Hybrid Process and Machine Learning Framework for Greenhouse Gas Mitigation in Legume Cropping Systems

Abstract

Agricultural soils provide around 25% of worldwide anthropogenic greenhouse gas emissions, with synthetic nitrogen fertilisers playing a substantial role in nitrous oxide emissions via nitrification and denitrification processes. Legume-based farming systems provide a sustainable solution by organically fixing atmospheric nitrogen and decreasing reliance on fertilisers. Nonetheless, their efficacy in reducing emissions across various management approaches remains inadequately investigated. This paper presents a Hybrid Process and Machine Learning Framework designed to anticipate nitrous oxide emissions and categorise emission severity levels in legume-based agricultural systems. Data from twenty-two years of field observations at the Kellogg Biological Station Long-Term Ecological Research Main Cropping System Experiment were examined. The dataset comprised 3,696 observations across four management treatments: conventional, no-till, reduced-input with legume cover crops, and biologically based organic systems within a corn–soybean–winter wheat rotation. Five machine learning algorithms were developed, and the three most effective models were integrated into an ensemble framework. The ensemble model demonstrated robust predictive ability, evidenced by an R-squared value of 0.91 and a root mean square error of 0.51 g N ha⁻¹ day⁻¹. The accuracy and F1-score of the framework were 0.83 for the classification of emissions to low, medium and high categories. Emissions reduced by nearly 45% below the traditional systems in legume-based systems on average. The proposed framework provides a strategic and effective approach to support climate-smart agriculture and reduce GHGs.