Industries / Industrial equipment and services / Industrial services / Leak detection
Gas imaging cameras
Gas imaging cameras are infrared cameras tuned to wavelengths that hydrocarbon gases absorb, so that a leak, invisible to the eye, shows on the screen as moving smoke; the method is called optical gas imaging (OGI).
7
Milestones achieved, from 1985 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1985
Lasers make gas visible
Thomas G. McRae Jr. of Lawrence Livermore National Laboratory (LLNL) filed a patent in April 1983, granted on 26 November 1985, for what became backscatter absorption gas imaging (BAGI): a laser tuned to a wavelength the gas absorbs sweeps the scene, a detector records the light that bounces back, and a gas cloud shows on the video picture as a shadow. By 1993 Thomas Kulp and colleagues at LLNL reported BAGI instruments working at 6 to 110 m, using carbon dioxide lasers at 9 to 11 micrometers (µm) for many organic vapors or infrared helium-neon lasers at 3.39 µm, used mainly to locate methane leaks, and were building a 300 m version for imaging from aircraft.
Credit
Thomas G. McRae Jr., Thomas J. Kulp and colleagues, LLNL
2005
Leaks seen without a laser
On 8 June 2005 FLIR Systems announced GasFindIR, a passive infrared camera, one that needs no light source of its own, filtered to a narrow band where hydrocarbon vapors absorb; FLIR called it the first commercially viable infrared camera able to detect volatile organic compounds (VOCs). Battelle tested it for EPA's Environmental Technology Verification program in late 2008, in a BP laboratory at Naperville, Illinois, and at Dow Chemical plants in Freeport, Texas. In the laboratory the smallest leak it showed ranged from 0.28 g/h of propane or pentane in still air to more than 70 g/h of ethylene in a 5 mile an hour (8 km/h) wind, and distance and wind both raised the limit. Battelle noted that a FLIR employee ran the camera, that the two other observers sometimes disagreed with the operator, and that the camera, priced at $64,950, was not intrinsically safe for use in explosive atmospheres.
LessonA gas camera has no single detection limit; it depends on the gas, the distance, the wind and the person looking.
Credit
FLIR Systems; Battelle, for EPA
2008
The camera allowed by rule
EPA's final rule of 22 December 2008 let plants covered by its equipment leak rules find leaks with an OGI camera instead of a Method 21 analyzer, the portable probe instrument. The camera must be checked each day against a known flow of gas from a cylinder, video of that check and of each survey must be kept for 5 years, and the final rule added one Method 21 survey a year. Commenters had argued that the camera could not see leaks under 60 g/h and that side-by-side comparisons showed it found fewer leaks; EPA based its finding that the two programs were equivalent on a Monte Carlo model of gas plant data from the early 1990s, because the 25-year-old data behind the original rules was no longer available.
LessonEPA allowed the new method but kept the old one as a yearly check while its record on small leaks was disputed.
Credit
US EPA
2016
Camera surveys become required
EPA's rule of 3 June 2016 (subpart OOOOa) required new and modified US oil and gas well sites to be surveyed for leaks twice a year and compressor stations four times a year, using OGI. Any gas the camera shows must be repaired within 30 days, and the repaired part checked again within 30 days, with OGI or with Method 21 at 500 ppm, which may also be used instead of the camera.
Credit
US EPA
2018
Distance sets the limit
A team led by Arvind Ravikumar and Adam Brandt ran single-blind tests of a FLIR camera on controlled methane releases and found a median detection limit, the leak seen half the time, of about 20 g/h at 6 m, an order of magnitude above the 1.4 g/h estimated earlier. Both the median and the 90% limits rose with imaging distance following a power law. They also modeled that once a survey's median limit reaches about 100 g/h, better sensitivity adds almost nothing to what periodic surveys can cut.
Credit
Arvind Ravikumar, Jingfan Wang, Adam Brandt and colleagues, in Environmental Science & Technology, January 2018
2020
The surveyor matters
Daniel Zimmerle and colleagues at CSU ran 488 tests over 10 months at an 8-acre (3.2 ha) outdoor site built to resemble field gas operations, with professional surveyors from 16 oil and gas companies and 8 regulators using their own cameras and methods. Detection rates were much lower than in earlier studies that focused on the camera: the leak needed for a 90% chance of detection was an order of magnitude larger. Surveyors from companies that had surveyed more than 551 sites found 1.7 times as many leaks as the others, and the experienced ones slowed down and looked from more viewpoints.
LessonA camera survey is only as good as its operator, so tests of gas cameras have to include the people who use them.
Credit
Daniel Zimmerle, Timothy Vaughn, Clay Bell and colleagues, in Environmental Science & Technology, August 2020
2024
Camera leak rates checked
Quantitative OGI (QOGI) software estimates how fast a leak is escaping from the camera's video. In single-blind tests at an outdoor test site, Chiemezie Ilonze and colleagues took 357 measurements of 26 controlled releases, between 0.1 and 2.9 kg/h of methane in compressed natural gas, with the FLIR QL320 system. 75% were within a factor of 3 of the true rate, single errors ran from −90% to +831%, and averaging the views from several camera positions narrowed the range to −79% to +297%. Results were better against a clear sky and in winds of 1 mile an hour (1.6 km/h) or less.
LessonOne camera estimate of a leak rate can be off by a factor of ten, so estimates are better taken from several positions and treated as ranges.
Credit
Chiemezie Ilonze, Jingfan Wang, Arvind Ravikumar and Daniel Zimmerle, in Sensors, June 2024
Watch for
AI inside a handheld camera
Raythink Technology of Yantai, China, says its RG630 handheld OGI camera recognizes methane and other gases with built-in AI, colors each plume by estimated concentration, sounds an alarm, and meets the camera requirements of EPA's Appendix K survey protocol.
Achieved whenan independent controlled-release test publishes the camera's detection rates with its AI alarm and without it, beside a trained surveyor using the same camera.
Claimed by
Raythink, March 2026
Open
01
AI and surveyors, same clips
A public reading test in which certified OGI surveyors and AI leak detectors watch the same blind video clips, with and without leaks of known size, so that people and software can be compared directly.
The briefCredit · unclaimed02
Check what the camera saw
Software that reads the video of an OGI survey and the site's list of components and reports, for each component, how long it was held in view and from how many angles, flagging any that were missed or watched for less than Appendix K requires: two angles, and 10 seconds per angle for a scene of 10 or fewer components or 2 seconds per component in a busier scene.
The briefCredit · unclaimed
Open 01 · the brief
AI and surveyors, same clips
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public reading test in which certified OGI surveyors and AI leak detectors watch the same blind video clips, with and without leaks of known size, so that people and software can be compared directly.
Why it is still open
Stanford's GasNet work trained AI on about 1 million frames of leak video recorded at METEC, but its authors wrote that they "cannot draw a conclusion" on how it compares with people, because the surveyors in earlier tests used a more sensitive camera mode under other conditions. EPA has since approved fixed, automated OGI cameras whose deep-learning software flags leaks, such as CleanConnect's LeakFinder in May 2025, rated to detect 5 kg/h 90% of the time, and Raythink sells AI in a handheld camera, but searches in November 2026 found no comparison of software and surveyors on the same clips.
Built with what exists
GasVid, the Stanford team's labeled METEC video set, if its authors share it, or new clips recorded at a controlled-release test site; the published GasNet and VideoGasNet models and open plume-segmentation code such as Gasformer on GitHub; a web page that plays clips in random order and logs each answer; ten surveyors trained to Appendix K.
First step
Ask the Stanford authors for GasVid, cut 50 held-out clips, and have two volunteers and GasNet read them.
Done looks like
400 clips of 10 seconds, 200 with leaks from 5 to 2,000 g/h at 4.6 to 15.6 m and 200 without, each read by 10 surveyors and at least 3 models; detection rates by leak size and distance, with 95% confidence intervals, for each reader; clips, answers and code published.
Open 02 · the brief
Check what the camera saw
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Software that reads the video of an OGI survey and the site's list of components and reports, for each component, how long it was held in view and from how many angles, flagging any that were missed or watched for less than Appendix K requires: two angles, and 10 seconds per angle for a scene of 10 or fewer components or 2 seconds per component in a busier scene.
Why it is still open
Appendix K asks each operator to record a 5-minute check video each survey day showing dwell times and angles, and the Clean Air Task Force's June 2026 survey guidance leaves the final check, that every component on the route map was surveyed, to people reviewing logs. Opgal advertises GPS route maps to "confirm complete component coverage", Inspectionlogic holds a 2019 US patent for prompting a Method 21 operator to hold the probe for the required time, and EPA's approved Xplorobot method logs each component by computer vision for its own laser tool, but searches in November 2026 found no open checker for any camera's video, measured against a hand-coded answer.
Built with what exists
A mock pad of 60 unpressurized scrap valves, flanges and fittings; any OGI camera or a thermal camera with a visible-light channel, since no gas is needed to test coverage; Ultralytics YOLO models fine-tuned to find components; OpenCV for tracking; CVAT for hand-coding the answer.
First step
Set out 20 scrap components in a yard, record three surveys with a thermal camera, and hand-code each component's dwell time.
Done looks like
30 recorded surveys, each with 5 components skipped on purpose; dwell time within 2 seconds of the hand-coded answer for 95% of component views; at least 95% of skipped or short-viewed components flagged, with no more than 5% false flags; code and videos published.