Infrastructure · Climate · Computer science
MAPL-EMIT automates global methane-plume mapping from hyperspectral satellite data
A peer-reviewed methane detection model plus a released global plume database, trained model, synthetic dataset and inference library.
Summary
MAPL-EMIT was trained on 3.6 million physics-based synthetic plumes injected into real EMIT radiance scenes. Against NASA's expert-annotated plume set, the model recovered 84% of labeled plumes and produced about 50% more plausible detections across roughly 1,100 EMIT granules. It also mapped plumes at 24 of the 25 highest-emitting landfills and released its outputs and artifacts for research use.
AI role
A vision transformer uses full-spectrum radiance and spatial context to quantify methane enhancement, delineate overlapping plumes and estimate their source locations.
Narrative role
Turns a manually intensive remote-sensing task into reusable scientific infrastructure with global-scale outputs. The result expands observable methane sources rather than merely improving a laboratory benchmark.
Caveat
Additional plausible detections are not all independently verified emissions. False positives remain a problem in complex terrain, confidence filtering trades recall against precision, and training depends heavily on simulated rather than labeled real plumes.