PlumeBed: A Multispectral Satellite Methane Plume Detector Enabled by Transfer Learning of a Multi‐Source Hyperspectral Data Set

Abstract Satellite‐based detection of methane super‐emitters in oil and gas fields is critical to inform methane mitigation actions. Multispectral satellite instruments such as Sentinel‐2 offer frequent global coverage, making them suitable for monitoring methane super‐emitters worldwide. However, automatically detecting methane emissions from the vast amount of noisy multispectral satellite data remains challenging. Recent studies have shown that deep learning is promising for this task, but it requires a large set of representative training samples, which are still limited. Hyperspectral data, particularly from airborne sources, are relatively mature and have accumulated some data sets, for example, from Carbon Mapper. Here, we develop PlumeBed, which consists of a synthetic image generation module and a domain adversarial neural network (DANN) module. The synthetic image generation module synthesizes training data by combining Carbon Mapper methane plumes and Sentinel‐2 background noises. The DANN module is then trained to detect methane plumes from Sentinel‐2 images. Evaluation against testing data sets compiled from previously reported super‐emitters shows that the PlumeBed detector achieves an average macro‐F1 score of 0.86, outperforming the conventional deep learning frameworks such as ResNet‐50. We further apply PlumeBed to a previously unseen region in the Dauletabad gas field of Turkmenistan. This application unveils 14 methane super‐emitters based on 1‐year of Sentinel‐2 data. Our study demonstrates that utilizing airborne hyperspectral data through transfer learning is promising to efficiently detect methane super‐emitters in the global‐coverage multispectral satellite data.