Reducing the impact of model scale on simulated, gridded switchgrass yields

Results of gridded ecosystem simulations of bioenergy crops are used for estimating economic viability, environmental impacts, and potential land use change. Gridded model uncertainty propagates through these uses, thus we propose a simple method for estimating regional, spatial model error from sparse field data. We apply this method to the Agricultural-BioGeochemical Cycles (Agro-BGC) model to examine and reduce the model uncertainty associated with grid scale for simulated switchgrass yields in a 6° latitude × 5° longitude (∼300,000 km2) region covering Illinois, United States of America. Based on three evaluation sites, changes in yield with scale result from complex intra-model interactions driven by a combination of meteorological rather than soil or terrain variables. Spatial bias of the regional mean significantly increases with increasing cell size for 11 of 15 measurement dates. This bias is primarily due to grid scale, thus bias correction of output yield reduces the model uncertainty associated with grid scale. The corresponding Root Mean Squared Error and Bias-Corrected RMSE (RMSEBC) have effectively negligible trends with inconsistent signs. The range of RMSEBC for 2-year Average Mature August Yield (AMAY) is 267-285 g C m−2 across 3- to 3600-arcsec resolution (∼90 m-∼100 km) with biases from 9 to 61 g C m−2. AMAY bias significantly increases with increasing cell size. Spatial bias of the regional mean is relatively consistent for resolutions ≤1200 arcsec (∼33 km) (AMAY bias <3%), and larger AMAY biases (4-13%) at coarse resolutions indicate poorly characterized spatial heterogeneity. Including the 68% confidence interval around bias-corrected values, AMAY ranges from 0 to 1116 g C m−2 across a 150-arsec grid, which is similar to the range reported for 24 eastern United States field sites. Spatial bias of the regional mean yield can vary across grid resolution by as much as 31% of the observed regional mean and can dramatically affect calculations dependent on the resolution of the estimate. We conclude that grid scale profoundly affects model accuracy such that regional studies must match evaluation and simulation scales and should utilize multi-scale analyses to determine robustness of results.