Low-cost sensor systems for air quality monitoring are typically calibrated by collocation next to a reference station and a machine-learning regression that compensates for cross-sensitivities and environmental interferences. Tancev and Pascale (Sensors, 2020) traced the well-documented collapse of such field-calibrated devices after spatial or temporal relocation to a sampling bias: because pollutant concentrations are mutually correlated and environmental conditions are non-uniformly sampled at any single site, every regression algorithm learns relationships that do not transfer, and the failure is a property of the data rather than the algorithm. They proposed calibration data generated in controlled laboratory experiments over designed, balanced (orthogonal) combinations of pollutant concentrations, temperature, and humidity as the remedy, and explicitly left its empirical proof to future research, together with the linked requirements of scalable automated laboratory workflows, drift-versus-non-stationarity disentanglement for recalibration scheduling, traceable uncertainty certification, and quantitative characterization of how much data calibration mappings need to generalize. Subsequent literature established that the metrological infrastructure for automated orthogonal laboratory calibration now exists (Tancev et al. 2022; Horender et al. 2021), that class-based performance evaluation entered European standardization (CEN/TS 17660-1:2021), that hybrid laboratory-field protocols and multi-location data-integration mitigations were demonstrated (Russell et al. 2022; Casari et al. 2025), and that drift detection remains an active front (Tancev 2021; Allka et al. 2026). The central declared question — whether designed balanced laboratory calibration materially reduces relocation error relative to practical field collocation under matched deployments — remains empirically unanswered, and existing evidence that naive laboratory calibration transfers poorly sharpens rather than settles it.

Background

Air quality monitoring increasingly supplements regulatory reference stations with networks of low-cost sensor systems — battery- or mains-powered devices built around electrochemical or metal-oxide gas sensors and optical particle counters that cost orders of magnitude less than certified analyzers. Raw signals from these sensors drift with age and respond not only to the target pollutant but also to other gases (cross-sensitivity) and to environmental factors such as temperature and relative humidity (RH). The standard remedy is field calibration: the device is colocated next to a reference station — an instrument implementing the certified reference measurement method — for weeks to months, and a regression model (linear regression, random forests, or neural networks) is trained to map the sensor signals to the reference concentrations. An alternative is laboratory calibration, in which the device is exposed to synthetic gas mixtures or test aerosols under controlled temperature and humidity in a climate chamber, with mixture compositions traceable to the International System of Units (SI) through calibrated reference generators.

A repeatedly reported failure mode of field calibration is relocation: a model calibrated at one site or in one season deteriorates when the device is moved to a new location or operated into a new season. Prior work had documented this empirically — De Vito et al. (2020) examined the robustness of field calibration for smart air quality monitors, and Vikram et al. (2019) showed that gas-phase sensor calibrations transfer unreliably to new locations and can only partially be repaired with data from multiple sites, while Levy Zamora et al. (2023) optimized how long collocation must last. Tancev and Pascale (2020), analyzing a public one-year data set from an urban street canyon — a street flanked by buildings on both sides, which traps traffic emissions — (De Vito et al.) with correlation matrices, principal component analysis (a standard technique that exposes the dominant correlated directions in a data set), and model-inspection techniques, traced the origin of the relocation problem to the data itself: pollutant concentrations are mutually correlated at any given site, so the joint probability distribution of pollutants and environmental conditions is sampled in a biased, non-representative way, and the calibration model incorporates spurious relationships. This holds regardless of the regression algorithm used — an instance of sampling bias, the same phenomenon known from machine learning at large.

Their proposed remedy is to generate balanced calibration data in the laboratory using design of experiments (DoE) — the statistical theory of choosing informative combinations of experimental factors — so that pollutant concentrations, temperature, and humidity are varied orthogonally, i.e., without mutual correlation, and the whole feature space is covered uniformly. However, they noted that this requires substantial infrastructure and, to be affordable for large fleets of devices, the calibration workflow must be automated and parallelized. They further identified three linked obstacles: sensor aging drift must be separated from genuine temporal change (non-stationarity) of the pollutant and environmental regimes so that recalibration can be scheduled optimally; an objective, traceable procedure to certify the measurement uncertainty of machine-learning-calibrated devices — comparable to the data quality objectives of the European Ambient Air Quality Directive 2008/50/EC — did not exist; and it remained unquantified how much and what kinds of data (laboratory or field) calibration mappings need in order to generalize spatially and temporally. The closing sentence of the paper leaves the central claim — that balanced laboratory data allow the sensor-environment relationships to be learned reliably with little bias — explicitly "to prove for future research".

Problem Statement

The problem, as posed by Tancev and Pascale (2020) and left to future research, is to make the calibration of low-cost air-quality sensor systems robust to relocation — spatial (deploying the device at a new site) and temporal (operating across seasons and sensor lifetime) — and in particular to establish empirically whether this is achieved by the laboratory route they propose. Concretely, the linked questions are: (1) prove or disprove, in matched real-world deployments, that calibration mappings learned from laboratory-generated balanced data — designed orthogonal combinations of pollutant concentrations, temperature, and humidity — learn the sensor-environment relationships with little sampling bias and thereby materially reduce relocation error relative to practical field-collocation protocols (the error metric, the matched field-collocation baseline, and the margin that counts as material are to be defined and justified explicitly); (2) develop and demonstrate the scalable, automated laboratory calibration workflows (parallelized climate-chamber experiments, automated gas-mixture generation and data acquisition) that make such balanced calibration feasible and affordable for large numbers of devices; (3) quantify how sensor-aging drift can be disentangled from temporal non-stationarity of pollutant and environmental regimes, so that recalibration intervals and drift-correction methods can be optimally scheduled; (4) construct objective, experimentally validated procedures to certify the uncertainty and transferability of machine-learning-calibrated low-cost sensors, including traceable laboratory-to-field validation protocols sufficient to meet regulatory data-quality objectives; and (5) quantify systematically, across multiple real-world locations and sensor types, how much and which kinds of additional data (laboratory or field) are needed for calibration mappings that generalize spatially and temporally.

References

  1. Georgi Tancev, Céline Pascale (2020), "The Relocation Problem of Field Calibrated Low-Cost Sensor Systems in Air Quality Monitoring: A Sampling Bias", Sensors 20(21):6198, doi:10.3390/s20216198, https://doi.org/10.3390/s20216198
  2. Georgi Tancev, Andreas Ackermann, Gerald Schaller, Celine Pascale (2022), "Efficient and Automated Generation of Orthogonal Atmospheres for the Characterization of Low-Cost Gas Sensor Systems in Air Quality Monitoring", IEEE Transactions on Instrumentation and Measurement 71:3133620, doi:10.1109/TIM.2022.3198747, https://doi.org/10.1109/TIM.2022.3198747
  3. Stefan Horender, Georgi Tancev, Kevin Auderset, Konstantina Vasilatou (2021), "Traceable PM2.5 and PM10 Calibration of Low-Cost Sensors with Ambient-like Aerosols Generated in the Laboratory", Applied Sciences 11(19):9014, doi:10.3390/app11199014, https://doi.org/10.3390/app11199014
  4. Hugo Savill Russell, Louise Bøge Frederickson, Szymon Kwiatkowski, Ana Paula Mendes Emygdio, Prashant Kumar, Johan Albrecht Schmidt, Ole Hertel, Matthew Stanley Johnson (2022), "Enhanced Ambient Sensing Environment—A New Method for Calibrating Low-Cost Gas Sensors", Sensors 22(19):7238, doi:10.3390/s22197238, https://doi.org/10.3390/s22197238
  5. Georgi Tancev (2021), "Relevance of Drift Components and Unit-to-Unit Variability in the Predictive Maintenance of Low-Cost Electrochemical Sensor Systems in Air Quality Monitoring", Sensors 21(9):3298, doi:10.3390/s21093298, https://doi.org/10.3390/s21093298
  6. Georgi Tancev, Federico Grasso Toro (2022), "Stochastic Online Calibration of Low-Cost Gas Sensor Networks With Mobile References", IEEE Access 10:13901-13910, doi:10.1109/ACCESS.2022.3145945, https://doi.org/10.1109/ACCESS.2022.3145945
  7. Sharad Vikram, Ashley Collier-Oxandale, Michael H. Ostertag, Massimiliano Menarini, Camron Chermak, Sanjoy Dasgupta, Tajana Rosing, Michael Hannigan, William G. Griswold (2019), "Evaluating and improving the reliability of gas-phase sensor system calibrations across new locations for ambient measurements and personal exposure monitoring", Atmospheric Measurement Techniques 12(8):4211-4239, doi:10.5194/amt-12-4211-2019, https://doi.org/10.5194/amt-12-4211-2019
  8. Saverio De Vito, Elena Esposito, Nuria Castell, Philipp Schneider, Alena Bartonova (2020), "On the robustness of field calibration for smart air quality monitors", Sensors and Actuators B: Chemical 310:127869, doi:10.1016/j.snb.2020.127869, https://doi.org/10.1016/j.snb.2020.127869
  9. Misti Levy Zamora, Colby Buehler, Abhirup Datta, Drew R. Gentner, Kirsten Koehler (2023), "Identifying optimal co-location calibration periods for low-cost sensors", Atmospheric Measurement Techniques 16(1):169-179, doi:10.5194/amt-16-169-2023, https://doi.org/10.5194/amt-16-169-2023
  10. Martina Casari, Laura Po, Federica Rollo (2025), "Scalable Cross-Location Calibration of Low-Cost Air Quality Sensors Using Heterogeneous Data", 2025 IEEE International Conference on Big Data (BigData), pp. 2110-2117, doi:10.1109/BigData66926.2025.11402149, https://doi.org/10.1109/BigData66926.2025.11402149
  11. Federica Rollo, Matteo Angelinelli, Martina Casari, Laura Po, Giorgio Pedrazzi, Roberta Turra (2026), "Beyond traditional models: Foundation models for accurate particulate matter prediction", Expert Systems with Applications 331:133099, doi:10.1016/j.eswa.2026.133099, https://doi.org/10.1016/j.eswa.2026.133099
  12. Xhensilda Allka, Pau Ferrer-Cid, Jose M. Barcelo-Ordinas, Jorge Garcia-Vidal (2026), "W-UDDR: A Unified Framework for Automated Drift Detection in IoT-based Air Quality Monitoring Systems", ACM Transactions on Internet of Things 7(3):1-20, doi:10.1145/3803419, https://doi.org/10.1145/3803419
  13. Jelle Hofman, Borislav Lazarov, Christophe Stroobants, Evelyne Elst, Inge Smets, Martine Van Poppel (2024), "Portable Sensors for Dynamic Exposure Assessments in Urban Environments: State of the Science", Sensors 24(17):5653, doi:10.3390/s24175653, https://doi.org/10.3390/s24175653
  14. CEN (2021), "Air quality — Performance evaluation of air quality sensor systems — Part 1: Gaseous pollutants in ambient air", CEN/TS 17660-1:2021, European Committee for Standardization (CEN), Brussels, doi:10.3403/30403653, https://doi.org/10.3403/30403653