地球觀測與應用研究
MAPL-EMIT Trains Satellite Methane Detection on 3.6 Million Simulated Plumes, Releases Model and Global Dataset
Google and NASA JPL enable a vision Transformer to directly combine EMIT’s spectral and spatial context, simultaneously estimating concentrations, plume boundaries, and emission-source locations. The publicly released model recovers about 80% of expert-labeled cases, but its higher sensitivity also introduces risks of surface confusion and false positives.
Google and NASA’s Jet Propulsion Laboratory have unveiled MAPL-EMIT, transforming satellite methane searches—which previously relied heavily on manual interpretation—into a deep-learning pipeline that can run in batches. The system processes hyperspectral radiance from the EMIT instrument aboard the International Space Station. With pixels measuring about 60 meters and a swath roughly 80 kilometers wide, EMIT can capture methane signatures at the facility scale, although minerals, surface materials, and noise can produce similar signals.
The model uses a Swin-S vision Transformer. Instead of applying a simple matched filter pixel by pixel, it uses both the full spectrum and nearby spatial patterns to identify plumes dispersing downwind. Each inference run jointly handles three tasks: estimating methane enhancement for every pixel, separating adjacent or overlapping plumes, and locating the plume head or emission source. Because real-world labeled data is scarce, the researchers injected 3.6 million physically simulated plumes generated with a Lagrangian puff model into real EMIT scenes, covering a range of wind conditions, terrain, and emission rates.
Results from the PNAS version published by Google show that the model recovered 84% of expert-labeled plumes across approximately 1,100 EMIT granules and identified about 50% more plausible candidates. It also detected 24 of the world’s 25 highest-emitting landfills. The global database is now available through Earth Engine, while the model weights and synthetic dataset are hosted on Kaggle. The Apache 2.0-licensed inference code supports tiled inference, deduplication, and spectral validation, producing Parquet files, plume masks, concentrations, integrated mass, and emission-rate estimates based on ERA5 wind fields.
This is still not a system for automatically assigning responsibility for emissions. Training relies primarily on synthetic plumes, so real atmospheric conditions and instrument drift may create a domain gap. High recall can also amplify false positives over complex surfaces. Users should retain spectral-fit confidence scores, seek repeat observations, and conduct ground-based validation rather than treating a single model output as regulatory evidence.