Deep Learning Ensemble for Methane Emissions Detection in Satellite Imagery

Methane emissions are a significant contributor to global warming and their effective detection and monitoring is essential to achieve NetZero goals. The role of methane in global warming is widely acknowledged, prompting initiatives such as the Global Methane Pledge and Oil and Gas Methane Partnership 2.0 to drive meaningful environmental action. Although methane reductions in the energy sector are often considered low-hanging fruit, significant technological challenges remain in detecting and quantifying methane emissions accurately under manageable monitoring costs.

Satellite imagery has emerged as the most promising and cost-effective solution to this issue. However, detecting methane emissions below 1 metric ton/hour continues to be a challenge.

This work outlines a novel methodology of an AI-based system for detecting methane emission locations using Sentinel-2 multispectral satellite imagery. The proposed technology employs at its core a bespoke ensemble of AI models to identify methane emission signals and the distinctive shapes of methane plumes, which are concealed in the short-wave infrared (SWIR) satellite bands. Our system can accurately distinguish methane signal and methane plume shape from the background noise with detection limit of 300 kg/hour. This capability is built upon a unique dataset of real-world methane leak events and leverages public-domain, multispectral satellite data.