Abstract
Crop yield can be affected by crop water use and vice versa, so when trying to simulate one or the other, it can be important that both are simulated well. In a prior inter-comparison among maize growth models, evapotranspiration (ET) predictions varied widely, but no observations of actual ET were available for comparison. Therefore, this follow-up study was initiated under the umbrella of AgMIP (Agricultural Model Inter-Comparison and Improvement Project). Observations of daily ET using the eddy covariance technique from an 8-year-long (2006-2013) experiment conducted at Ames, IA were used as the standard for comparison among models. Simulation results from 29 models are reported herein. In the first "blind" phase for which only weather, soils, phenology, and management information were provided to the modelers, estimates of seasonal ET varied from about 200 to about 700 mm. Subsequent three phases provided (1) leaf area indices for all years, (2) all daily ET and agronomic data for a typical year (2011), and (3) all data for all years, thus allowing the modelers to progressively calibrate their models as more information was provided, but the range among ET estimates still varied by a factor of two or more. Much of the variability among the models was due to differing estimates of potential evapotranspiration, which suggests an avenue for substantial model improvement. Nevertheless, the ensemble median values were generally close to the observations, and the medians were best (had the lowest mean squared deviations from observations, MSD) for several ET categories for inter-comparison, but not all. Further, the medians were best when considering both ET and agronomic parameters together. The best six models with the lowest MSDs were identified for several ET and agronomic categories, and they proved to vary widely in complexity in spite of having similar prediction accuracies. At the same time, other models with apparently similar approaches were not as accurate. The models that are widely used tended to perform better, leading us speculate that a larger number of users testing these models over a wider range of conditions likely has led to improvement. User experience and skill at calibration and dealing with missing input data likely were also a factor in determining the accuracy of model predictions. In several cases different versions of a model within the same family of models were run, and these within-family inter-comparisons identified particular approaches that were better while other factors were held constant. Thus, improvement is needed in many of the models with regard to their ability to simulate ET over a wide range of conditions, and several aspects for progress have been identified, especially in their simulation of potential ET.