Carbon dynamics, emission mitigation, and yield optimization in farmlands: A machine learning framework for multi-variable prediction

Machine learning (ML) has become a promising approach in agro-ecosystem applications to simulate different fertilization management practices (FMP) and their impact on crop yield, soil organic carbon (SOC) concentration, and greenhouse gas (GHG) emissions. However, existing ML-based studies often focus on predicting single variables and lack a systematic framework, limiting model reliability and practical application. This study addresses these gaps by proposing a systematic, modular framework for ML-based multi-variable prediction in agro-ecosystems. Utilizing a comprehensive field-measured dataset, we developed a multi-variable system to predict crop yield, SOC accumulation, and GHG emissions (N₂O, CH₄) for three staple crops in China. The study evaluates a range of preprocessing techniques and algorithms on model performance and demonstrates the model’s application in the North China Plain (NCP) to identify FMPs that balance productivity with carbon mitigation. Our results show that k-Nearest Neighbors missing data imputation, Local Outlier Factor outlier detection, and the Random Forest ML algorithm combined to deliver the best performance for yield (R² = 0.83), SOC (R² = 0.91), and N₂O emissions (R² = 0.83). These models outperformed other candidates across all environmental, soil, and management subgroups. Notably, models with similar accuracy when evaluated with measured variables can show substantial variability in predicting the effects of FMPs compared to conventional practices, highlighting the importance of robust model selection for providing reliable guidance in optimizing FMPs. Partial Dependence Plot (PDP) analysis revealed distinct phases in SOC accumulation, underscoring the need for input datasets with broad temporal coverage to capture both short-term dynamics and long-term trends in SOC dynamics. Overall across the case study area, we identify that 50% Manure-N substitution can reduce global warming potential by 29.5% for maize and 19.5% for wheat while increasing SOC concentration more than fivefold over 30 years.