fig. 5a; Table 6). Similarly, and as expected, when modified plant parameters ( Fig. 6; Table 4) were introduced into the APSIM Lucerne module, prediction of switchgrass DM yield at the same location was improved, as indicated by the increase CCC (0.11 to 0.96) and the reduction in SB (57 to 4%) ( Table 6).
The APSIM Lucerne module showed excellent accuracy for predicting the accumulated DM yield at IN locations used for switchgrass model calibration (Fig. 10a) as evidenced by the values of 0.93 for the CCC and 0% for the SB (Table 7). The APSIM Lucerne module also predicted DM yields when validated using yield data from trials conducted at northern locations (Fig. 11a), but model accuracy at southern locations was unsatisfactory. In fact, the CCC = 0.95 from comparisons using data from northern locations contrasted with the CCC = 0.01 for comparisons using data from southern locations (Table 7). Remarkably, SBs obtained for northern and southern locations were similar, 0 vs. 2%. The observed switchgrass DM yield during validation ranged from 5329 kg DM ha À1 in SD to 10668 kg DM ha À1 in IN, with the average discrepancy in DM yield being 1%. The modelled DM yield ranged from 1391 kg ha À1 in VA to 10786 kg ha À1 in TN. The better DM yield predictions in the northern locations were in IL, TN, NE, IA and SD. In contrast, in NY, IN, and ND the switchgrass DM yield was simulated with less precision (Table 7). By comparison, the modelled DM yields at southern locations were, on average, 10% less than observed values. When data were clustered by ecotype at northern locations, the DM yield was better predicted for upland ecotypes (0.96 for the CCC and 0% for the SB) than for lowland ecotypes (0.64 for the CCC and 9% for the SB). The variation of DM yield was well predicted by the re-parameterized model irrespective of stage of establishment of the crop.
Results of regression analysis of observed DM yields on accumulated annual rainfall revealed a poor fit at southern US locations (R 2 = 0.18 and slope regression À4.43; Fig. 8a). In contrast, the rainfall regression at northern locations showed a greater R 2 value (0.43) than the southern locations, and a positive slope (6.18) (Fig. 8b).

Miscanthus
As with the Lucerne plant module, the most sensitive parameter of the Sugarcane plant module was y_rue. However, unlike y_rue, the model was not sensitive to changes in y_extinct_coef (Fig. 3d). Model sensitivity to modification of y_rue (Fig. 3b,d) varied depending on  soil type. The largest change in DM yield (increment of 15-32%) occurred when y_rue was increased from 1.8 to 4 g MJ À1 (Fig. 3b). Reducing y_rue from 1.8 to 1.25 g MJ À1 resulted in 8-19% lower DM yields when compared to initial model conditions (Fig. 3b).
The sensitivity analysis carried out to identify possible effects of soil parameters (Fig. 4) on Miscanthus DM yield showed differential responses depending on soil type and parameter. When PAWC was increased, changes in DM yield were higher for the loamy soil (18%) than for the silty and sandy soils (6% and 3% respectively; Fig. 4b). Similarly, when PAWC was decreased, the reductions in DM yield were greater for the loamy soil (17%) than for the silty and sandy soils (12% and 5%, respectively; Fig. 4b). When initial OC was increased by 50% from default values, predicted increases in DM yield on sandy soil (48%) were higher than the silty soil (17%) and the loamy soil (4%; Fig. 4h). In contrast, the model exhibited low sensitivity of DM yield to soil pH, KL and XF with changes in DM yield predicted to be no >5%, 9% and 12% for the respective parameters (Fig. 4j,d,f).
The original APSIM Sugarcane model with the default plant parameters could not accurately predict Miscanthus LAI and accumulated DM yield. Summary statistics comparing observed to predicted LAI with the re-parameterized model using data from the Water Quality Field Station in West Lafayette, IN demonstrated improvement in LAI predictions as indicated by the high CCC (0.69) and low SB (<30%) (Fig. 5b; Table 9). Similarly, and as expected, when modifications of plant parameters ( Fig. 9; Table 5) were introduced into the model, the prediction of Miscanthus DM yield, at the same location was improved as indicated by the excellent CCC (0.94) and low SB (<30%) ( Table 9).
The re-parameterized APSIM Sugarcane module showed excellent accuracy for predicting Miscanthus Fig. 4 Relative change in dry matter yield of switchgrass and Miscanthus vs. relative change of soil parameters using APSIM Lucerne and APSIM Sugarcane modules, respectively, for three contrasting soil textures. The soil parameters analysed were PAWC, plant available water capacity (a, b); KL, water extraction coefficient (c, d); XF, root exploration factor (e, f); OC, initial organic carbon (g, h) and pH (i, j). The value 1 on the x-axis corresponds to the default values used in the sensitivity analysis. Broken lines indicate the baseline parameter and no changes in dry matter yield, respectively. DM yield at IN locations used for model calibration (Fig. 10b) as evidenced by the values of 0.92 and 13% for the CCC and SB respectively (Table 10). The model validation was acceptable for most locations (0.65 and 0% for CCC and SB, respectively; Fig. 11b). However, the model accuracy at KY and NJ was unacceptable as indicated by the low CCC values of 0.38 and 0.46, respectively (Table 10). However, the SB obtained during validation was similar and <30%. The observed DM yield of Miscanthus for validation ranged from 8398 kg DM ha À1 in VA to 33980 kg DM ha À1 under irrigated conditions in CA (Fig. 11b). The modelled DM yield for calibration was 9% higher than observed DM yield. This difference, however, was negligible (0.5%) when compared to the observed and modelled DM yield association done for validation (Table 10). The modelled DM yield ranged from 13883 kg DM ha À1 in VA to 20518 kg DM ha À1 in NE.

Discussion
The main objective of this study was to evaluate the ability of APSIM to simulate the growth and DM yields of switchgrass (using the re-parameterized Lucerne module) and Miscanthus (using the re-parameterized Sugarcane module) at several locations across the US. The modelling approach was based on an exhaustive sensitivity analysis of plant and soil parameters using the one-at-a-time method followed by a detailed model calibration using field data from experiments in IN, and ending with a model validation using data from numerous US locations. Results indicate that these re-parameterized APSIM Lucerne and Sugarcane modules can accurately simulate growth and yield of switchgrass and Miscanthus respectively. Further considerations, specific to each crop, are discussed below.

Switchgrass
The original APSIM Lucerne module was developed and extensively tested in many environments for its ability to predict the phenology and DM yield of lucerne (Robertson et al., 2002;Dolling et al., 2005;Chen et al., 2008;Pembleton et al., 2011;Moot et al., 2015;Ojeda et al., 2016). However, in its original format with thermal parameters for a C3 species, the module is not able to adequately simulate switchgrass DM yield. Therefore, several modifications in plant module parameters were needed to improve the prediction of switchgrass DM yield. The range of modelled DM yield in this study for northern locations (5392 to 10 668 kg ha À1 ) was coincident with modelled DM yields of the upland ecotype (Wang et al., 2015) in the marginally saline soil of northeast Fort Collins, CO (5200 to 9600 kg ha À1 ), as well as with the observed DM yield described by Schmer et al. (2008) on marginal cropland on ten farms in ND, SD Table 6 Summary statistics indicating the cumulative improvement that resulted from re-parameterization of the APSIM Lucerne model for predicting LAI (n = 11) and dry matter yield (n = 20) of switchgrass at Water Quality Field Station, West Lafayette, IN. The parameters modified were y_tt, thermal time requirements needed to attain specific phenological stages; y_rue, radiation use efficiency; transp_eff_cf, transpiration efficiency coefficient; y_stress_photo, temperature response of photosynthesis and y_extinct_coef, extinction coefficient. CCC, SB and MSV are the concordance correlation coefficient, bias of the simulation from the measurement and mean square variation, respectively and NE (5200 to 11 100 kg ha À1 ) and by Wullschleger et al. (2010) for 25 upland cultivars in the northern US. The re-parameterization was based on sensitivity of DM yield when parameters were modified to values obtained in published studies (Table 4). In addition, differential effects of soil parameters on switchgrass DM yield were observed (Fig. 4). Soil water availability is one of the key soil parameters that explained most of the differences in switchgrass growth and yield (Fig. 4a). Similarly, the low PAWC due to high sand contents in the soil (Saxton & Rawls, 2006) reduced the canopy expansion decreasing the light interception and photosynthesis, thus, reduced plant growth (Durand et al., 1995) in tall fescue. In addition, our results showed highest sensitivity of DM yield to parameter changes in the sandy soil that had the lowest PAWC. Although the re-parameterized model substantially improved the prediction of DM yield at northern locations, a poor DM yield prediction at southern US locations was found (Figs 10 and 11a). This poor validation was associated with difficulty in accurately estimating PAWC at southern locations (Fig. 7), specifically estimates of LL and DUL by the HYDRAULIC PROPERTIES CALCU-LATOR Software (Saxton & Rawls, 2006) (Table 8). This was evidenced by the statistical analysis performed on observed and estimated LL and DUL at ten soil series near selected southern and northern locations evaluated in this study (Table 8). For example, the over and under estimation of PAWC at OK and VA (Fig. 7) respectively, would explain the over and under prediction in DM yield at both locations. In contrast, good agreement was found between observed and estimated LL and DUL in two soil series at northern sites in IN and IL (Table 8; Fig. 7). This observation suggests a new line of research that should be addressed to clarify to what extent the under or over estimation of these soil water parameters affects the outcome of predicted DM yield in APSIM.
Previous modelling efforts for predicting DM yield of switchgrass were reported. While Grassini et al. (2009) demonstrated similar trends in the DM yield predictions (CCC = 0.77), these results were obtained based on a limited number of observations (8) from two northern US environments (Ames, IA and Mead, NE). Additionally, the accuracy of the ALMANAC model (Kiniry et al., 2005) and APSIM to predict DM yield were similar differing by <7%. However, yield values reported by these authors was nearly double what we observed in our study (ca. 17000 vs. 8000 kg DM ha À1 ) despite comparable dryland conditions. While ALMANAC accounted for 47% of the variability in observed DM yields (Kiniry et al., 2005), when the CCC was calculated from the published results, both models (APSIM and ALMANAC) were poor predictors of DM yield (CCC<0.50) at southern locations (with the exception of Stephenville TX). These authors also observed high year-to-year variability in measured yields at southern locations in the US (TX, LA and AR) and reported that this was not closely associated with variation in rainfall. The lack of fit for the southern locations was evaluated here using our complete dataset. The results showed that the southern locations showed poor fits for observed DM yield as a function of accumulated annual rainfall (Fig. 8a), in contrast with northern locations (Fig. 8b). An additional explanation for the low fit between observed and modelled DM yield at these locations is that the observed DM yields used to validate the model in TX, AR, and LA were derived from the mean of nine cultivars Table S3) in each location. The absence of genotypic parameters for each cultivar of switchgrass used by these authors, did not allow us to re-parameterize/calibrate/validate the model at the cultivar level. Although the model predicted DM yield of upland switchgrass cultivars better than that of lowland cultivars, the limited number of observations and locations evaluated for lowland ecotypes in this study did not allow us to demonstrate differences in APSIM accuracy by ecotype.  (Saxton & Rawls, 2006) in predicting the soil water parameters of ten soil series from different states.

Miscanthus
Accurate prediction of Miscanthus DM yield using the APSIM Sugarcane module required fewer model re-parameterizations when compared to changes made in the APSIM Lucerne module parameters to predict DM yield of switchgrass. Sugarcane (Saccharum officinarum) shares phenological and physiological attributes with Miscanthus due to their close polyphyletic relationship at the subtribe level (Hodkinson et al., 2002), so it is not surprising that this APSIM module predicted the DM yield of Miscanthus. As with switchgrass, the main plant parameter modified was the RUE. Model DM yield prediction improved when y_rue was increased (Fig. 9a). Unlike switchgrass, no change occurred in Miscanthus DM yield prediction from changes in y_extinct_coef in the three soils evaluated (Fig. 3d). Similarly, DM yield was not sensitive to change in y_extinct_coef using SWAT in IN (Trybula et al., 2014). In addition, Davey et al. (2016) demonstrated that the time period in which increases in the y_extinct_coef value had a greatest impact on light interception, and consequently on DM yield, is at the beginning of the growing season before In a recent study (Zhao et al., 2014) the modelled root biomass of wheat (Triticum aestivum L.) was improved trough re-parameterization of the ratio_root_shoot parameter using APSIM in China. Likewise, biomass partitioning between roots, rhizomes and shoots for Miscanthus has been parameterized for the WIMOVAC model using data from Beale & Long (1997) and this trait has been validated using data from Europe (Miguez, 2007). Based on the mentioned studies, and using data collected by Burks (2013) and Trybula et al. (2014) in West Lafayette, IN, the ratio_root_shoot parameter was changed in APSIM for all stages, which led to an accurate prediction of DM yield ( Fig. 9b; CCC = 0.94).
Similar to switchgrass, sensitivity analysis demonstrated definite trends associated with soil PAWC changes. However, this response differed from switchgrass in that DM yield was greater for the loamy soil than the silty and sandy soils. While the cause is not clear at this time, one plausible explanation is genotypic differences in root exploration between species depending on soil type (Monti & Zatta, 2009). These authors found that Miscanthus roots were more concentrated in the top layers of the soil profile as compared with switchgrass, which led to the crop water capture was close and negatively related to root distribution.
Most Miscanthus studies for US locations have predicted peak autumn yield (17 500-48 000 kg DM ha À1 ) and assumed adequate soil moisture and nutrient availability (Heaton et al., 2004;Khanna et al., 2008;Jain et al., 2010;Mishra et al., 2013). However, our predicted DM yields from the validation work for this same region (17 000 kg ha À1 at IN to 20 500 kg ha À1 at NE) are lower because Miscanthus was grown under water and/or nutrient-limiting dryland conditions, (except in CA).
The APSIM Sugarcane module was able to satisfactorily predict Miscanthus DM yields at IN locations (calibration) and for most locations evaluated (validation) with the exception of NJ and KY. As was discussed for switchgrass, the poor ability of the model to predict DM yield of Miscanthus at these two locations was associated with the inaccurate estimation of PAWC. An over-estimation of DUL was found in the Maury soil series at KY (Fig. 7c). In fact, the estimated soil water parameters were not in a satisfactory range as compared with the observed values (0.80 for CCC and 30% for SB) at this site when compared with the soils at IL, IN and NY (0.96-0.98 for CCC and 0-14% for SB). Similarly, a poor fit was found between observed and estimated DUL in the soil layers from 0.5 to 1.5 m in the Holmdel soil series in NJ (Fig. 7g); however, the prediction level of the LL and DUL using the HYDRAULIC PROPERTIES CALCULATOR Software developed by Saxton & Rawls (2006) was acceptable in this soil (0.92 for CCC and 13% for SB).
APSIM model: a promising tool to simulate DM yield for switchgrass and Miscanthus in several US environments This work was the first attempt to re-parameterize two current APSIM plant modules (Lucerne and Sugarcane) for predicting the DM yield of switchgrass and Miscanthus. Such re-parameterization was conducted based on an extensive literature review and using detailed experimental datasets. We initially focused on the re-parameterization of plant and soil modules and on predicting the direction and the magnitude of the DM yield responses.
The study demonstrates: • The simulation of switchgrass DM yield in northern locations of the US using the re-parameterized APSIM Lucerne module had greater accuracy than in southern ones. The improved predictions were associated with a strong, positive association between DM yield and accumulated annual rainfall.
• The original version of the APSIM Sugarcane module can be used to accurately simulate the growth and yield of Miscanthus in a broad range of geographies and ecosystems within the US that differ in local weather, soil characteristics, and crop management.
• The predictions of the DM yield for Miscanthus improved substantially when the physiological parameters (rue and ratio_root_shoot) of the model were modified.
• PAWC parameterization in a soil profile was critical for explaining DM yield differences for both crops.
This study represents an advance with respect to previous ones to simulate switchgrass and Miscanthus because: (i) the DM yield predictions were carried out with the same model (ii) the re-parameterization was started from two existing APSIM plant modules, (iii) the modelled DM yields have been compared against independent datasets, which include contrasting cultivars of switchgrass and environments, and (iv) the average errors associated with the predictions of DM yield at northern locations of switchgrass and Miscanthus were extremely low for both the calibration and the validation (26-57 kg ha À1 and 1557-86 kg ha À1 , respectively). To improve the APSIM accuracy under these environments additional agronomic studies are needed, since only a limited number of locations were utilized for each species. In addition, as our study was based on APSIM calibrations at IN locations, further calibrations of the model using data obtained from other environments is recommended. Nevertheless, these re-parameterized APSIM modules hold promise as tools for predicting switchgrass and Miscanthus yields in several US environments. Table S1. Summary of climate datasets used for evaluate the NASA-POWER data. Table S2. Means, standard deviation and variation (VAR) of soil water parameters by texture class used by sensitivity analysis of plant available water capacity (PAWC). Table S3. Summary of switchgrass datasets used for the calibration/validation of APSIM. Table S4. Summary of Miscanthus datasets used for the calibration/validation of APSIM. Table S5. Description of actual and estimated soil parameters by location used for the calibration/validation of APSIM. Table S6. Crop genotype parameterization for the calibration/validation of APSIM. Table S7. Ratio_root_shoot by phenological stage of Miscanthus with the range of values calculated based on the aerial and root biomass obtained from two field experiments at Water Quality Field Station at Purdue University (Burks, 2013;Trybula et al., 2014). The average ratio_root_shoot value by stage was used to re-parameterize the Sugarcane APSIM module.