PNNL

Abstract

Modeling biophysical processes is a complex endeavor because of large data requirements and uncertainty in model parameters. Rather than being considered absolute, model predictions should incorporate, when possible, analyses of their uncertainty and sensitivity. The study incorporated uncertainty analysis on EPIC (Environmental Policy Impact Calculator) predictions of corn (Zea mays L.) yields and soil organic carbon (SOC) using the Generalized Likelihood Uncertainty Estimation (GLUE). An automatic parameter optimization procedure was developed at the conclusion of sensitivity analysis, which was conducted using the extended Fourier Amplitude Sensitivity Test (FAST) for nine model parameters for the two EPIC components investigated. The analyses were based on an experimental field under 34-yr continuous corn with 5 N treatments at Arlington Agricultural Research Station in Wisconsin. The observed mean corn yields fell well within the 5% and 95% confidence limits. The width of 90% confidence interval bands for corn yields ranged from 0.31 to 1.6 Mg ha-1, while predicted and observed means were 3.26 - 6.37 Mg ha-1 and 3.28 - 6.4 Mg ha-1, respectively, for the 5 treatments. The 90% confidence width for SOC was 0.97-2.13 g kg-1, while predicted means and observed SOC were 17.4-22.3 g kg-1 and 19.2 to 22.9 g kg-1, respectively. Parameter estimations were provided for the most influential and uncertain ones. The optimal parameter set gave an R2 of 0.96 for mean corn yield predictions and 0.89 for yearly SOC. EPIC was dependable and accurate, from a statistical point of view, in predicting mean corn yields and SOC dynamics.