Improvements of AI-driven emission estimation for point sources applied to high resolution 2-D methane-plume imagery

Anthropogenic methane (CH 4 ) sources have had a considerable impact on the Earth’s changing radiation budget since pre-industrial times. Localized sources such as those resulting from the fossil fuel industry and waste treatment have been shown to make up a substantial fraction of the emission total, and CH 4 plumes from such sources are detectable through airborne and space-based hyperspectral imaging techniques. Here, we further develop a machine learning technique to estimate CH 4 emission rates from such plume images without the need for auxiliary data such as local wind speed information. We directly build upon the idea of previous research which used a convolutional neural network (CNN) called MethaNet and a library of large-eddy-simulations (LES) of turbulent CH 4 plumes as our synthetic data environment. Here we suggest appropriate error metrics and changes to the training procedure that reduce systematic biases present in previous studies. Our improved setup has a mean absolute percentage error (MAPE) of 10% for sources with flux rates above 40 kg h −1 , a Pearson correlation coefficient of 98% and is capable of providing meaningful error estimates for its predictions. This is a significant improvement to MethaNet and other studies and can be used as an efficient method for point source quantification in the future. • Improvements to deep learning driven point source emission estimation for methane. • Removing biases present in previous deep learning driven studies. • Providing meaningful uncertainty estimates for the predicted emission rates. • Significantly improved model performance compared to previous studies. • Analysis pipeline revealing current limitations of deep learning-based approaches.