关键词: 玉米秸秆, 农艺性状, 机器学习, 特征筛选, 回归模型

Abstract:

Based on 263 mature corn straw samples, Pearson correlation coefficient (PCC), variance inflation factor (VIF), grey relational analysis (GRA) and recursive feature elimination (RFE) were employed to optimize model input variables, with stem hierarchy−centered trait combinations being identified and optimal machine learning models that integrate agronomic traits being explored for efficient, cost−effective, and rapid non−destructive field estimation of corn straw dry weight. Experimental results identified eight optimal traits through rigorous evaluation of operational feasibility and predictive performance: ear stem diameter, ear height, basal long axis, ear leaf area, cob diameter, cob length, basal short axis and plant height. Machine learning evaluations demonstrated the Extra trees regressor (ETR) achieved superior predictive accuracy (R²=0.92, RMSE=13.52g) compared to eXtreme gradient boosting (XGBoost) implementations. Feature importance and SHAP interpretability analysis revealed stem-related traits as dominant predictors, with basal long axis and ear stem diameter collectively contributing 41.7% of predictive influence. The optimized trait combination offers a reliable, non−destructive approach for estimating corn straw biomass, providing a theoretical foundation and technical framework for the sustainable utilization of corn straw resources. 

Key words: Corn straw, Agronomic traits, Machine learning, Feature selection, Regression model