Abstract
Climate-smart agriculture emphasizes improving crop production while mitigating greenhouse gas (GHG) emissions. In recent years, machine learning (ML) models have been increasingly used to unlock complex problems in agriculture, such as an enhanced understanding of drivers regulating GHG emissions. However, these models have not been evaluated widely, specifically quantifying GHG emissions in arid and semi-arid cropping systems. We evaluated soil GHG emissions in irrigated crop rotations with and without cover cropping using field-based measurements and five ML models, that is, decision tree, random forest (RF), gradient boosting (GB), bagging regressor, and XGB booster. The ML model was trained and evaluated using >1600 data points of daily carbon dioxide (CO2), and nitrous oxide (N2O) measurements from four cover cropping practices (no-cover crop and grass-brassica, grass-legume, and grass-brassica-legume cover crops) along with environmental (air temperature, precipitation), soil (moisture, temperature), and crop and management (cover crop carbon content, irrigation) variables and CO2 and N2O emissions was predicted. The RF model predicted CO2 emissions (R-2 up to 0.68) more accurately, and N2O emissions were predicted better by the RF and GB models than other models with R-2 up to 0.56 and 0.55, respectively. Irrigation was the most critical driver of CO2 emissions, and air temperature was the main driver of N2O emissions. ML models can effectively estimate GHG emissions using simple predictors, such as field management and environmental parameters, and help in designing climate-smart management in arid and semi-arid regions.