Artificial intelligence for methane detection: from continuous monitoring to verified mitigation

TL;DR

The paper presents MARS-S2L, a large-scale, operational methane plume detector for public Sentinel-2 and Landsat imagery, leveraging a 16-channel UNet augmented with cloud and wind data and physics-based plume simulations. Trained on a global, expert-validated dataset of 87,929 images spanning 1,315 emitter sites, the model achieves high recall with low false positives and is deployed in near-real-time with analyst-driven verification and government/operator notifications. Over 16 months, it identified over a thousand emissions, triggering mitigations across multiple countries and demonstrating real-world impact, including the cessation of persistent emitters. The work also outlines ongoing improvements (hyperspectral sensors, new platforms, and broader GHG monitoring) and provides a public data portal and code to enable reproducibility and scale.

Abstract

Methane is a potent greenhouse gas, responsible for roughly 30\% of warming since pre-industrial times. A small number of large point sources account for a disproportionate share of emissions, creating an opportunity for substantial reductions by targeting relatively few sites. Detection and attribution of large emissions at scale for notification to asset owners remains challenging. Here, we introduce MARS-S2L, a machine learning model that detects methane emissions in publicly available multispectral satellite imagery. Trained on a manually curated dataset of over 80,000 images, the model provides high-resolution detections every two days, enabling facility-level attribution and identifying 78\% of plumes with an 8\% false positive rate at 697 previously unseen sites. Deployed operationally, MARS-S2L has issued 1,015 notifications to stakeholders in 20 countries, enabling verified, permanent mitigation of six persistent emitters, including a previously unknown site in Libya. These results demonstrate a scalable pathway from satellite detection to quantifiable methane mitigation.

Figures (32)

• Figure 1: Characteristics of the MARS-S2L dataset We compile a large dataset of emissions for training and evaluating machine learning models. In total, 87,929 images containing 5,534 emissions over 1,315 distinct emitters globally are included in the dataset. Top emitter locations and for the test (a) and training (b) sets. Bottom Time series of monthly number of images, plumes, and distinct locations. Shaded areas indicate the split of the dataset in train, validation, and test subsets.
• Figure 2: MARS-S2L enables large scale processing of Sentinel-2 and Landsat images. Top: Images required to be reviewed to identify different amount of plumes when ordered by model probability. This indicates that incorporating AI enables analysts to review around 22 times fewer images, making the monitoring feasible to be tackled by a small team. Bottom: False positive rate and cumulative recall of the MARS-S2L model compared to CH4Net and MBMP.
• Figure 3: Multi-regional evaluation of model performance. False positive rate (FPR) and cumulative recall of the MARS-S2L model compared with multi-band multi-pass (MBMP) thresholding and CH4Net. Results are stratified by flux rate over the 12 selected geographical areas. MARS-S2L provides skilful identification of emissions showing a false positive rate under 10% and recall around 80% for all plumes and up to 90% for plumes of more than 5 t/h.
• Figure 4: Model notifications during sixteen months of operational deployment. (a) Map with MARS-S2L notified detections in red. During this period, MARS-S2L detected 1,015 emissions across 20 countries that triggered a formal notification to governments and operators. Examples of some of these notifications are shown for events in Mexico, Turkmenistan, Argentina, the US and Uzbekistan (b). All plume detections are validated by IMEO analysts prior to notification.
• Figure 5: MARS-S2L predictions for a mitigated emitter in Hassi Messaoud, Algeria A timeseries of emissions with quantification is shown in (a), showing the mitigation of this source on the 14th of October 2024. Examples of MARS-S2L predictions compared to expert-annotated ground-truth are shown in (b). MARS-S2L successfully detects emissions from this source until its cessation.