MAPL-EMIT finds about 50% more methane plumes in satellite scenes

By Sophie Lindqvist

In tests covering 1,084 observation scenes from NASA’s EMIT imaging spectrometer aboard the International Space Station, Google and NASA researchers’ Methane Analysis and Plume Localization with EMIT model, or MAPL-EMIT, identified about 50% more plausible methane plumes than human analysts, according to a Sept. 1 study in the Proceedings of the National Academy of Sciences.

The comparison expanded the pool of potential emission sources for follow-up, but it did not make every new detection a verified leak. The Earth Engine catalog separates candidates by confidence, a distinction that directly affects how teams allocate field-verification resources.

The study targets methane released from oil and gas infrastructure, coal operations and landfills. Methane traps more than 80 times as much heat as carbon dioxide over 20 years, which the authors frame as a reason to prioritize large leaks for near-term emissions cuts.

How does EMIT reveal methane?

NASA built the Earth Surface Mineral Dust Source Investigation instrument, or EMIT, to study mineral dust. Mounted on the International Space Station, it scans an 80-kilometer-wide swath at 60-meter spatial resolution and records 285 narrow spectral bands.

EMIT also captures wavelengths at which methane absorbs sunlight, making the gas visible to the instrument even though ordinary visible-light images cannot show it. MAPL-EMIT combines methane’s infrared absorption pattern with plume shape, helping separate emissions from misleading surface features.

Why do conventional filters miss some plumes?

Conventional matched filters examine the absorption signature one pixel at a time and estimate background from many pixels. Bright roofs and other surfaces can resemble methane’s spectral pattern, introducing errors, while pixel-by-pixel processing lacks the spatial context needed to trace a plume downwind from its source.

Deep learning can combine spectral information and shape, but the available labels limited that approach. An unnamed narrator in a Google Research video on the study said: “Reliably detecting both the spectral signature of methane and the plume shape requires deep learning and lots of labeled plumes.”

Only a few thousand labeled plumes existed, so the researchers used physics-based atmospheric simulation to expand the training set.

How did simulation create training data?

The team built a Lagrangian puff model that tracks how gas puffs drift and spread in turbulent wind. The simulation generated 3.6 million plumes, including overlapping plumes from neighboring facilities.

Researchers used methane absorption data from the HITRAN spectroscopic database to insert synthetic signals into about 235,000 real EMIT image tiles, each 256 by 256 pixels. This gave MAPL-EMIT known methane distributions over varied real-world backgrounds.

What does MAPL-EMIT estimate?

The system uses a Swin-v2 vision transformer to analyze spectral information and neighboring pixels together. It estimates methane enhancement above background, outlines individual plumes, predicts their origins and checks candidate detections against the expected methane spectrum.

On simulated test data, MAPL-EMIT placed estimated sources an average of about 100 meters from their true origins, which the researchers said was close enough to indicate an individual facility. The reported localization result therefore comes from simulation, not a reported field trial.

Separating plumes that merge downwind can distinguish neighboring facilities that otherwise appear to share one source. For R&D teams, that capability can narrow the set of sites selected for ground measurement.

How reliable are the Earth Engine detections?

The Earth Engine plume catalog applies two confidence tiers:

• High confidence: plumes observed at least three times.
• Medium confidence: candidates without a temporal match.

Human review of random samples found false-positive rates of about 3% to 5% for high-confidence plumes and 50% to 55% for medium-confidence candidates. Medium-confidence detections require additional screening against infrastructure, spectral evidence and local wind before investigators treat them as source leads.

The source does not state the number of random samples reviewed or provide confidence intervals for those false-positive rates.

Can the catalog establish an emission rate?

No. The output measures methane enhancement in the atmospheric column, not an emission rate in kilograms per hour. Calculating a rate requires wind data and further calculations, so detections can direct investigation and support comparisons over time without, by themselves, providing a mass-flow estimate.

Researchers and operators can use the catalog to select sources for follow-up, inspect repeated observations and compare model predictions with field measurements. The source material provides no field-comparison result, leaving that verification step for users.

What can research teams reuse?

The release separates operational access from model development materials:

• Earth Engine: the plume catalog for screening and repeated-observation review.
• GitHub: inference code.
• Kaggle: model weights and synthetic training data.

Neither the study summary nor the Google Research video identifies the research’s funding organization, leaving grant and portfolio reviewers without a named sponsor. Researchers can now use the released materials to test whether repeated satellite detections, infrastructure checks, spectral evidence and local wind can convert medium-confidence candidates into traceable source investigations.