Development and optimization of an Agro-BGC ecosystem model for C4 perennial grasses

Extrapolating simulations of bioenergy crop agro-ecosystems beyond data-rich sites requires biophysically accurate ecosystem models and careful estimation of model parameters not available in the literature. To increase biophysical accuracy we added C₄ perennial grass functionality and agricultural practices to the Biome-BGC (BioGeochemical Cycles) ecosystem model. This new model, Agro-BGC, includes enzyme-driven C₄ photosynthesis, individual live and dead leaf, stem, and root carbon and nitrogen pools, separate senescence and litter fall processes, fruit growth, optional annual seeding, flood irrigation, a growing degree day phenology with a killing frost option, and a disturbance handler that simulates nitrogen fertilization, harvest, fire, and incremental irrigation. To obtain spatially generalizable vegetation parameters we used a numerical method to optimize five unavailable parameters for Panicum virgatum (switchgrass) using biomass yield data from three sites: Mead, Nebraska, Rockspring, Pennsylvania, and Mandan, North Dakota. We then verified simulated switchgrass yields at three independent sites in Illinois (IL). Agro-BGC is more accurate than Biome-BGC in representing the physiology and dynamics of C₄ grass and management practices associated with agro-ecosystems. The simulated two-year average mature yields with single-site Rockspring optimization have Root Mean Square Errors (RMSE) of 70, 152, and 162 and biases of 43, −87, 156 g carbon m⁻² for Shabbona, Urbana, and Simpson IL, respectively. The simulated annual yields in June, August, October, December, and February have RMSEs of 114, 390, and 185 and biases of −19, −258, and 147 g carbon m⁻² for Shabbona, Urbana, and Simpson IL, respectively. These RMSE and bias values are all within the largest 90% confidence interval around respective IL site measurements. Twenty-four of twenty-six simulated annual yields with Rockspring optimization are within 95% confidence intervals of Illinois site measurements during the mature fourth and fifth years of growth. Ten of eleven simulated two-year average mature yields with Rockspring optimization are within 65% confidence intervals of Illinois site measurements and the eleventh is within the 95% confidence interval. Rockspring optimized Agro-BGC achieves accuracies comparable to those of two previously published models: Agricultural Land Management Alternatives with Numerical Assessment Criteria (ALMANAC) and Integrated Farm System Model (IFSM). Agro-BGC suffers from static vegetation parameters that can change seasonally and as plants age. Using mature plant data for optimization mitigates this deficiency. Our results suggest that a multi-site optimization scheme using mature plant data from more sites would be adequate for generating spatially generalizable vegetation parameters for simulating mature bioenergy crop agro-ecosystems with Agro-BGC.