Wetlands are responsible for ~30% of global methane (CH₄) emissions and introduce some of the largest uncertainties in the global CH₄ budget. Comparison of CH₄ emissions simulated by land surface models show that uncertainties arise from differences in model parameterization and wetland extents. Data-driven upscaling may help constrain model uncertainties by providing independent emissions estimates. Although eddy covariance flux measurements of carbon dioxide have been upscaled to estimate global gross primary production, similar products for CH₄ fluxes are lacking. To address this gap, we use data from 47 FLUXNET-CH₄ freshwater wetland sites with machine learning to 1) evaluate data-driven model performance in predicting global CH₄ fluxes; 2) identify useful classes of predictors and important single predictors from a large suite of remote sensing, climatic, topographic, and biometeorological covariates; and 3) predict monthly CH₄ emissions from freshwater wetlands at regional to global scales.

Globally, the model performed well (R² = 0.6, MAE = 26.2 nmol m^-2 s^-1, Bias = 1.6 nmol m^-2 s^-1). Normalized errors were largest at swamps (n = 6), four of which are distinctive tropical sites, and smallest at marshes (n = 10). Mean seasonal cycles were reproduced well (R² > 0.7) at two-thirds of sites, although interannual anomalies were not accurately reproduced. Gridded climatology and tower-measured biometeorology were the most useful covariate classes for predicting CH₄ fluxes, land cover (including inundation and vegetation cover) was intermediate; and soil, relief, and vegetation greenness were least useful. The most important individual predictors included nighttime land surface temperature, potential radiation, enhanced vegetation index, air temperature, and latent heat flux. Several static predictors were also useful, including percent agricultural land use and slope. Preliminary estimates of average (2001-2012) annual CH₄ emissions scaled by wetland area were 151 Tg CH₄ globally, in close agreement with recent bottom-up model estimates (Saunois et al. 2020), with 96 Tg (~64%) from the tropics, and 31 Tg (~21%) from >45°N, in agreement with a recent northern upscaling effort (31-38 Tg; Peltola et al. 2019). We acknowledge the FLUXNET-CH₄ contributors for the data provided in these analyses.