Optimization of crop yield prediction models based on stratified soil properties

• Reveal that the impact of deep soil properties on crop yield formation is comparable to that of topsoil, clarifying their critical role in yield regulation. • Clarify that multicollinearity is a core issue requiring key attention and regulation when incorporating layered soil properties into grain yield prediction models. • Propose a zonal integration model, which significantly enhances the response accuracy and stability of the model to deep soil properties through zonal feature engineering. Accurate crop yield prediction facilitates the optimization of agricultural management practices, thereby supporting governmental departments in trade policy formulation and national economic planning. Contemporary crop yield prediction models exhibit overreliance on topsoil parameters, neglecting three-dimensional soil interactions while demonstrating suboptimal feature extraction efficiency, which fundamentally limits accuracy improvement. The primary objectives of this study are to analyze the relationship between stratified soil properties and maize yield in the black soil of Northeast China, and to propose a zonal-integrated modeling framework that enhances model feature extraction capability and incorporates three-dimensional soil-crop interactions to improve crop yield prediction accuracy. We employed Pearson correlation analysis to quantify stratifed soil-yield relationships, developed the partition-prioritized modeling (PPM) framework using ensemble learning, and conducted comparative evaluations against baseline models, such as multiple linear regression, random forest, gradient boosting regression, and XGBoost to assess PPM's efficacy. Key findings revealed significant vertical differentiation of soil attributes was observed, with deep soil properties (60-100 cm) demonstrating comparable explanatory power to topsoil (0-20 cm) in yield formation. Direct integration of stratified soil properties induced collinearity and the curse of dimensionality in predictive models. The proposed PPM framework effectively resolved these challenges by leveraging stratified soil properties, achieving a minimum 11.76% improvement in prediction accuracy compared to baseline models. These results not only confirm the critical role of soil vertical heterogeneity in yield modeling but also provide a methodological breakthrough for multi-strata soil data utilization in agricultural forecasting systems.