British researchers have created a tool that automatically detects methane clouds from orbit using machine learning with hyperspectral data. This innovation could pinpoint the locations of methane super emitters, enabling quicker takedown of these greenhouse gas sources.
Although climate strategies primarily target CO2, reducing methane emissions is essential for slowing the rise in temperatures. Methane traps heat at a rate about 80 times higher than CO2 in the short term, yet it stays in the atmosphere for only about seven to 12 years compared with the centuries CO2 can endure.
Acting swiftly to curb methane from human activities is expected to directly slow global warming and improve air quality. Researchers estimate that attainable methane reductions could prevent nearly 0.3 degrees Celsius of warming over the next two decades.
The new system has demonstrated its effectiveness due to a collaboration involving Oxford University.
Until recently, mapping methane clouds from aerial imagery was inefficient, with data processing being time consuming. Methane gas is transparent to both the human eye and the spectral ranges used by most satellites. Even when sensors operate in the appropriate range to detect methane, noise and laborious manual methods have hindered plume identification.
The new system is far more efficient
A pioneering machine learning tool developed by researchers from Oxford in the United Kingdom overcomes these hurdles by spotting methane clouds in hyperspectral satellite data. It works by detecting narrower spectral bands than common multispectral satellites, which helps tune the methane signature and filter out noise. The tradeoff is the sheer volume of data, which makes processing rely on artificial intelligence to handle the load.
Researchers validated the model with 167,825 hyperspectral mosaics, each covering about 1.64 square kilometers, captured by NASA’s AVIRIS sensor over the Four Corners region of the United States. The algorithm was then applied to data from additional hyperspectral satellites in orbit.
On average the model achieves over 81 percent accuracy in identifying large methane plumes and is about 21.5 percent more accurate than the most detailed systems previously available. It also delivers a notably lower false positive rate, cutting such errors by roughly 41.83 percent compared with earlier methods.
To encourage ongoing methane research, the annotated dataset and the code used for the model have been released on the project page. The team is exploring whether the model can run directly on satellites, enabling other satellites to make follow-up observations as part of the NIO.space initiative. The goal is to support faster, coordinated sensing across multiple satellites in space-based networks.
Lead researcher Vít Růžička, a PhD student in the Department of Computer Science at the University of Oxford, explained that the process could start with priority alerts on Earth. These alerts would include coordinates of identified methane sources, allowing a constellation of satellites to cooperate autonomously. The initial weak detection could serve as a warning to focus additional imagery on the location for follow-up confirmation and action.
Reference work: DOI: 10.1038/s41598-023-44918-6
Additional notes specify the environmental research team and project scope. The environmental department contact is no longer included in this rewrite. The work illustrates a scalable approach to monitoring methane emissions with hyperspectral data and AI, aligning with efforts to improve air quality and mitigate climate impacts. The methodology also has implications for other atmospheric gas detections and satellite-enabled environmental surveillance. [Citation: University of Oxford; NASA AVIRIS; NIO.space initiative]