Effects of automatic multi-objective optimization of crop models on corn yield reproducibility in the U.S.A

This study presents a detailed analysis relating to the effectiveness and efficiency of the multi-objective complex evolution (MOCOM-UA) algorithm on the reproducibility of corn yield for the period 1999–2010 in the United States of America. The algorithm was coupled with the Environmental Policy Integrated Climate (EPIC) model and is capable of solving the multi-objective calibration problem for crop models. Understanding how model parameters are determined from past research work is especially difficult because parameter determination is usually performed by manual trial and error methods and still remains a black box. Single-objective functions are often inadequate to properly measure all characteristics of the observed datasets and do not usually provide parameter estimates that are considered acceptable. In contrast, a multi-criteria optimization would allow the analysis of the trade-offs among different criteria, and would indicate the limitations of the current crop model structure better. Therefore, a multi-objective algorithm was used in this study to identify the critical model parameters and to glean a better understanding of model uncertainties. The uncertainty analysis of the model was performed by comparing simulation results obtained using randomly generated initial parameters and inputs with the attainable calibrated parameters and inputs. Three model parameters (biomass to energy ratio (BE), harvest index (HI), potential heat units (PHU)) and two inputs (planting dates (PDAY), planting density (PDENS)) were optimized on a grid-scale to capture a more realistic spatial pattern and temporal variability of corn yield. The conclusions derived from this study are: (1) the critical parameters needed to reproduce corn yield with EPIC are BE, HI, and PHU; (2) the MOCOM-UA calibration procedure is highly effective in improving the spatial agreement and temporal variability of corn yield and for reducing the width of the uncertainty band; (3) by using the calibrated parameter sets, measurement errors of spatial-averaged ensemble mean yield are significantly improved; (4) simulation using the EPIC model showed better performance in the major corn area with relatively accurate input datasets, but performed poorly in the minor corn area with lower yield and corn area ratio. This study highlights the projected contribution of the multi-objective optimization procedure in identifying critical model parameters in reducing the width of the uncertainty band, and in improving the spatial agreement and temporal variability of corn yield. Future work should focus on the effect of optimizing more appropriate parameters such as those that affect water availability or soil fertility status on crop yield reproducibility.