Wetlands are the largest natural methane (CH₄) source to the atmosphere, contributing 25%-40% to global CH₄ emissions. With global warming, wetland CH₄ emissions are expected to increase in the future, which will in return further amplify the warming through increased radiative effect. Wetland CH₄ emissions are also the most uncertain component of the global CH₄ budget, with considerable discrepancies among bottom-up biogeochemistry models, top-down atmospheric inversion models, and data-driven machine learning models. The complex nature of wetland CH₄ processes presents a challenge for modeling at in situ scales as well as upscaling at large scales.

In this study, we combined causal inference, machine learning, and FLUXNET-CH4 measurements to model the temporal dynamics of CH₄ flux intensity at 31 sites covering four major wetland ecosystems, including bogs, fens, marshes, and wet tundra. We focused on model performance, model structure robustness, and time-lagged responses of CH₄ emission to environmental and biological factors. This work provided causality constrained machine learning models that could serve as surrogate modules with efficient parameterization and high accuracy in methane biogeochemistry models. Further, we employed the in situ CH₄ emission model and upscaled fluxes to the regional scale (above 45 degree, Northern Hemisphere). We found that large uncertainties were primarily associated with wetland inundation maps (from static to dynamic wetland maps) and secondarily depended on regional gridded products of environmental forcings. Finally, we benchmarked the upscaled products with existing estimates from bottom-up biogeochemistry models and data-driven machine learning models.