Application of XGBoost model and multi-source data for winter wheat yield prediction in Henan Province of China

Accurate prediction of winter wheat yield is essential for precision agricultural management. Traditional methods, which primarily rely on vegetation indices, often fall short in achieving high prediction accuracy. In this study, we proposed an innovative approach by integrating multi-source data, including vegetation, geographical, and soil features, with 61 features extracted across growth stages. Using these features, the XGBoost regression model was applied to predict winter wheat yield in Henan Province, China. The model was trained on data from 2016 to 2020 and validated using the 2021 dataset. The XGBoost model achieved a mean absolute error (MAE) of 371.36 kg/hm2, a root mean square error (RMSE) of 516.97 kg/hm2, and a coefficient of determination (R2) of 0.85. These results significantly outperformed other models, such as random forest (RF), gradient boosting decision tree (GBDT), and Lasso. Notably, the XGBoost model provided high accuracy 2–3 growth stages before harvest, enabling earlier yield prediction than traditional methods. To enhance interpretability, SHAP (shapley additive explanations) was employed to quantify the influence of each input variable on yield and determine the direction of influence. This study underscores the potential of multi-source data and advanced machine learning techniques for accurate and interpretable winter wheat yield prediction, providing a valuable solution for precision agriculture.