Deep Learning-Based quantifications of methane emissions with field applications

Tackling methane emissions is critical for mitigating climate change, emphasizing the need to identify, quantify, and mitigate emission sources. Methane emissions can be detected using airborne techniques equipped with suitable sensors; however, a method is needed to precisely estimate methane source flux rates from these measurements. We have developed a convolutional neural network (CNN) model using a dataset generated from Large-Eddy Simulation (LES) mimicking the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS-NG) measurements with different source flux rates (5–100 kg/hr) and varying wind speeds (1–10 m/s). Our proposed deep neural network architecture is simple, ensuring efficient training on synthetic data. Additionally, we have optimized model hyperparameters and architecture, improving accuracy and enabling precise quantification of source flux rates. We test the trained model on synthetic 2D images, demonstrating excellent overall predictability with mean absolute percentage errors of 2.15 % for the training dataset, 5.02 % for the validation dataset, and 4.84 % for the test dataset in predicting the source flux rates in kg/hr using the synthetic images. The error of predicting source flux rates using conventional flux inversion methods is around 25–50 %, whereas the machine learning method of predicting source flux rate achieves an accuracy of more than 17 %, which is the current state-of-the-art in this process. Validation against field data also showed excellent results, with a low error (∼5%) in estimating source flux rates in the Permian Basin oil and gas wells. Validation against field measurements indicates that this method can reduce the uncertainties associated with currently used conventional source flux inversions without relying on background information like wind speeds. Sometimes, the uncertainties of field measurements are as significant as the actual emission rates, especially for low emissions. This machine learning model can serve as a screening tool to prioritize top emitters for plugging orphaned and abandoned wells in large geographical areas, such as city boundaries covering more than 1,000 miles.