Methane (CH₄) is a strong greenhouse gas that accelerates climate change and offsets mitigation efforts. Although wetlands are considered to be the largest natural CH₄ source, estimates of global wetland CH₄ emissions vary widely among approaches taken by bottom-up process-based biogeochemical models (102–182 TgCH₄ yr⁻¹) and top-down atmospheric-inversion methods (159–200 TgCH₄ yr⁻¹). Importantly, the discrepancy between bottom-up and top-down wetland CH₄ emission estimates has remained wide, ~17 TgCH₄ yr⁻¹ in 2016 and ~30 TgCH₄ yr⁻¹ in 2020, despite recent advances in CH₄ observations and simulations. Here, we use ecosystem-scale measurements, multi-model ensemble from bottom-up and top-down approaches, and machine-learning upscaling estimates with ILAMB (the International Land Model Benchmarking System) to constrain uncertainty in modeled global wetland CH₄ emissions. We find the discrepancy between bottom-up and top-down global wetland CH₄ emission estimates increased by 49%, 2%, and decreased by 65% with constraints posed by ecosystem-scale measurements, multi-model ensemble, and machine-learning upscaling estimates, respectively. Our analyses highlight the importance of improving wetland heterogeneity representation in current observations and reveal the large sensitivity of the best-available multi-model ensemble to underlying approaches. Our results also demonstrate that traditional model benchmarking is strongly limited by data representativeness, and motivate the use of machine-learning methods to reduce uncertainty in sparse observations to facilitate model development. We acknowledge the FLUXNET-CH₄ contributors and Global Carbon Project CH₄ modeling groups for the data provided in these analyses.