Industries / Energy, water and waste / Environmental services / Air quality monitoring
Low-cost and mobile sensors
Low-cost and mobile sensors spread measurement across a city, either as small devices, often a few hundred dollars or less, that estimate pollution with simple light-scattering or chemical parts, or as instruments carried by cars and people, so that readings come from many places instead of one reference station.
6
Milestones achieved, from 2013 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2013
Gas cells sense trace levels
Electrochemical cells are small sensors in which a gas reacts at an electrode and makes an electric current, and they had been used to measure gases in parts per million. A University of Cambridge team, with authors from the sensor maker Alphasense and Cambridge City Council, showed that when suitably configured and operated they can measure the parts-per-billion levels of carbon monoxide, nitric oxide and nitrogen dioxide found in city air. They ran a fixed network around Cambridge and handheld units carried by pedestrians and cyclists, and the readings varied so much that the authors concluded city air "cannot be fully characterised using sparse, static networks".
Credit
M. I. Mead, R. L. Jones and colleagues, University of Cambridge; Alphasense
2014
A public test bench
In 2014 the South Coast Air Quality Management District, a California air agency, set up AQ-SPEC to tell the public how commercial "low-cost" sensors perform in real use. In one early field test, from June to August 2015, three Air Quality Egg monitors costing about $200 each ran beside the agency's reference instruments at Rubidoux. Particle data from all three were invalid and carbon monoxide readings were out of range. Nitrogen dioxide matched the reference poorly (R², a 0-to-1 measure of how well two sets of readings rise and fall together, of 0.34 to 0.40) and read high, while ozone followed the reference closely (R² about 0.85) but read well below it.
LessonA sensor can rise and fall with the true level and still give the wrong number, so a test has to report bias as well as correlation.
Credit
South Coast AQMD
2015
Citizens build a dust network
Stuttgart often broke European limits for fine particles, yet the city had only four official stations and their filter-weighed results took four to six weeks to appear, according to OK Lab Stuttgart, a volunteer civic-tech group. From early 2015 its members built a particle sensor kit costing about 30 to 40 euros. A crowdfunding campaign started in January 2016 to pay for 300 sensors reached its goal in March, and by 2019 the group counted more than 8,700 sensors worldwide in its project, luftdaten.info. In 2018 researchers at the Karlsruhe Institute of Technology and the TROPOS institute, studying the project's SDS011 laser sensor, wrote that earlier side-by-side tests had shown good agreement with reference instruments at 20% to 50% relative humidity and low particle levels, but worse results in other conditions, especially high humidity.
LessonCheap light-scattering sensors need checking and correction for humidity before their numbers are set beside official ones.
Credit
Jan Lutz, David Lackovic and OK Lab Stuttgart (Code for Germany); Matthias Budde and colleagues
2017
Street-by-street pollution maps
A team led by Joshua Apte of the University of Texas at Austin, with co-authors from Aclima and the Environmental Defense Fund, fitted Google Street View cars with fast-response instruments and drove every street in a 30 km² area of Oakland, California, again and again. Their maps of yearly daytime nitric oxide, nitrogen dioxide and black carbon (soot), at a 30 m scale, showed stable patterns, with levels differing 5 to 8 times within single city blocks.
LessonPollution can change sharply within a block, so one central station can misstate what people on a given street breathe.
Credit
Joshua Apte and colleagues; Aclima; Environmental Defense Fund
2021
Targets for sensor makers
In February 2021 the US Environmental Protection Agency (EPA) published test protocols and target values for fine-particle and ozone sensors, and in February 2024 added coarse particles (PM10), nitrogen dioxide, carbon monoxide and sulfur dioxide, all for "non-regulatory supplemental and informational monitoring". In Europe the standards body CEN issued CEN/TS 17660-1 in 2021, a test program that sorts gas sensor systems into classes consistent with the EU's "indicative measurements" and "objective estimation", categories of data that EU law holds to looser accuracy than fixed measurements: for nitrogen dioxide, an uncertainty of 25% and 75% respectively, against 15% for fixed stations. The UK's MCERTS certification scheme now requires that test program for low-cost gas sensor systems.
LessonEPA says sensors on the market "often have performance issues or trouble measuring these pollutants", and its testing is voluntary, with no certification or endorsement.
epa.gov19january2025snapshot.epa.govevs.eepublications.europa.eugov.uk
Credit
US EPA; CEN; UK MCERTS
2023
A city network switched off
Project Eclipse, run by Microsoft's research arm in partnership with the City of Chicago, placed air sensors on bus shelters across the city from July 2021. In April 2023 Microsoft ended it and said the sensors would be taken down. The city's health department said it could not carry the work on "because it doesn't own the intellectual property for the air sensors". Two years of data stayed online.
LessonA city that does not own its sensors' hardware and software can lose its network when a partner's pilot ends.
Credit
Microsoft Research; Chicago Department of Public Health; reported by Dillon Bergin for MuckRock and WBEZ
Watch for
AI calibration for Africa
AirQo, which has run sensor networks in 14 cities, says it is scaling up models trained "to periodically sharpen the raw sensor data to match the reference-grade data", starting in Kampala, Uganda, and Nairobi, Kenya, and that it is building a public registry of reference monitors across the continent. The post gives no error figures.
Achieved whenerror figures for the calibrated sensors, checked against reference monitors whose data were not used to train the models, are published and confirmed by someone outside AirQo, such as a city air agency, and the registry is public.
Claimed by
Deo Okure of AirQo, on the Clean Air Fund's site, July 2026
Open
01
Test NO2 sensors to targets
An independent, public test of at least five commercial nitrogen dioxide (NO2) sensor systems, run to EPA's February 2024 base-testing protocol beside a regulatory NO2 monitor, with results reported in EPA's fillable templates and all raw data published.
The briefCredit · unclaimed02
Score sensor siting from photos
A siting score for citizen air sensors, made by applying a ten-item checklist drawn from EPA's siting guide (height, free airflow, nearby vents, doors, walls, grills) to photos supplied by sensor owners, and a test of whether the score explains why a sensor disagrees with its neighbors.
The briefCredit · unclaimed
Open 01 · the brief
Test NO2 sensors to targets
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent, public test of at least five commercial nitrogen dioxide (NO2) sensor systems, run to EPA's February 2024 base-testing protocol beside a regulatory NO2 monitor, with results reported in EPA's fillable templates and all raw data published.
Why it is still open
EPA's testing is voluntary and carries no certification, and EPA says sensors on the market often have trouble measuring these gases. The only report found that scores a sensor against EPA's NO2 targets is a maker's own: QuantAQ wrote in April 2024 that its MODULAIR "meets or exceeds" the targets for carbon monoxide and NO2. South Coast AQMD's AQ-SPEC table gives NO2 field results under its own protocol, with a field R² as low as 0.02 to 0.29 for one device.
Built with what exists
Three units each of five or more NO2 sensor systems on sale today; space beside a state or local agency's NO2 reference analyzer; EPA's 2024 supplemental report for NO2, carbon monoxide and sulfur dioxide sensors, with its checklists and templates; Python or R for the metrics.
First step
Read the base-testing checklist in EPA's 2024 supplemental report, then ask one air agency for space and data access beside its NO2 analyzer.
Done looks like
A filled EPA template for every model showing, target by target, whether it was met, with raw data and code public, and the test repeated at a second site for at least two models.
Open 02 · the brief
Score sensor siting from photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A siting score for citizen air sensors, made by applying a ten-item checklist drawn from EPA's siting guide (height, free airflow, nearby vents, doors, walls, grills) to photos supplied by sensor owners, and a test of whether the score explains why a sensor disagrees with its neighbors.
Why it is still open
EPA's guide recommends a height of 3 to 6 ft and free airflow and says photos "may assist you with data interpretation later", but sets no distances from walls or vents. At EPA's 2023 sensor quality workshop, researchers from Washington University in St. Louis asked how regulatory siting rules "map onto low-cost sensor (LCS) devices", found particle data at one residential site insensitive to obstructions but biases from mounting in more complex settings, and listed guidance as a possible later step. Two searches found no study that scores citizen sensors' siting from photos.
Built with what exists
Photos volunteered by PurpleAir or Sensor.Community hosts, with their consent; a general AI model that reads images and answers the checklist; the PurpleAir API or the Sensor.Community data archive for readings; Python.
First step
Write the checklist from EPA's siting guide and ask 30 sensor hosts in one city, through a community forum, for two photos each.
Done looks like
300 sensors scored, with the model's checklist answers matching two trained human raters on at least 90% of items, and a published analysis, with data and code, of how much the score explains each sensor's average difference from its three nearest neighbors.