EGUsphere EGUsphere EGUsphere EGUsphere Copernicus Publications Göttingen, Germany 10.5194/egusphere-2026-1871 Machine learning-based emission rate estimates of global methane super-emissions Roberts Clayton https://orcid.org/0000-0002-5184-7485 1 Maasakkers Joannes D. https://orcid.org/0000-0001-8118-0311 1 de Jong Tobias A. https://orcid.org/0000-0002-5211-8081 1 Schuit Berend J. https://orcid.org/0000-0003-1768-3592 1 2 Sharma Shubham 1 Huegens Theo 1 van den Berg Anne-Wil 3 Houweling Sander https://orcid.org/0000-0002-6189-1009 1 4 Aben Ilse 1 4 SRON Space Research Organisation Netherlands, Leiden, The Netherlands GHGSat Inc., Montreal, Canada Meteorology and Air Quality group, Wageningen University & Research, Wageningen, The Netherlands Department of Earth Sciences, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands 19 05 2026 2026 1 43 Copyright: © 2026 Clayton Roberts et al. 2026 This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/ This article is available from https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1871/ The full text article is available as a PDF file from https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1871/egusphere-2026-1871.pdf

Methane, the second most important greenhouse gas, has a global warming potential more than 80 times that of carbon dioxide over a 20-year period. Given its decadal atmospheric lifetime, reducing anthropogenic methane emissions is critical for limiting near-term warming. The TROPOspheric Monitoring Instrument (TROPOMI) provides daily global methane satellite observations, enabling rapid detection of super-emitters. Here, we develop ML-SPERE, a machine-learning framework based on a convolutional neural network trained on simulated TROPOMI methane observations and meteorological data to estimate emission rates for super-emitters. ML-SPERE outperforms the Integrated Mass Enhancement (IME) method on simulated plumes that incorporate real TROPOMI backgrounds and missing spatial data, reducing the median absolute percentage error from 42.4% to 24.3% for well-observed methane plumes. ML-SPERE estimates also do not exhibit the low wind-speed dependent biases present in IME estimates. Applied to TROPOMI observations of a 200-day well blowout in Kazakhstan, ML-SPERE shows better agreement with inverse modeling results and estimates from high-resolution point-source imagers than TROPOMI IME estimates do. Global spatial patterns of methane emissions inferred from ML-SPERE and the IME method for all super-emitters found by TROPOMI in 2021 are broadly consistent, with notable regional differences in northern Russia (where transient pipeline may not be well characterized by either method), the Congo Basin (where IME estimates are potentially inflated due to the large spatial extent of plumes), and southeastern Australia (where IME estimates are potentially negatively biased owing to predominantly low wind speeds). Mean estimated emission rates for this dataset aggregated by estimated source sector remain similar between both methods. Overall, improved performance on simulated plumes and consistency with independent estimates for real-world observations demonstrate the utility of ML-SPERE for quantifying TROPOMI methane super-emitters.