loading page

Robust Multi-Campaign Imaging Spectrometer Methane Plume Detection using Deep Learning
  • +4
  • Jake Lee,
  • Brian Bue,
  • Michael Garay,
  • Andrew Thorpe,
  • Riley Duren,
  • Daniel Cusworth,
  • Alana Ayasse
Jake Lee
Jet Propulsion Laboratory, California Institute of Technology

Corresponding Author:jake.h.lee@jpl.nasa.gov

Author Profile
Brian Bue
Jet Propulsion Laboratory, California Institute of Technology
Michael Garay
Jet Propulsion Laboratory, California Institute of Technology
Andrew Thorpe
Jet Propulsion Laboratory, California Institute of Technology
Riley Duren
Carbon Mapper, Inc, University of Arizona
Daniel Cusworth
Carbon Mapper, Inc, University of Arizona
Alana Ayasse
Carbon Mapper, Inc, University of Arizona


Identification of global methane (CH4) sources is critical to the quantification and mitigation of this greenhouse gas. Future imaging spectrometer missions, such as Carbon Mapper, will provide global, spatially resolved observations that will make it possible to accurately map methane sources. However, the sheer data volume of these missions make manual source identification infeasible, and expected artifacts in matched filter methane plume identification challenge simple thresholding. Recent works have demonstrated the feasibility of Convolutional Neural Networks (CNNs) for plume detection; however, in the past, these models have suffered from high false positive rates and were limited in their training and evaluation to individual flight campaigns.
We have assembled quality-controlled tiled datasets from three Airborne Visible/Infrared Imaging Spectrometer Next Generation (AVIRIS-NG) campaigns: a 2020 California campaign (“COVID”), a 2019 Texas Permian Basin campaign (“Permian”), and another 2018 California campaign (“CACH4”). These datasets are notable for their diversity of surface conditions, spatial resolutions, and source types (oil & gas, energy, waste, livestock). Labeled methane sources in these datasets have been manually verified, and flightlines with systematic artifacts have been filtered out. We trained a GoogLeNet CNN classifier model on each of these campaigns to evaluate intra- and inter- campaign performance. We also trained a model on all three campaigns and evaluated its performance on each dataset. We observed an F1 performance of 0.7 or greater for each model trained and evaluated on its own dataset. We also observed that the model trained on all three datasets often outperforms individual models on multiple metrics. Finally, we converted the model into a fully convolutional network (FCN) for methane plume saliency map generation. We plan to extend this work to datasets acquired by the Global Airborne Observatory (GAO) and prepare a model for deployment for the Carbon Mapper orbital data product pipeline.
11 Dec 2022Submitted to AGU Fall Meeting 2022
03 Jan 2023Published in AGU Fall Meeting 2022