Reducing False Positives in Remote Sensing and Analysis of Atmospheric Methane

Abstract

Industrial methane (CH₄) emissions constitute a significant percentage of the global atmospheric methane inventory. CH₄’s warming effects are estimated to be 86 times those of carbon dioxide over a 20-year time scale. Reducing these emissions is critical to addressing climate change, and because many sources of methane are anthropogenic, we have a direct ability to mitigate the emissions if we have a good understanding of the emission sources. Recent advances in remote sensing have greatly improved the ability to locate and quantify methane emissions at their sources. However, much of the data analysis to confirm atmospheric CH₄ plumes in satellite data is manual, and analysts are often presented with false positive detections that require their review. The review process is laborious and can be improved by introducing a system that helps to flag likely false positives. This paper looks specifically at the data processing pipeline of Carbon Mapper, one of the dominant methane detection missions. Using open-source Python libraries scikit-learn and scikit-image, a machine learning model is developed based on qualities of suspected methane plumes such as shape, texture, and orientation relative to recorded local wind data. It is shown that this method is effective at recognizing false positives, though it also categorizes many true positives as false positives. It is believed that with a higher volume of verified false positives, the model’s performance would improve, as it was trained on a ~9:1 ratio of real methane plumes to false positives. The analysis suggests that even if not deployed as a binary classifier, the model may be successful as part of a decision-support system.