Prediction of greenhouse gas emissions from agricultural fields with and without cover crops

The significant contribution of greenhouse gas (GHG) emissions to global climate change and stratospheric ozone depletion has been calling the attention to assess the effect of agricultural management on them. Although machine learning (ML) methods have been widely used for the quantification of various inherent and dynamic soil properties, the accuracy of these techniques in the prediction of agricultural soil GHG emissions remains unclear. Therefore, this study aims at evaluating the performance of six different ML methods including simple linear regression, Cubist, support vector machines (SVM) with three different kernel functions, and random forest (RF) for the prediction of CO2 and N2O fluxes from plots managed with and without cover cropping. The input parameters include typical meteorological data as well as soil temperature and soil water content. Results show that the daily flux of CO2 is positively correlated with soil temperature and solar radiation and negatively with soil water content, whereas daily N2O flux is positively correlated with soil water content. For plots without cover cropping, Cubist and SVM with polynomial kernel function outperformed in the prediction of the testing dataset with RMSE of 14.18 kg ha–1 d–1 and 11.52 g ha–1 d–1 for daily CO2 and N2O emissions, respectively. Where cover cropping was considered, SVM with radial basis function and Cubist were the best models in the prediction of daily CO2 and N2O fluxes with respective RMSE values of 15.71 kg ha–1 d–1 and 14.12 g ha–1 d–1. Considering the ranking of the models for all scenarios and both GHGs, SVM with nonlinear kernel functions and Cubist method surpassed other ML techniques