A data-efficient deep transfer learning framework for methane super-emitter detection in oil and gas fields using the Sentinel-2 satellite

Abstract

Efficiently detecting large methane point sources (super-emitters) in oil and gas fields is crucial for informing stakeholder decisions about mitigation actions. Satellite measurements by multispectral instruments, such as Sentinel-2, offer global and frequent coverage. However, methane signals retrieved from satellite multispectral images are prone to surface and atmospheric artifacts that vary spatially and temporally, making it challenging to build a detection algorithm that applies everywhere. Hence, laborious manual inspection is often necessary, hindering widespread deployment of the technology. Here, we propose a novel deep-transfer-learning-based methane plume detection framework. It consists of two components: an adaptive artifact removal algorithm (low-reflectance artifact detection, LRAD) to reduce artifacts in methane retrievals and a deep subdomain adaptation network (DSAN) to detect methane plumes. To train the algorithm, we compile a dataset comprising 1627 Sentinel-2 images from six known methane super-emitters reported in the literature. We evaluate the ability of the algorithm to discover new methane sources with a suite of transfer tasks, in which training and evaluation data come from different regions. Results show that DSAN (average macro F1 score 0.86) outperforms four convolutional neural networks (CNNs), MethaNet (average macro F1 score 0.70), ResNet-50 (average macro F1 score 0.77), VGG16 (average macro F1 score 0.73), and EfficientNet-V2L (average macro F1 score 0.78), in transfer tasks. The transfer learning algorithm overcomes the issue of conventional CNNs, which is their performance degrades substantially in regions outside regions with training data. We apply the algorithm trained with known sources to an unannotated region in the Algerian Hassi Messaoud oil field and reveal 34 anomalous emission events during a 1-year period, which are attributed to three methane super-emitters associated with production and transmission infrastructure. These results demonstrate the potential of our deep-transfer-learning-based method in contributing towards efficient methane super-emitter discovery using Sentinel-2 across different oil and gas fields worldwide.