Machine Learning Techniques for the Retrieval of Methane from the Sentinel-5 Precursor Satellite

Rose Fenwick,  University of Leicster,  rof3@le.ac.uk (Presenter)
Hartmut Boesch,  University of Leicester,  hartmut.boesch@le.ac.uk
Ivan Tyukin,  University of Leicester,  i.tyukin@le.ac.uk
Robert Parker,  University of Leicester,  rjp23@le.ac.uk

Methane is the second most important anthropocentric greenhouse gas. Accurate and timely observations of its global distribution from satellites is an important prerequisite for monitoring of its emission sources and their and their time evolution.
The TROPOMI on board the Sentinel5 Precursor satellite provides shortwave-infrared radiances that allow to infer Methane mixing ratio via a physics-based retrieval algorithm. Owing to the vast data amount collected by TROPOMI, application of traditional physics-based methods is becoming very challenging due to the large computational effort involved.
Machine learning techniques have the potential to replicate such physics-based retrievals; a relatively simple task on the face of it, but the challenge is to produce results which are sufficiently accurate for highly complex and variable conditions
In this presentation, we describe the method of retrieval of methane using Machine Learning Techniques.
We show result from a neural network trained to predict the concentration of methane trained with the operational TROPOMI product. We also discuss classification methods that can be used to classify the problem before training. The analysis of these results and a discussion of the reliability and accuracy concludes the presentation.

Poster: Poster_Fenwick__92_25.pdf 

Presentation Type: Poster

Session: 1.5d Retrieval algorithms and methods for inter-instrument and product Cal/Val

Session Date: Monday (6/14) 12:00 PM