Convolutional Neural Networks for Detection of Point Source Methane
Plumes in Airborne Imaging Spectrometer Data
Abstract
To address discrepancies between bottom-up and top-down inventories of
methane emissions, the detection and quantification of methane point
source emissions is of critical concern. Multiple airborne imaging
spectrometer campaigns have identified the heavy-tailed distribution of
point source methane emissions. The quantification of point source
plumes is a two-part problem requiring the detection and delineation of
plumes, followed by an accurate accounting of their total methane
enhancement. Plume detection and delineation has often relied on manual
or statistical methods, but automated methods taking into account plume
morphology are essential as the volume of imaging spectrometer data
rapidly increases. Fully convolutional neural networks (FCNNs) represent
a robust solution to this problem allowing for the detection and
delineation of methane point source emissions with minimal analyst
input. This work demonstrates the applicability of FCNNs for accurate
quantification of methane point source emissions by training a model on
data from a 2019 Permian Basin survey by the Next Generation Airborne
Visible InfraRed Imaging Spectrometer (AVIRIS-NG). FCNNs were trained
using plumes that were manually interpreted from matched filter
retrievals of methane enhancements. Our methodology was able to
accurately detect and delineate methane plumes, and did so with fewer
false positives than statistical methods. Given the anticipated
satellite imaging spectrometer missions capable of global mapping of
point sources, automated deep learning methods will be necessary to deal
with methane plume detection in very large volumes of data.