Why Low-Cost Air Quality Sensor Networks Fail, and How to Make Them Defensible

For most of the history of air quality monitoring, the constraint was cost. A reference-grade monitoring station is expensive to buy, expensive to house and expensive to run, so operators bought one, sited it where it was convenient, and accepted that a single point would have to represent an entire mine boundary, industrial estate or town. Low-cost optical sensors have removed that constraint. For the price of one reference station an operator can now deploy dozens of nodes and see the spatial structure of a dust or emissions problem for the first time.

That is a genuine advance, and it is why we deploy these networks. But it comes with a failure mode that is now widespread. A community group, a mine or a municipality installs a network, publishes the readings, and then finds the entire dataset dismissed the first time a regulator, a consultant or an opposing expert examines it. The sensors were not wrong so much as unaccompanied. Nobody built the scientific scaffolding that turns a raw signal into a defensible measurement.

This article sets out the four reasons these networks fail, and the practice that prevents each one.

01Failure one: drift, and the assumption of permanence

Optical particle sensors do not hold their calibration indefinitely. The light source ages, the detector response shifts, dust accumulates on internal optics and the relationship between the raw scatter signal and the reported mass concentration slowly changes. A node that was accurate at commissioning can be reading materially high or low a year later, and nothing in the data stream announces this. The readings still arrive, still look plausible and still populate the dashboard.

The practice that prevents it is scheduled co-location. Each node is placed alongside a reference or reference-equivalent instrument before deployment to establish its correction, and returned or revisited on a defined cycle to check whether that correction still holds. Drift is then a measured quantity with a documented history rather than an unknown. When somebody asks how you know the network was accurate in the third quarter, the answer is a calibration record, not an assurance.

02Failure two: humidity, and particles that are not there

Optical sensors size particles by how they scatter light. Water is very good at scattering light. As relative humidity rises, hygroscopic particles absorb water and swell, so the sensor sees larger particles and reports a higher mass concentration, even though the dry particulate mass has not changed at all. On a humid morning an uncorrected network can report an exceedance that is, in physical terms, fog.

This is the single most common reason sensor data is rejected. It is also entirely tractable. Correction models that use co-located humidity and temperature to adjust the reported concentration are well established in the literature, and applying one is standard practice in a properly run network. The point is not that the raw signal is useless. It is that the raw signal is not the measurement, and reporting it as though it were invites a challenge that will succeed.

An uncorrected sensor reading is an observation about light. A corrected, calibrated, uncertainty-bounded reading is an observation about air. Only one of them survives cross-examination.

03Failure three: siting that answers the wrong question

A network is only as good as its geometry. Nodes clustered where power and mounting points happened to be convenient will produce a great deal of data about a few arbitrary locations. Nodes placed without reference to the prevailing wind will miss the plume entirely for most of the year, or sit permanently inside it and misrepresent the general condition.

Good siting starts with the wind rose and the source inventory, not the site plan. Where are the sources, where does the air go, which receptors matter, and what is the question the network exists to answer? Boundary compliance, community exposure, source apportionment and control verification are four different questions, and they imply different node placements. Deciding the question first, then designing the geometry around it, is what separates a monitoring network from a collection of sensors.

• Upwind nodes to establish the incoming background against which the site is judged
• Downwind nodes on the receptor bearings that actually matter for the community
• Source-proximate nodes where apportionment or control verification is the objective
• At least one reference-grade anchor point for ongoing correction of the wider array
• Co-located meteorology, because a concentration without a wind vector explains nothing

04Failure four: treating precision as accuracy

Low-cost sensors are often highly precise and only moderately accurate. They will report a value to two decimal places, every minute, with impressive internal consistency, while sitting some distance from the true concentration. Precision creates a powerful impression of rigour, and it is easy to publish a dashboard that looks authoritative while the underlying uncertainty is never stated.

The discipline here is the same one that governs all exposure assessment. A reported concentration should carry its uncertainty, and a comparison against a guideline should be expressed as a probability rather than a verdict. Saying that the 24 hour mean was 48 micrograms per cubic metre against a guideline of 50, and stopping there, is not an honest account if the measurement uncertainty is wide enough to place the true value on either side of the line. Stating the interval, and the confidence with which the guideline was or was not exceeded, is both better science and a far stronger position to defend.

05What a well built network makes possible

It is worth being clear that the answer to these failure modes is not to abandon low-cost sensing and return to a single reference station. Density genuinely reveals things that a single point cannot. A properly designed network resolves the gradient across a boundary, identifies which of several sources is driving an exceedance, distinguishes a haul road from a tailings beach, shows when in the day and in what wind conditions the problem occurs, and gives a community something more useful than an annual average.

That resolution feeds directly into the rest of the exposure science. Network output validates and refines dispersion modelling. It supports source apportionment. It provides the continuous record against which a control measure can be shown to have worked. In our practice it also feeds Risk Exposome360°, where ambient data sits alongside occupational and governance signals in a single computed picture of exposure risk.