Can new mobile technologies enable fugitive methane reductions from the oil and gas industry?

Page created by Marshall Lane
 
CONTINUE READING
Can new mobile technologies enable fugitive methane reductions from the oil and gas industry?
LETTER • OPEN ACCESS

Can new mobile technologies enable fugitive methane reductions from
the oil and gas industry?
To cite this article: Thomas A Fox et al 2021 Environ. Res. Lett. 16 064077

View the article online for updates and enhancements.

                               This content was downloaded from IP address 46.4.80.155 on 13/09/2021 at 05:56
Can new mobile technologies enable fugitive methane reductions from the oil and gas industry?
Environ. Res. Lett. 16 (2021) 064077                                                       https://doi.org/10.1088/1748-9326/ac0565

                              LETTER

                              Can new mobile technologies enable fugitive methane reductions
OPEN ACCESS
                              from the oil and gas industry?
RECEIVED
29 January 2021
                              Thomas A Fox1,∗, Chris H Hugenholtz, Thomas E Barchyn, Tyler R Gough, Mozhou Gao
REVISED                       and Marshall Staples
14 May 2021
                              Centre for Smart Emissions Sensing Technologies, University of Calgary, Calgary, Alberta, Canada
ACCEPTED FOR PUBLICATION      1
26 May 2021                     Present address Highwood Emissions Management thomas@highwoodemissions.com.
                              ∗
                                Author to whom any correspondence should be addressed.
PUBLISHED
14 June 2021                  E-mail: thomas.fox@ucalgary.ca

                              Keywords: fugitive methane, leak detection and repair, oil and gas, screening technology, LDAR-Sim
Original content from
this work may be used
under the terms of the
Creative Commons
Attribution 4.0 licence.      Abstract
Any further distribution      New mobile platforms such as vehicles, drones, aircraft, and satellites have emerged to help identify
of this work must
maintain attribution to
                              and reduce fugitive methane emissions from the oil and gas sector. When deployed as part of leak
the author(s) and the title   detection and repair (LDAR) programs, most of these technologies use multi-visit LDAR (MVL),
of the work, journal
citation and DOI.             which consists of four steps: (a) rapidly screen all facilities, (b) triage by emission rate, (c)
                              follow-up with close-range methods at the highest-emitting sites, and (d) conduct repairs. The
                              proposed value of MVL is to identify large leaks soon after they arise. Whether MVL offers an
                              improvement over traditional single-visit LDAR (SVL), which relies on undirected close-range
                              surveys, remains poorly understood. We use the Leak Detection and Repair Simulator (LDAR-Sim)
                              to examine the performance and cost-effectiveness of MVL relative to SVL. Results suggest that
                              facility-scale MVL programs can achieve fugitive emission reductions equivalent to SVL, but that
                              improved cost-effectiveness is not guaranteed. Under a best-case scenario, we find that screening
                              must cost < USD 100 per site for MVL to achieve 30% cost reductions relative to SVL. In scenarios
                              with non-target vented emissions and screening quantification uncertainty, triaging errors force
                              excessive close-range follow-up to achieve emissions reduction equivalence. The viability of MVL
                              as a cost-effective alternative to SVL for reducing fugitive methane emissions hinges on accurate
                              triaging after the screening phase.

                              1. Introduction                                                     new LDAR technologies and methods are emerging
                                                                                                  (Fox et al 2019, Ravikumar et al 2019).
                              The oil and gas (O&G) industry is among the largest                     Emerging LDAR technologies differ markedly in
                              sources of anthropogenic methane, a potent green-                   measurement scale, deployment mode, and inform-
                              house gas. To reduce fugitive (defined here as unin-                ation product, leading to a range of niche use cases
                              tentional loss of containment) methane emissions                    that are not yet well defined (Fox et al 2019a).
                              from O&G, regulators are increasingly mandating the                 Many new technologies consist of methane sensors
                              use of leak detection and repair (LDAR) programs                    deployed on mobile platforms: vehicles (Robertson
                              (US EPA 2016, AER 2018, ECCC 2018). In jurisdic-                    et al 2017, Caulton et al 2018), drones (Golston et al
                              tions without methane regulations, voluntary LDAR                   2018, Barchyn et al 2019), aircraft (Englander et al
                              programs are common. Most LDAR programs rely on                     2018, Schwietzke et al 2018), and satellites (Jacob
                              handheld organic vapour analyzers (OVAs) or optical                 et al 2016, Varon et al 2018). Data gathered by LDAR
                              gas imaging (OGI) cameras, which are functional,                    technologies can be used to attribute emissions to
                              familiar, and recommended by regulators (US EPA                     various scales, including regions, facilities, equip-
                              2016). Most OVA and OGI programs use single visit                   ment, or components. Ultimately, component-scale
                              LDAR (SVL), where all O&G facilities within a pro-                  detection is required for diagnosis and repair, but
                              gram scope are surveyed at a specified frequency.                   most emerging technologies measure at equipment
                              However, rapid innovation is underway, and diverse                  or facility scales. A new strategy called ‘screening’ has

                              © 2021 The Author(s). Published by IOP Publishing Ltd
Can new mobile technologies enable fugitive methane reductions from the oil and gas industry?
Environ. Res. Lett. 16 (2021) 064077                                                                    T A Fox et al

therefore emerged, which uses rapid mobile surveys         is large (Ravikumar et al 2019). More accurate quan-
to identify a subset of the highest-emitting facilit-      tification is often possible with longer surveys—but
ies in greatest need of LDAR (Fox et al 2019a). The        longer surveys are more expensive. When triaging
motivation behind screening is to reduce the cost          decisions are inaccurate because of vented emissions
of exhaustive SVL by rapidly targeting the highest         and quantification error, the ranked list of facility-
emitters. Studies in Canada and the United States          level emissions is inaccurate, and facilities with lower
have shown that most methane leaks are small, with         fugitive emissions can get visited at the expense of
5% of sources accounting for approximately 50% of          those with higher fugitive emissions. In the worst-case
emissions (Brandt et al 2016). The rationale behind        scenario, follow-up inspectors could be sent to facil-
screening is that smaller leaks can be deprioritized or    ities with no fugitive emissions, while facilities with
overlooked if large leaks can be found quickly.            very large leaks are overlooked.
     Screening must be combined with close-range                The objective of this paper is to compare the emis-
follow-up inspections to diagnose and repair indi-         sions reduction performance and cost-effectiveness
vidual leaks. We define multi-visit LDAR (MVL) pro-        of emerging MVL programs and conventional SVL
grams as consisting of four steps: screening, triaging,    programs. Using an agent-based modeling frame-
follow-up, and repair. First, rapid screening assesses     work called LDAR-Sim, we generate a broad range
all facilities, detecting and often quantifying emis-      of emissions equivalence scenarios (the conditions
sions. Second, triaging is a decision-making step that     under which MVL and SVL achieve identical fugit-
uses the results from screening to determine which         ive emission reductions), spanning different screen-
facilities should receive follow-up inspections. Most      ing survey frequencies, follow-up requirements,
MVL programs consist of a single layer of screening,       quantification errors, empirical leak-size distribu-
but screening and triaging can be repeated if multiple     tions, and vented emissions. These equivalence scen-
methods of increasing precision are used (e.g. satellite   arios are then used to explore the cost-effectiveness
then aircraft). Third, facilities that were flagged dur-   of MVL under different conditions. We show
ing triaging are inspected at close-range to confirm       that equivalent and cost-effective MVL programs
whether a leak exists and to identify repair require-      require low-cost screening and provide target met-
ments. Fourth, if a leak exists and cannot be immedi-      rics to help identify successful technologies and
ately repaired, an additional visit may be required. In    programs.
general, the goal of MVL is to reduce time-consuming
and expensive SVL at every facility by combining           2. Methods
cheaper screening with targeted follow-up inspection
at a subset of screened facilities. Screening frequency    Equivalence and cost-effectiveness of MVL are evalu-
in MVL should be higher than inspection frequencies        ated using LDAR-Sim (Fox et al 2021). LDAR-Sim is
in SVL because (a) most screening methods are less         an open-source modeling framework used to evalu-
sensitive than close-range methods, and (b) follow-        ate and compare LDAR programs. In a virtual asset
up occurs at only a subset of facilities, while SVL pro-   field, LDAR workers search for leaks at O&G facil-
grams are exhaustive.                                      ities. LDAR workers apply technology modules that
     Certain conditions are favourable for MVL. First,     can detect leaks or quantify facility-scale emissions.
screening costs per facility should be lower than SVL      The deployment of LDAR workers is managed under
because (a) screening is applied more frequently, and      a program definition, which defines the number of
(b) after screening, follow-up surveys may still be        workers and types of technology they are using. New
required. Second, leak-size distributions with heav-       leaks and vented emissions appear stochastically in
ier tails (more small leaks, fewer large leaks) favour     the asset field, drawn from empirical distributions.
MVL because the benefit of finding large leaks             The model proceeds forward through time, tracking
increases. Third, deployment conditions and tech-          the performance of different LDAR programs. Empir-
nology performance should enable accurate triaging.        ical inputs describing specific deployment regions,
When ranking facilities by emission rate, quantifica-      target facilities, monitoring technologies, work prac-
tion error and the presence of confounding sources         tices, and regulations are required to operate the
may affect triaging. For example, vented emissions         model. Outputs describe anticipated emissions mit-
(defined here as methane emissions that occur by           igation and a cost analysis of the simulated program.
design, such as from pneumatic controllers or meth-        To demonstrate equivalence, simulations can com-
ane slip during combustion) at upstream facilities         pare different LDAR programs.
are not classified as leaks. Because vented emissions           Generally speaking, three measurement scales are
are not candidates for repair, they are not targeted       discussed in the context of LDAR in O&G: facil-
by LDAR programs. However, vented emissions are            ity, equipment, and component. Facilities consist of
measured by facility-scale screening technologies and      one or more pieces of equipment, and each piece
can confound the identification of fugitive emissions.     of equipment can consist of dozens or hundreds of
Further, quantification error from screening methods       individual components. Although leaks arise from

                                              2
Environ. Res. Lett. 16 (2021) 064077                                                                    T A Fox et al

components, some methods only measure at equip-           scenarios are important as regulators often approve
ment of facility scales. In LDAR-Sim, inspection          alternative LDAR programs based on achieving emis-
agents can measure emissions at the component             sions reduction equivalence with prescribed SVL pro-
scale, simulating handheld instruments like OGI, or       grams. Alternative programs will not be approved if
at the facility scale, simulating screening methods.      they do not achieve equivalence, but those respons-
In this study we assume that facility-scale screen-       ible for conducting LDAR (e.g. O&G producers) want
ing methods are unable to discern fugitive and ven-       to minimize costs and are unlikely to reduce emis-
ted emissions, which are aggregated into a single         sions beyond what is required. Equivalence scenarios
measurement. Some screening methods can measure           balance these opposing requirements. The concept of
emissions at intermediate scales, localizing sources to   minimizing the cost of equivalence is also useful bey-
pieces of equipment (e.g. liquid tanks) or equipment      ond a regulatory context, as operators with volun-
groups but not individual components (Ravikumar           tary emissions reduction targets are likely to seek cost-
et al 2019). However, modeling at intermediate scales     effective solutions.
is complex and we leave these analyses for future             This study consists of identifying equivalence
work.                                                     scenarios and evaluating their cost-effectiveness.
     In LDAR-Sim, an MVL workflow consists of the         First, we run LDAR-Sim to estimate emissions under
following four steps: (a) screening agents estimate       annual and triannual SVL programs that rely on OGI.
emission rates of each facility in the LDAR program;      These SVL programs represent conventional regulat-
(b) a triaging procedure ranks all screened facilit-      ory and voluntary LDAR programs. We then run
ies by emission rate and flags those that meet a          ensembles of MVL simulations to identify equival-
follow-up threshold; (c) follow-up OGI agents visit       ence scenarios for two leak size distributions, two
flagged facilities to identify individual leaks and tag   follow-up threshold definitions, with and without
them for repair; (d) leaks are repaired, lowering emis-   vented emissions, and with and without quantific-
sions. We use the following terminology: entire facil-    ation error. We then evaluate the cost-effectiveness
ities are ‘flagged’ during screening and individual       of equivalence scenarios for generic MVL programs.
leaks are ‘tagged’ for repair. Simulated screening        All analysis code, empirical inputs, and output data
methods measure aggregate emissions, so a site with       are available in the supplementary material and are
many small leaks might look the same as a site            accompanied by a reproducibility guide.
with one large leak. Even a facility with no leaks
may be flagged for follow-up if vented emissions are      2.1. Triaging procedures
high, or if mistakes are made during ranking due to       Triaging combines ranking facilities by emission rate
quantification error. However, follow-up surveys at       and deciding which facilities to flag. Once facilit-
these mis-ranked flagged facilities can still result in   ies have been ranked by emission rate, an expli-
tags of smaller leaks that contributed only a small       cit rule must be used to distinguish between (a)
fraction to total site-level emissions.                   high-emitting facilities that are flagged for follow-up
     The number and timing of follow-up inspec-           inspections by close-range methods so that leaking
tions is important for determining fugitive emissions     components can be tagged and queued for repair,
in LDAR programs. In MVL, emissions reductions            and (b) low-emitting facilities that will not receive
are positively correlated with screening survey fre-      close-range inspection and repair. Here, facilities are
quency and amount of follow-up. Additional screen-        flagged if their estimated emission rate during screen-
ing surveys will find newly generated leaks faster,       ing is above a threshold emission rate. We introduce
reducing the time that a large leak is left emit-         two types of follow-up threshold: static and dynamic.
ting. More follow-up surveys will find more leaks.            A static follow-up threshold is a constant value
For example, an LDAR program with 500 facilities          derived from a baseline emissions distribution; it is
may have triannual screening and dispatch follow-         the emission rate that corresponds with a desired tar-
up inspectors to the top 10% of emitters. Each            get proportion of highest emitting facilities (static tar-
year, 1500 facility-scale screening inspections will      get). Table S1 (available online at stacks.iop.org/ERL/
be performed, resulting in 150 flags and follow-          16/064077/mmedia) shows the static targets used in
up inspections. To double the number of follow-           this study, the corresponding proportion of total
up inspections to 300, three options are available:       emissions they represent in the distribution, and their
(a) double the number of screening surveys to             static follow-up thresholds. For example, to target
3000, (b) double the percentage of facilities receiv-     the top 2% of leaks, static thresholds of 0.34 and
ing follow-up to 20%, or (c) some combination of          0.54 g s−1 are required for empirical distributions A
(a) and (b).                                              and B, respectively (inputs are described in the fol-
     To better examine the economics of LDAR, we          lowing section). Although we define static thresholds
vary screening frequencies and follow-up to isolate       using a leak-size distribution, facility-scale fugitive
specific scenarios that achieve emissions reduction       emissions measurements could also be used when
equivalence with SVL. These ‘equivalence scenarios’       available. The MVL methods modeled in this study
are used as the base of cost comparisons. Equivalence     measure at the facility scale, which may aggregate

                                             3
Environ. Res. Lett. 16 (2021) 064077                                                                      T A Fox et al

multiple leaks. Therefore, a static threshold of 2%         surveys, as these are most representative of pre-
does not mean that 2% of facilities will receive follow-    LDAR conditions. Compared to D1 , D2 is larger
up—far more than 2% of facilities will be flagged as        (n = 969 vs. 281) and better matches distributions
there are often multiple leaks and vented emissions.        seen elsewhere, where typically 5% of leaks rep-
We use facility-scale emissions as a starting proxy for     resent 50% of emissions (table S1) (Brandt et al
facilities with anomalously high leak rates. Facilities     2016). D1 represents a highly skewed distribution in
are flagged if their summed emission rate exceeds the       which 2% of sources account for ∼54% of emis-
follow-up threshold.                                        sions. Only one company was surveyed in D2 , whereas
    Dynamic thresholds depend on the relative emis-         D1 represents 63 companies with broad geographical
sions among facilities at the time of screening. Dur-       range.
ing a survey, a screening method visits each facil-              Fugitive emissions are simulated under a range of
ity in a program, quantifies the emissions of each          follow-up targets (listed in table S1). The same follow-
facility, ranks facilities according to emission rate,      up targets are used to establish static and dynamic
and sends a follow-up crew to a specified propor-           thresholds. All simulations are run on 500 randomly
tion of the highest-emitting facilities (e.g. top 10%).     selected facilities in Alberta. Locations of facilities do
The dynamic target defines what proportion of the           not matter to this simulation because weather and
highest emitters should be flagged for follow-up (e.g.      travel distances are ignored to generalize the study,
0.1 for the top 10%). Note that the dynamic threshold       but these parameters are important in applied situ-
that follows from the dynamic target is implicit and        ations. Similarly, method-specific parameters (e.g.
does not need to be calculated. Compared to static          detection limits, reporting and repair delays) and
thresholds, dynamic thresholds are conceptually sim-        labour availability are not required to identify equi-
pler but may not be realistic for all screening meth-       valence scenarios. For each static or dynamic target,
ods. An advantage of dynamic follow-up is that pro-         ten simulations are run over six years (our analysis
gram cost should be relatively constant, as the same        excludes year 1). Following previous studies, we use
number of facilities always receive follow-up. How-         a daily leak production rate (LPR) of 0.0065 leaks per
ever, if aggregate program emissions are very high,         site, but sensitivity to LPR is explored later (Kemp et al
dynamic thresholds are not guaranteed to meet mit-          2016, Fox et al 2021).
igation goals relative to a baseline. Similarly, if emis-        Quantification is only required for technologies
sions are lower than expected, dynamic thresholds           that use triaging—not close-range technologies used
may lead to unnecessary follow-up. A disadvantage           to detect and diagnose individual leaks. In some
of dynamic thresholds is that all facilities in a group-    treatments, vented emissions and screening quanti-
ing must be surveyed before follow-up decisions can         fication errors are introduced. Vented emissions are
be made, whereas decisions can be made on-the-fly           introduced following the methodology described pre-
when using static thresholds.                               viously (Fox et al 2021). Facility-scale quantifica-
                                                            tion error (E) remains poorly constrained for LDAR
2.2. Equivalence scenario modeling                          screening methods, and likely depends on the work
The parameters and empirical inputs used in this            practice used, dispersion modeling approaches, and
study are based on an LDAR-Sim case-study demon-            a range of method-specific environmental factors.
stration for Alberta, Canada (Fox et al 2021). Alberta      For example, quantification uncertainties for ground
produces approximately 0.3 billion m3 of marketable         vehicles are reported to range from 50% to 350%
natural gas and 0.5 million m3 of crude oil per day         (Fox et al 2019a). More recently, blind controlled
from bituminous sands and a network of ∼176 000             release experiments found quantification estimates
conventional and unconventional oil and gas wells           from mobile LDAR technologies to be within a factor
(AER ST37, AER ST98). As of January 2020, close-            of two ∼35% of the time, and within an order or
range (component-scale) LDAR must be performed              magnitude ∼82% of the time (Ravikumar et al 2019).
at tens of thousands of facilities up to three times per    Rather than attempt to replicate E for any given
year using OGI cameras or OVAs (AER 2018, Johnson           method, we present three hypothetical scenarios. In
and Tyner 2020). These new regulations are among            E1 , facility-scale screening quantification has zero
the first globally that allow producers to develop and      uncertainty. In E2 , an error term is drawn from a
implement ‘alternative’ LDAR programs, which can            normal distribution with a mean of 0 and a stand-
consist of combinations of technologies and work            ard deviation of 2.2, such that ∼35% of observations
practices that demonstrate equivalence (AER 2018).          fall within ±1 of the true value (factor of two). In
     We establish equivalence scenarios using two           E3 , the error term is drawn from a normal distribu-
empirical leak-size distributions: (D1 ) the Clearstone     tion with a mean of 0 and a standard deviation of 7.5,
Engineering dataset (as used in Fox et al 2021),            such that ∼82% of observations fall within ±10 times
and (D2 ) recently published data from fieldwork            the true value (order of magnitude). As the error term
in Alberta (Ravikumar et al 2020). Both distribu-           departs from zero, it shifts the true emission rate (QT )
tions distinguish between fugitive and vented emis-         away from the estimated rate (QE ). In E1 , the true rate
sions. In D2 , we use only leaks from initial LDAR          equals the estimated rate, and all follow-up decisions

                                               4
Environ. Res. Lett. 16 (2021) 064077                                                                                         T A Fox et al

   Figure 1. Ensemble of simulations across a range of dynamic follow-up targets and annual surveys (colours). As the dynamic
   target increases, more facilities receive follow-up, more leaks are found, and fugitive emissions are lower. Increasing survey
   frequency similarly results in lower fugitive emissions as infrequent large emissions are found and repaired sooner. Intersections
   between SVL programs (black horizontal lines) and screening programs (colour lines) denote equivalence. For example, to be
   equivalent with triannual OGI-based SVL, a screening program that screens all facilities nine times per year must send OGI
   follow-up to the top 12% of highest-emitting facilities. Shaded regions are two standard deviations ensemble variability in
   emissions rate. This example shows simulations under D1 , E1 , and without vented emissions (the most favourable situation for
   screening technologies to be economically viable). Fugitive emission rates are averaged over five year simulations.

are optimal. In E2 and E3 , discrepancies between the                 the market. One way to estimate costs is to simulate
two rates reduce follow-up effectiveness as facilities                typical screening programs in LDAR-Sim by platform
are mis-ranked.                                                       class. However, parameterization of these programs is
                                                                     difficult, as little is known about daily costs, the num-
             0,                     if E = E1
                                                                      ber of facilities possible to screen per day, and general
        E=      Normal (0, 2.2) , if E = E2 .     (1)
                                                                     performance metrics of any specific solution.
                Normal (0, 7.5) , if E = E3
                                                                          Instead of forward simulating the cost of screen-
                {
                    QT + QT · E, if E ⩾ 0                             ing programs, we examine the amount of money that
          QE =                               .    (2)                 would be available for a screening technology within
                    QT / |E − 1| , if E < 0
                                                                      a full program budget. This addresses the question
2.3. Cost-effectiveness modeling                                      of ‘How much do screening programs need to cost
For each survey frequency, we extract optimum                         to result in savings relative to a single-visit OGI pro-
follow-up targets by linearly interpolating between                   gram?’ This inverse question is easier to answer as
the nearest programs above and below SVL pro-                         fewer assumptions about screening effectiveness are
gram emissions (figure 1). Equivalent MVL programs                    required.
should offer cost savings over SVL programs to be                         For any equivalence scenario, we know: (a) the
adopted by the O&G industry. However, estimating                      SVL survey frequency (FSVL ), (b) the dynamic tar-
the cost-effectiveness of screening programs is chal-                 get proportion (τ ), and (c) the screening survey fre-
lenging as the market is nascent and well-established                 quency (Fscreen ). If we assume that follow-up OGI
workflows have not been rigorously demonstrated.                      inspections are identical to OGI inspections under
Many companies exist, but each has a unique product                   SVL, the per-facility cost of OGI (COGI ) should
or service that combines a platform (e.g. aircraft,                   also be the same. Equivalent cost occurs when the
drone, satellite, vehicle), its sensors, and some work                cost of OGI (left side of equation) equals those
practice. Costs are likely to change as these com-                    costs of both screening and follow-up (right side of
panies continue to develop and identify their role in                 equation):

                                                       5
Environ. Res. Lett. 16 (2021) 064077                                                                      T A Fox et al

 FSVL · COGI · (1 − φ) = Fscreen · Cscreen + COGI · τ (3)    diminishing returns in emissions reductions. Emis-
                                                             sions reduction equivalence occurs when simulated
where φ is the discount in MVL program costs                 emissions for an SVL program equal the emissions of
desired, relative to SVL. Solving for Cscreen , the money    an MVL program. In theory, when the MVL and SVL
available for screening after accounting for the cost of     programs have the same survey frequency, τ must
follow-up:                                                   equal 1 to achieve equivalence (i.e. all facilities will
                                                             receive follow-up and screening becomes redundant).
                FSVL · COGI · (1 − φ) − COGI · τ             Therefore, each screening survey frequency above
    Cscreen =                                    .    (4)
                              Fscreen                        the SVL survey frequency will have a correspond-
                                                             ing τ required for equivalence. For example, if the
To simulate the risk and costs of new program devel-         SVL program has triannual inspections at all facil-
opment, we use a discount requirement (φ). The               ities, screening once or twice per year will not be
discount can be used to calculate Cscreen for MVL pro-       equivalent, and screening three times per year will
grams that must be less expensive than a correspond-         only be equivalent when 100% of screened facilities
ing SVL program. For example, if a producer requires         receive follow-up (τ = 1). The only sensible equival-
the MVL program to be 30% cheaper than SVL to jus-           ence scenarios occur when τ < 1, which requires a
tify the effort and risk of adoption, φ = 0.3 is used to     higher screening survey frequency than the SVL pro-
calculate Cscreen . For cost equivalence, φ = 0. Aver-       gram. As screening survey frequency increases, the
age OGI costs are generally between USD 250 and              corresponding τ required for equivalence decreases.
450 per facility (EDF 2016). The Cscreen calculated               The simplest definition of an equivalence scen-
here is conceptualized as the money available for per        ario is a screening survey frequency and correspond-
facility screening, or a contract bid price that a pro-      ing dynamic (τ ) or static target under a set of model-
ducer could offer to screening solution providers. If        ing assumptions. All equivalence scenarios developed
a screening technology cannot perform screening at           in this study are shown in figure 2. In panels A
this price, it is not cost effective. If a screening tech-   and B, equivalence scenarios closer to the top-right
nology can perform the screening for less than Cscreen       require more work (i.e. higher screening frequency
additional cost savings are available in the program.        and follow-up) relative to those in the bottom left.
     To illustrate Cscreen , consider an MVL program         Achieving equivalence with a triannual SVL pro-
that applies screening two times per year with τ = 0.1       gram requires either more screening or more follow-
(i.e. the top 10% of emitting facilities are flagged         up than achieving equivalence with annual SVL. A
and receive follow-up after each round of screen-            similar pattern exists between programs with and
ing). To keep the scenario simple, we exclude con-           without vented emissions. Here, screening techno-
founding aspects of vented emissions and quantific-          logies are unable to differentiate fugitive from ven-
ation error. The SVL program consists of 500 facil-          ted emissions, meaning that facility-scale emissions
ity surveys using OGI with a total cost between USD          estimates include both. The fugitive emission signal
125 000 and 225 000. The MVL program would res-              can therefore be lost in the noise of vented emissions,
ult in 1,000 facility screening surveys, with follow-        which impacts facility ranking during triaging and
up OGI inspection at 100 facilities. The OGI follow-         results in follow-up crews being sent to the wrong
up cost for 100 facilities is between USD 25000 and          facilities—ones with high vented emissions but not
45000. The money available for screening, Cscreen , is       necessarily high fugitive emissions. To achieve equi-
between USD 100 and 180/facility when φ = 0, or              valence, the τ must increase to account for triaging
between USD 62.5 and 112.5/facility when φ = 0.3.            errors, in some cases more than doubling the required
There are scenarios in which Cscreen can be negat-           number of follow-up inspections.
ive (e.g. τ or φ approach 1); in these scenarios, the             Triaging is also impacted by quantification error.
cost of follow-up surveys alone exceeds the cost of          With no screening quantification error (E1 ), facil-
the SVL program. In other words, screening solu-             ities can be correctly ranked according to emission
tion providers must work for free or pay to per-             rate before follow-up inspectors are dispatched. As
form the service—a situation that is unlikely to             quantification error increases to E2 and E3 , facilit-
occur.                                                       ies are ranked not according to their true emissions
                                                             rate (QT ), but according to the estimated rate (QE ),
3. Results and discussion                                    effectively ‘shuffling’ the ranking. Whereas the impact
                                                             of vented emissions is relatively consistent for differ-
3.1. Equivalence scenarios                                   ent survey frequencies, quantification error is more
Each MVL and SVL program results in a fugitive               problematic for MVL programs with lower screening
emission rate averaged over a five year simulation.          survey frequencies. This occurs because quantifica-
Figure 1 provides an example of how increasing either        tion errors are a percentage of the true emission rate,
the number of screening surveys or the amount of             whereas vented emissions are drawn from an empir-
follow-up (τ ) can lead to lower fugitive emissions.         ical distribution and are assumed independent of the
Increasing either screening surveys or τ results in          fugitive emissions at a facility. When screening survey

                                                6
Environ. Res. Lett. 16 (2021) 064077                                                                                           T A Fox et al

   Figure 2. Catalogue of equivalence scenarios under various conditions: (A) dynamic targets assuming leak-rate distribution D1
   (more skewed), (B) dynamic targets assuming leak-rate distribution D2 (less skewed), and (C) static targets assuming D1 . Each
   point represents an equivalence scenario with either annual (orange) or triannual (blue) OGI-based SVL programs. In general,
   more screening (either surveys or follow-up) is required to achieve equivalence with SVL programs that have higher survey
   frequencies. Solid lines assume that all emissions are fugitives, while dashed lines assume the additional presence of vented
   emissions. Quantification errors increase from E1 (perfect facility-scale quantification) to E3 (order or magnitude error) and are
   explained in the main text. Note the different interpretations of dynamic targets (τ ; proportion of top-emitting facilities requiring
   follow-up) and static targets (proportion of the fugitive leak-size distribution to target).

frequency is high, less time passes between surveys,                   requires a larger number of follow-up inspections to
which allows less time for leaks to accumulate. When                   achieve equivalence.
QT is close to zero, high quantification errors have less                  Figure 2(C) shows equivalence scenarios using
absolute impact on QE . Future work should investig-                   static targets, and must be interpreted differently
ate if QE should be estimated as a percentage of QT or                 from panels A and B. Both vented emissions and
as an absolute difference.                                             quantification error result in lower required targets.
    Comparing panels A and B in figure 2 illus-                        When vented emissions are added to fugitive emis-
trates the impact of leak-size distribution. Equival-                  sions at a facility, it makes it easier for the static
ence scenarios are achieved with less work for screen-                 threshold to be surpassed. However, this results in
ing programs under D1 , which is more heavily skewed                   much more work, which is not immediately evident
(figure 2(A)). When a smaller number of large leaks                    as a higher static target does not correspond intuit-
comprise the majority of total emissions, strategies                   ively with follow-up requirements like it does with a
that focus on these leaks benefit. Figure 2(B) shows                   dynamic target. Lower static targets are required to
equivalence scenarios under the D2 leak-rate distri-                   achieve equivalence as quantification error increases
bution, which is less heavy-tailed and consequently                    due to the heavy-tailed shape of the facility-scale

                                                        7
Environ. Res. Lett. 16 (2021) 064077                                                                                         T A Fox et al

   Figure 3. Maximum screening program costs (Cscreen ) under a range of equivalence scenarios. High Cscreen values indicate more
   cost-effective MVL programs. Estimates are made assuming OGI survey costs of USD $250/facility (lower limit of each line) and
   USD $450/facility (upper limit of each line). Blue lines show required screening costs when screening program target costs are
   equal to those of the SVL program (φ = 0). Orange lines show required screening costs for screening programs that offer a 30%
   cost reduction relative to the SVL program (φ = 0.3). The best-case scenario for MVL assumes perfect screening quantification
   (E1 ), leak rates from D1 , and no vented emissions. The worst-case scenario assumes E3 , D2 , and presence of vented emissions.
   Negative Cscreen values indicate that screening technology providers must pay to perform services, a situation unlikely to exist in
   practice.

emissions distribution. More facilities will fall below                best-case scenarios for MVL, which assume no vented
a given follow-up threshold than above, meaning that                   emissions, E1 , and D1 (extreme skew), Cscreen rarely
quantification error will push lower-emitting facil-                   exceeds USD 100/facility. To reduce costs relative to
ities above the threshold more often than higher-                      the SVL program by 30%, screening costs must not
emitting facilities below the threshold. This results in               exceed USD ∼50–90/facility. The worst-case scenario
more work being done than necessary, leading to a                      is more realistic; it accounts for vented emissions,
lower static target.                                                   quantification error (E3 ), and assumes D2 , which bet-
     Previous studies have shown the influence of LPR                  ter approximates emissions elsewhere (Brandt et al
on simulation results (Kemp et al 2016, Fox et al                      2016). Every screening program under the worst-
2021). We present simulations under a range of LPRs                    case scenario is more expensive (i.e. negative val-
to determine whether results in this study are sensit-                 ues) than SVL before screening is paid for. In other
ive to LPR (figure S1). Fugitive emissions vary greatly                words, triaging is so inefficient that, to achieve equi-
between programs, but because emissions increase for                   valence, more follow-up surveys must be conducted
both the SVL and MVL programs, equivalence scen-                       than would be required under the exhaustive but less
arios are robust to differences in LPR.                                frequent SVL program. To illustrate how Cscreen can
                                                                       be negative, consider a hypothetical LDAR program
3.2. Cost-effectiveness                                                with 100 facilities. An SVL program that requires
In figure 3, we estimate Cscreen for φ = 0 (blue) and                  annual OGI will require 100 inspections. A pro-
0.3 (orange) where COGI = USD 250/facility (lower                      posed MVL program with biannual surveys will have
limit) and USD 450/facility (upper limit). Under                       Cscreen = 0 if 50 follow-up surveys are required after

                                                        8
Environ. Res. Lett. 16 (2021) 064077                                                                         T A Fox et al

each survey (τ = 0.5), as the total number of follow-         challenge for mobile LDAR methods that acquire only
up surveys (100) equals the number of required sur-           a ‘snapshot’ of emissions (Alden et al 2020, Cardoso-
veys under the SVL program. Therefore, Cscreen for            Saldaña et al 2020). Here, simulations account for
the biannual screening program can only be posit-             temporal variability of vented emissions, but not of
ive when τ < 0.5. If screening quantification error           fugitive emissions. If fugitive emissions are episodic,
reduces triaging effectiveness by introducing triaging        then problems arise for MVL including (a) detect-
errors, τ may increase above 0.5, causing Cscreen to          ing an episodic fugitive and dispatching a follow-
drop below zero.                                              up inspector who is unable to identify the source,
     In general, the screening programs simulated in          and (b) failing to detect an episodic source that was
this study are more cost-effective when: (a) COGI is          not emitting at the time of screening. For MVL to
higher; (b) equivalence scenarios with fewer screen-          be effective, the facilities that receive follow-up must
ing surveys are used (i.e. those with higher τ and            be the highest-emitting facilities over time and must
more follow-up); (c) vented emissions are lower; (d)          be clearly distinguishable from low-emission facilit-
quantification error is low; and (e) leak-size distri-        ies. Given that screening is a snapshot in time, the
butions are more skewed. Results also suggest that            implicit assumption is that instantaneous emissions
more money is available for screening when SVL pro-           are representative of the long-term average, which
gram requires more OGI surveys. This is likely due to         may not be the case (Johnson et al 2019). Third,
the relative burden of going from one to two surveys          follow-up OGI could be more expensive than OGI
being greater than the burden of going from three to          used in SVL because flagged facilities may be spaced
four.                                                         further apart, which increases travel time and cost.
     Our results suggest that MVL methods may                 Fourth, this study assumes that screening surveys are
struggle to be cost-effective compared to conven-             spaced evenly apart throughout the year and that they
tional OGI-based SVL (figure 3). Vented emissions             are not impacted by adverse environmental condi-
and quantification errors can lead to incorrect rank-         tions, seasonal variability, or logistical constraints.
ing of facilities by emission rate, which increases the       However, all screening methods have specific opera-
required number of follow-up inspections to achieve           tional limitations that may further impact perform-
equivalence. Simulations suggest that MVL will be             ance, such as clouds for satellites, roads for vehicle sys-
more cost-effective if triaging errors can be reduced.        tems, snow for LiDAR, and wind and precipitation for
While negative Cscreen values are clearly unprofitable        drones (Fox et al 2019a). Fifth, we assume that screen-
(screening providers would have to pay to provide             ing technologies have minimum detection limits suf-
their service), it is difficult to infer profitability when   ficient to accurately rank at least as many facilities as
screening must be USD 50–100 per site. Screening              require follow-up. However, the most cost-effective
technologies are variable in approach as the industry         scenarios require a small number of screening surveys
is nascent and is repurposing technologies used out-          with considerable follow-up (τ up to 0.8 for triannual
side the O&G industry. For example, satellites are            SVL equivalence). When τ is higher, the correspond-
expensive to build, launch, and operate, but are not          ing minimum detection limit required by the screen-
constrained by access to services such as lodging for         ing method is lower.
crews and road access. Vehicle-based systems are con-              Despite these challenges, there may be oppor-
strained by driving time but are more scalable as hard-       tunities for screening technologies. First, our results
ware and deployment are less expensive than satel-            apply when fugitive emissions are the only target of
lites. These two examples are ends of a spectrum;             reductions. New regulations and voluntary reduction
many variables affect the cost of screening surveys,          initiatives by industry may use screening technolo-
which may be unprofitable in some contexts while              gies not specifically for LDAR, but to understand total
profitable in others.                                         facility-scale emissions. If so, vented emissions will go
     The scenario with parameters least favourable for        from being ‘noise’ to part of the ‘signal’ and will not
MVL viability (‘worst-case scenario’) considered in           complicate triaging. Second, vented emissions may
this study may overly optimistic due to five model-           decline significantly over the coming decades as reg-
ing assumptions. First, our analysis does not consider        ulators and industry work towards mitigation tar-
the presence of false positive and false negative detec-      gets. Clear steps can be taken to reduce vented emis-
tions for screening technologies, which could intro-          sions, such as conversion from high- to low-bleed
duce additional triaging errors. False positive and           pneumatics, installation of vapour recovery units on
false negative detections exist among screening meth-         storage vessels, and best practices for well comple-
ods but are more common when the emission rate                tions and manual liquid unloadings. These efforts
is close to zero (Ravikumar et al 2019). False negat-         should increase the relative strength of the fugitive
ives are less likely to occur with the largest and most       signal, assuming no change in leak production rates
consequential emission sources as detection probab-           and LDAR, but may require more sensitive techno-
ility increases with emissions rate (Ravikumar et al          logies to detect smaller emissions. Third, screening
2019). Second, new work is increasingly pointing to           technologies in this study measure only at the facil-
temporal variability of methane emissions as a major          ity scale. In reality, many screening technologies may

                                                 9
Environ. Res. Lett. 16 (2021) 064077                                                                      T A Fox et al

measure at the equipment scale, and the additional           previous studies (e.g. Ravikumar et al 2019). While
granularity could possibly improve MVL. Fourth, we           systematic bias should not impact the relative rank-
do not consider the cost of repair, as it is assumed that    ing of facility emission rates, this topic warrants fur-
all leaks are eventually repaired, whether the result of     ther study.
an LDAR program or by an operator. However, SVL                   More research is required to understand the ways
may lead to higher repair costs, especially as a larger      in which mobile screening technologies might con-
number of smaller leaks are targeted. Fifth, our res-        tribute to methane monitoring and mitigation. Con-
ults are sensitive to the skewness of leak-size distribu-    siderable testing is underway to better understand
tions, with heavier tails favouring MVL. Production          how a broad variety of different technologies might
regions, facility types, or other contexts that exhibit      contribute to reducing methane emissions (Bell et al
extreme heavy tails may therefore benefit the most           2020, Zimmerle et al 2020). Better MVL perform-
from MVL.                                                    ance can result from stricter venting limits, improved
     It may also be possible to reduce triaging errors       sensors and quantification algorithms, and a better
with better ranking algorithms. The simple static            understanding of how, where, and when to use MVL.
and dynamic thresholds used here are intuitive yet           Close-range systems may also continue to evolve to
crude; more complex models may be developed                  have better performance than those considered in this
that incorporate facility-specific quantification errors     study. It should be noted that this study only applies
and knowledge. For example, if estimated emissions           to facility-scale MVL, and that additional analyses
exceed production, it may be possible to identify an         are needed to evaluate equipment-scale screening,
error and require additional measurements. How-              continuous monitoring, single-visit screening mod-
ever, complex follow-up schemes that are not codified        els, and other approaches to alternative LDAR.
can be difficult to regulate. It is important to note that
these results are interpreted in the context of LDAR.
                                                             4. Conclusion
Another option is attempting to predict vented emis-
sions from equipment inventories and operator logs
                                                             This study used LDAR-Sim—an agent based numer-
to calibrate high-venting sites and improve ranking.
                                                             ical LDAR model—to explore scenarios involving
MVL technologies may also provide additional value
                                                             screening technologies. Our results suggest that (a)
beyond the immediate goal of mitigating fugitive
                                                             emissions-reduction equivalence is possible between
emissions. For example, facility-scale emissions data,
                                                             SVL and MVL programs, and (b) cost-effective equi-
even if inaccurate, may be desirable for producers
                                                             valence scenarios exist but may be difficult to achieve.
for reporting, tracking progress, or estimating prob-
                                                             Circumstances that impact performance also impact
ability of non-compliance with facility-scale venting
                                                             cost-effectiveness, as more surveys are required to
limits. Some screening technologies can also produce
                                                             achieve equivalence. We have shown that these cir-
ancillary data for facilities, such as mapping products
                                                             cumstances include vented emissions, quantification
from vehicles, drones, and aircraft. Screening can also
                                                             error, and the empirical leak-size distribution used
be performed passively if measurement systems are
                                                             for modeling.
deployed on existing vehicles.
                                                                  Our simulations suggest the largest impediment
     Reducing quantification error for screening tech-
                                                             to cost-effective MVL is the confounding presence
nologies may also improve MVL but is a major chal-
                                                             of vented emissions. Placing tight regulatory focus
lenge. Atmospheric variability is a significant source
                                                             on leaks only, instead of all methane emissions
of error (Caulton et al 2018). Accurate wind meas-
                                                             from a facility, may limit uptake of screening. The
urement is also critical in emissions modeling. Some
                                                             primary issue is scale—to find leaks on a facil-
mobile methods have reduced quantification error
                                                             ity with extensive vented emissions requires fine-
by increasing measurement times (Brantley et al
                                                             scale measurements, which favours SVL. In terms
2014) or by installing wind sensors on site. How-
                                                             of chemistry and environmental impact, vented and
ever, increased measurement times or requiring wind
                                                             fugitive methane at any one facility are usually
measurements at every site come with a cost pen-
                                                             identical. Facility-scale screening technologies could
alty. Additional work will be required to understand
                                                             be more effective if the fugitive-vented emissions
whether the trade-off between these factors and quan-
                                                             dichotomy was replaced with a more general focus on
tification error can be sufficiently reconciled to enable
                                                             identifying and resolving the highest-emitting facil-
cost-effective screening. Possible workarounds may
                                                             ities. However, this shift may hinge on improve-
exist. For example, low-sensitivity screening methods
                                                             ments in the quantification accuracy of screening
may opt for τ = 1, eliminating the need for triaging
                                                             technologies.
and thus avoiding incorrect ranking. However, min-
imum detection limits must be consistent for screen-
ing methods to ensure that all facilities emitting above     Data availability statement
a specified amount receive follow-up. It should be
noted that this study did not consider systematic bias       Supplementary information contains: (a) all model
in quantification errors, which have been observed in        code, data inputs, and outputs; (b) a comprehensive

                                               10
Environ. Res. Lett. 16 (2021) 064077                                                                                          T A Fox et al

reproducibility guide to enable transparent reproduc-                   Fox T A, Barchyn T E, Risk D, Ravikumar A P and
tion of study results; (c) a sensitivity analysis for leak                   Hugenholtz C H 2019a A review of close-range and
                                                                             screening technologies for mitigating fugitive methane
production rate.
                                                                             emissions in upstream oil and gas Environ. Res. Lett.
    The data that support the findings of this study are                     14 053002
openly available at the following URL/DOI: https://                     Fox T A, Gao M, Barchyn T E, Jamin Y L and Hugenholtz C H
github.com/tarcadius/Fox_etal_2021_ERL.                                      2021 An agent-based model for estimating emissions
                                                                             reduction equivalence among leak detection and repair
                                                                             programs J. Clean. Prod. 282 125237
Acknowledgments                                                         Fox T A, Ravikumar A R, Hugenholtz C H, Zimmerle D,
                                                                             Barchyn T E, Johnson M R, Lyon D and Taylor T 2019 A
                                                                             methane emissions reduction equivalence framework for
We acknowledge Alberta Innovates for a Technology                            alternative leak detection and repair programs Elementa Sci.
Futures scholarship to TAF and NSERC for a Vanier                            Anthropocene 7 30
scholarship to TAF.                                                     Golston L M, Aubut N F, Frish M B, Yang S, Talbot R W,
                                                                             Gretencord C, McSpiritt J and Zondlo M A 2018 Natural gas
                                                                             fugitive leak detection using an unmanned aerial vehicle:
ORCID iD                                                                     localization and quantification of emission rate Atmosphere
                                                                             9 333
Thomas A Fox  https://orcid.org/0000-0002-0066-                        Jacob D J, Turner A J, Maasakkers J D, Sheng J, Sun K, Liu X,
4048                                                                         Chance K, Aben I, McKeever J and Frankenberg C 2016
                                                                             Satellite observations of atmospheric methane and their
                                                                             value for quantifying methane emissions Atmos. Chem.
References                                                                   Phys. 16 14371–96
                                                                        Johnson D, Heltzel R and Oliver D 2019 Temporal variations in
AER 2018 Directive 060 upstream petroleum industry flaring,                  methane emissions from an unconventional well site ACS
      incinerating, and venting (available at: https://static.aer.ca/        Omega 4 3708–15
      prd/2020-10/Directive060.pdf)                                     Johnson M R and Tyner D R 2020 A case study in competing
Alden C B, Wright R J, Coburn S C, Caputi D, Wendland G,                     methane regulations: will Canada’s and Alberta’s contrasting
      Rybchuk A, Conley S, Faloona I and Rieker G B 2020                     regulations achieve equivalent reductions? Elementa Sci.
      Temporal variability of emissions revealed by continuous,              Anthropocene 8 7
      long-term monitoring of an underground natural gas                Kemp C E, Ravikumar A P and Brandt A R 2016 Comparing
      storage facility Environ. Sci. Technol. 54 14589–97                    natural gas leakage detection technologies using an
Barchyn T E, Hugenholtz C H and Fox T A 2019 Plume detection                 open-source virtual gas field simulator Environ. Sci. Technol.
      modeling of a drone-based natural gas leak detection system            50 4546–53
      Elementa Sci. Anthropocene 7 41                                   Ravikumar A P, Roda-Stuart D, Liu R, Bradley A, Bergerson J,
Bell C S, Vaughn T, Zimmerle D, Helmig D and Lamb B 2020                     Nie Y, Zhang S, Bi X and Brandt A R 2020 Repeated leak
      Evaluation of next generation emission measurement                     detection and repair surveys reduce methane emissions over
      technologies under repeatable test protocols Elementa Sci.             scale of years Environ. Res. Lett. 15 034029
      Anthropocene 8 32                                                 Ravikumar A P, Sreedhara S, Wang J, Englander J, Roda-Stuart D,
Brandt A R, Heath G A and Cooley D 2016 Methane leaks from                   Bell C, Zimmerle D, Lyon D, Mogstad I and Ratner B 2019
      natural gas systems follow extreme distributions Environ.              Single-blind inter-comparison of methane detection
      Sci. Technol. 50 12512–20                                              technologies–results from the stanford/EDF mobile
Brantley H L, Thoma E D, Squier W C, Guven B B and Lyon D                    monitoring challenge Elementa Sci. Anthropocene 7 37
      2014 Assessment of methane emissions from oil and gas             Robertson A M, Edie R, Snare D, Soltis J, Field R A,
      production pads using mobile measurements Environ. Sci.                Burkhart M D, Bell C S, Zimmerle D and Murphy S M 2017
      Technol. 48 14508–15                                                   Variation in methane emission rates from well pads in four
Cardoso-Saldaña F J and Allen D T 2020 Projecting the temporal               oil and gas basins with contrasting production volumes and
      evolution of methane emissions from oil and gas production             compositions Environ. Sci. Technol. 51 8832–40
      sites Environ. Sci. Technol. 54 14172–81                          Schwietzke S, Harrison M, Lauderdale T, Branson K, Conley S and
Caulton D R et al 2018 Quantifying uncertainties from                        George F C 2018 Aerially-guided leak detection and repair: a
      mobile-laboratory-derived emissions of well pads using                 pilot field study for evaluating the potential of methane
      inverse Gaussian methods Atmos. Chem. Phys. 18 15145–68                emission detection and cost-effectiveness J. Air Waste
ECCC 2018 SOR/2018-66 regulations respecting reduction in the                Manage. Assoc. 69 71–88
      release of methane and certain volatile organic compounds         USEPA 2016 Oil and natural gas sector: emission standards for
      (upstream oil and gas sector) (available at: https://pollution-        new, reconstructed, and modified sources (available at:
      waste.canada.ca/environmental-protection-registry/                     www.govinfo.gov/content/pkg/FR-2016-06-03/pdf/2016-
      regulations/view?Id=146)                                               11971.pdf)
EDF 2016 Finding, fixing leaks is a cost-effective way to cut oil       Varon D J, Jacob D J, McKeever J, Jervis D, Durak B O A, Xia Y
      and gas methane emissions (available at: www.edf.org/sites/            and Huang Y 2018 Quantifying methane point sources from
      default/files/content/ldar_fact_sheet_updated_                         fine-scale satellite observations of atmospheric methane
      with_citations.pdf)                                                    plumes Atmos. Meas. Tech. 11 5673–86
Englander J G, Brandt A R, Conley S, Lyon D R and Jackson R B           Zimmerle D, Vaughn T, Bell C, Bennett K, Deshmukh P and
      2018 Aerial interyear comparison and quantification of                 Thoma E 2020 Detection limits of optical gas imaging for
      methane emissions persistence in the Bakken formation of               natural gas leak detection in realistic controlled conditions
      North Dakota, USA Environ. Sci. Technol. 52 8947–53                    Environ. Sci. Technol. 54 11506–14

                                                        11
You can also read