How AI finds methane plumes from space, and where it can be wrong

The satellite measures light. The model looks for a pattern.

Methane does not have to appear as a visible cloud for a satellite to detect it. It absorbs particular wavelengths of light. NASA's EMIT instrument, aboard the International Space Station, records a detailed spectrum for each patch of ground.

MAPL-EMIT uses those measurements to look for methane above the normal background, trace the shape of a plume and estimate where it starts. The work comes from researchers at Google, NASA's Jet Propulsion Laboratory and the Carnegie Institution for Science.

Google described the system this month. The useful question is not whether AI can make a striking emissions map. It is how to tell a credible detection from a convincing-looking mistake.

Why neighbouring pixels help

A patch of terrain can produce a signal that resembles methane. Reading each pixel alone makes that ambiguity harder to resolve. MAPL-EMIT also considers the surrounding scene: does the pattern look like gas spreading downwind, or like a feature on the ground?

The authors trained it using 3.6 million simulated plumes added to real satellite observations. Simulation supplies examples with known locations and shapes, including overlapping plumes that would be difficult to label at that scale by hand.

Their July manuscript reports that the model recovered 84% of NASA's expert-annotated plume complexes across 1,084 satellite data granules, or observation blocks. It also identified additional plausible plumes. The authors say they did not investigate every new detection.

They tested other evidence too, including airborne observations and seven controlled releases with usable coincident EMIT observations. The model detected five of those seven releases. These are the researchers' results, not a universal detection rate.

The confidence label changes the meaning

The public Earth Engine catalogue defines high-confidence plumes as detections recurring in at least three separate observations. It reports roughly 3–5% false positives in a human-reviewed random subset of this group.

For the medium-confidence group, without that match across observations, the documented false-positive estimate is about 50–55%. Those entries are candidates for further filtering, not equally secure evidence. These estimates apply to the documented categories and review, not every place or future observation.

A bright mark on a map therefore needs its context. Repetition, the spectral signal, local wind and the presence of plausible infrastructure can help assess it. For someone investigating a particular facility, the date and confidence category belong beside the image, not in a footnote.

A detection is not an emissions rate

The catalogue's pixel values describe extra methane in the atmospheric column. They do not directly say how many kilograms a source emits per hour. Estimating that flow needs additional methods and wind information, adding uncertainty.

Coverage is limited to where and when EMIT collected data. Cloud, shadow and dark surfaces can degrade the signal. A missing marker is not a clean bill of health for a site.

The authors also warn that performance on future observations depends on sensor calibration and changes over time. A model that works on an archive still needs checks as the instrument and conditions change.

What this makes possible, and what still has to happen

The released code supports running the model, removing duplicate detections and checking candidate plumes. The repository explicitly says it is not an officially supported Google product. Publishing these tools gives other researchers something concrete to examine; it does not mean Model Current has validated the software in operation.

Our assessment is that the value lies in narrowing a large search. A candidate can direct attention to a place worth investigating. Confirmation, a response from the operator and repeat measurement would still be needed to establish that emissions were reduced.

The model, manuscript and data documentation are public. The performance figures remain author-reported, and the catalogue is neither complete nor error-free. How often its detections lead to verified repairs and lasting reductions is a separate question that the release does not answer.