Finding a gas that ordinary photographs cannot show
Google and NASA's Jet Propulsion Laboratory are using AI to search space-based observations for methane plumes. Google's September 9 announcement highlights MAPL-EMIT, a research framework whose detailed explanation appeared on September 1. The company has made a plume database and viewing app available alongside model and code resources.
The attraction is practical. Before an operator can investigate an unexpected emission, someone has to identify where to look. Our assessment is that an automated shortlist could make that search more manageable. But a map marker is the beginning of an investigation, not its conclusion, and the underlying documentation gives important reasons for caution.
Sources: Google: MAPL-EMIT announcement, September 9; Google Research: Model, training and evaluation
This begins with a physical sensor, not a chatbot
EMIT is an imaging spectrometer installed on the International Space Station. NASA originally designed it to study the mineral composition of arid regions that produce airborne dust. It separates incoming light into many wavelengths, allowing scientists to study spectral signatures rather than only the red, green and blue information familiar from a photograph.
Google's research explanation describes using these measurements to distinguish methane's signal from a complicated background. Surface materials, shadows and overlapping sources make the problem harder than drawing a circle around a visible cloud. The model examines both the spectrum and the surrounding spatial pattern, including how a plume spreads across neighboring pixels.
Sources: NASA JPL: EMIT mission and instrument; Google Research: Model, training and evaluation
The training plumes and the observed plumes are different
The researchers describe training with 3.6 million physics-simulated methane plumes inserted into real EMIT scenes. Simulations supply examples that would be difficult to label at that scale using only observed emissions. That training total must not be reported as millions of newly discovered leaks.
Google reports 84% recall on expert-annotated plumes. Recall asks how many known examples the system recovered: it is not the probability that any new detection is correct. For an illustrative set of 100 known plumes with the same measured recall, about 84 would be recovered and 16 missed. This arithmetic explains the metric; it is not a new evaluation by Lumacta.
Two labels with very different error profiles
The Earth Engine catalog calls detections seen in at least three different observations high-confidence. It reports roughly 3–5% false positives in a human-reviewed random subset of that group. For medium-confidence detections, which lack a match across observations, the reported false-positive rate is about 50–55%.
Medium-confidence candidates need further checks against infrastructure, spectra and wind. A marker alone is insufficient evidence for a public accusation.
Sources: Earth Engine: Confidence levels and dataset limitations
A plume map is not a continuous emissions meter
The same catalog distinguishes methane enhancement from an emission rate. Its pixel values describe an increase in atmospheric methane along the viewing column, not kilograms released per hour. Estimating a flow rate requires additional methods and wind information. Coverage is also limited to where and when EMIT collected usable observations; clouds, dark surfaces and shadows can degrade results.
Our practical takeaway: check both coverage and observation date before acting on a mapped candidate.
Sources: Earth Engine: Confidence levels and dataset limitations
The useful outcome happens back on the ground
A hypothetical response chain would be: review the candidate, check the site and timing, arrange appropriate follow-up, address a confirmed problem, then measure again. That chain separates detection from repair. More dots on a map can improve visibility without, by themselves, demonstrating an environmental benefit.
The research repository provides tools for processing scenes and exporting plume information, but states that it is not an officially supported Google product. Our conclusion: this is a valuable open research resource when used with its confidence levels and limitations intact. The outcome to watch is independently confirmed emissions reduced, not simply a larger count of AI detections.
Sources & Methods
Checked September 11, 2026. The September 9 announcement is separated from the September 1 technical post. Performance and confidence figures are the producers' reported results, not our measurements. The 100-plume example is illustrative arithmetic. We did not independently inspect a facility, validate a plume or measure emissions reductions.
- Google: MAPL-EMIT announcement, September 9 — Primary source
- NASA JPL: EMIT mission and instrument — Instrument background
- Google Research: Model, training and evaluation — Research explanation
- Earth Engine: Confidence levels and dataset limitations — Dataset documentation
- Google Research: MAPL inference repository — Research code