A colored plume begins with missing light
Yes, satellites can detect methane releases. The useful qualification is that they measure a physical signal under particular conditions, rather than taking an ordinary photograph of visible gas. A bright plume on a map is a visualization of inferred methane, not its natural color.
NASA’s Jet Propulsion Laboratory describes the underlying measurement through a spectral fingerprint. Methane absorbs particular shortwave-infrared wavelengths. Sunlight travels through the atmosphere, reflects from the surface, and returns toward an instrument. Extra methane along that path weakens the returning light in a characteristic pattern.
An imaging spectrometer measures light at many wavelengths across neighboring areas. Analysts look for an enhancement relative to background and map its shape. That is why the ground matters as well as the air: the surface supplies the reflected light used for the measurement. JPL’s AVIRIS examples demonstrate the principle with airborne instruments; they explain the physics without being results from the satellite trial discussed below.
Detection existed before this test
NASA reported in 2022 that EMIT, an imaging spectrometer on the International Space Station originally designed for mineral mapping, had detected methane plumes at major emission sources. Its spectral measurements could distinguish methane’s absorption pattern, and its observations showed that an instrument with another original purpose could also locate gas emissions.
Those demonstrations answer whether detection is possible. They do not tell us how often every system succeeds, how many false alarms it produces, or how accurate each flow estimate is. A striking image can establish capability while leaving reliability under different conditions unresolved.
A fair evaluation therefore needs observations when methane is present and observations when it is absent. It also needs an answer known independently of the satellite’s own analysis.
Giving analysts a hidden answer sheet
A study published in Nature on 14 October 2026 compared satellite estimates with controlled releases in Arizona over a year. It included 481 methane releases and 265 blank controls, with 15 analysis teams. The published paper reports 12 satellites; Stanford’s accompanying news account describes 11. The trial counts and measured release range come from the paper, rather than treating the two satellite counts as interchangeable.
Measured rates ranged from 9.5 to 1,183.7 kilograms per hour. These are rates of release, not the total quantity emitted in each event. The single-blind design kept the answers from analysts, while the release operators necessarily knew what they had done.
An important boundary is that the approximate source location was known. Assessing a specified location is a different task from discovering an unknown source anywhere in a large landscape. The result should not be presented as a test of exhaustive worldwide leak finding.
Which systems performed well?
Stanford’s report describes clear-day detection rates of 92–100% for dedicated methane instruments, 62–88% for instruments repurposed for methane sensing, and 29–37% for general land scanners. The broad scanners collect less spectral information useful for isolating methane, although their coverage and archives remain valuable.
The Nature abstract also reports deterioration when ground-measured cloud cover exceeded 20%. That is a result under the experiment’s conditions, not a universal boundary at which every satellite suddenly stops working.
These percentages answer a conditional question about tested observations. They should not become a promise of continuous coverage, success through cloud, or detection of every smaller release. Nor should the instruments be treated as having identical capabilities simply because all produced images from orbit.
An unusable observation provides a particularly important distinction: no usable reading is not a measurement of zero emissions. A blank area on a map needs an explanation before it becomes evidence of clean air.
A plume is not a flow meter
Detection asks whether extra methane is present. Quantification asks how quickly the source is releasing it. The second task adds another inference.
Varon and colleagues’ 2018 research on fine-scale plume observations shows why wind and turbulence matter. They tested methods using simulated plumes and found that some common approaches were poorly suited to the small, turbulent shapes in instantaneous observations. Methods based on integrated plume mass or flux across plume sections performed better, but uncertainty remained, especially without local wind information.
Think of a photograph of smoke from a chimney. The material visible in the frame depends on both supply and transport. A stronger wind changes its shape and concentration even if the source is unchanged. Converting a methane map into kilograms per hour likewise requires more than counting colored pixels.
For purpose-built instruments, the 2026 paper reports a mean absolute percentage error of 44% in quantification. This does not mean 44% of plume detections were false. It describes the size of errors in estimating flow, with overestimates and underestimates counted as errors rather than allowed to cancel.
What the 79% AI figure means
Stanford reports that 79% of the trial’s false positives came from commercial teams using AI-based approaches on land-scanning satellite data. The denominator is all false positives in the study.
It is not the percentage of AI alerts that were wrong. Consider a hypothetical trial with 100 false alarms, 79 from one method. Suppose that method issued 1,000 alerts in total:
| Question | Calculation | Answer in this example |
|---|---|---|
| What share of all false alarms came from the method? | 79 ÷ 100 | 79% |
| What share of the method’s own alerts were false? | 79 ÷ 1,000 | 7.9% |
Both percentages can be true at once. The first denominator counts false alarms across methods; the second counts every alert issued by one method. Without the second denominator, the first percentage cannot tell us that method’s false-alert fraction.
The example is arithmetic, not an additional result from the trial. Keeping the denominator visible prevents a warning about tested methods from becoming an unsupported verdict on every use of AI.
An alert still needs follow-through
The study discloses support from Saudi Aramco and Stanford’s industry-funded Natural Gas Initiative, among other funders. It also discloses authors’ roles on Carbon Mapper’s Science and Measurement Committee. Those connections merit scrutiny; they do not by themselves establish that the measurements are wrong.
The practical value of a well-supported detection is that it gives investigators a location to check. Repeated observations and measurements closer to the source can help determine what is happening. Detecting an emission is still different from demonstrating that a repair reduced it.
The same discipline runs through the difference between correlation and causation: ask which question the evidence actually answers. Here, a strong detection result earns confidence in a measurement under tested conditions. It does not earn certainty about every location, every hour, or the success of a subsequent fix.
