Google Research has released MAPL-EMIT, a deep learning framework that automates the detection and tracking of methane emissions from space. The system leverages NASA's EMIT instrument aboard the International Space Station, which uses advanced hyperspectral imaging to capture the unique chemical signatures of methane gas invisible to standard satellite cameras. By analyzing how methane disperses across landscapes using vision transformer technology, MAPL-EMIT achieves 84% recall on expert-annotated plumes while maintaining high signal-to-noise ratios compared to existing detection methods.
The technology addresses a critical climate imperative: methane is 30 times more potent than carbon dioxide over a century, and reducing emissions offers a fast-action pathway to limit global warming. The model enables stakeholders to identify and track point-source emissions from oil and gas infrastructure, agricultural facilities, and landfills—the most cost-effective targets for emissions reduction. Google's release supports the Global Methane Pledge, which over 125 countries have committed to, targeting a 30% reduction in emissions by 2030.
To accelerate broader climate action, Google is releasing the trained MAPL-EMIT model, a global plume database, synthetic training data, and an open-source inference library. The initiative aligns with Google Earth AI, the company's broader geospatial AI program designed to transform planetary data into actionable environmental intelligence. The research was published in the Proceedings of the National Academy of Sciences (PNAS).
Key Points
MAPL-EMIT deep learning model automates methane detection and source estimation from satellite hyperspectral data with 84% recall
Vision transformer architecture analyzes full spectra and spatial context to distinguish true methane plumes from background landscape features
Model enables cost-effective tracking of facility-scale emissions globally to support Global Methane Pledge climate targets
Google releasing trained model, global plume database, synthetic training dataset, and open-source inference library to scientific community