Sensitivity Buys Sources, Revisit Buys Mass: What a Methane Detection Limit Is Actually For

An operator asks a methane monitoring vendor one question first, and it is almost always the same question: what is your detection limit. The answer arrives in kilograms per hour, it is small, and it is the number the vendor put on the datasheet because it is the number that gets asked for. Everything after that is a negotiation about how small.

That conversation has the wrong first term in it. A detection limit is a property of the instrument. What a monitoring programme is bought to deliver is a share of an emissions budget, and the share depends on the population of sources at least as much as on the instrument. For methane point sources the population is known, it is published, and its shape says that a factor of twenty in sensitivity is worth roughly one extra overpass.

The procurement conversation is real, and the majors are all having it

This is not a hypothetical about future tenders. TotalEnergies has been flying a drone-mounted sensor over its own upstream estate since 2022. The company describes AUSEA as "a miniature dual sensor mounted on a drone, capable of detecting methane and carbon dioxide emissions, while at the same time identifying their source", tested at sites in Nigeria, Italy, the Republic of the Congo and the Netherlands before being extended to all upstream oil and gas sites it operates [1]. Its own technology page states that "All of the Company's upstream sites are currently subject to an AUSEA detection campaign at least once a year", and that in 2025 the programme reached approximately 700 mission days across operated, non-operated and third-party assets [2]. Separately, the company said at COP29 that it would install continuous, real-time detection equipment on all its operated upstream assets, with the plan fully implemented by the end of 2025 [3].

Eni is on the reporting side of the same problem. It joined the UN Environment Programme's Oil and Gas Methane Partnership 2.0 in 2020, received the Gold Standard Pathway award in 2023, achieved Gold Standard reporting in 2024 by reaching the highest levels of data quality, and was recognised again in the 2025 report for identifying and quantifying emissions across non-operated assets and for delivering leak detection and repair training to national oil company personnel with UNEP support [4].

They are not alone in the partnership. The OGMP 2.0 member list dated 20 February 2026 carries, in its upstream segment, ADNOC, Bp, Eni, Equinor, Repsol, Shell and TotalEnergies, among others [5]. Seven operators, one reporting framework, and a shared measurement problem.

The measurement supply side has moved in the same period. The IEA's Global Methane Tracker 2025 reports more than 25 satellites in orbit providing methane insights, with MethaneSAT and Tanager-1 becoming operational in 2024, against energy-related methane emissions still above 120 million tonnes annually and around 200 billion cubic metres released by the fossil fuel sector in 2024 [6]. So a buyer now has a genuine choice of instruments at very different sensitivities, flown at very different frequencies, and has to rank them.

The population that decides the answer

The best public description of what those instruments are pointed at is the California Methane Survey. Duren and colleagues surveyed more than 272,000 infrastructure elements with an airborne imaging spectrometer across five campaigns from 2016 to 2018 and detected, geolocated and quantified emissions from 564 strong methane point sources, estimating net California point-source emissions at 0.618 Tg per year with a 95% confidence interval of 0.523 to 0.725, of which the oil and gas sector accounts for 26% [7].

Three things that paper prints decide the procurement question, and none of them is a detection limit.

The first is the heavy tail. The caption to Figure 2a states that 10% of the point sources are responsible for 60% of the detected point-source emissions [7]. The second is where the mass sits in absolute terms: the survey and the EPA Greenhouse Gas Reporting Program agree that 99% of point-source emissions come from facilities that emit at least 25 kg per hour [7]. The third is time. Many sources are highly intermittent, with a median persistence of 0.20 for the whole population, a mean of 0.33 and a range of 0.02 to 1.0 [7]. A source at the median persistence is emitting on one overflight in five.

The paper's own detection-limit figures come from the same Figure 2 caption: typical limits for this class of infrared imaging spectrometer range "from 2-10 kg CH4 h-1 for the typical 3-km flight altitudes used in this study to 100 kg CH4 h-1 for an equivalent satellite in low Earth orbit" [7]. That is a factor of ten to fifty between the airborne and the orbital case, from one instrument family, stated by the same authors. It is the cleanest sensitivity comparison available, and it is the one worth doing the arithmetic on.

Two printed numbers are enough to pin the distribution

You do not need to guess a source-size distribution. Two of the figures above pin a lognormal completely, with no free parameters left over.

Write the source rate as Q, lognormal with log-scale location mu and log-scale width sigma. The quantity a monitoring programme is paid for is not the count of sources above a threshold but the mass they carry, and the mass-weighted version of a lognormal is another lognormal with the same width and its location shifted by sigma^2. So the share of emitted mass carried by sources above a rate T is

$$M(T) \;=\; 1 - \Phi\!\left(\frac{\ln T - \mu - \sigma^{2}}{\sigma}\right)$$

FormulaLaTeX

Copy LaTeX

while the share of the source population above the same rate is the plain tail,

$$S(T) \;=\; 1 - \Phi\!\left(\frac{\ln T - \mu}{\sigma}\right)$$

FormulaLaTeX

Copy LaTeX

Now apply the first anchor. The top decile of sources by rate sits above mu + sigma * Phi^{-1}(0.90), and the mass above that point is Phi(sigma + Phi^{-1}(0.10)). Setting that to 0.60 fixes the width on its own:

$$\Phi\!\left(\sigma + \Phi^{-1}(0.10)\right) = 0.60 \;\;\Longrightarrow\;\; \sigma = \Phi^{-1}(0.60) - \Phi^{-1}(0.10) = 1.5349$$

FormulaLaTeX

Copy LaTeX

Apply the second. If 99% of the mass sits above 25 kg per hour, then M(25) = 0.99, and with sigma already known that fixes the location:

$$\Phi\!\left(\frac{\mu + \sigma^{2} - \ln 25}{\sigma}\right) = 0.99 \;\;\Longrightarrow\;\; \mu = \ln 25 + 2.3263\,\sigma - \sigma^{2} = 4.4337$$

FormulaLaTeX

Copy LaTeX

A median source rate of exp(4.4337), about 84 kg per hour, and a log-scale width of 1.53. Both anchors are recovered by numerically integrating the fitted density in the unit test that ships with the exhibit below, by a route that never touches the shifted-location identity used to solve for them, so the fit cannot certify itself.

What twenty times the sensitivity actually buys

Now run the two published detection limits through it. Take 5 kg per hour as the middle of the airborne band and 100 kg per hour as the orbital figure, a factor of twenty apart.

At 5 kg per hour the instrument sees 96.7% of the source population and 99.96% of the mass. At 100 kg per hour it sees 45.6% of the population and 92.3% of the mass.

Read those four numbers next to each other. Twenty times the sensitivity buys 7.7 percentage points of mass. On source count it more than doubles the detected population, 96.7 against 45.6%, so 52.9% of what the fine instrument finds is invisible to the coarse one. The same instrument change is worth almost nothing on one axis and more than half of everything on the other, and which axis a buyer is on is a question about the contract, not about the sensor. This is the quantitative form of the paper's own remark that detection limits could be relaxed by a factor of ten and still identify 90% of super-emitters [7].

If the deliverable is a source register, a repair work list, a per-site attribution, then sensitivity is the binding term and there is no substitute for it. If the deliverable is a reconciled emissions inventory, a reported total, an OGMP 2.0 Level 5 number, then sensitivity is nearly exhausted well above the airborne limit and the money is better spent elsewhere.

Persistence is a factor of five, and it is bought back geometrically

Elsewhere is revisit, and persistence is why.

At the reported median persistence of 0.20, a source is emitting on one overflight in five. A single pass therefore attributes at most one fifth of the annual budget however sensitive the instrument is. That is the whole of the airborne advantage gone, and more: 99.96% of mass in view becomes 20.0% of the annual budget on one pass.

Treat successive passes as independent trials at the same persistence, which is a modelling choice and not a result in the paper. A source above the threshold is then caught at least once with probability 1 - (1-p)^N. Because the annual mass of a source is its active rate times its persistence, the persistence factor cancels between the detected mass and the total, and what is left is

$$D(T, N, p) \;=\; \left[1 - (1-p)^{N}\right] M(T)$$

FormulaLaTeX

Copy LaTeX

Here D is the share of the annual point-source mass budget attributable to sources the survey saw at least once. Set the two instruments against each other. One airborne pass at 5 kg per hour delivers p * M(5), which is 20.0% at the median persistence. The satellite at 100 kg per hour delivers 18.5% on its first pass and 33.2% on its second. It loses on one pass and wins on two. The exact crossing is where

$$N^{\ast} \;=\; \frac{\ln\!\left(1 - \dfrac{p\,M(5)}{M(100)}\right)}{\ln(1-p)} \;=\; 1.09$$

FormulaLaTeX

Copy LaTeX

One point zero nine passes. That is the entire exchange rate. Twenty times the sensitivity, priced in revisits, is worth about one extra overpass at the median persistence of this population.

The exhibit

The exhibit opens on the state that makes the argument. The threshold lever is at 100 kg per hour, the revisit lever is at one, the persistence lever is on the paper's reported median of 0.20, and the amber readout says 1.09.

Drag the revisit lever from 1 to 2 and watch the aqua ring pass the white dashed rule. That rule is what a single airborne pass at 5 kg per hour delivers, and the amber disc sits where the coarser instrument's curve crosses it. Twenty times worse on the datasheet, ahead by the second pass.

Then drag the threshold lever and watch the two readouts on the left move at completely different rates. Between 1 and 100 kg per hour the mass readout falls from 100.00% to 92.3% while the source readout falls from 99.8% to 45.6%. The rug of sampled sources under the density is the same fact drawn as objects: dots go dim in large numbers long before the mass number moves.

The persistence lever is where the argument can be broken, and the exhibit lets a reader break it. Push persistence to 1.00, where every source is emitting on every pass, and the break-even readout says "none" and the amber disc disappears. The disc and that readout are one decision, so the disc is on the plot exactly when the readout prints a number. At full persistence a single airborne pass already collects 99.96% of the budget, the coarser instrument tops out at 92.3%, and no number of revisits closes a gap that revisits do not address. Revisit buys back what intermittency takes away, and when intermittency takes nothing away there is nothing to buy back. That is the boundary of the claim, and it is worth knowing where it is before quoting the 1.09.

Where a peer will push back, correctly

Three objections, all of them fair, and one of them serious.

The first is the independence assumption. Persistence in the paper is a measured detection frequency across revisits, not a stationary Bernoulli parameter, and real intermittency is correlated in time: a compressor station with a leaking bypass valve is not resampling a coin every overflight. Correlated activity makes revisits worth less than the geometric expression says, which pushes the break-even up. How far up is not something this fit can tell you.

The second is the anchor at 25 kg per hour. The paper states that 99% of point-source emissions come from facilities emitting at least that rate, and a facility can hold more than one source: the survey quantified 564 distinct sources at 250 facilities [7]. Applying a facility-level anchor at source level puts the location slightly too high, which slightly overstates how much of the mass sits at large rates and therefore slightly flatters the coarse instrument. The direction of that error is known even though its size is not.

The third is the serious one. The whole calculation assumes persistence is independent of source size, and the paper's own data says it is not. The largest methane emitters in California are a subset of landfills, which the abstract describes as exhibiting persistent anomalous activity [7]. If the biggest sources are also the most persistent, then the mass a single pass sees is larger than p * M(T) and the airborne single-pass figure of 20.0% is too low. That correction favours the sensitive instrument and cuts against the finding. It is the first thing to test with an operator's own revisit history, and the exhibit deliberately does not model it, because modelling it would require a size-persistence joint distribution that this paper does not publish.

None of the three changes the shape of the argument, which is that sensitivity and revisit buy different quantities. All three change the number.

The same lesson in a different vocabulary, from our own delivery

We have never flown a methane survey. Every figure above about methane is either transcribed from a public source or arithmetic on a distribution fitted to two figures that source prints.

What we have done, from our own delivery experience, is spend a roughly twenty-month subsurface-AI engagement with a mid-sized Middle East carbonate operator learning the same lesson in image space, twice, the second time by getting it wrong first.

The first version is about thresholds and objects. On fracture picking we applied a morphological path-opening operator across 39 path-length thresholds, and the whole point of the sweep was that the threshold is a monotone control on object count: raise it and short structures drop out, lower it and they come back, and each fracture appears at the threshold matching its own connected length (Path-Opening Across 39 Thresholds). Lowering a detection threshold buys you objects. That is the same sentence as S(T) above.

The second version is about which quantity the client is actually paying for, and it cost us more to learn. On the vug work the deliverable was a vug ratio, an area fraction per depth interval, and the only comparator we had was an interpreter's own vug ratio in the incumbent software: one number per interval, no per-vug boundaries. Reviewers asked for intersection-over-union, and intersection-over-union is not a weak metric against that comparator, it is undefined, because there are no reference regions to intersect (Vug Ratio, Not IoU). The metric is a property of the labels and of the deliverable, not of the model. We reported the ratio because the ratio was the thing being bought.

A methane detection limit is the same class of question wearing different units. Sensitivity is an object-count control. Mass is the deliverable in an inventory contract and object count is the deliverable in a repair contract, and a buyer who has not decided which contract they are writing will buy the number on the datasheet.

What to ask a methane vendor

The short version fits in four questions, and none of them is about the detection limit.

What is the deliverable, a source register or a reconciled total. What is the revisit interval at the offered price, and is it fixed or opportunistic. What persistence do you assume, and is it measured on this asset class or taken from a paper. And what share of the annual budget does the offer deliver, computed on this operator's own source-size distribution rather than on California's.

An operator that can answer the fourth question already has the number that matters. One that cannot is going to be sold sensitivity, because sensitivity is what is printed.

Key takeaways

1. 1A detection limit is a property of an instrument. The share of an emissions budget a programme delivers is a property of the source population, and for methane point sources that population is published: Duren et al. surveyed more than 272,000 infrastructure elements and quantified 564 strong point sources, with 10% of them carrying roughly 60% of point-source emissions.
2. 2Two figures that paper prints pin a lognormal source-size law with no free parameters: the top decile carrying 60% of the mass fixes the log-scale width at 1.5349, and 99% of the mass sitting above 25 kg per hour then fixes the location at 4.4337.
3. 3Moving from the paper's 5 kg per hour airborne limit to the 100 kg per hour it quotes for an equivalent satellite in low Earth orbit, a factor of twenty, costs 7.7 percentage points of detected mass and 52.9% of the detected source count. Sensitivity buys source count; it barely touches mass.
4. 4At the reported median persistence of 0.20 a single pass attributes at most one fifth of the annual budget however sensitive the instrument is. Treating passes as independent, the coarser instrument matches one airborne pass after 1.09 revisits, so twenty times the sensitivity is worth about one extra overpass.
5. 5The claim has a stated boundary. At persistence 1.00 there is no crossing at all, because revisit only buys back what intermittency takes away. And the paper's own largest emitters, a subset of landfills, are the most persistent, so the assumption that persistence is independent of source size is the first thing to test against an operator's own revisit history.

Limitations

The distribution is a two-parameter fit to two rounded published figures, not a fit to the underlying 564 measurements, which are available but were not used here. The first anchor is stated as 60% in the Figure 2a caption and as "roughly 60%" in the abstract, and the fitted width inherits that rounding; a true value of 58 or 62% moves every downstream figure. The second anchor is facility-level and is applied here at source level, in the direction described above. The lognormal itself is the family the paper fits in its own Figure 2b, but a fitted family is a modelling choice and the far tail is where a point-source budget lives.

The persistence model is a single population-wide Bernoulli parameter applied uniformly to every source size, applied to passes assumed independent. Both simplifications are stated where they are used and neither is defensible as a description of a real asset. The 5 kg per hour airborne reference is the middle of a published 2 to 10 kg per hour band that the paper says depends on surface brightness and assumes 5 metres per second surface winds; the 100 kg per hour orbital figure carries the same wind assumption and depends on spatial resolution. Neither is a datasheet figure for any specific commercial instrument, and no figure in this piece is attributed to any operator's own detection performance.

Everything here is a California distribution. The paper is explicit that its sectoral mix is specific, that landfills dominate at 41% with dairies and oil and gas at 26% each, and that whether similar distributions occur in other key regions is a hypothesis requiring further surveys [7]. An upstream oil and gas estate in the Permian, the Sultanate of Oman or offshore West Africa has a different mix and possibly a different tail, and the exchange rate between sensitivity and revisit is a function of that tail. The method transfers; the 1.09 does not.

References

[1] TotalEnergies. Methane Emissions Reduction: TotalEnergies Implements a Worldwide Drone-Based Detection Campaign. Press release, 16 May 2022. AUSEA described as a miniature dual sensor mounted on a drone, capable of detecting methane and carbon dioxide emissions while at the same time identifying their source; tested at sites in Nigeria, Italy, the Republic of the Congo and the Netherlands; extended to all upstream oil and gas sites the company operates. The release gives no detection-limit figure. https://totalenergies.com/news/press-releases/methane-emissions-reduction-totalenergies-implements-worldwide-drone-based

[2] TotalEnergies. AUSEA: the innovative technology for detecting and reducing methane emissions. Company technology page. States that all of the company's upstream sites are currently subject to an AUSEA detection campaign at least once a year, and that in 2025 the total reaches approximately 700 mission days across operated, non-operated and third-party assets. https://totalenergies.com/projects/innovation-and-rd/ausea-innovative-technology-reducing-our-methane-emissions

[3] TotalEnergies. COP29: TotalEnergies deploys continuous, real-time methane emissions detection equipment on all its operated upstream assets. Press release, November 2024. States the continuous detection plan would be fully implemented by end-2025, and names the GranMorgu FPSO in Suriname. https://totalenergies.com/news/press-releases/cop29-totalenergies-deploys-continuous-real-time-methane-emissions-detection

[4] Eni. Eni Confirms UN Gold Standard for Methane Emissions Reporting. News item, 23 October 2025. Joined OGMP 2.0 in 2020; Gold Standard Pathway in 2023; Gold Standard reporting in 2024; recognised again in the 2025 report for identifying and quantifying emissions across non-operated assets and for LDAR training delivered to National Oil Company personnel with UNEP support. https://www.eni.com/en-IT/media/news/2025/10/ns-eni-confirm-un-gold-standard-methane-emission-reporting.html

[5] UNEP Oil and Gas Methane Partnership. List of OGMP 2.0 Member Companies, as of 20 February 2026. Upstream segment includes ADNOC (Abu Dhabi National Oil Company), Bp, Eni, Equinor, Repsol, Shell and TotalEnergies. Company names are quoted as the list prints them. https://www.ogmpartnership.org/sites/default/files/documents/2026-02/List_of_OGMP2.0_Member_Companies_20.02.pdf

[6] International Energy Agency. Global Methane Tracker 2025, key findings. Energy-related methane emissions above 120 Mt annually; around 200 bcm released by the fossil fuel sector in 2024; around 70% abatable with existing technologies and around 35 Mt at no net cost; more than 25 satellites in orbit providing methane insights, with MethaneSAT and Tanager-1 operational in 2024; Carbon Mapper analysis of more than 2,000 plumes finding roughly one quarter of sources detected at oil and gas facilities to be recurrent. https://www.iea.org/reports/global-methane-tracker-2025/key-findings

[7] Duren, R. M., Thorpe, A. K., Foster, K. T., Rafiq, T., Hopkins, F. M., Yadav, V., Bue, B. D., Thompson, D. R., Conley, S., Colombi, N. K., Frankenberg, C., McCubbin, I. B., Eastwood, M. L., Falk, M., Herner, J. D., Croes, B. E., Green, R. O., and Miller, C. E. California's methane super-emitters. Nature 575, 180-184 (2019). DOI 10.1038/s41586-019-1720-3. More than 272,000 infrastructure elements surveyed; 564 distinct sources quantified at 250 facilities; net point-source emissions 0.618 Tg per year (95% CI 0.523 to 0.725); landfills 41%, dairies 26%, oil and gas 26%; Figure 2a caption states 10% of point sources are responsible for 60% of detected point-source emissions and that the California numbers are not adjusted for persistence; Figure 2b shows lognormal fits; the Figure 2 caption gives typical detection limits from 2 to 10 kg CH4 per hour at the study's 3 km flight altitudes to 100 kg CH4 per hour for an equivalent satellite in low Earth orbit; median persistence 0.20, mean 0.33, range 0.02 to 1.0; 99% of point-source emissions come from facilities emitting at least 25 kg per hour; detection limits could be relaxed by a factor of ten and still identify 90% of super-emitters. https://doi.org/10.1038/s41586-019-1720-3