Projecting maize yield in the North China Plain under climate extremes: A hybrid crop model-machine learning approach

<title>Abstract</title> <p>Process-based crop models are effective tools for assessing the impact of climate change on crop yields, but they have limitations in accurately describing the effects of extreme climate events on yields. We constructed a hybrid simulation framework that integrates a process-based model (APSIM) with machine learning (ML) models, i.e., random forest (RF) and light gradient boosting machine (LGBM). We applied two feature selection techniques, i.e., stepwise regression (SR) and genetic algorithm (GA), to identify the most informative extreme climate indicators (ECIs) for specific growth stages. These selected ECIs, along with the outputs of the APSIM, were then incorporated into the RF and LGBM to evaluate the impacts of extreme climate on maize yields in the North China Plain. Overall, the hybrid model significantly improved the accuracy of yield predictions, and APSIM + RF+GA was the best model for estimating yields, explaining 94% and 93% of the observed yield variations for spring and summer maize, respectively. Compared to APSIM alone, the hybrid model improved simulation accuracy by 30% and 31% for spring and summer maize, respectively. Overall, the APSIM + RF+GA projected significantly greater yield losses than the standalone APSIM, highlighting its superior ability to capture the severe impacts of extreme climate events. For future simulation under SSP585_2080s scenario, the hybrid model predicted yield losses of 43.6% for spring maize and 26.8% for summer maize without the carbon dioxide (CO₂) fertilization effect. With the CO₂ considered, the hybrid model still predicted yield losses of 25.0% for spring maize and 24.6% for summer maize. Although elevated future CO₂ concentrations exerted a significant positive effect on maize yield, this benefit was far from sufficient to offset the negative impacts of climate change. Overall, the hybrid model developed in this study addresses the limitations inherent in standalone process-based models and demonstrates stronger suitability for forecasting future crop yields.</p>