A Stacked Ensemble Machine Learning Framework for Predicting Nitrogen Reduction Impact on Agricultural N₂O Emissions

Abstract

To accurately predict the impact of nitrogen reduction strategies on greenhouse gas emissions in agricultural systems and support informed decision-making for a green, low-carbon transition, this study proposes a multi-model ensemble prediction framework using integrated machine learning techniques. To address the limitations of traditional emission assessment methods—such as their inability to handle high-dimensional data, nonlinear interactions, and regional heterogeneity—three models are employed: eXtreme Gradient Boosting (XGBoost), Random Forest, and Light Gradient Boosting Machine (LightGBM). These models are combined using a Stacking ensemble approach to enhance predictive accuracy and robustness. Comprehensive performance evaluations and variable importance analyses were conducted. The proposed framework achieved an R² of 0.901 and a Root Mean Square Error (RMSE) of 0.301 on the N₂O Global dataset, significantly outperforming benchmark models such as Natural Gradient Boosting (NGBoost) (R²: 0.829; RMSE: 0.382) and Categorical Boosting (CatBoost) (R²: 0.864; RMSE: 0.341). It also demonstrated strong adaptability on the Rothamsted and DayCent datasets. SHapley Additive exPlanations (SHAP) analysis identified nitrogen application rate, rainfall frequency, and soil type as the most influential factors affecting emissions. Scenario simulations showed that a 20% reduction in nitrogen application could decrease emissions to 3.124 kg N₂O-N/ha/year. When combined with optimized management practices—such as cover cropping and improved tillage scheduling—emissions were further reduced to 2.487 kg N₂O-N/ha/year. Accordingly, the proposed method provides a practical exploration for enhancing the applicability and interpretability of models. It offers valuable insights for agricultural environmental modeling and emission reduction policy support, demonstrating both theoretical significance and practical relevance.