Seyed Arash Khosravani Shariati, Ali Abbasi
Accurate crop yield prediction and understanding its underlying factors facilitate better food supply management and more informed decision-making. To forecast crop yield, the majority of previous studies have utilized vegetation indices and meteorological data. However, other important factors are often overlooked. Moreover, the temporal influence of input variables has been underexplored in prior research. To fill these gaps, we integrated a diverse range of satellite-based data, including vegetation indices and actual evapotranspiration (ET a ), with climate and soil information. Then, the input variables were narrowed down using a feature selection approach to provide the most relevant variables for predictive models. Three machine learning algorithms, Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Linear Regression (LR), were trained to forecast winter wheat yield across Oklahoma and Kansas counties. The models were trained on 2014–2021 data and tested on 2022–2023 yields. According to the results, XGBoost emerged as the most accurate algorithm in both test years. It achieved an R² of 0.71 (RMSE = 0.46 t ha −1 ) in 2022 and an R² of 0.63 (RMSE = 0.60 t ha −1 ) in 2023 when using the selected feature set. For most models, particularly in 2022, using the selected features instead of the entire set improved the accuracy. We also found that ET a is a promising factor in yield prediction, as it was selected multiple times across the growing season in the feature selection process. Additionally, correlation analysis showed that April and May, which are two to three months before harvest, were the most sensitive months in shaping the final yield. • The influence of variables on wheat yield was assessed using correlation analysis and feature selection. • XGBoost performed better than LR and RF. • Feature selection enhanced model accuracy in most cases compared to using all available variables. • ET a provided useful information for wheat yield forecasting.