Google and NASA AI finds more methane plumes in satellite data

An AI model developed by Google and NASA researchers identified about 50% more plausible methane plumes than human analysts across 1,084 scenes from NASA’s EMIT instrument. The results could help investigators find more potential emissions sources, though a detection is not the same as a confirmed leak.

Google DeepMind Watch analysis

What happened

The model, called MAPL-EMIT, uses methane’s infrared signature alongside the shape of a plume to distinguish likely emissions from misleading features on the ground. The study was published in the Proceedings of the National Academy of Sciences on 1 September; R&D World reported the findings on 26 October.

The researchers also released a plume catalogue through Google Earth Engine, along with inference code, model weights and synthetic training data. The catalogue separates high-confidence detections, observed at least three times, from medium-confidence candidates without a repeat observation.

Why it matters

Satellite instruments can reveal methane that ordinary visible-light images cannot, but finding it reliably is tricky. MAPL-EMIT’s use of both spectral patterns and plume shape is designed to address a weakness in older methods, which can mistake bright roofs or other ground features for methane.

The confidence labels matter. Human review found false-positive rates of about 3% to 5% for the high-confidence group, but 50% to 55% for medium-confidence candidates. Those lower-confidence results need further checks, not a celebratory pin on a map. The output also estimates methane enhancement in the atmosphere, not an emissions rate; calculating that requires wind data and further work.

Our read

This is a useful example of AI widening the search rather than replacing the follow-up. Finding more plausible plumes gives researchers and operators more leads to investigate, while the catalogue’s confidence levels help show where the evidence is stronger and where it still needs checking.

The practical next step is to compare detections with infrastructure, wind conditions and field measurements. A bigger list of possible leaks is valuable only if people can work out which ones are real.

What to watch

• Whether follow-up measurements confirm more of the newly identified plumes.
• How often medium-confidence detections survive further screening.
• Whether researchers use the released tools and data to extend monitoring beyond the tested scenes.

Discussion spark: Should satellite AI systems prioritise finding every plausible methane plume, even if that creates many false alarms, or return fewer detections with greater confidence?

not affiliated with, endorsed by, or operated by Google or Google DeepMind