Operational machine learning for remote spectroscopic detection of CH$_{4}$ point sources

TL;DR

We address automated detection of anthropogenic methane plumes in hyperspectral satellite data within a fully operational MARS pipeline. The approach combines a U-Net–based detector trained on a large, multi-sensor dataset (EMIT, PRISMA, EnMAP) with ensemble predictions to suppress false positives, demonstrating robust cross-sensor generalization via zero-shot transfer and targeted fine-tuning. Key contributions include the largest public dataset of annotated methane plumes, an ensemble detection framework that substantially reduces false alerts, and successful deployment processing thousands of scenes resulting in 1,351 verified leaks and 479 stakeholder notifications, plus mitigation case studies. This work advances AI-assisted methane monitoring and lays groundwork for scalable global surveillance as new hyperspectral missions come online.

Abstract

Mitigating anthropogenic methane sources is one the most cost-effective levers to slow down global warming. While satellite-based imaging spectrometers, such as EMIT, PRISMA, and EnMAP, can detect these point sources, current methane retrieval methods based on matched filters still produce a high number of false detections requiring laborious manual verification. This paper describes the operational deployment of a machine learning system for detecting methane emissions within the Methane Alert and Response System (MARS) of the United Nations Environment Programme's International Methane Emissions Observatory. We created the largest and most diverse global dataset of annotated methane plumes from three imaging spectrometer missions and quantitatively compared different deep learning model configurations. Focusing on the requirements for operational deployment, we extended prior evaluation methodologies from small tiled datasets to full granule evaluation. This revealed that deep learning models still produce a large number of false detections, a problem we address with model ensembling, which reduced false detections by over 74%. Deployed in the MARS pipeline, our system processes scenes and proposes plumes to analysts, accelerating the detection and analysis process. During seven months of operational deployment, it facilitated the verification of 1,351 distinct methane leaks, resulting in 479 stakeholder notifications. We further demonstrate the model's utility in verifying mitigation success through case studies in Libya, Argentina, Oman, and Azerbaijan. Our work represents a critical step towards a global AI-assisted methane leak detection system, which is required to process the dramatically higher data volumes expected from new and current imaging spectrometers.

Figures (12)

• Figure 1: Stratification of EMIT data by sectors and countries.
• Figure 2: Spatial distribution of samples in the created datasets and their division into train (red), validation (green, only for EMIT) and test (blue) subsets.
• Figure 3: Temporal distribution of samples in the EMIT dataset. Training, validation and test subsets are also shown, full tiles are downloaded for selected samples from the test subset. Note that for no plume tiles (with fluxrate equal to 0), we jitter these points around x-axis for better visualisation.
• Figure 4: Comparison of the explored methane enhancement products computed for a sample scene from the EMIT full tile test split. For easier comparison, all matched filter products were scaled to the same visualisation range (0 to 4000 ppm$\times$m units corresponding to the mixing ratio length). While some artifacts copy the structure of the real scene (e.g. rivers and mountain ranges) in all products, the Mag1c product seems to have most prominent confounders.
• Figure 5: HyperMARS model illustration. We use the U-Net model architecture with the MobileNetV3 encoder and a variety of input features further described in the text. Depending on the used configuration of input products, the model has around 6.69M trainable parameters.