Science & Research

This study tackles one of the biggest challenges in hyperspectral remote sensing: detecting clouds and cloud shadows. By developing and benchmarking deep learning models, the team significantly improved cloud and shadow segmentation performance for both MethaneSAT and its airborne partner, MethaneAIR. These improvements enhance the reliability of methane retrievals worldwide, strengthening MethaneSAT’s capacity to support actionable climate solutions.

Using TROPOMI and VIIRS satellite instruments to measure nitrogen oxide (NOx) emissions from 44 major oil and gas regions around the world, the study finds that commonly used emission inventories significantly underestimate NOx emissions from these activities -- in some cases by more than 70% -- meaning a major source of air pollution is being undercounted. It also show that NOx emissions often occur alongside methane emissions, highlighting important links between air quality and climate impacts from oil and gas operations.

Demonstrates MethaneSAT's capabilities to survey high-emitting methane sources across global oil and gas basins. Analysis of global hotspots reveals persistent sources in Turkmenistan’s South Caspian and the US Permian Basin, alongside major super-emitters in Venezuela, Iran, and the Appalachian basin. Authors also identify significant emissions in West Siberia, offshore Gulf of Mexico, and from the waste sector. These results highlight MethaneSAT’s utility in mapping regional methane hotspots and super-emitters worldwide.

Using observations from MethaneAIR from regions responsible for 70% of 2023 U.S. onshore oil and gas production, this study estimates total oil/gas methane emissions to be ~8 Tg/yr—around five times higher than U.S. EPA estimates and equivalent to a 1.6% methane loss rate. Performance varies widely by basin, with highly productive regions like the Permian, Appalachian, and Haynesville-Bossier showing the largest total emissions, while older basins such as the Uinta and Piceance exhibit higher loss rates. Regional comparisons showed good agreement across total emissions quantified by MethaneAIR and other empirical and remote sensing estimates.
 

Introduces the Spectral Channel Attention Network (SCAN), a deep learning approach designed to improve cloud and cloud shadow detection in hyperspectral imagery from MethaneSAT and MethaneAIR. By dynamically weighting individual spectral bands based on their physical relevance, rather than treating all wavelengths equally, SCAN outperforms traditional U-Net and transformer-based attention models on MethaneSAT data, improving F1-scores and shadow detection accuracy. When combined with spatial models in an ensemble framework, the approach achieves state-of-the-art performance, directly strengthening the reliability of satellite-based methane retrievals worldwide.