loading page

Improving Deep Learning Methods  for Robust Methane Plume Detection using Alternative Input Representations
  • +4
  • Anagha Satish,
  • Brian Bue,
  • Jake Lee,
  • Andrew Thorpe,
  • Daniel Cusworth,
  • Alana Ayasse,
  • Riley Duren
Anagha Satish
California Institute of Technology, Jet Propulsion Laboratory, California Institute of Technology

Corresponding Author:[email protected]

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


Methane (CH4) is a prominent greenhouse gas responsible for about 20% of all atmospheric radiative forcing. As we notice trends in increasing global temperatures, understanding and detecting these emissions has become increasingly important. This requires the creation of robust greenhouse gas plume detectors. Previous work at the NASA Jet Propulsion Laboratory has shown Convolutional Neural Networks (CNN) to be an appropriate solution to map methane sources from future imaging spectrometer missions, such as Carbon Mapper. However, current models suffer from a high rate of false positives due to false enhancements in the detected images.
We have compiled datasets from two Airborne Visible/Infrared Imaging Spectrometer Next Generation (AVIRIS-NG) California campaigns. We then trained a GoogleNet CNN Classifier model on each campaign. The baseline current model uses a Unimodal column-wise matched filter (CMF). This results in a model known to be sensitive to false enhancements, such as water/water vapor, bright/dark surfaces, or confuser materials with similar absorption wavelengths to methane. We first note improvements between the Unimodal CMF model and a new Surface-Controlled CMF model, whose dataset matches that of the Unimodal CMF model, but removes enhancements not matching the absorption wavelength of methane. From this, we note minimal improvement (1% increase in F1 score). We then experiment with various auxiliary products measuring albedo (rgbmu, SWALB), vegetation (NDVI, ENDVI), and water (h2o, NDWI) indices designed to combat issues known to produce false enhancements. After training on these new input representations for both campaigns, we noticed a significant improvement in the multi-channel model’s results. We observe an increase in the F1 score for classifying positive tiles from 0.78 to 0.86 when trained using auxiliary albedo indices, showing promise for future use of auxiliary products in improving methane plume detectors.
17 Jan 2024Submitted to ESS Open Archive
02 Feb 2024Published in ESS Open Archive