Oceanic emissions represent a highly uncertain term in the natural atmospheric methane (CH4) budget, due to the sparse sampling of dissolved CH4 in the marine environment. Here we overcome this limitation by training machine-learning models to map the surface distribution of methane disequilibrium (∆CH4). Our approach yields a global diffusive CH4 flux of 2-6TgCH4yr-1 from the ocean to the atmosphere, after propagating uncertainties in ∆CH4 and gas transfer velocity. Combined with constraints on bubble-driven ebullitive fluxes, we place total oceanic CH4 emissions between 6-12TgCH4yr-1, narrowing the range adopted by recent atmospheric budgets (5-25TgCH4yr-1) by a factor of three. The global flux is dominated by shallow near-shore environments, where CH4 released from the seafloor can escape to the atmosphere before oxidation. In the open ocean, our models reveal a significant relationship between ∆CH4 and primary production that is consistent with hypothesized pathways of in situ methane production during organic matter cycling.