A decade ago, atmospheric methane was monitored mostly through ground-based sampling networks and aircraft campaigns — sparse, expensive, and incapable of attributing emissions to specific facilities. Today, a small number of satellite-based services using machine learning to interpret remote-sensing data are detecting facility-level methane plumes from orbit. The implications for compliance, finance, and corporate reporting are substantial, but the technology is more uneven than the marketing suggests.
The relevant question is no longer whether satellites can detect emissions. They can. The questions are: what can they detect, what can't they detect, and how reliable are the detections that get reported.
The capability landscape
Several distinct services now operate at meaningful scale, each with different strengths.
GHGSat operates a constellation of small satellites with high spatial resolution but limited revisit frequency. Its strength is facility-level attribution — resolving plumes to specific wellheads, compressor stations, or storage tanks. Its limitation is that it sees each site infrequently; a plume that emerges and dissipates between passes goes undetected.
Kayrros aggregates data from public satellite missions (Sentinel-5P, others) and applies ML to interpret atmospheric column measurements. The advantage is high-frequency global coverage; the constraint is that the underlying data has lower spatial resolution, which limits attribution precision for smaller emitters.
Climate TRACE takes a broader approach, integrating satellite data with non-satellite signals (shipping data, power plant outputs, agricultural land use) to produce facility-level emissions inventories across many gases and sectors, not just methane. Its outputs are best understood as estimates with uncertainty bounds rather than point measurements.
MethaneSAT, an Environmental Defense Fund initiative, was designed specifically for systematic methane detection from oil and gas operations. It pairs higher spatial resolution than the public-satellite-based services with broader coverage than the small-constellation services.
These services are not strictly competitive. They occupy different points on the spatial-resolution / temporal-frequency / coverage tradeoff. Sophisticated users combine multiple services to compensate for individual blind spots.
Where ML adds value, and where it adds noise
Machine learning enters the workflow at two distinct points: signal extraction from raw satellite data, and attribution of detected plumes to source facilities.
Signal extraction is where ML has produced the clearest gains. Identifying weak atmospheric concentration anomalies against noisy backgrounds is the kind of pattern recognition problem ML excels at. Manual interpretation of methane absorption spectra at scale isn't feasible; ML pipelines have made it routine.
Attribution is harder, and where false positives matter. Plume detection alone doesn't identify a source — wind patterns, terrain, and overlapping facilities complicate attribution. ML models trained on labelled plumes with known sources can attribute reasonably well in clear cases, less well when multiple potential sources are co-located, and poorly when the source category isn't well-represented in training data.
The honest assessment: published detection thresholds need to be read alongside published false positive rates, and the false positive rates aren't always disclosed with the same prominence.
Detection thresholds in practice
Public benchmarks for methane detection from satellite generally describe minimum detectable emission rates under favorable conditions. Specialized small-constellation satellites can detect plumes at rates that would have been undetectable a decade ago — on the order of single-digit metric tons per hour under good viewing conditions for some platforms.
Two caveats matter for users.
First, “favorable conditions” isn't most conditions. Cloud cover, low solar angles, certain surface types (snow, water, urban), and atmospheric turbulence all degrade detection. A real-world detection threshold averaged across observation conditions is higher than the headline best-case figure.
Second, what's detectable from orbit isn't the same as what's relevant for emissions accounting. Many emissions sources sit below the detection threshold of any current satellite — smaller leaks aggregate to large totals but don't trigger plume-level detection. This matters for inventory completeness: satellites detect the largest emitters well, the long tail less well.
How the data is being used
Three categories of users are integrating satellite emissions monitoring into operational workflows.
Operators and compliance teams use the data to verify their own facility performance and detect anomalies between scheduled inspections. This is the most operationally direct use: known emissions detected by a third party often surface as operational issues internally first, allowing faster response.
Investors and lenders use the data as a verification check on reported emissions. Discrepancies between corporate disclosures and satellite-derived estimates create due diligence questions, particularly for upstream oil and gas, where methane intensity is a key climate-risk metric.
Regulators and NGOs use it for enforcement and accountability. EPA's methane Super-Emitter Program references satellite-detected plumes. NGO campaigns increasingly cite specific facility-level detections rather than aggregate inventories.
The accountability vector is the one with the highest stakes for emitters. A detection published with a named operator and a coordinate is materially different from an inventory statistic.
What's still genuinely uncertain
The space is improving rapidly enough that confident long-term claims are unwise. Three areas remain genuinely uncertain:
Attribution under co-located emissions. Dense oil-and-gas operations or landfill complexes with adjacent facilities still produce attribution ambiguities. Improvements in spatial resolution help but don't fully solve this.
Non-methane gases. Current capabilities are strongest for methane because of its specific absorption signature. Detection of CO2, N2O, and other gases at facility-level resolution is less developed and remains an active research area.
Verification of satellite-derived estimates against ground truth. Comprehensive ground-truth validation is expensive and limited. Validation campaigns to date generally support the headline accuracy claims of major services, but with caveats and conditions that don't always translate cleanly into operational confidence intervals.
The implication for buyers of emissions data
For investors, lenders, corporates, and regulators evaluating satellite-derived emissions data, the right framing is augmentation rather than replacement. Satellite monitoring provides an independent verification layer that operator-reported data cannot replicate, but it doesn't fully substitute for facility-level direct measurement.
Used together — satellite for systematic anomaly detection, ground measurement for high-precision quantification — the two methodologies cover each other's blind spots. Either alone produces an incomplete picture.
The space is one of the fastest-improving areas in environmental data infrastructure. Capabilities that didn't exist five years ago are now operational. Buyers should plan for the technology to keep improving and the disclosure expectations that follow it to keep tightening.