The study combined GHGSat-organized campaigns with independent third-party experiments conducted from 2021 through 2026, providing a statistically grounded baseline for targeted observations of known point sources. The detection analysis included 125 release events with known outcomes: 94 were correctly detected and 31 were missed. Sixty-two samples came from single-blind campaigns, in which independent organizers controlled ground operations and reference measurements while GHGSat was kept unaware of the release rate and whether a release was present.

The other 63 were pseudo-blind releases organized with partners but processed using the same inspection and quantification procedures used in operational work. Experiments occurred at six controlled-release facilities with prairie, arid and more vegetated surface settings. No-release observations were also included, but the available false-positive and true-negative sample was too small to parameterize those outcomes in parallel with the true-positive and false-negative model.

The resulting performance estimate therefore describes the probability of detecting a real emission conditionally, rather than a marginal probability that depends on how many observed sites actually contain sources. GHGSat’s satellites measure methane using a wide-angle Fabry–Pérot interferometer that records spectrally resolved, backscattered sunlight in the shortwave infrared and targets methane absorption near 1.6 µm. In target mode, multiple partially overlapping camera frames are collected during a 20-second acquisition window while the satellite is maneuvered so that each ground point remains in view longer than it would under nadir-only pointing.

After downlink, a proprietary processing chain produces level-2 products containing resolved albedo, methane-column enhancement and uncertainty, followed by level-4 emission identification and quantification. Analysts distinguish a plume from noise and artifacts by assessing its general morphology, alignment with the wind and any relationship to features in the albedo layer. Once a plume is identified, its emission rate is estimated from the integrated mass enhancement, calculated by summing column-density enhancements over the visible plume mask, together with an effective transport speed and an effective plume length.

The researchers represented detection probabilistically rather than as a hard threshold. Their preferred model combined metered emission rate, wind speed, retrieval noise and effective spatial resolution in a dimensionless, signal-to-noise-like predictor, then used a log-normal inverse link to map that predictor to detection probability. The formulation was based on a mass-balance heuristic, but the authors replaced plain wind speed with a fitted effective wind to account for methane transport by diffusion and small-scale eddies even when the reported wind approaches zero.

Emission rate, wind and retrieval noise emerged as the dominant drivers of detection. Spatial resolution added little on its own, but improved model fit when included with the other three variables, consistent with its ability to determine how strongly a plume signal is diluted within a pixel. The fitted model required a signal-to-noise ratio of 1.72 ± 0.39 for a 50% probability of detection.

The threshold is conditional on observing conditions and therefore cannot serve as a universal performance figure. Under the same median controlled-release conditions, candidate model forms produced 50% detection limits ranging from 89.5 ± 6.1 to 102.8 ± 6.0 kg h−1, indicating that the principal result was reasonably stable to the chosen formula. The unmodified heuristic, which did not allow for effective transport at zero reported wind, fitted less well.

A likelihood-ratio test gave a statistic of 8.58 for the linear wind-offset model relative to that heuristic, corresponding to p = 0.003, so the improvement was greater than expected from adding one free parameter alone. The four tested inverse-link functions lay within 1.1 AICc units of one another, meaning the observations did not meaningfully discriminate among them. The researchers selected the linear wind-offset predictor with a log-normal link because the predictor is a product and ratio of positive uncertain quantities, for which a log-normal distribution is physically expected.

Bootstrap analysis used 1,000 resamples of the 125 detected and missed releases to estimate uncertainty in fitted parameters and derived thresholds. The 90% detection limit at the median test conditions was 139.3 ± 11.1 kg h−1. A time-resolved convergence analysis could be initiated only after the first missed detections appeared in mid-2022.

Following early fluctuations, the 50% threshold settled near its final value, and its uncertainty narrowed as more releases accumulated. The addition of a substantial set of single-blind Stanford samples in 2025 did not materially shift the 50% estimate and produced only a modest improvement in the 90% estimate. Quantification accuracy was evaluated separately in 92 true-positive samples because no rate estimate is produced for missed plumes.

Two additional detections were excluded from this analysis: their concentration retrievals had been corrupted by artifacts or high cirrus cloud, even though analysts could discern that an emission was present. With operationally reported model winds, the ordinary least-squares slope relating estimated to metered rates was 0.93 ± 0.03 and R² was 0.92. Substituting a co-located anemometer wind increased the slope to 0.96 ± 0.02 and R² to 0.95.

Adjusting the anemometer reading to a 10 m reference height under a logarithmic wind profile moved the slope to 0.98 ± 0.03, while R² eased slightly to 0.93. Log-scale parity plots revealed greater scatter and a tendency to underestimate small releases near the detection limit; estimates for larger releases were more tightly grouped and more symmetric around the one-to-one line. Two linked corrections addressed much of the low-rate underestimation.

A local background correction estimated the baseline methane concentration in an annulus around the plume, beginning more than three pixels and ending less than 13 pixels from the plume mask, and subtracted its mean from the column enhancements used to calculate integrated mass enhancement. This reduced the standard deviation of the logarithmic estimated-to-metered error ratio from 0.19 to 0.15 and raised its mean from −0.08 to −0.05. The second correction empirically recalibrated the effective wind speed used in the quantification equation.

Combined background and wind corrections raised the mean logarithmic error ratio to +0.03 while preserving the reduced scatter of 0.15. In this dataset, the combined procedure recovered previously underestimated low-rate sources without degrading performance at higher rates. A sensitivity analysis found that varying the width of the background annulus and its buffer distance changed the global calibration slope by only about ±2% over a broad range of settings.

Quantification proved comparatively insensitive to the choice among operationally available wind products. Rates calculated with ERA5, the ECMWF Integrated Forecasting System, the NOAA High-Resolution Rapid Refresh and the OpenWeather aggregator produced slopes through the origin of 0.91–0.93, with R² values from 0.89 to 0.93. Across the full dataset, geometric mean estimated-to-metered ratios for the six wind sources ranged from 0.64 to 0.82 before correction, indicating overall underestimation.

The height-adjusted anemometer was least biased at 0.82, followed by the as-reported baseline at 0.81, while ERA5 had the largest negative bias at 0.64. The authors associated this pattern with the spatial and temporal averaging of sub-grid wind variability: the finer 3 km HRRR grid was less biased than the 9 km IFS grid, while the approximately 30 km ERA5 grid was most biased. Applying the fitted detection model to the constellation’s global observing catalogue gave a median 50% detection limit of 159.7 kg h−1, about 1.6 times the controlled-release value, because median global winds and retrieval noise were less favorable.

Several limitations define the study’s scope. Most releases occurred at flat, arid or prairie sites, including settings representative of many oil and gas production basins but not necessarily landfills, coal mines or urban environments with different topography, land cover and meteorology. All releases were point sources, so the findings cannot simply be transferred to extended area sources such as open-pit mines or broad landfill surfaces.

The sample also was insufficient to fit separate detection and quantification models for each instrument payload; the reported results characterize the constellation, while instrument-specific noise and resolution are included as observation-level predictors. Finally, controlled releases primarily characterize target-mode processing, in which analysts had approximate knowledge of potential source locations. Survey-mode observations of unknown sources involve different inspection conditions and cannot be described by the same threshold.

The results therefore establish a transparent, conditions-dependent baseline for targeted point-source detection and quantification, rather than a single operating limit applicable to every global observation or emission type.

Subject of Research: Controlled-release validation of GHGSat methane satellite detection and emission quantification

Article Title: Performance validation of the GHGSat methane constellation via controlled releases

Article References: Ramier, A., Deglint, H., Deslieres, A., Formenti, L., Girard, M., Jervis, D., Marshall, D. J., MacLean, J.-P. W., McKeever, J., McKell, A., Strupler, M., & Tarrant, E. (2026). Performance validation of the GHGSat methane constellation via controlled releases. Atmospheric Measurement Techniques, 19(20), 6643-6660. https://doi.org/10.5194/amt-19-6643-2026

Image Credits: AI Generated

DOI: 10.5194/amt-19-6643-2026

Keywords: GHGSat, methane satellites, controlled releases, probability of detection, emission quantification, remote sensing, retrieval noise, wind speed, spatial resolution, atmospheric methane, Performance, validation