Automatic detection of methane emissions in multispectral satellite imagery using a vision transformer

Area of Science:

• Environmental Science
• Remote Sensing
• Artificial Intelligence

Background:

• Methane emissions significantly contribute to global warming.
• Current methane monitoring methods face limitations in quantification completeness due to trade-offs between coverage, resolution, and accuracy.
• Satellite-based detection requires balancing spectral resolution with data coverage and accuracy.

Purpose of the Study:

• To develop a deep learning tool for overcoming spectral resolution trade-offs in multi-spectral satellite data for methane detection.
• To achieve global coverage with high temporal and spatial resolution for methane emission monitoring.
• To significantly improve the state-of-the-art in automated methane emission detection.

Main Methods:

• Utilized deep learning algorithms to process multi-spectral satellite data.
• Developed a methane detection tool leveraging enhanced spectral resolution capabilities.
• Validated the model's performance against airborne methane measurement campaigns.

Main Results:

• The deep learning model overcomes spectral resolution limitations inherent in multi-spectral satellite data.
• Achieved global coverage with high temporal and spatial resolution for methane detection.
• Demonstrated the capability to detect methane point sources down to 0.01 km² plumes (200-300 kg CH₄ h⁻¹ sources) using Sentinel-2 data.
• Showcased an order of magnitude improvement over existing state-of-the-art detection methods.

Conclusions:

• Deep learning offers a powerful solution to enhance methane emission monitoring.
• The developed tool represents a significant advancement towards automated, high-resolution, global methane emission detection.
• Enables frequent (every few days) and precise tracking of methane sources, crucial for climate change mitigation efforts.