Wetlands are responsible for 18-35% of global methane (CH4) emissions and account for the largest single source of uncertainty in the global CH4 budget. Upscaling of eddy covariance CH4 flux measurements using machine learning can provide new and independent bottom-up emissions estimates to help constrain wetland CH4 source uncertainties. Here, we join 119 site-years of eddy covariance CH4 flux data from 43 globally distributed freshwater wetland sites with geospatial predictor data to train a random forest upscaling model. Across all sites, the model was moderately accurate in reproducing wetland annual mean CH4 flux and seasonality (Nash-Sutcliffe Efficiency (NSE) = 0.54 and 0.53, respectively) with low bias. The model was best at predicting site mean CH4 flux and seasonality at temperate (NSE = 0.66 and 0.63) and boreal (NSE = 0.29 and 0.54) sites and could generalize well across different wetland classes, with best individual site performance (NSE > 0.7) at a bog (SE-Deg), marsh (US-LA2), and wet tundra site (US-Bes). At a majority of sites, the model coefficient of determination (R2) exceeded 0.5 and errors were smaller than site CH4 flux variance. The model was less accurate at the few (5) tropical sites. The upscaling model used globally-gridded predictor data and a recent wetland extent product (WAD2M) to create the first data-driven global freshwater wetland CH4 emissions product for 2001-2018. Predicted emissions showed good agreement with four test datasets. Global emission estimates of 147 ± 42.7 Tg CH4 y-1 agreed closely with a recent bottom-up process model ensemble, indicating that data-driven upscaling can provide plausible global estimates, and mid-to-high latitude upscaling emissions estimates agreed well with past inversions. In contrast, tropical emissions diverged from past bottom-up and inversion-based modeling efforts, with high emissions in the semi-arid tropics and low emissions in the humid tropics, however, tropical regions are associated with larger CH4 flux and wetland area uncertainties. Finally, a global representativeness analysis was used to provide guidance for where upscaling approaches will benefit from expanded in situ flux measurements in the tropics. We acknowledge the FLUXNET-CH4 contributors and Global Carbon Project CH4 modeling groups for the data used in this study.