Machine Learning-Based Prediction of N2O Emissions from Tea Plantations and Identification of Driving Factors for Sustainable Nitrogen Management

Abstract

Tea plantations are high-input agricultural systems and have been recognized as hotspots of soil nitrous oxide (N2O) emissions; however, the key controlling factors of these emissions and their quantitative prediction remain insufficiently understood. We compiled 115 field-observation records from 26 published studies into a multi-factor database covering climate, soil properties, and fertilization management, and compared five machine learning models—multiple linear regression (MLR), ridge regression, support vector regression (SVR), random forest (RF), and gradient-boosting regression trees (GBRTs)—using 5-fold cross-validation, combined with Spearman correlation and feature-importance analyses. Annual N2O emissions varied widely (0.40–73.20 kg·hm−2·a−1; mean 9.85 kg·hm−2·a−1), and the mean direct emission factor (EFd, 2.04%) far exceeded the IPCC default value. Emissions were significantly positively correlated with total nitrogen (TN) input but negatively correlated with mean annual temperature (MAT) and mean annual precipitation (MAP). GBRT performed best, effectively capturing nonlinear multifactor interactions; TN input and soil pH were the dominant predictors, followed by rainfall. However, feature importance rankings were method-dependent: the RF/SHAP analysis ranked MAT first rather than fifth, reflecting the different algorithmic mechanisms of the two approaches. The GBRT-based model provides a useful tool for estimating tea-plantation N2O emissions (LOOCV R2 = 0.668) and quantitative support for sustainable nitrogen management and targeted greenhouse gas mitigation strategies in tea production systems.