TRS-CatBoost: An Interpretable Corn Yield Prediction Model Based on an Integrated Learning Approach

As one of the three major crops in the world, accurate prediction of maize yield can provide a reliable scientific basis for its planting decisions and trading activities. At present, many studies still use traditional machine learning methods for prediction, which still has some shortcomings in terms of training efficiency, accuracy, and interpretability. In this paper, an interpretable corn yield prediction model based on an integrated learning approach is proposed for the first time and validated using multi-source environmental feature data. Firstly, the original data are filtered out by feature screening using recursive feature elimination (RFE) to filter out the feature factors affecting the prediction accuracy; subsequently, the processed data are input into CatBoost model for training, and the model is optimized using Bayesian optimization algorithm based on the tree structure (TPE); finally, the model is combined with the SHAP interpretable algorithm to make a global interpretation of the model and determine the important factors affecting maize yield. In this paper, corn yield data from an EU project are used, and the results show that the TRS-CatBoost-based corn yield prediction model performs well, with a 36.97 reduction in RMSE over XGBoost, which is the best-performing of the remaining four integrated models optimized using TPE (Random Forest, Gradient Boosting Tree, Extreme Gradient Boosting Tree, and Lightweight Gradient Boosting Machine). The TRS-CatBoost model proposed in this paper is more in line with the real situation of corn yield, which not only improves the accuracy of corn yield prediction but also enhances the use value and acceptability of the integrated model in agricultural production.