Google Research has unveiled MAPL-EMIT, a deep-learning framework that automates the detection and mapping of methane emissions from satellite imagery. The system, built on a Swin-S vision transformer architecture, processes hyperspectral data from NASA's EMIT instrument aboard the International Space Station, achieving an 84% recall rate on expert-annotated methane plumes. By analyzing how methane disperses across landscapes alongside surrounding spatial context, the model significantly reduces false positives compared to traditional matched-filter methods, effectively turning invisible greenhouse gases into visible, actionable data.
The tool targets critical methane point sources in the waste, agriculture, and energy sectors—areas where scientists identify the most cost-effective mitigation opportunities. Google is releasing the complete framework to the scientific community, including a global plume database via Earth Engine, trained models, synthetic training data on Kaggle, and open-source inference libraries on GitHub. The research aligns with the Global Methane Pledge, through which 125 countries have committed to reducing emissions by 30% by 2030, and demonstrates AI's role in accelerating climate action by enabling real-time monitoring at global scale.
Key Points
MAPL-EMIT uses vision transformer architecture to detect methane plumes with 84% recall, outperforming existing satellite-based enhancement methods
Google is releasing trained models, global plume database, and inference libraries as open-source resources for the scientific community
The system processes NASA's EMIT satellite data to locate methane point sources across oil/gas, agricultural, and waste facilities globally
The tool supports the Global Methane Pledge, helping 125+ countries track progress toward 30% emissions reduction by 2030