MethaNet: an AI-powered approach to quantifying methane point-source emission from high-resolution 2-D plume imagery.

Abstract:

Methane is one of the most important anthropogenic greenhouse gas with impact on the Earth’s atmospheric radiative budget and the tropospheric air quality. Despite its well-appreciated significance, emission quantification of local and regional methane sources has been proven challenging. Localized point sources contribute to the majority of methane emissions from fossil fuel production and usage. Quantifying these point sources thus holds key to understanding the regional budgets and source category distributions. Recent advancements in airborne remote sensing instruments enable retrievals of methane enhancements at an unprecedented resolution of 1–5 m at regional scales. Here, we developed an algorithm using Convolutional Neuron Network (CNN) to predict methane point-source emission from high-resolution 2-D plume images by the next-generation Airborne Visible/Infrared Imaging Spectrometer (AVIRIS-NG). We built this on top of our previous study that shows methane plume morphology can inform the wind speed near the plume. The knowledge of wind speed together with the integrated total methane enhancement in each plume can yield the total emission estimate. On average, the error estimate from our algorithm based on randomly generated plumes is approximately 20 % for an individual estimate and less than 6 % for an aggregation of 30 plumes. This shows a noticeable improvement from previous methods that rely on in-situ measurements or reanalysis weather data. Our results support the basis for the applicability of this technique to quantifying point sources in an automated manner over large geographical areas, benefiting future field campaigns as well as upcoming satellite missions.