Coupling Process-Based Models and Machine Learning Algorithms for Predicting Yield and Evapotranspiration of Maize in Arid Environments

Crop yield prediction is critical for investigating the gap and potential adaptations to environmental management factors in arid regions. models (CMs) are powerful tools predicting water use, but they still have some limitations uncertainties; therefore, combining them with machine learning algorithms (MLs) could improve predictions reduce uncertainty. To that end, DSSAT-CERES-maize model was calibrated one location validated others across Egypt varying agro-climatic zones. Following that, dynamic (CERES-Maize) used long-term simulation (1990–2020) of maize grain (GY) evapotranspiration (ET) under a wide range factors. Detailed outputs from three growing seasons field experiments Egypt, as well CERES-maize outputs, were train test six (linear regression, ridge lasso K-nearest neighbors, random forest, XGBoost), resulting more than 1.5 million simulated scenarios. Seven warming years (i.e., 1991, 1998, 2002, 2005, 2010, 2013, 2020) chosen 31-year dataset MLs, while remaining 23 models. The Ensemble (super learner) XGBoost outperform other GY ET maize, evidenced by R2 values greater 0.82 RRMSE less 9%. broad practices, when averaged all locations 31 simulation, not only reduced hazard impact also increased ET. Moving beyond interpreting Lasso XGBoost, using global local SHAP values, we found most important features maximum temperatures, minimum temperature, available content, soil organic carbon, irrigation, cultivars, texture, solar radiation, planting date. Determining assisting farmers agronomists prioritizing such over order increase resource efficiency values. combination CMs ML tool use regions, which particularly vulnerable climate change scarcity.