Evaluation of wheat yield in North China Plain under extreme climate by coupling crop model with machine learning

Assessing the impact of climate extremes on crop production is an important prerequisite for exploring agronomic practices to deal with changing climate. Process-based crop models are effective tools to assess the effect of climate change on crop yield, but cannot accurately express the impact of extreme climate events on crop yield. In this study, we developed a series of hybrid models by incorporating the APSIM model outputs and the most informative growth stage-specific extreme climate indices (ECIs) selected by two feature selection techniques (stepwise regression, SR and genetic algorithm, GA) into two machine learning algorithms (random forest, RF and light gradient boosting machine, LGBM) to evaluate impacts of climate extremes on wheat yields in the North China Plain (NCP). The results showed that the RF model outperformed the LGBM model in estimating wheat yield regardless of input variables. Applying feature selection to two machine learning algorithms can greatly reduce computational cost without significantly affecting model accuracy. The APSIM+RF-GA hybrid model was the optimal model for estimating wheat yield with explained 93 % of the observed yield variation and the accuracy of the model is improved by 33 % compared with the APSIM model alone. Extreme low temperature events before flowering and extreme high temperature events after flowering are the main extreme climate events causing the loss of wheat yield. In addition, we evaluated the impact of future climate change on wheat yield using the APSIM+RF-GA hybrid model and the APSIM model, respectively. Yields projected using a single APSIM model increased at all stations but yields projected using APSIM+RF-GA model decreased at 12.5–28.1 % of stations in the NCP under future climate scenarios. Compared to the APSIM+RF hybrid model, the future yield projected using single APSIM model might be overestimated by 12.7–19.2 % because of underestimating the yield loss caused by climate extremes. The increase of heat stress after flowering and frost stress during floral initiation to flowering were the main factors for future yield loss. Using the machine learning algorithm to make an external modification to the outputs of the APSIM model could improve the accuracy of yield estimation under extreme climate conditions and the method is more suitable for projecting future crop yield. This study is conducive to developing adaptation strategies to alleviate the negative impacts of future climate extremes on crop production.