A dynamic agricultural prediction system for large-scale drought assessment on the Sunway TaihuLight supercomputer

Abstract

Crop models are widely used to evaluate the response of crop growth to drought. However, over large geographic regions, the most advanced models are often restricted by available computing resource. This limits capacity to undertake uncertainty analysis and prohibits the use of models in real-time ensemble forecasting systems. This study addresses these concerns by presenting an integrated system for the dynamic prediction and assessment of agricultural yield using the top-ranked Sunway TaihuLight supercomputer platform. This system enables parallelization and acceleration for the existing AquaCrop, DNDC (DeNitrification and DeComposition) and SWAP (Soil Water Atmosphere Plant) models, thus facilitating multi-model ensemble and parameter optimization and subsequent drought risk analysis in multiple regions and at multiple scales. The high computing capability also opens up the possibility of real-time simulation during droughts, providing the basis for more effective drought management. Initial testing with varying core group numbers shows that computation time can be reduced by between 2.6 and 3.6 times. Based on the powerful computing capacity, a county-level model parameter optimization (2043 counties for 1996–2007) by Bayesian inference and multi-model ensemble using BMA (Bayesian Model Average) method were performed, demonstrating the enhancements in predictive accuracy that can be achieved. An application of this system is presented predicting the impacts of the drought of May–July 2017 on maize yield in North and Northeast China. The spatial variability in yield losses is presented demonstrating new capability to provide high resolution information with associated uncertainty estimates.