An interdisciplinary machine learning-based approach for predicting corn grain yield under biofertilizer applications

Predicting grain yield is a critical aspect of agriculture that assists farmers and planners in managing resources more efficiently and enhancing productivity. This study investigates and compares eight machine learning algorithms for predicting corn grain yield (Zea mays L. ) using 73 plant and soil features, including 32 primary features and 41 engineered interaction features. Experiments were conducted over two years at the research farm of Ferdowsi University of Mashhad, and the data were processed using random splitting (70% training, 15% validation, 15% testing) and standardization (StandardScaler). Initially, a stepwise backward regression model identified 13 key features (e. g. , Canopy Temp₃, % P plant) with an adjusted R 2 of 58. 53%. Subsequently, among fifteen common algorithms, ANFIS, Transformer, ANN, SVR, LightGBM, XGBoost, Enhanced DNN, and SVM were shortlisted and evaluated. Metrics including R2, RMSE, MAE, and Willmott’s d indicated that ANFIS (R 2 = 0. 555), Transformer (R2 = 0. 545), and ANN (R 2 = 0. 518) performed best, while SVM (R 2 = 0. 325) was the weakest. SHAP plots and Decision Plots revealed that interactions such as Leaf Area IndexDry Matter Yield and Canopy Temp₄ play a key role in prediction. Correlation analysis identified two main clusters (physiological and yield-related), and the skewed data distributions confirmed the necessity for nonlinear models. Results demonstrated that neural network models (especially those with Attention mechanisms and TensorFlow) better model complex ecophysiological relationships. Features like canopy temperature and the interaction between plant nitrogen content and root colonization percentage were identified as primary variables, highlighting the importance of nutrient uptake and plant physiological responses. This study showed that combining interaction features with advanced machine learning algorithms can, on one hand, improve prediction accuracy and contribute to enhancing crop production system efficiency toward sustainable agriculture goals, and on the other hand, provide the foundation for precision agronomic management, identification of effective ecophysiological pathways in final yield formation, and adaptation to climate change.