Unveiling the Responses of Maize Yield to Management Practices in the Mollisols of Northeast China: A Machine Learning Perspective

Abstract

Soil properties, fertilization, tillage, and straw management practices vary widely in the Mollisol region of Northeast China. Such multivariate complexity masks the individual and interactive effects of these factors on maize yield and limits the ability to define high -yield ranges or effective practice combinations. To address this issue, we applied four interpretable machine learning algorithms (XGBoost (XGB), random forest, extra trees, and multi-layer perceptron) to predict the effects of soil properties, fertilization types, tillage types, and cropping systems on maize yield and reveal the complex relationship between the factors affecting maize yield. Results showed that the XGB model provided the best prediction accuracy with R2 (0.8562). N fertilizer, P fertilizer, soil organic carbon (SOC), pH, and available nitrogen (AN) were the important factors affecting maize yield. The high-predicted-yield ranges were identified as N fertilizer 180–220 kg/hm2, P fertilizer 40–80 kg/hm2, SOC content of 12–18 g/kg, pH of 6.5–7.5, and AN content of 120–150 mg/kg. Higher maize yields were associated with combined chemical fertilizer and straw return, deep tillage, and monocropping. Under these optimal conditions, maize yields reached 10,000–12,000 kg/hm2. Key interactive effects included moderate SOC content in combination with chemical fertilizer and straw return or deep tillage, moderate N and P fertilization application, and AN exceeding 120 mg/kg together with available phosphorus above 40 mg/kg, collectively contributing to higher maize yield. These findings provide a practical, evidence-based reference for precision soil management aimed at stabilizing yield and improving soil fertility in the Mollisol region.