Using machine learning to enable national methane emissions inference from large satellite datasets

The growing volume of methane measurements from space provides new opportunities for evaluating and improving countries' self-reported emissions. Surface emissions can be estimated from atmospheric observations using inverse modelling systems, which often rely on Lagrangian Particle Dispersion Models (LPDMs) to simulate how methane is transported through the atmosphere. Ensembles of particles are transported backwards in time from the measurement point, to define source-receptor relationships (“footprints”), which reflect the sensitivity of a measurement to all potential upwind sources within the domain. However, LPDM-based techniques are computationally costly, struggling to scale to the size of modern satellite datasets and limiting the amount of data that can be used for emissions inference. Previously, we presented a machine learning-driven LPDM emulator that can approximate satellite footprints using only meteorology and topography, and demonstrated its use over the South American continent, achieving speed-ups of over three orders of magnitude compared to the LPDM. We integrated the emulator into an emissions inference pipeline to estimate Brazil’s methane emissions from GOSAT observations in 2016 and 2018, and found that the emulator-based estimates were consistent with those obtained using the more expensive physics-based LPDM. Here, we show preliminary results of applying the emulator to other regions with high natural methane emissions, like North Africa and India. We compare the emulator’s performance across the selected time periods and geographical domains as well as the estimated emissions. Furthermore, we discuss solutions to improve performance and reduce the training data needed, like transfer learning across domains.