Google and NASA Release AI 'MAPL-EMIT' for Methane Leak Detection

Google Research and NASA JPL released MAPL-EMIT, a deep learning model that automatically detects methane leaks by analyzing EMIT hyperspectral instrument data from the International Space Station (ISS). This study was published in PNAS 2026 and achieved the result of discovering approximately 1.5 times more methane plumes compared to traditional manual expert methods.

Technical Approach and Training Data

MAPL-EMIT uses an architecture combining Swin Transformer V2 and U-Net. Unlike existing methods that suffered from false positives due to similar surface spectra, this model takes raw radiance as input and simultaneously estimates concentration, plume boundaries, and source locations. Training utilized a 3.6 million synthetic plume dataset generated using Lagrangian puff models and Simplex noise, designed to enable accurate detection based on physical laws (HITRAN2020, Beer-Lambert law) without expert labeling.

Performance and Validation Results

In comparison with actual observation data (NASA L2B), MAPL-EMIT detected a total of 2,210 plumes, approximately 1.4 times more than the 1,580 in existing L2B data. Notably, it captured 2.9 times more plumes in the low-concentration range (0-50ppm-m) compared to existing methods. It detected 24 (96%) of the world's 25 largest methane-emitting landfills, and showed high signal-to-noise ratio (SNR 8.60) and shape accuracy in comparison with airborne observation data (AVIRIS-3). In controlled release experiments, it recorded 0 false positives during non-release periods, demonstrating reliability.

Public Assets and Usage

The model and dataset are released as open source for research and commercial use.

• GitHub: Inference library (Apache 2.0)
• Kaggle: TensorFlow SavedModel and 3.6 million synthetic plume dataset
• Earth Engine: Global plume database (CC BY 4.0)

Google plans to expand coverage by 30-50 times in conjunction with the next-generation spectrometer (SBG) mission scheduled for launch in 2028. However, uncertainty exists in the conversion between the concentration (ppm-m) output by the model and actual emission rates (kg/hr), and final mitigation actions require human judgment.