Methane is a short-lived but very potent greenhouse gas, and a large share of oil and gas, coal and landfill emissions comes from a small number of unintended leaks and venting events. Satellite spectrometers see methane absorption in the shortwave infrared, and ML models turn those spectra into plume masks, emission rates and facility-level attributions. The result is independent, repeatable evidence for leak-repair campaigns, regulatory reporting and import verification that is otherwise only available from sparse ground surveys.
Problem & value
- Today: operators rely on periodic handheld or vehicle leak detection (LDAR), bottom-up emission factors and self-reporting. Large, intermittent releases ("super-emitters") are easily missed, and importers, regulators and investors have no independent check.
- AI value: satellites cover remote assets continuously. Deep-learning segmentation and physics-based retrieval find plumes in imagery that humans cannot inspect at scale, quantify them and link them to a named facility. This lets responders prioritise repairs.
- Regulatory pull in the EU: Regulation (EU) 2024/1787 on methane emissions reduction in the energy sector requires measurement, reporting and verification (MRV) and leak detection and repair for EU oil, gas and coal operations. Importers must show that supplying countries and companies apply MRV equivalent to EU rules, with equivalence evidence due from 2027 onward. The Commission also uses satellite data to identify large emission events and to notify Member States, who must make operators investigate and report [1][2].
- Global alerting: UNEP's International Methane Emissions Observatory (IMEO) runs the Methane Alert and Response System (MARS), which turns satellite detections into notifications to governments and operators [3].
- Reduction potential figures are published by the IEA and UNEP; use their numbers rather than vendor claims when building a business case.
Variants
- Oil and gas operator: screen own assets, validate LDAR, find unreported events.
- Importer / gas buyer: verify supplier claims under the EU import MRV rules (LNG, pipeline gas, crude).
- Coal mines: ventilation-shaft and drainage-station emissions.
- Landfills and waste sites: plume detection at sites with persistent point-source emissions.
- Regulator / public authority: national follow-up of super-emitter notifications; inspection targeting.
- Investors and ESG analysts: portfolio-level emission screening, see DATA-012 and LANG-037.
- Other greenhouse gases: CO2 point sources (power plants, cement) from missions such as OCO-3, PRISMA and the upcoming Copernicus CO2M, plus regional inversion for national inventories.
- EnMAP-based mapping: the German EnMAP hyperspectral mission is a research-grade EU data source for methane retrievals.
How it works
- satellite scene (SWIR hyperspectral or Sentinel-2/Landsat bands)
- radiometric prep, cloud/glint/surface masking
- methane enhancement retrieval (matched filter or ΔXCH4 from spectra)
- ML plume detection/segmentation (CNN/U-Net/ViT)
- false-positive rejection
- wind lookup (ERA5 / forecast)
- emission rate by integrated mass enhancement or cross-sectional flux
- facility attribution (asset registry geometry, plume origin)
- expert review of candidate plumes
- alert / notification / MRV record
- Retrieval: the matched-filter family (and iterative physics-based retrievals) estimates a per-pixel methane enhancement from the SWIR absorption features around 2300 nm. Sentinel-2 SWIR bands give coarse spectral information but wide coverage and frequent revisits.
- Learned detection: segmentation networks trained on labelled enhancement maps (for example the STARCOP work on hyperspectral data and several Sentinel-2 plume papers) suppress false positives caused by surface artefacts and find weaker plumes [4][5].
- Quantification: requires wind speed and a plume-mask-based mass estimate (integrated mass enhancement, IME, or cross-sectional flux). Wind uncertainty often dominates the error.
- Human-in-the-loop: analysts verify candidate plumes before any alert goes to an operator or government, because a wrong attribution has contractual and reputational consequences.
Data preparation
The core data are satellite scenes (EnMAP, EMIT, Sentinel-2, TROPOMI), which are public or licensed and need no cleaning by the user. The hard part is the operator side: asset registers, facility coordinates, maintenance and LDAR records. These typically live in GIS, an asset-management system (SAP PM or similar), spreadsheets and PDF inspection reports on file shares.
Typical gaps are wrong or outdated facility coordinates, assets without unique IDs, leak-repair records only as scanned paper or free text, inconsistent units (kg/h versus t/year) and wind or time stamps without a common time zone. Without a clean asset register, a detected plume cannot be attributed to a responsible operator.
Preparation steps: (1) build a geo-referenced facility register with stable IDs and polygons (QGIS, PostGIS, OSM as cross-check); (2) extract repair and inspection history from the maintenance system and digitise paper reports, see PREP-005 and PREP-006; (3) harmonise units and time stamps in tidy tables, see PREP-011; (4) define an owner for each data source, see PREP-004; (5) collect a small validation set of confirmed leaks with domain experts, see PREP-018. Plan roughly half of the effort for data work, because the models and satellite feeds are bought or open. Run the data readiness check in GUIDE-003 first.
Implementation paths
Which path for whom
- Enterprise: Path A, because mature emission-intelligence providers deliver validated plume data and attribution as a service, and verification evidence must be defensible to auditors and regulators.
- Developer: Path C, because open data (Sentinel-2, EMIT, TROPOMI) and open retrieval code allow building and validating own pipelines on GPU hardware.
- Public: Path C, because authorities benefit from open data, reproducible methods and sovereign EU hosting (e.g. Copernicus infrastructure) instead of dependence on one commercial provider.
Path A — Buy (recommended)
- Commercial emission-intelligence providers: GHGSat (Canada; own methane satellites with facility-level resolution), Kayrros (France, EU; methane monitoring from multi-source satellite data), Orbio Earth (Germany, EU; Sentinel-2-based methane monitoring), Carbon Mapper (non-profit; Tanager satellites and public data portal), Planet and Kuva Space (Finland, EU; hyperspectral data). Several offer API feeds and importer-oriented reports.
- Public alert services: UNEP IMEO MARS notifications for governments and eligible operators [3].
- Choose A when: you need credible, audit-ready results fast. Check data-licensing terms and which missions the vendor actually uses, as detection limits differ strongly.
Path B — Build on cloud APIs
- Data access: Copernicus Data Space Ecosystem (EU) for Sentinel-2 and Sentinel-5P, NASA EMIT and Carbon Mapper open data portals, EnMAP data via the German Aerospace Center (DLR) EOWEB, Google Earth Engine for exploration, ERA5 wind from the Copernicus Climate Data Store.
- Compute: EU-region cloud GPUs; orchestrate with STAC catalogues and workflow tools (Airflow, n8n).
- Choose B when: you want your own screening layer and use vendor data only for confirmation.
Path C — Self-host (open source)
- Stack: Python with rasterio, xarray, STAC clients (pystac-client), PyTorch segmentation models, open matched-filter implementations from NASA JPL and the EMIT team, and the open STARCOP and similar research code [4].
- Data: open Sentinel-2, EMIT and EnMAP scenes. Use ERA5 for wind.
- Hardware: HW3 for training and batch inference over large areas (scene tiles are large). Inference on single scenes fits a smaller GPU.
- Effort: needs remote-sensing and atmospheric-physics expertise.
Path D — Custom training
Worth it when your assets or regions differ from public benchmarks (for example particular land cover, off-nadir geometry, or a new sensor). Needs labelled enhancement maps with confirmed plumes and confirmed non-plumes, ideally with controlled-release or aircraft validation data. Synthetic plume injection into real scenes is standard for augmentation.
State of the art (2026)
- Missions: GHGSat constellation, NASA EMIT on the ISS, EnMAP, PRISMA, Carbon Mapper Tanager satellites, and Sentinel-2/Landsat for broad coverage. MethaneSAT (EDF) launched in 2024 for area-wide mapping, but the mission was reported lost in 2025; check current status before relying on its archive [6]. Sentinel-5P TROPOMI provides regional screening at coarse resolution.
- Models: deep-learning detectors on matched-filter output outperform thresholding on hyperspectral data, and Sentinel-2 detection of large plumes is now routine in research and commercial services [4][5]. Foundation models for Earth observation are used as feature extractors, but physics-based retrievals remain the basis of quantification.
- Controlled-release tests by independent groups are the accepted way to check detection limits and quantification error; request such results from any vendor.
- Limits: detection thresholds vary by sensor (coarse sensors find only large emitters), and bright or heterogeneous surfaces, clouds, sun glint, snow and water cause false positives or blind spots. Offshore and high-latitude sites are hard. Single overpasses capture only a snapshot of intermittent sources, so annual totals cannot be extrapolated naively. Wind error dominates rate uncertainty. Satellites complement, not replace, ground and aircraft measurement.
Shared resources & combinations
- Same remote-sensing stack: VIS-021 land use and change detection, VIS-038 public-authority compliance checks, VIS-049 EUDR and forest carbon MRV, VIS-061 InSAR ground motion — shared STAC pipelines, tiling, GPU inference.
- Combinations: ANOM-013 (in-network leak detection) gives ground-truth for pipelines; DATA-012 (CSRD reporting) and DATA-021 (product carbon footprint) consume the emission estimates; SCI-010 (climate and earth-system modelling) uses them for inversions.
- Hardware: HW3 shared with other remote-sensing batch workloads.
Effort, cost & KPIs
- Path A: weeks to onboard; subscription or per-site/per-report pricing.
- Path B: 3–6 months for a screening pipeline plus validation.
- Path C/D: 6–12 months with specialist staff; labelled data is the bottleneck.
- Cost drivers: commercial high-resolution tasking, GPU time for large areas, expert review.
- KPIs: detection recall and false-positive rate against controlled releases, emission-rate error, time from overpass to alert, share of alerts confirmed on the ground, repair lead time, coverage and revisit rate of assets.
People & change
Start with management and the department lead for operations or compliance, who define the goal (regulatory reporting, import verification or leak repair) and the budget. Involve the data owners of the asset register and maintenance system early, because attribution depends on their data. IT must integrate alerts into the ticketing or maintenance workflow; legal and compliance check reporting duties and contract terms with data providers, and suppliers or partner operators should be informed when their sites are monitored.
Works council: normally not required, because satellite data concern installations, not employees, so §87(1) no. 6 BetrVG is not triggered. It becomes relevant if alerts are used to evaluate individual staff; then involve it, and inform it under §90(1) no. 3 when work processes change. No personal data, so DPO and DPIA are usually not needed, but confirm this when site cameras or staff data are linked.
AI literacy: Art. 4 AI Act requires measures that support staff literacy; a short briefing for analysts and field crews on detection limits (clouds, bright surfaces, detection thresholds) is enough, see GUIDE-006. Pick one asset class or region as pilot, see GUIDE-009. Typical concern: false alarms and unnecessary site visits; address them with confidence scores and human review before any dispatch or regulatory report. See also GUIDE-008.
Risks, regulation & compliance
- EU AI Act: minimal risk. Analysis of non-personal environmental data; Annex III does not apply. Operators of safety-relevant gas infrastructure should still keep AI out of safety-critical control loops. AI literacy (Art. 4) applies.
- Methane Regulation (EU) 2024/1787: defines operator LDAR, MRV and reporting duties and import-equivalence rules. Satellite evidence supports but does not replace the prescribed measurement methods [1].
- GDPR: generally not relevant (no persons at these resolutions). Check for personal data if landfill or facility imagery is combined with owner data.
- Data licences and export control: commercial imagery licences limit redistribution; very high-resolution data may be subject to national rules.
- Reputational and legal risk: a false attribution to a named operator can lead to disputes. Mitigate with analyst sign-off, documented uncertainty, wind provenance and archived scenes.
Getting started
- Define the question (own assets, supplier verification, regulator screening) and the minimum emission rate you must detect.
- Pull 12 months of open Sentinel-2/EMIT/EnMAP data for a pilot set of assets and check what is visible.
- Ask two vendors (Path A) for controlled-release results and run a blind test on known events.
- Set up the review workflow, facility registry and uncertainty reporting.
- Link alerts to LDAR work orders and to CSRD or importer MRV documentation.
Sources
- Regulation (EU) 2024/1787 on methane emissions reduction in the energy sector — https://eur-lex.europa.eu/eli/reg/2024/1787/oj
- European Commission — Methane emissions in the energy sector (regulation overview) — https://energy.ec.europa.eu/topics/oil-gas-and-coal/methane-emissions_en
- UNEP International Methane Emissions Observatory — Methane Alert and Response System — https://www.unep.org/topics/energy/methane/international-methane-emissions-observatory/methane-alert-and-response-system
- Ruzicka et al. — Semantic segmentation of methane plumes with hyperspectral machine learning models (STARCOP), Scientific Reports 2023 — https://www.nature.com/articles/s41598-023-44918-6
- Varon et al. — Satellite-based quantification of methane point sources with Sentinel-2, Atmospheric Measurement Techniques 2021 — https://amt.copernicus.org/articles/14/2771/2021/
- MethaneSAT (Environmental Defense Fund) — https://www.methanesat.org/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Copernicus Data Space Ecosystem — https://dataspace.copernicus.eu/