N-BPMSNet: an NDMI-guided bitemporal network for methane plume detection and segmentation from sentinel-2 multispectral observations
File(s)
Accepted version
Author(s)
Mason, Philippa
Type
Journal Article
Abstract
Mitigating methane emissions is critical for addressing global warming, and accurate point-source detection is crucial for identifying super-emitters and targeted mitigation. With its fine spatial resolution, global coverage, and open accessibility, Sentinel-2 offers strong potential for large-scale methane monitoring. Current detection methods, whether traditional retrieval-based approaches or AI-based models, suffer
from frequent false positives, low signal-to-noise ratios, reliance on manual verification, and limited generalization across heterogeneous surfaces, leading to unreliable plume identification and segmentation. To address these challenges, we propose normalized difference methane index-guided bi-temporal plume
methane segmentation network (N-BPMSNet), a Sentinel-2-based methane plume detection and segmentation framework. The model integrates temporal, spectral, spatial, and differential
information through specialized modules that guide the network to focus on subtle plume regions and accurately extract them from complex and interfering backgrounds. A large-scale real world Sentinel-2 dataset from 44 point sources across the USA,
Algeria, and Turkmenistan is also constructed, comprising 11494 samples with annotated methane plumes and nonplume scenes across various land cover types and emission facilities for model training, validation, and evaluation. N-BPMSNet has been tested
and achieved an F1-score of 0.8858 and an AUC of 0.9856 in the test dataset, outperforming existing approaches and advanced segmentation networks. Out-of-domain generalization is further evaluated using supplementary datasets from Khuzestan Province, Iran, and an O&G region in Algeria, demonstrating
robust performance in unseen regions. Case studies confirm its robustness to confounding factors such as flare smoke, cloud shadows, and offshore emissions, while an inference speed on the order of 0.1 s per 3 × 3 km tile highlights its scalability for operational monitoring. Collectively, the experimental results
demonstrate that N-BPMSNet achieves accurate, automated, and scalable plume segmentation under diverse environmental and surface conditions, making this method an important contribution to the detection of methane plumes from freely available data sources.
from frequent false positives, low signal-to-noise ratios, reliance on manual verification, and limited generalization across heterogeneous surfaces, leading to unreliable plume identification and segmentation. To address these challenges, we propose normalized difference methane index-guided bi-temporal plume
methane segmentation network (N-BPMSNet), a Sentinel-2-based methane plume detection and segmentation framework. The model integrates temporal, spectral, spatial, and differential
information through specialized modules that guide the network to focus on subtle plume regions and accurately extract them from complex and interfering backgrounds. A large-scale real world Sentinel-2 dataset from 44 point sources across the USA,
Algeria, and Turkmenistan is also constructed, comprising 11494 samples with annotated methane plumes and nonplume scenes across various land cover types and emission facilities for model training, validation, and evaluation. N-BPMSNet has been tested
and achieved an F1-score of 0.8858 and an AUC of 0.9856 in the test dataset, outperforming existing approaches and advanced segmentation networks. Out-of-domain generalization is further evaluated using supplementary datasets from Khuzestan Province, Iran, and an O&G region in Algeria, demonstrating
robust performance in unseen regions. Case studies confirm its robustness to confounding factors such as flare smoke, cloud shadows, and offshore emissions, while an inference speed on the order of 0.1 s per 3 × 3 km tile highlights its scalability for operational monitoring. Collectively, the experimental results
demonstrate that N-BPMSNet achieves accurate, automated, and scalable plume segmentation under diverse environmental and surface conditions, making this method an important contribution to the detection of methane plumes from freely available data sources.
Date Issued
2026-04-29
Date Acceptance
2026-04-27
Citation
IEEE Transactions on Geoscience and Remote Sensing, 2026, 64
ISSN
0196-2892
Publisher
Institute of Electrical and Electronics Engineers
Journal / Book Title
IEEE Transactions on Geoscience and Remote Sensing
Volume
64
Copyright Statement
Copyright © 2026 IEEE. This is the author’s accepted manuscript made available under a CC-BY licence in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy)
License URL
Publication Status
Published
Article Number
4106615
Date Publish Online
2026-04-29