Detecting methane emissions in TROPOMI data using machine learning
Joannes Dyonisius Maasakkers, SRON Netherlands Institute for Space Research, j.d.maasakkers@sron.nl (Presenter)
Berend Schuit, SRON Netherlands Institute for Space Research, berendjschuit@gmail.com
Gourav Mahapatra, SRON Netherlands Institute for Space Research, g.mahapatra@sron.nl
Sudhanshu Pandey, SRON Netherlands Institute for Space Research, s.pandey@sron.nl
Allard de Boeij, SRON Netherlands Institute for Space Research, allard@deboeij.eu
Alba Lorente, SRON Netherlands Institute for Space Research, a.lorente.delgado@sron.nl
Sander Houweling, Vrije Universiteit Amsterdam, s.houweling@sron.nl
Daniel J. Varon, Harvard University, danielvaron@g.harvard.edu
Dylan Jervis, GHGSat Inc, dyj@ghgsat.com
Jason McKeever, GHGSat Inc, jtmckeev@ghgsat.com
Itziar Irakulis-Loitxate, Universitat Politècnica de València, iiraloi@doctor.upv.es
Luis Guanter, Universitat Politècnica de València, lguanter@fis.upv.es
Daniel H. Cusworth, JPL, daniel.cusworth@jpl.nasa.gov
Ilse Aben, SRON Netherlands Institute for Space Research, i.aben@sron.nl
With atmospheric methane concentrations increasing at record pace, identifying “super-emitters” that provide great opportunities for emission reduction is of vital urgency. Since its launch in 2017, the Tropospheric Monitoring Instrument (TROPOMI) aboard ESA’s Sentinel-5P satellite has been acquiring a vast amount of methane data across the world, including numerous emission plumes. We now use that data to train a set of neural networks to automatically detect methane plumes in TROPOMI observations as soon as they become available. We follow a two-step approach where we first use a convolutional neural network to detect plume-like features in the TROPOMI data. Subsequently, we use a second fully-connected neural network that incorporates supporting data to analyze whether these features are actually the result of methane emissions. The pair of networks can process a new orbit of data within a minute, something that would be impossible without applying artificial intelligence. After detection, sites of interest can be investigated further using long-term data. We show several applications of the setup, including an overview of the largest emissions detected in 2020, which are related to various sources including wetlands, coal mines, landfills, and oil/gas production. Furthermore, we present a detailed look into short-duration emission events at Russian natural gas compressor stations. Finally, we show how detections can be used in conjunction with high-resolution instruments such as GHGSat, Sentinel-2, and PRISMA, to pinpoint the exact facilities responsible for the observed emission plumes.
Poster: Poster_Maasakkers__74_25.pdf
Presentation Type: Poster
Session: 3.5b Observations to quantify hot spots and local/urban emissions
Session Date: Wednesday (6/16) 12:00 PM