WetCH4: Machine learning-based Upscaling of Wetland Methane Fluxes at High Latitudes

Wetlands are the largest natural source of methane (CH₄) emissions globally. Northern wetlands CH₄ emissions may accelerate under intensified warming in the Arctic and subsequent permafrost thaw. Although top-down atmospheric inversions of wetland CH₄ emissions differ from bottom-up estimates of process models, the two estimates are not completely independent as atmospheric inversions would take process-based estimation as priors in the model. Data-driven upscaling instead provides independent estimates of large-scale methane fluxes. To upscale wetland CH₄ fluxes at high latitudes (>45° N), we developed a machine learning framework (WetCH₄) based on eddy covariance observations, reanalysis, and remote sensing constraints from Soil Moisture Active Passive (SMAP) soil moisture and MODIS reflectance. Our results modeled from 155 site-years across 26 sites of wetland EC data with accuracies of a mean R² of 0.46 and 0.62 and the mean absolute error (MAE) of 19.4 nmol m^-2 s^-1 and 21 nmol m^-2 s^-1, for daily and monthly fluxes, respectively. With a dynamic wetland inundation dataset, WetCH₄ estimated average annual CH₄ emissions of 22.3 ± 2.5 Tg CH₄ yr^-1 from Arctic and boreal wetlands for 2016-2022, with 86% emitted during the May-October period. The estimated magnitude and seasonal cycle of CH₄ emissions aligned with estimates of process-based models. Furthermore, a comparison of our estimated fluxes with inversions of aircraft-observed concentration data in the North Slope area underscores the significance of accurately delineating wet-emitting surfaces in this region to reduce estimation uncertainty. In summary, WetCH₄ provides daily wetland CH₄ fluxes from 2016 to 2022, facilitating investigations into the global CH₄ budget and wetland responses to climate change through a robust, data-driven perspective.