Despite methane’s important role as a greenhouse gas, the contribution of individual sources to rising methane concentrations in Earth’s atmosphere is poorly understood. This is in part due to the lack of frequent measurements on a global scale, required to accurately quantify fugitive methane sources. Future missions such as Earth Surface Mineral Dust Source Investigation (EMIT), Surface Biology and Geology (SBG), and Carbon Mapper promise to provide global, spatially resolved spectroscopy observations that will allow us to map methane sources. However, the detection and attribution of individual methane sources is challenged by retrieval artifacts and noise in retrieved methane concentrations. Additionally, manual methane plume detection is not scalable to global space-borne observations due to the sheer volume of data generated. A robust automated system to detect methane plumes is needed. We evaluated the performance and sensitivity of several methane plume detection methods on 30m to 60m hyperspectral imagery, downsampled from airborne campaigns with AVIRIS-NG. To aid the training of the plume detection models, we explored supplementing downsampled airborne imagery with Large Eddy Simulations (LES) of methane plumes. We compared baseline methods such as thresholding and random forest classifiers, as well as state-of-the-art deep learning methods such as convolutional neural networks (CNNs) for classification and conditional adversarial networks (pix2pix) for plume segmentation.

Methane’s high heat trapping potential has made it a priority for quantification and mitigation efforts worldwide. Ground-based surveys and in-situ measurement techniques to quantify natural and fugitive methane emission sources are time-consuming, expensive, and often lead to sparse measurements. Failure to accurately quantify emissions at the point-source scale have thus led to poorly constrained emission estimates. Airborne imaging spectrometers such as the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS-NG) and the Global Airborne Observatory (GAO) have been employed to map the often stochastic and intermittent point-source emissions from a diverse set of source types including oil and gas, dairy, etc. A matched filter is applied to the methane-absorption relevant spectral features of the instrument’s radiance cube. Machine learning models are then trained to recognize methane plumes from these column-matched filter methane maps. However, current Convolutional Neural Network (CNN) models suffer from a high false-positive rate and poorly generalize to new scenes. False-positive detections are primarily due to methane absorption-mimicking surface spectroscopic features, as well as a lack of training data. To supplement the available training data, we utilize Large Eddy Simulations (LES) of methane point-source emissions to train a Convolutional Neural Network (CNN) on a plume-classification task. We observe a significant distribution shift between LES and AVIRIS-NG plumes, primarily caused by high LES plume enhancements. Through a series of image transforms verified through an adversarial approach using a discriminator network, we minimize the distribution shift between synthetic LES plumes and plumes observed by AVIRIS-NG and GAO. CNNs trained on a mixture of LES and real-world plumes, and tested on flightlines from multiple campaigns exhibit an error reduction compared to previous models. The reduction in false-positive plume detections demonstrates that supplementing the limited training data of real methane plumes with LES provides an avenue to make automatic detection more robust for future airborne and spaceborne missions such as SBG, EMIT, and Carbon Mapper.