Machine learning based estimations of methane plume flux rates with base case comparisons

Abstract

This thesis is concerned with applying machine learning methods to the problem of quantifying methane emissions; an important step in the mitigation of methane and its negative effects to climate change. The use of machine learning for estimating emissions has proven promising for its ability to streamline emission estimations by forgoing the need for in situ ancillary measurements. Given 2-D column integrated enhancement images of methane plumes, we apply various methods to estimate flux rate. As a base case we estimate flux rate with two flux inversion methods: cross-sectional method and the integrated mass enhancement method. We then benchmarked a machine learning method called MethaNet by experimenting with various training/testing partitioning schemes and then compared the results with the flux inversion base cases. This thesis finds that machine learning methods can be more accurate than flux inversion methods. However, more tests should be done with a more statistically ample data set to further confirm our findings.