Crop Model Ensemble Averaging: A Large But Underappreciated Uncertainty Source for Global Crop Yield Projections Under Climate Change

Abstract

Using an ensemble of crop models have been encouraged for global crop yield projections, which would, however, introduce additional uncertainty from the choice of ensemble averaging methods. Here, we use seven ensemble averaging methods including simple average, regression, Support Vector Regressor, Bayesian model average (BMA), Random Forest (RF), Artificial neural network (ANN) and Long-short term memory to derive the ensemble mean of eight process-based crop models for global maize yield projections. Results show that the choice of ensemble averaging methods has a large impact on the projection of long-term mean yield and year-to-year yield variability, with a range of −19.79%–16.62% and −47.92%–55.39% for the globe, respectively. Regionally, the largest uncertainties from the choice of ensemble averaging methods are observed in Indonesia and Canada. Further uncertainty decomposition analysis shows that ensemble averaging methods contributes to 39%–87% of total uncertainties for global yield projections, which is even higher than climate models. These results imply that although using an ensemble of crop models is valuable for informing risk-based policy-makings, how we choose to combine and derive the best estimates of crop model ensembles has large influence on the assessment outcomes. This study highlights an important but not well recognized uncertainty source for global yield predictions which arises from the choice of ensemble averaging methods.