Google’s Jet Propulsion Laboratory and NASA introduced MAPL EMIT, an advanced deep learning model, on 9 September 2026. It detects, quantifies and locates global methane emissions from NASA EMIT satellite instruments. Google’s announcement. Google says the model detects 50 percent more methane than experts, and can identify more than 23,000 more plumes worldwide, including those at 24 out of 25 landfills that emit the most.
In a document called a study in the Proceedings of the National Academy of Sciences The PNAS article was written by Google Research researchers and JPL scientists, as well as appearing in an article published at the Journal of Physics A. Google Research technical post Vishal Batchu, a research scientist and Vishal batchu, a research engineer, drafted a letter dated 1 September 2026. Batchu was also the author of Google’s 9th September announcement.
EMIT Methane Monitoring Instrument
Google describes methane in its announcement as a powerful greenhouse gas with a warming potential 30 times greater than carbon dioxide over 100 years. In the study it is stated that methane’s global warming capacity over 20 years will be 81 to 86% greater than CO2, and that the gas has been responsible for about 25% of warming caused by humans since the beginning of the Industrial Revolution. The Global Methane Pledge has committed more than 125 nations to reduce methane emission by 30% by 2030. According to the study, human activity is responsible for 60% of the global methane production, which includes waste (19%) agriculture (44%) and energy (37%).
EMIT is the Earth Surface Mineral Dust Source Investigation. It operates on the International Space Station. Originally, it was designed to map the mineral composition of arid areas. Google Research explained that this instrument recorded 285 spectral band spanning from 381- 2,493nanometers with 60-meter spatial resoluton and an area of 80 kilometers. By recording hundreds distinct bands for each pixel, it reveals chemical fingerprints on an invisible gas. It contrasts the configuration of this study with that of global mappers, such as TROPOMI. TROPOMI uses a 2,600-kilometer swath at a resolution 5.5 x 3.5 km to track background concentrations.
The Training of a Vision Transformer on Simulated Plumes
It is a vision converter that works from end-to-end. MAPL EMIT processes all the pixels within a scene to determine methane enhancements. Model combines Swin-v2-S Transformer encoders of approximately 30,000,000 parameters with convolutional decoders in a UNet-like architectural design. The model performs 3 tasks at once: Quantifying the methane enhance in every pixel; Delineating and mapping each plume shape and boundary and Localizing each emission. This includes the separation of overlapping plumes and neighboring facilities. The existing methods, according to the study, do not adequately capture this scenario.
The team used a line-byline radiative transfer method with HITRAN data to inject 3.6 million artificial plumes into EMIT scenes. They divided the 235,000 tiles in EMIT training sets into validation and testing and then trained their model over 32 Google TPUs for approximately 96-hours. Google Research says the synthetic data exposes the model to a variety of emission rates and terrains as well as atmospheric conditions. It improves its generalization from real observations.
Validation of Synthetic Benchmarks and Real World Benchmarks
According to real-world comparisons, MAPL-EMIT detected 84% of NASA EMIT L2B complexes that were manually annotated across 1,084 EMIT Granules. It also identified roughly 1.5x as many plausible plumes compared with human analysts. The 3,672 initial detectors for those granules were reduced to 2,209 when the filter was applied to plumes on land detected in at least 13 out of 16 overlapped strided conclusions.
In order to reduce the number of false positives, the researchers ran their model on 20,000 grains where they expected minimal plumes. After filtering the study reported 1,513 detections. Of these, 355 were deemed likely to be genuine by spectral fitting evaluation. The remaining 1,158 detections that are not confirmed, occur mostly in dense forest or hilly terrain, and amount to an estimate of 0.06 false plumes for each granule.
This model detected two of three resolvable plumes with no false positives, compared to coincident AVIRIS-3 airborne observations. The study also notes that previous work showed the EMIT L2B products detected potential plumes in 17 of these 25 sites. MAPL-EMIT, compared to coincident AVIRIS-3 observations, detected two out of three resolvable smog plumes. There were no false positives, and the signal-to noise ratio was 8.60, compared with 1.21 when using the resampled AVIRIS-3 matched filters and 0.40 when using the standard L2B matched filters. Stanford conducted controlled-release tests from August 1, 2020, through December 31, 2025 near an Arizona field site. The model accurately detected 5 of 7 releases, and there were no false positives.
The study reports that the model reliably captures plumes corresponding to leak rates of 250 to 500 kg/h at a nominal wind speed of 2.5 m/s. This is an approximate 2-to-4x improvement over existing models. The study reported that the model captures reliably plumes equivalent to leak rates ranging from 250 to 500 kilograms per hour at a nominal speed of 2.5 meters/s. That’s an approximately 2 to 4x improvement over current models.
Restrictions, Publications and Next Steps
MAPL shows significantly more false positives compared to expert-vetted catalogues like the EMIT L2B plumb complexes. The authors state that false positives are still a challenge, especially in complex terrain, and MAPL-EMIT shows substantially more false positives than expert-vetted catalogs such as the EMIT L2B plumplexes.
Google also released the synthetic plume dataset and trained model on Kaggle along with the paper. An inference library was published on GitHub. The study appeared in PNAS, Volume 123, Issue 36, dated 9/8/2026.
In the future, researchers will work to extend this pipeline into multigas and multiple satellite retrievals and towards end-toend estimation of emission rates. The result would be direct leak-rate assessments without post-hoc wind integration. Google Research reports that NASA plans to boost coverage of its next generation imaging spectrometers from 30 to 50%.

