Observation-conditioned probability of detection for satellite methane point sources

Methane is a key target for rapid mitigation because of its large warming potential and short life in the atmosphere. The Methane Alert and Response System (MARS), under the UNEP's International Methane Emissions Observatory (IMEO), supports mitigation efforts through satellite-based methane monitoring. In 2025, MARS notified 3738 methane plumes from the oil and gas sector to governments and companies using public multi-spectral (Sentinel-2 and Landsat) and hyper-spectral (EMIT, PRISMA and EnMAP) detections. An important aspect of MARS is the collection of satellite observations where no methane plumes are detected, which can be used as evidence of effective mitigation efforts reported by companies. However, non-detections can also result from unfavourable observing and environmental conditions, such as strong winds, retrieval artifacts, or differences in the sensitivity of instruments. Therefore, interpreting non-detections properly requires considering the observation-specific probability of detection (PoD), which depends on wind conditions, observation geometry, the on-board satellite instrument, and image noise. Here, we quantify how these factors influence the PoD and develop an operational parametric model to efficiently evaluate MARS observations. 

We assess detection performance across a wide range of realistic conditions in oil and gas regions by sampling scenes from the MARS archive covering diverse wind speeds, solar/viewing geometries, and noise regimes that vary with surface albedo and time. Representative synthetic methane plumes simulated with the WRF-LES model are injected into the top-of-atmosphere (TOA) radiance of satellite images. For each scene, multiple plume realizations at different flux rates are processed using the MARS operational detection pipeline to determine the detection frequency. We then fit a logistic PoD curve as a function of flux, with the slope and midpoint related to observation conditions for each instrument. Results show that wind speeds and noise levels are the dominant drivers affecting the slope and shift of the sigmoid PoD curve in most cases. We compare this parametric model on independent testing scenes and provide average probability of detection estimates for different instruments on major oil and gas basins. This parametric model will support the decision-making process in MARS in future potential mitigation actions.