Google, NASA JPL Launch Methane-Plume AI With 84% Recall and 50% More Plausible Detections
Updated
Updated · Google Research · Sep 1
Google, NASA JPL Launch Methane-Plume AI With 84% Recall and 50% More Plausible Detections
2 articles · Updated · Google Research · Sep 1
Summary
MAPL-EMIT automates global methane-plume detection, enhancement estimates and source localization from NASA EMIT satellite data, with Google and NASA JPL releasing the model, plume database and tools publicly.
84% recall against expert-annotated plumes and about 50% more plausible plume detections across roughly 1,100 EMIT granules suggest the system can pick up weaker emissions than existing matched-filter methods.
3.6 million synthetic plumes were injected into real EMIT scenes to train the vision-transformer model, helping it separate true wind-blown methane plumes from lookalike surface signals and untangle overlapping emissions in dense industrial areas.
24 of the world's 25 top-emitting landfills were successfully mapped, though false positives remain a challenge; outputs are therefore tagged with confidence scores so users can balance sensitivity against error risk.
Methane has about 30 times the warming potential of CO2 over 100 years and has driven roughly 25% of human-caused warming, making faster facility-scale detection increasingly important as future NASA spectrometers expand coverage by 30 to 50 times.