Google Research and NASA’s Jet Propulsion Laboratory released MAPL-EMIT on September 9, 2026, a deep-learning model mapping global methane plumes at a 60-meter resolution. Trained on 3.6 million physics-simulated plumes, the vision transformer detects 50% more plumes than human analysts, identifying 23,000 additional plumes globally, including at 24 of the 25 largest-emitting landfills.
Methane is a potent greenhouse gas. According to the PNAS study published by researchers at Google Research and JPL, methane’s 20-year global warming potential is 81 to 86 times that of carbon dioxide, and its 100-year potential sits at 30 times. The gas accounts for roughly 25% of human-induced warming since the start of the industrial era, boasting an atmospheric lifetime of about nine years.
Traditional monitoring tools often face a severe spatial resolution constraint. To monitor background methane levels, European instruments like TROPOMI combine a spatial resolution of 5.5 by 3.5 kilometers with a swath width of approximately 2,600 kilometers. That scale is simply too coarse to isolate individual facilities or pinpoint leaks. Conversely, NASA’s Earth Surface Mineral Dust Source Investigation (EMIT) instrument, operating aboard the International Space Station, records 285 spectral bands spanning 381 to 2,493 nanometers at a 60-meter spatial resolution with an 80-kilometer swath. Recording hundreds of distinct spectral bands per pixel exposes the chemical fingerprints of an otherwise invisible gas.
The Swin-v2-S Architecture and Training Pipeline
Building an effective computer vision model for orbital hyperspectral imaging requires overcoming a distinct data scarcity problem. Because no massive, labeled dataset of real-world methane plumes existed at scale, the Google and JPL team built a physics-based simulation pipeline. Utilizing Lagrangian puff models—which replicate how airborne particles disperse—they generated 3.6 million synthetic plumes and embedded them into actual EMIT scenes via line-by-line radiative transfer calculations using HITRAN spectral data.
The resulting architecture, formally titled Methane Analysis and Plume Localization with EMIT, is an end-to-end vision transformer. Vishal Batchu and Michelangelo Conserva detailed the technical mechanics in a Google Research technical post. The system integrates a U-Net-inspired architecture that pairs a convolutional decoder with a Swin-v2-S transformer encoder containing roughly 30 million parameters.
Instead of processing data sequentially or analyzing isolated tiles, the network processes the complete EMIT radiance spectrum to retrieve methane enhancements across all pixels in a scene jointly. It executes three tasks simultaneously:
- Quantifying the methane enhancement value in every individual pixel.
- Delineating each distinct plume’s shape and geographical boundaries.
- Localizing each emission source, which includes separating overlapping plumes from neighboring industrial facilities.
The researchers partitioned 235,000 EMIT tiles into training, validation, and test sets. They trained the model on 32 Google TPU chips for approximately 96 hours. According to the study, this synthetic pre-training exposed the model to widely varied emission rates, difficult surface terrains, and shifting atmospheric conditions, sharply improving its generalization capabilities when facing messy, real-world observations.
Real-World Benchmarks Versus Human Analysts
When evaluated against established real-world benchmarks, the system demonstrated striking efficacy. Across an evaluation dataset comprising 1,084 EMIT granules, MAPL-EMIT successfully detected 84% of the plume complexes that had been hand-annotated by NASA EMIT L2B experts, as detailed in the study.
The model mapped 23,000 additional plumes globally, finding emission events at 24 of the 25 largest-emitting landfills tracked by researchers. Human analysts poring over vast quantities of satellite granules routinely miss scattered or overlapping plumes masked by ambient atmospheric noise or complex topography. By automating the extraction of these spectral features via deep learning, the pipeline turns raw hyperspectral cubes into actionable intelligence almost instantly.
Europe’s Methane Regulation ordered a satellite monitoring tool and a super-emitter alert system, but its scope previously covered only limited baseline structures.
What This Means for Enterprise Infrastructure
Energy, waste, and agricultural sectors—which split the 60% human-driven share of global output into energy (37%), agriculture (44%), and waste (19%)—can no longer rely on self-reported estimates or delayed audits.
