New Publication Towards AI-driven emission estimation for point sources of methane
Thomas Plewa and coworkers developed a machine learning method to estimate methane emission rates from satellite and aircraft images of exhaust plumes emanating from localized sources such as leakages in the oil & gas industry. Such images are typically provided through hyperspectral imaging spectrometers.
The study team used a modified ResNet-50 architecutre to estimate methane fluxes from synthetic high-resolution airborne imagery. The ResNet was trained on plumes simulated with a large-eddy code at 5 m resolution with varying emission rates and wind speeds. The model predicts flux directly from images, using a negative Gaussian log-likelihood loss to improve accuracy and provide uncertainty estimates. It achieved high performance (MPE 3.24%, MAPE 13.86%, r = 0.98) and outperformed previous methods like MethaNet. Analysis revealed systematic biases at very high wind speeds and low flux rates, suggesting the model infers wind information from plume shape but struggles under extreme conditions. The work introduces training and tuning improvements, a robust evaluation pipeline, and reliable per-prediction uncertainties, laying the groundwork for application to real flight data in future studies.
Plewa, T., Butz, A., Frankenberg, C., Thorpe, A. K., and Marshall, J.: Improvements of AI-driven emission estimation for point sources applied to high resolution 2-D methane-plume imagery, Remote Sensing of Environment, 331, 115002, , 2025.