Industries / Energy, water and waste / Environmental services / Air quality monitoring
Reference stations
Reference stations are fixed sites where air agencies run tested and certified instruments around the clock, and their readings are the ones that laws, court cases and health warnings rely on.
7
Milestones achieved, from 1952 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1952
Smog traced to exhaust
Los Angeles had its first recognized smog episodes in the summer of 1943. In June 1952 the chemist Arie Haagen-Smit published "Chemistry and Physiology of Los Angeles Smog", and the California Air Resources Board credits him with discovering the nature and causes of photochemical smog, smog made when sunlight acts on hydrocarbons from gasoline and on nitrogen oxides, both in car exhaust.
LessonThe city first blamed a nearby butadiene plant, but the smog stayed after the plant shut, so the cause had to be measured rather than assumed.
Credit
Arie Haagen-Smit
1961
A national survey begins
In 1961 the UK set up the National Survey, which measured black smoke and sulfur dioxide at around 1,200 sites. Defra, the UK environment department, calls it the world's first coordinated national air pollution monitoring network, and its records show both pollutants falling steeply as cleaner fuels and new laws took hold.
LessonAs smoke and sulfur dioxide fell, the work moved to traffic pollutants: the UK started an automatic urban network in 1987, and from 2008 it could run fewer sulfur dioxide and carbon monoxide monitors.
Credit
the UK government
1970
Light measures nitrogen oxides
In May 1970 Arthur Fontijn, Alberto Sabadell and Richard Ronco described measuring nitric oxide by the faint light it gives off when it reacts with ozone, which they put forward for continuous monitoring of air pollutants. In 1976 the US Environmental Protection Agency (EPA) made this "gas phase chemiluminescence" the principle of its reference method for nitrogen dioxide, citing their paper first: the analyzer turns nitrogen dioxide into nitric oxide in a converter, measures the light, and subtracts the nitric oxide already present.
LessonThe converter also turns some other nitrogen compounds, such as peroxyacetyl nitrate (PAN), into nitric oxide, so the method can count them as nitrogen dioxide, and EPA recommends another method where they are high.
Credit
Arthur Fontijn, Alberto Sabadell and Richard Ronco; EPA
1997
Fine particles get a standard
EPA's first national standards for particles, published on 30 April 1971, covered total suspended particulates, everything the sampler caught. In 1987 EPA switched to PM10, particles 10 micrometers across or smaller, and on 18 July 1997 it added standards for PM2.5, particles 2.5 micrometers or smaller: 15.0 micrograms per cubic meter as an annual average and 65 over 24 hours. Later reviews tightened PM2.5 to 35 over 24 hours (2006), 12.0 a year (2013) and 9.0 a year (2024), and on 26 June 2026 the federal appeals court in Washington upheld the 9.0 limit after EPA itself had asked the court to strike it down.
Credit
EPA; the US Court of Appeals for the D.C. Circuit
2016
Cotton in the samplers
In February and March 2016, staff of two district environmental protection bureaus in Xi'an, China, entered air quality stations with a copied key and stuffed cotton gauze into the air samplers to lower the readings. Experts at the national environment ministry's air monitoring department in Beijing noticed abnormal changes in Xi'an's data that March, and in June 2017 a court sentenced seven people, including both bureau directors, to between 15 and 22 months for falsifying computer data.
LessonRemote review of the data stream caught tampering that happened inside locked stations.
Credit
China's Ministry of Environmental Protection (detection); a Xi'an court
2024
Equivalent monitor reads high
In July 2016 EPA approved Teledyne's T640 and T640X, continuous optical particle monitors, as federal equivalent methods (FEMs), instruments accepted as matching the federal reference method (FRM), which weighs particles caught on a filter. States ran about 400 of them by 2023, and agencies found they read about 20% higher than filter samplers beside them, and more in wildfire smoke. In May 2024 EPA applied a correction to all T640 PM2.5 data in its national database back to 2017. In March 2026 the Association of Air Pollution Control Agencies, which represents 53 state and local agencies, wrote that the correction "did not adequately reduce the bias" and that agencies are invalidating T640 data and moving away from the instruments.
LessonApproval as "equivalent" did not guarantee equal readings in routine use; filter samplers run side by side and independent audits are what showed the gap.
Credit
state and local monitoring agencies and EPA's Performance Evaluation Program, which found the bias; Teledyne and EPA, which wrote the correction
2024
Europe tightens toward WHO
On 22 September 2021 the World Health Organization (WHO) lowered almost all its air quality guideline levels. The EU's recast Ambient Air Quality Directive (2024/2881), published on 20 November 2024, moves EU law closer: by 1 January 2030 the annual limit for PM2.5 falls from 25 to 10 micrograms per cubic meter, and for nitrogen dioxide from 40 to 20. It also requires "monitoring supersites", at least one urban background station per 10 million people that also measures ultrafine particles, black carbon and ammonia, and gives people a right to claim compensation when authorities, on purpose or through negligence, break its rules on air quality plans and short-term action plans and that harms their health. Member states must bring it into national law by 11 December 2026.
Credit
the European Parliament and Council; WHO
Watch for
AI checks station data
Station readings are validated by people before they become official. ENVEA, a French maker of monitoring systems, presented an AI "pre-validation" module for its ADVAL software, built on Microsoft's open-source LightGBM library and trained on 300 million data points, with the goal of validating 90% of data automatically so that people review only 10%. In its own test on nitric oxide data, the model marked as invalid 87% of the hours that human validators had invalidated.
Achieved whenan air agency publishes results from routine use, or an outside group scores the module against human validators' decisions on data it was not trained on.
Claimed by
ENVEA, in slides posted in November 2025
India's 40% target
India's National Clean Air Programme (NCAP) asks 130 cities to cut PM10 by up to 40% by 2025–26 against 2017–18. In a written reply to the Lok Sabha, reported on 1 December 2025, Minister of State for Environment Kirti Vardhan Singh said 103 of the 130 cities had lower PM10 in 2024–25 than in 2017–18, 64 of them by more than 20% and 25 by more than 40%. The Centre for Research on Energy and Clean Air (CREA) found in January 2026 that 28 of the 130 cities still had no continuous monitoring station, and that of 100 cities with enough data, 23 had met the 40% target and 23 had higher PM10 than at the start. The Health Effects Institute counted continuous stations rising from 78 in 38 cities in 2017 to 346 in 101 cities in 2024.
Achieved whenfinal 2025–26 figures from stations with complete data are published and an independent review reaches the same count of cities meeting the target.
Claimed by
India's Ministry of Environment, Forest and Climate Change, December 2025
Open
01
A public test for validation
An open benchmark for the hand-checking of station data. Take hourly readings from European reference stations twice, as first reported in near real time and as later validated by the network's staff, and mark every hour the validators deleted or changed. Publish the paired data, a scoring script and a baseline model, so that any automatic checker can be scored against human decisions on a year it has not seen.
The briefCredit · unclaimed02
Check station siting from above
A remote check of US monitoring sites against the federal siting rules. For each site, measure from lidar, aerial photos and street-level images the distance from the air inlet to tree driplines (the edge of the canopy) and to nearby obstacles, and flag sites that appear to break the rules, for the agency to verify on the ground.
The briefCredit · unclaimed
Open 01 · the brief
A public test for validation
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open benchmark for the hand-checking of station data. Take hourly readings from European reference stations twice, as first reported in near real time and as later validated by the network's staff, and mark every hour the validators deleted or changed. Publish the paired data, a scoring script and a baseline model, so that any automatic checker can be scored against human decisions on a year it has not seen.
Why it is still open
Validation is still largely manual; ENVEA's March 2026 slides say it "requires several hours of meticulous analysis each day" plus a weekly review. ENVEA's AI module is trained on its partners' data and scored only by ENVEA. Empa's GAW-QC dashboard, published in Atmospheric Chemistry and Physics in October 2025, flags outliers in four gases without using validators' decisions as labels. Météo-France's team for the Copernicus Atmosphere Monitoring Service used both the near-real-time and the validated European data in February 2026, but to classify stations, not to learn validators' edits. A search found no public benchmark of this kind.
Built with what exists
The European Environment Agency's air quality download service, which serves station data as Parquet files; Python with pandas or DuckDB; LightGBM, the open-source library ENVEA's module uses, as a first baseline; Zenodo or Hugging Face to host the data.
First step
Download one country's near-real-time and validated nitrogen dioxide data for the same year, line up the hours, and count how many the validators removed or changed.
Done looks like
Paired near-real-time and validated hourly data for nitrogen dioxide, ozone, PM10 and PM2.5 from at least 500 stations over at least two years, with every changed or deleted hour labeled; a held-out final year; a baseline whose precision and recall on invalid hours are published; and a score from at least one outside team.
Open 02 · the brief
Check station siting from above
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A remote check of US monitoring sites against the federal siting rules. For each site, measure from lidar, aerial photos and street-level images the distance from the air inlet to tree driplines (the edge of the canopy) and to nearby obstacles, and flag sites that appear to break the rules, for the agency to verify on the ground.
Why it is still open
The rules say an inlet must be at least 10 m from the dripline of trees, at least twice an obstacle's protruding height away from that obstacle, and have unobstructed airflow in an arc of at least 270 degrees. Each state's annual network plan must state whether every monitor meets them, and auditors check by hand: the California Air Resources Board's site survey procedure (revision 4, June 2020) has the auditor inspect the site and measure with a 50 m tape and a rangefinder. A search found no remote or automated check.
Built with what exists
EPA's AirData site list with coordinates; US Geological Survey 3D Elevation Program lidar; USDA NAIP aerial photos; the Google Street View Static API; PDAL and QGIS for point clouds; an open object detector such as Grounding DINO to find inlet masts in images.
First step
Pick 20 sites in one state whose annual network plan reports tree distances, and measure the same distances from USGS lidar.
Done looks like
Every ozone and PM2.5 site in one EPA region measured; for at least 50 sites with on-site figures in network plans or audit reports, the remote distances agree within 2 m for at least 90%; and the list of flagged sites published with images.