Google AI Model MAPL-EMIT Tracks Methane From Space

Google Research, in collaboration with NASA's Jet Propulsion Laboratory, has developed MAPL-EMIT, a deep-learning system designed to automate the detection, quantification, and source estimation of methane plumes. Built on a Swin-S vision transformer architecture, the model processes hyperspectral data from NASA's Earth Surface Mineral Dust Source Investigation (EMIT) instrument on the International Space Station. Unlike traditional pixel-by-pixel analysis, this framework evaluates the full spectrum of light alongside surrounding spatial context to distinguish actual wind-blown gas plumes from surface materials with similar spectral signatures.

To train the transformer without a massive real-world dataset, researchers created 3.6 million synthetic methane plumes using physics-based Lagrangian puff models and injected them into real EMIT scenes. When tested on real-world satellite data, MAPL-EMIT achieved an 84% recall rate on expert-annotated plumes. It identified approximately 50% more plausible plumes across roughly 1,100 EMIT granules compared to NASA's L2B dataset. Furthermore, the model successfully mapped emissions at 24 of the world's 25 top-emitting landfills, demonstrating its high sensitivity to weaker, localized point sources.

For climate scientists and environmental practitioners, MAPL-EMIT solves three critical tasks simultaneously: measuring the precise amount of methane per pixel, segmenting overlapping plumes, and tracing gas back to its exact source. It leverages EMIT's 80-kilometer field of view, 60-meter spatial resolution, and 7.4-nanometer spectral sampling. This offers a much finer spatial resolution than global mappers like TROPOMI, which has a 2,600-kilometer swath but a coarser 5.5-by-3.5-kilometer resolution. To support open research, Google has released the trained model and synthetic plumes on Kaggle, an inference library on GitHub, and the global plume database on Earth Engine.