Skip to content
Method

Provenance: how to audit an environmental sensor reading

The questions to ask any supplier before you cite their environmental data: instrument, timing, calibration chain, uncertainty, operating envelope, and reproducibility.

You are about to put a number in a business case. It says 34 µg/m³, or a condition score, or a heat loss ranking on a particular building. Someone senior will eventually ask where it came from, and someone hostile may ask later still.

A value and a value you can defend are different objects. The difference is not the sensor. It is whether the number arrives with enough attached to it that a third party can reconstruct how it came to exist.

What follows is a set of questions you can take to any supplier of environmental data. They apply to reference networks, to satellite products, to low-cost sensor fleets, and to us.

1. What measured this — an instrument, or a model?

The first thing to establish is whether you are looking at an observation or an estimate. Both are legitimate. Confusing them is not.

The international observation model, ISO 19156:2023 (second edition, April 2023, replacing the 2011 version), is built around exactly this distinction: every observation names a procedure, an observed property, a feature of interest and a result. If a supplier cannot tell you which of their values are measured, which are interpolated between measurements, and which are modelled from other datasets entirely, the layer is not auditable, however good the map looks.

Ask for the split, per value. Not per dataset.

2. When did it happen — and when was it recorded?

These are two different times and good systems carry both. The OGC SensorThings API v1.1 separates phenomenonTime (when the thing being measured actually occurred) from resultTime (when the result was produced), and adds validTime for how long a result should be treated as current.

This matters more than it sounds. A reading reprocessed under a revised calibration six months later has a new resultTime and the same phenomenonTime. If your supplier's system cannot express that, you cannot tell a corrected number from a new one — and you cannot explain, a year on, why the figure in your report no longer matches the figure in the system.

3. Where was it, and how well is that known?

Position has its own uncertainty, and for mobile measurement it is often the dominant one. A reading placed on the wrong side of a junction is wrong in a way no amount of sensor accuracy fixes.

Ask what the positional uncertainty is, how it is derived, and what happens in the places where it degrades — urban canyons, tunnels, dense tree cover. Ask what feature the reading is attached to, and whether that attachment is asserted or inferred.

4. Calibrated against what, and how recently?

This is the question most often answered with a brochure. The right answer has a shape.

Metrological traceability is defined in the international vocabulary of metrology as the property of a result that can be related to a reference "through a documented unbroken chain of calibrations, each contributing to the measurement uncertainty" (VIM 2.41). Unbroken and documented are the load-bearing words. UKAS TPS 41 (edition 6, December 2022) sets out how that is interpreted in the UK.

So: what reference, held by whom, calibrated by whom, to what uncertainty, on what date. Then the harder follow-up — how does the calibration survive contact with the field?

For low-cost air quality sensing in particular, the honest literature is unambiguous that it often does not. In a nine-month network study, deSouza et al. (2022, Atmospheric Measurement Techniques) found that calibrations derived from short co-location periods "were not transferable to other time periods, because the conditions during the co-location were not representative of broader operating conditions" — and that more complex correction models which performed best at the co-location site did not necessarily transfer to the rest of the network, in some cases performing worse than no correction at all.

The practical questions that follow: how long was the co-location, in what season, against what reference, and was the correction validated at sites that were held out of fitting? A supplier who co-located once, in one place, in one season, has a number that is defensible there and then.

5. What is the stated uncertainty, and how was it derived?

"Accurate to within X" is not an uncertainty statement. A usable one names a coverage interval and a method.

The GUM (JCGM 100:2008) is the reference: a combined standard uncertainty multiplied by a coverage factor, conventionally k = 2 for approximately 95 % confidence. Regulatory air quality uses the same convention. The data quality objectives in Annex I of Directive 2008/50/EC — which UK assessment still works to, via regulation 7 of the Air Quality Standards Regulations 2010 — cap uncertainty at 95 % confidence at 15 % for fixed measurements of SO₂, NO₂, NOₓ and CO, 25 % for PM and lead, and loosen to 25 %, 50 % and 30 % respectively for indicative measurements. Objective estimation is allowed 75–100 %.

Those tiers are useful to you as a buyer even outside a statutory context. They tell you the class of claim being made. CEN/TS 17660-1:2021 classifies sensor systems for gaseous pollutants against those same tiers using prescribed laboratory and field tests, and adds a relaxed class for non-regulatory use. CEN/TS 17660-2:2024 does the same for particulate matter. Ask which class a system has been tested to, by whom, and whether the test report is available — not whether the word "compliant" appears in the marketing.

6. What happens to readings taken outside the operating envelope?

Every instrument has conditions in which it should not be believed. Optical particle counters at high humidity. Thermography without a sufficient indoor–outdoor temperature difference, in sunlight, or on a wet façade — the reason BS EN ISO 6781-1:2023 specifies conditions and operator competence rather than just a camera. A road survey at the wrong speed.

Ask three things. What is the envelope, stated numerically. What happens to a reading taken outside it — is it flagged, corrected, downweighted, or silently kept. And can you filter on that flag yourself. A dataset that quietly interpolates over its own bad conditions will look better and be worth less.

7. What quality regime sits behind the supplier, not just the sensor?

Instrument certification and organisational competence are separate things. UK highways offers a clean example of the second: SCANNER survey vehicles must hold a valid accreditation certificate obtained through annual tests, and supplier data is quality assured annually by TRL under the Road Condition Management Group. The same guidance states plainly that SCANNER surveys cannot identify potholes as a specific defect — and that regime is now changing. That is what a mature regime looks like: recurring external checks, and published limitations.

For ambient particulates, the equivalent instrument-level route is MCERTS certification against the Environment Agency's performance standards for indicative ambient particulate monitors, assessed by an accredited certification body.

Ask which regime applies, whether the supplier is in it, and if not, what they do instead.

8. Can you get from the number back to the instrument?

The final test is reproducibility. Take one value out of the API. Can you walk it back to the sensor unit, its calibration record, the pass on which it was captured, the corrections applied, and the version of the processing that produced it?

The vocabulary for this already exists. ISO 19115-1:2014 carries lineage as sources and process steps; ISO 19157-1:2023 covers how data quality is described and reported; and W3C PROV-O (Recommendation, 30 April 2013) gives a general model of Entity, Activity and Agent for recording what was derived from what, by which process, under whose responsibility. SensorThings carries resultQuality on the Observation itself.

None of these are exotic. If a supplier's answer is "we can export a spreadsheet", the chain does not exist — it is being reconstructed for you on request, which is not the same thing.

Where Make Sense stands on its own questions

We think a buyer should apply all eight to us, and we would rather write them down than be asked.

Make Sense is building a vehicle-mounted multi-sensor pod for air quality, road surface condition and building heat loss. The first vehicle system is in build. Nothing is operating, so there is nothing measured to report. We publish no accuracy figures, because we have not validated anything and a figure without a validation behind it is decoration.

We hold no certifications. We are not MCERTS certified and not PAS 2161 approved. If your use requires either, we are not currently the right supplier, and we would say so on a call rather than after a procurement.

What we can say is about design intent, and it is the reason this page exists. Provenance and confidence are being built as properties of each reading — instrument identity, capture time and processing time, positional uncertainty, calibration state at the moment of capture, condition flags — rather than as a report generated afterwards. Retrofitting a provenance chain onto a system that did not record it is the one thing that genuinely cannot be done later, so it is being done first.

Ask us the eight questions when there is something to answer them about. Until then, use them on everyone else.

If you would find it useful to pull this apart, or to tell us which of these questions your own governance process actually asks, get in touch.

Sources

Last reviewed 12 August 2026

Related
Talk to us about this