Abstract
<jats:p>Satellite remote sensing has become a critical tool for monitoring, measuring, and mitigating methane emissions globally. Reliable attribution of methane plumes to their source facility is needed to enable emission mitigation action, robust emissions accounting, and regulatory oversight. Although experts can manually attribute plumes to individual facilities, automated approaches are required for consistency, scalability, and transparency. Here, we provide a methodology to automatically estimate the source location of a plume as a spatial probability. Plume origin estimation is an essential component of attributing plumes to their source facility. We train a neural network to estimate a confidence region for the location of the plume source. The model achieves good accuracy, with 65.1% predicted origin points falling within 100m (4 pixels) of an expert-annotated point, and a median distance error of 64.2m (∼ 3 pixels). We demonstrate that the predicted origin probability is well-calibrated by comparing the percentage of expert origins which fall into different confidence region bins. We apply this methodology to methane plumes in GHGSat satellite imagery, but the framework could be adapted to other sensors and types of plumes.</jats:p>