Methane 84 times more potent than carbon dioxide
Combining hyperspectral satellites with deep learning for automatic detection
Proposing practical AI guidelines distinguishing ‘rapid detection’ and ‘precision analysis’
Ensuring reliability with explainable AI
Demonstrating scalability for private satellites ‘[Edaily Reporter Kim Hyun-ah] AI technology capable of automatically detecting methane (CH₄) leaks—a key contributor to global warming—from space has been developed. Unlike previous approaches where humans needed to analyze satellite images, AI identifies methane leak points in the atmosphere, potentially revolutionizing the paradigm of satellite-based environmental monitoring.’
The research was led by Prof. Lim Jung-ho’s team at the Department of Civil Urban and Environmental Engineering at Ulsan National Institute of Science and Technology (UNIST).
Combining deep learning with hyperspectral satellite data, Prof. Lim’s team has built a deep learning-based image segmentation model using hyperspectral satellite data collected from NASA’s International Space Station (ISS) observation sensor EMIT. This enabled the development of a technique to automatically detect ‘plume’ shaped methane emissions in the atmosphere.
Methane is invisible but has a unique characteristic of absorbing specific infrared wavelengths. By designing the AI to learn from hyperspectral data, which includes hundreds of wavelength pieces of information, the model can recognize both the physical absorption characteristics and spatial distribution of methane, surpassing simple image analysis.
This model has successfully pinpointed methane leaks at major oil and gas production sites in countries including the US, China, Turkmenistan, and Algeria, as well as from waste treatment plants and coal mining sites.
The researchers’ approach included two methods: one directly using satellite radiance data for rapid detection without complicated preprocessing, suitable for screening large areas quickly, and the other using methane concentration enhancement data, which, despite more complex processing, allows for precise estimation of emission areas and volumes.
The study distinguishes AI usage based on detection purpose, allowing industries and regulatory entities to choose between ‘rapid detection’ and ‘precision verification’ models based on their needs.
To overcome the black-box limitation of traditional AI satellite analysis, the team employed explainable AI methods such as Grad-CAM and Integrated Gradients to ensure transparency in the analysis process, visualizing that AI indeed bases its decisions on methane’s infrared absorption properties.
Moreover, the model maintained high performance not only with NASA data but also with data from the private satellite Tanager-1, demonstrating its applicability across various hyperspectral satellites beyond specific satellites or sensors.
Supported by the Ministry of Environment and the Ministry of Education, the study, involving co-first authors Yang Se-young and Kim Ye-jin, was published in the international journal npj Climate and Atmospheric Science.
Prof. Lim stated, “Methane has a short atmospheric lifespan, making prompt location and containment crucial,” and noted that previous prolonged analysis and verification processes delayed responses. He added, “This technology builds a dense monitoring system for global methane leaks, combining hyperspectral satellites with AI, potentially serving as a key infrastructure for future global carbon regulations and emission verification systems.”