Air pollution data from citizen-operated sensors improved by novel quality control system

Low-cost air quality sensors have become popular in the last decade, enabling more widespread and participatory monitoring of air quality by communities. Researchers and policymakers would like to use this citizen-gathered data to reveal patterns of pollution, especially pollution caused by PM_2.5 particles (smaller than 2.5 μm).

However, low-cost sensors do not always provide accurate data that can be integrated with existing air quality networks and technical systems. Data quality issues caused by challenges in sensor calibration, differing environmental conditions, degradation over time, and other factors are a key concern limiting citizen science data being used in research projects. For example, researchers have noted that sensors are typically calibrated to detect particulates in controlled conditions, but conditions in the field such as humidity levels may interfere with readings. 

Such accuracy issues prevent low-cost sensors from being used to support official measures like the revised EU Ambient Air Quality Directive (2024/2881). One practice that seeks to address the problem is co-locating sensors with reference stations (official measurement stations that meet European Environment Agency (EEA) data quality standards) in the field, and recalibrating the sensors to align with local conditions. There are further issues, such as the fact that sensors are often operated by non-experts, who may not follow maintenance or operation protocols; and sensors from different manufacturers may require different calibration. To avoid the need for co-location and complicated calibration a quality control system is called for.

With this in mind, a new study suggests a robust, transparent, standardised Quality Control (QC) process that ‘corrects’ sensor data, based on nearby reference station data. The researchers present a framework named FILTER (Framework for Improving Low-cost Technology Effectiveness and Reliability), which can enhance the reliability of user-obtained/crowd-sourced PM_2.5 pollution data across multiple networks. The framework is applicable at both small and large geographical scales, and is designed to overcome the limitations of other similar frameworks previously developed (which for example work only at smaller scales, are sensor-specific, or need substantial calibration). 

The researchers applied FILTER to PM_2.5 data gathered between 2018 and 2023 from large-scale, citizen-operated sensors in Europe, which were stored in two databases called sensor.community and PurpleAir. These datasets comprised over 13.2 billion observations across 38,294 unique locations at sub-hour resolution. Observations were taken from more than 400 sensors, located within a maximum distance of 500 m from reference stations.

FILTER processed both raw and “corrected” data - calibrated to account for pollution type, local meteorological and atmospheric conditions, and standardised to account for differing approaches to geographical coordinates, using data from meteorological stations. 

Processing involved five quality control steps:

1. Range validity: checks that each measurement falls within a physically plausible range of PM_2.5 concentrations, between 0 and 1,000 μgm⁻³.
2. Constant value: flags any sensor continuously reporting the same value (within ≤0.1μgm⁻³) over an 8-hour rolling time window, as the sensor may be malfunctioning.
3. Outlier detection: identifies statistical outliers (extreme spikes or drops in PM_2.5 and deviation from averages in the EEA air quality monitoring network data).
4. Spatial correlation: assesses correlation with data from neighbouring sensors within a 30-kilometre radius over a 30-day window.
5. Spatial similarity: evaluates whether a sensor’s measurements are consistent and expected considering reference stations rather than sensors, within a 30-kilometre radius.

FILTER was able to provide corrected values for 52.5% of the original sensor measurements, bolstering the useful data in the study area from 224 measurements per km² sourced from reference stations, to 1,428 high quality measurements per km² including low-cost sensor data.

They noted that nearly two-thirds of readings came from urban areas. Countries with the most sensor locations were Germany (14,002), the Netherlands (4,541), Poland (2,574) and Belgium (2,484). Cities with the most sensors were Sofia (862), Berlin (792), Stuttgart (759), Brussels (683), and Dortmund (661).

The researchers note significant data loss between their quality control steps 4 and 5, due to a lack of reference points with which to verify low-cost sensor observations – if stopping at step 4, the spatial density of corrected measurements was nearly double (∼2,750 per km²). 

They state that data processed up to step 4 of their QC process appears reliable and therefore could be considered trustworthy for some applications, to allow a pragmatic balance between data availability and quality. They apply three quality tiers: ‘high-quality’ up to step 5; ‘good quality’ to step 4; and ‘other quality’, where quality cannot be assured.

They indicate two types of real-world application for FILTER-controlled sensor data: where relative PM_2.5 levels are sufficient (for instance in monitoring trends and fluctuations; ‘before and after’ pollution control measures, tracking diurnal patterns, or raising public awareness), and those where absolute PM_2.5 levels are essential (for e.g. regulatory compliance, health risk assessment, emission modelling or calculating Air Quality Index figures). 

Among the limitations of the approach are that it is designed for fixed outdoor low-cost sensors monitoring PM_2.5 only; it cannot account for measurements made at more than once per hour; and it assumes that data from nearby reference stations are suitable for data correction.

However, the researchers highlight that FILTER has proven able to improve the accuracy and reliability of PM_2.5 data gathered by low-cost sensors in Europe – and therefore could play a role in building a harmonised, quality-controlled dataset of air pollutant observations to support air quality research, public health assessments, and environmental policy.

Reference: 

Hassani, A., Salamalikis, V., Schneider, P, Stebel, K. and Castell, N. (2025) A scalable framework for harmonizing, standardization, and correcting crowd-sourced low-cost sensor PM2.5 data across Europe. Journal of Environmental Management Volume 380, Apr 2025, 125100. https://doi.org/10.1016/j.jenvman.2025.125100