<p>Accurate estimation of methane emissions from rice fields can provide essential information for formulating methane mitigation policies. Machine learning is increasingly being used to estimate methane emissions from rice fields. However the constraints of data representativenessand key feature selection hinder its application on a global scale. To enrich estimation methods and expand datasets, we developed a data-driven model to estimate global methane emissions from rice fields using 2839 observed data manually extracted from 305 articles. With an average R²value exceeding 0.6 alone with the lowest root mean square error and mean absolute error our results showed that the eXtreme Gradient Boosting had the best performance. The SHapley Additive exPlanations (SHAP) method identified the most important drivers:agricultural management, soil texture and geographical variations. The model estimation showed that the global average methane emissions from rice fields reached 24 .74 Tg yr^− 1in 2020 withan estimated range of 23 .82 to 26.74 Tg yr^− 1derived from five independently trained modelsand 90 .63% of these emissions were concentrated in Asia. Based on global observation data changing water management practices offersa promising approach for methane mitigation, and adopting alternate wetting and drying instead of continuous flooding can effectively reduce methane emissions by 10 .17%. The results proved that machine learning is an effective method to estimate methane emissions from rice fields at national, regional and global scales, and can serve as an important tool for rice field methane emission estimation and mitigation potential evaluation.</p> Graphical Abstract <p></p>