Mapping global methane emissions from space with deep learning
Google Research
Google says its new model can spot methane plumes from EMIT satellite data. It found 84% of expert-labeled plumes and could speed up fixes for a fast-warming gas.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Methane gets a lot less attention than carbon dioxide, but it punches hard. Over 100 years, it traps about 30 times more heat than CO2, and Google says it has driven roughly 25% of human-caused warming since the industrial era began. That makes the first cuts especially valuable: methane leaves the atmosphere faster, so trimming leaks now can cool things sooner.
That urgency is why space agencies and researchers are leaning on satellites. Google’s new system, MAPL-EMIT, works with data from NASA’s EMIT instrument on the International Space Station, which was built to study minerals but also picks up the chemical fingerprint of methane. The goal here is blunt and practical: find point sources at the facility scale, especially in oil and gas, agriculture and landfills.
The model itself is built as a Swin-S vision transformer, which means it does not stare at pixels one by one. It looks at the full spectrum and the scene around it, helping it tell a real plume from a patch of ground that merely looks suspicious in the spectrum. That matters because methane scenes can get messy fast, with overlapping clouds of gas from nearby sites and terrain that produces false alarms.
Training data was the next problem. There is no giant labeled archive of real methane plumes, so Google created 3.6 million synthetic ones and dropped them into real EMIT scenes. The simulations used Lagrangian puff models to mimic how gas moves and spreads, giving the system examples of different emission rates, different terrains and overlapping plumes. In the paper published in PNAS, the team says MAPL-EMIT reached 84% recall on expert-annotated plumes and produced a higher signal-to-noise ratio than existing matched-filter enhancement methods.
On real data, the system found about 50% more plausible plumes across roughly 1,100 EMIT granules, and it mapped plumes at 24 of the world’s 25 top-emitting landfills. Google is also releasing a global plume database on Earth Engine, plus the trained model and synthetic plumes on Kaggle and an inference library on GitHub. With NASA planning the next wave of imaging spectrometers, the bet is clear: the more space sees, the more automation matters.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI that earns its keep: narrow, technical, and aimed at a problem with an actual bill attached. The industry loves grand talk about planetary intelligence; methane mapping is better because it is boring, specific, and immediately useful. The real test now is whether operators fix the leaks instead of just admiring the heat map.
Read more about this at: Google Research