A multiple crop model ensemble for improving broad-scale yield prediction using Bayesian model averaging

Abstract

Process-based crop models are popular tools to evaluate the impact of climate change and agricultural management on crop growth. Accurate simulation of crop production over large geographic regions using an individual crop model remains challenging due to different sources of uncertainty. We present a Bayesian model averaging (BMA) method for a multiple crop-growth model ensemble to provide more reliable predictions of maize yields in Liaoning Province, northeastern China, which covers an area of 148,000km2 and has 2200,000ha of maize. We apply the photosynthesis-oriented WOFOST (WOrld FOod STudy) model, the water-oriented AquaCrop model and the nitrogen-oriented DNDC (DeNitrification and DeComposition) model to independently generate original predictions of county-level maize yields. The integrated prediction is achieved using a linear combination of the three ensemble members using BMA weights. This integrated approach results in more accurate and precise predictions than any individual model over the entire province. This is because the BMA framework effectively compensates for the uncertainty of individual model simulation and takes advantage of each competing model for reliable prediction. Furthermore, the interpretation of the BMA weight values is also strengthened by comparison with regional precipitation, fertilization and radiation data. We find these values adequately fit the regional limiting factors, e.g., the AquaCrop model generally has a high weight value in counties with frequent droughts, while WOFOST is the dominant member in areas with radiation deficit. Compared with the simple average method and median estimate, the results show that the BMA framework is powerful in computing the ensemble weights and interpreting the mechanism beyond the observed data.