City-scale calibration of a low-cost PM2.5 network for regulatory-compliant air-quality assessment

Robert Blaga; Nicoleta Stefu; Sneha Gautam; Chang-Hoi Ho

doi:10.1371/journal.pclm.0000780

Abstract

This study presents the calibration and performance analysis of a low-cost sensor (LCS) network for monitoring particulate matter with diameters ≤ 2.5 µm (PM_2.5) in Bucharest, Romania. The InfoAer network comprised 44 Clarity Node-S sensors deployed across the city. The performance of sensors was evaluated against reference measurements from the National Environmental Protection Agency (NEPA) regulatory monitoring stations. The manufacturer’s pre-calibration significantly underestimated PM_2.5 concentrations, particularly during the summer months, when meteorological conditions favor the formation of secondary aerosols. Nine spatial clusters with collocated InfoAer-NEPA measurements were identified, with one designated for calibration model development and the remaining eight for independent validation. Multiple seasonal calibration models were developed using temperature, relative humidity, and nitrogen dioxide as predictor variables in a multiple linear regression formulation. Calibration performance was compromised during hot, dry conditions, when PM_2.5 concentrations were typically low, likely due to reduced aerosol scattering efficiency and increased measurement uncertainty. Using only temperature and relative humidity as predictors, the optimal model, selected via a two-step calibration process, substantially improved measurement accuracy. Pearson correlation coefficients improved from 0.06 to 0.65 and from −0.28 to 0.89 for the dry and humid seasons, respectively. However, considerable inter-sensor variability in calibration performance was observed, indicating the need for additional meteorological or chemical parameters in future calibration algorithms. Application of the calibration model to the entire InfoAer network revealed significant air quality violations across Bucharest. On average, calibrated sensors recorded more than 60 days of exceedances of the European Union’s daily PM_2.5 limit value (25 μg m⁻³) per year, which far exceeded the permitted frequency of 35 exceedances per year. Prior to calibration, only 8 of 44 sensors (18%) exceeded this threshold; post-calibration analysis revealed violations at all monitoring locations. These results demonstrated the critical importance of proper LCS calibration for accurately assessing regulatory compliance and protecting public health.

Citation: Blaga R, Stefu N, Gautam S, Ho C-H (2026) City-scale calibration of a low-cost PM_2.5 network for regulatory-compliant air-quality assessment. PLOS Clim 5(2): e0000780. https://doi.org/10.1371/journal.pclm.0000780

Editor: Jingyu Wang, Nanyang Technological University, SINGAPORE

Received: September 16, 2025; Accepted: January 26, 2026; Published: February 13, 2026

Copyright: © 2026 Blaga et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The database is uploaded as part of the Supporting information.

Funding: The work of C-HH was supported by the National Research Foundation of Korea (NRF) funded by the Korean Government (MSIT; RS-2025-00555756) and the Ministry of Education (RS-2018-NR031078).

Competing interests: The authors have declared that no competing interests exist.

1. Introduction

Poor air quality is the most significant environmental threat to global public health. According to a 2024 report by the Energy Policy Institute at the University of Chicago, ambient air pollution reduces global life expectancy by 1.9 years on average and by more than 6 years in the most severely polluted regions [1]. Although overall pollution levels are lower in Europe, life expectancy is still reduced by more than 1 year in pollution hotspots, such as the Po Valley in Italy, the Upper Silesia Coal Basin in Poland, and the Western Balkans.

All United Nations member states have agreed to implement the 17 Sustainable Development Goals (SDGs) encompassed in the 2030 Agenda for Sustainable Development. Target 11.6.2 of the SDGs states that countries should “by 2030, reduce the adverse per capita environmental impact of cities, including by paying special attention to air quality”. High concentrations of particulate matter (PM), in particular, pose well-documented risks to human health that extend beyond environmental degradation. SDG target 3.9.1 requires that by 2030, countries should “substantially reduce the number of deaths and illnesses from hazardous chemicals and air, water, and soil pollution and contamination”. SDG 3.9.1 explicitly targets the reduction of mortality rates attributable to exposure to both household (indoor) and ambient (outdoor) air pollution.

In line with the commitments outlined in the SDGs, the European Union (EU) and the World Health Organization (WHO) have established updated air quality standards. The revised EU regulatory provision, which is scheduled to be implemented from 2030 onwards, establishes a daily PM with a diameter of ≤ 2.5 µm (PM_2.5) threshold of 25 μg m⁻³ with a tolerance of 18 exceedances per calendar year [2]. The updated WHO Air Quality Guidelines introduced in 2021 are more stringent, with a daily threshold of 15 μg m⁻³ and an allowable tolerance of 3 − 4 exceedances per year [3]. While the EU thresholds reflect pragmatic political considerations and implementation feasibility, the WHO guidelines are based on current epidemiological evidence regarding the health impacts of PM exposure. In particular, WHO guidelines are established such that concentrations below the threshold value are associated with no discernible adverse health effects attributable to PM_2.5 exposure. The epidemiological research conducted by Correia et al. (2013) showed that concentrations exceeding the daily WHO threshold were associated with a reduction in average life expectancy of around 0.35 years per 10 μg m⁻³ increment in PM_2.5 concentration [4]. Similarly, Ebenstein et al. (2017) reported that a 10 μg m⁻³ increase in PM10 (diameter ≤ 10 µm) concentrations corresponded to a decrease in life expectancy at birth of 0.65 years [5]. Measurement uncertainties of even 10 − 20 μg m⁻³ in PM concentrations can lead to substantial overestimation or underestimation of the associated health impacts. Therefore, accurate, high-resolution spatio-temporal PM data are essential for reliable health impact assessments.

Consequently, a primary limitation in epidemiological studies regarding air pollution comes from the lack of ambient air quality data with adequate spatial resolution and measurement accuracy. The aforementioned study [1], for example, relies on PM_2.5 concentrations derived from satellite-based aerosol optical depth retrievals [6]. While this approach provides global coverage, the accuracy of the measurements is compromised due to the inherently modelled nature of satellite-derived PM_2.5 estimates. Ground-based direct measurements would be preferable for validation and enhanced precision.

Over past decades, ground-based PM monitoring has been conducted primarily through regulatory networks operated by the National Environmental Protection Agency (NEPA). However, these networks have limited spatial coverage due to the high cost of reference-grade instrumentation. The high costs restrict accessibility to regulatory agencies and well-funded research institutions. In recent years, low-cost PM sensors (LCSs) based on light-scattering principles have proliferated and gained widespread popularity. Such devices are now marketed by numerous companies at prices as low as US$ 100–200, making them readily accessible to the general public [7]. Owing to their accessibility, LCS devices are transforming air pollution monitoring, and their use is expected to increase further in the coming years [8]. However, the data quality of LCS devices still falls short of reference-grade instruments. The most commonly reported biases in LCS data are (1) the artificial amplification of PM concentrations during humid periods due to the hygroscopic growth of aerosol particles, and (2) the underestimation of concentrations during dry periods, due to limitations in the detection of fine particles and related factors [9].

To achieve reference-equivalent status, which would permit their use in official quantitative air quality assessments, sensors must satisfy strict accuracy requirements. Regulatory agencies apply varying criteria; for example, the EU stipulates that data uncertainty should be below 25% and that at least 90% of the data be captured [10]. The expanded uncertainty must be evaluated at the 95% confidence level, a metric that incorporates the physicochemical properties of both aerosols and sensor mechanisms. For any new sensor to be certified as reference-equivalent, extensive in-situ testing with multiple collocated reference instruments is required [11]. Such testing protocols, however, are typically beyond the capabilities of ordinary LCS users.

The quality of LCS data can be improved by using calibration procedures that employ co-located reference-grade or reference-equivalent monitoring systems [12]. The calibration process typically does not raise LCS measurements to reference-equivalent levels. Nonetheless, the resulting data can serve as a valuable complement to regulatory monitoring networks, enhancing the spatial and temporal resolution of air quality assessments. The effectiveness of these calibration methods depends on the availability of mobile reference monitoring platforms or the use of established regulatory monitoring stations such as those operated under national environmental protection programs [13]. The latter approach has been adopted in the present study.

There are now many studies on LCS calibration [13,14–17], reflecting the growing recognition of their potential for air quality monitoring applications. However, the significant spatial variation in the physicochemical properties and compositional characteristics of aerosols presents a major constraint on the broader applicability of current calibration models. Furthermore, even within individual metropolitan areas, the cross-compatibility of calibration models across sensor manufacturers, types, and specifications has received limited scientific attention. Consequently, rigorous validation of model transferability is essential prior to implementation, with location-specific calibration models developed where feasible.

In this study, we show a comprehensive calibration of the LCS network deployed throughout Bucharest, Romania. The LCS network calibration was performed using data from co-located national regulatory monitoring stations, with sensor-reference station pairs identified based on spatial proximity criteria. Although machine learning approaches have demonstrated superior performance in sensor calibration applications, the complexity of their implementation and their computational requirements may limit access for independent researchers and practitioners. Therefore, a multiple linear regression methodology was adopted to ensure accessibility and practical applicability. Several calibration models were evaluated using regularly measured meteorological variables, such as relative humidity and temperature. The optimal calibration model, selected based on performance metrics, was then applied to the entire LCS network to analyze air quality trends and pollution patterns across time and space in the Bucharest metropolitan area.

2. Data and method

Historically, air quality monitoring in Romania has faced significant regulatory difficulties, with the European Commission initiating multiple infringement procedures against the country for failing to comply with environmental protection directives [18]. These enforcement actions have mentioned inadequate air pollution monitoring infrastructure in Bucharest [19], indicating critical shortcomings in the national monitoring framework. However, recent initiatives have substantially improved the monitoring landscape in the capital region. These include the expansion of the National Air Quality Monitoring Network (RNMCA) through the installation of additional NEPA stations beginning in 2021, and the subsequent deployment of a municipal LCS network by Bucharest City Hall in 2023. Despite these infrastructure improvements, rigorous data harmonization is essential to ensure compatibility between reference-grade and low-cost monitoring platforms. Calibrating LCS data is a vital step towards integrating these complementary monitoring systems and maximizing their collective utility for air quality assessment and management.

(a) Study area

This study was conducted in Bucharest (44.43°N, 26.10°E), the capital city of Romania, located in the southeastern region of the country. The local climate is humid continental (Köppen-Geiger classification: Cfa), featuring moderately humid winters with occasional temperature fluctuations and hot, predominantly dry summers. The metropolitan area is administratively divided into six municipal districts. For this investigation, LCS devices located within a 1.25 km radius of RNMCA stations are classified as co-located sensor pairs. Nine co-location clusters meeting these spatial criteria are identified and illustrated in Fig 1A. The figure shows the locations of these administrative boundaries, the main transportation infrastructure, and the deployment of the two sensor networks used in this study, including the delineation of co-location clusters established for calibration purposes.

Download:

Fig 1. Reference and low-cost PM sensor (LCS) networks were used in this study.

(a) Spatial distribution of monitoring nodes within the InfoAer LCS network (blue dots) and the Romanian National Air Quality Monitoring Network (RNMCA) reference stations (green squares) in Bucharest. Major transportation corridors are indicated in red lines, and minor roadways of all types are noted in grey lines. 1.25 km radius proximity zones centred on each RNMCA station are shown for co-location analysis, along with the ID of the RNMCA station for each cluster (B1 − B30). The base map data is obtained from OpenStreeMap (OSM, https://www.openstreetmap.org). OSM data was downloaded as an sp object using the osmdata R package, by querying the bounding box for ‘Bucharest’. The OSM data is made available under the Open Database License: http://opendatacommons.org/licenses/odbl/1.0/. (b) Intercomparison of daily PM_2.5 concentrations (μg m⁻³) measured at RNMCA stations using gravimetric reference methods and Derenda AMP2 automated monitoring systems.

https://doi.org/10.1371/journal.pclm.0000780.g001

(b) Low-cost PM sensor (LCS) data

The Bucharest municipal government operates a high-density air quality monitoring network comprising 44 Clarity Node-S devices [20]. These LCSs were provided to Bucharest City Hall by Vital Strategies (https://www.vitalstrategies.org/) as part of the collaboration protocol established by CGMB Decision No. 201/30.03.2022 for Bucharest’s participation in the Partnership for Healthy Cities initiative (https://cities4health.org/), funded by Bloomberg Philanthropies (https://www.bloomberg.org/). Network deployment was completed in late 2023. Real-time data visualization from this monitoring system is publicly accessible via the InfoAer platform on the municipal website (https://infoaer.pmb.ro/infoaer/). For this study, the network is referred to as the InfoAer network. The raw data recorded by the InfoAer network sensors were obtained via personal correspondence from the Air Quality Monitoring Service of the Environmental Department of Bucharest City Hall. The data was provided freely for research purposes.

The Clarity Node-S sensors employed in this study have a PM_2.5 measurement range of 0 − 1000 μg m⁻³, which covers the expected concentration levels at the study location. While the default temporal resolution of the Node-S platform is 15-minute intervals, the InfoAer network archives hourly-averaged PM_2.5 concentrations. The sensors incorporate factory pre-calibration protocols for PM_2.5 and nitrogen dioxide (NO₂) measurements. According to manufacturer specifications, the devices are expected to achieve the following measurement accuracies: a determination coefficient (R²) of 0.94 under optimal environmental conditions and R² > 0.7 under typical field conditions. The Node-S devices are certified to MCERTS CSA MC230425/00, indicating compliance with established performance standards for ambient air quality monitoring equipment. However, certification requirements explicitly mandate post-deployment field calibration against co-located reference-grade monitoring systems. Based on available documentation and municipal records, this mandatory field calibration procedure has not been implemented for the Bucharest network. This further justifies the calibration methodology presented in this investigation.

As part of the data quality assurance protocol, all hourly measurements exceeding 500 μg m⁻³ from the 44 PM_2.5 sensors were excluded from the analysis to eliminate potential instrumental artifacts and extreme outliers. Following EU standards for equivalency testing, which stipulate 90% data completeness for fixed monitoring stations and 50% for indicative measurements [11], stringent data availability criteria were established for sensor selection. For sensors designated for calibration model development, a threshold of 90% of the annual data was imposed, requiring valid PM_2.5 measurements for at least 328 days within the 365-day study period. For sensors allocated exclusively to model validation purposes, the threshold was reduced to 67% (244 days minimum) to ensure adequate representation across all seasonal and meteorological conditions while maintaining statistical robustness. In addition, all records containing non-physical temperature and relative humidity values were systematically removed during preprocessing. After applying these quality-control criteria, 42 sensors met the data-completeness requirements and were retained for analysis. The final consolidated dataset (S1 Data) comprises 358,151 individual records, each containing hourly measurements of PM_2.5 concentration, ambient temperature, relative humidity, and NO₂ concentration, accompanied by corresponding geospatial coordinates and sensor metadata.

(c) Reference data

The calibration of the LCS network was performed using reference measurements from the RNMCA, operated by the NEPA (see Fig 1A for its location). RNMCA data are publicly accessible through the national air quality portal (https://www.calitateaer.ro/). The RNMCA infrastructure contains 30 monitoring stations within Bucharest and the surrounding metropolitan region. Among these stations, only four employ gravimetric reference methods for PM_2.5 measurement, and only two are sufficiently close to LCS devices to enable meaningful calibration analysis. Post-2021 installations use Derenda APM2 automated monitoring systems for simultaneous PM_2.5 and PM₁₀ measurements. The Comte-Derenda APM2 instruments have achieved reference-equivalent certification under TÜV conformity assessment protocols, incorporating manufacturer-specified correction algorithms validated during certification [21].

The comparative analysis presented in Fig 1B indicates systematic differences between measurement methodologies: Derenda APM2 systems report slightly higher PM_2.5 concentrations than gravimetric reference methods (2024 annual averages: 16.9 μg m⁻³ vs. 12.8 μg m⁻³, respectively). Despite this observed bias, independent validation studies conducted by NEPA [22] have confirmed the reference-equivalent status of Derenda APM2 data for regulatory applications (personal communication with the public relations office of NEPA, February 2025, accessed through Law 544/2001 on free access to public information). Thus, this calibration study uses data from RNMCA stations within Bucharest’s administrative boundaries, equipped with post-2021 Derenda APM2 monitoring systems, as the reference standard for LCS calibration (S2 Data).

3. Air quality assessment

This section outlines the development and implementation of a calibration model for the InfoAer network. This model is then applied to conduct a comprehensive spatiotemporal analysis of air quality patterns across Bucharest. The calibration methodology employs a two-stage hierarchical approach to optimize model performance and assess its transferability across different monitoring locations. In the first calibration stage, various model formulations are independently fitted to data from each sensor cluster. Three representative model types are selected based on their ensemble performance metrics averaged across all available clusters. The second calibration stage involves designating a single cluster as the reference calibration site for fitting the three candidate models. The resulting models, with regression coefficients derived from the reference cluster, are rigorously validated across all remaining sensor clusters to assess their transferability and spatial generalizability. The calibration model with the most consistent performance across the entire cluster ensemble is selected as the final calibration algorithm. This validated model is then applied to the entire InfoAer network to generate calibrated PM_2.5 measurements for a comprehensive analysis of air quality across the study domain.

(a) Calibration of LCS data

Throughout this analysis, the term cluster denotes a paired dataset comprising measurements from the corresponding InfoAer sensor and its spatially proximate RNMCA reference station. The temporal scope of this study was restricted to the year 2024 for both monitoring networks to ensure data consistency and temporal alignment. Daily averaged PM_2.5 concentrations are computed for both datasets to reduce measurement noise and facilitate comparative analysis. Although recent studies have demonstrated the superior accuracy of machine learning approaches for PM_2.5 sensor calibration [23,24], these methodologies show significant implementation barriers for independent practitioners and research groups. Furthermore, machine learning models are susceptible to overfitting phenomena, which can substantially compromise their transferability across different spatial domains or temporal periods. The ‘black box’ nature of these models further makes them less compatible with the EU air quality framework, which requires a transparent model, estimations of error propagation, reproducibility, and traceability, which are all difficult with machine learning models. Hence, this study uses the more established multiple linear regression (MLR) framework to develop the calibration model. The developed regressions incorporate simple arithmetic transformations of temperature (T), relative humidity (RH), and NO₂ concentrations as predictor variables. Dew point temperature was initially considered an additional meteorological predictor, but it proved to have limited explanatory power. Thus, it was excluded from the final model formulation. The general form of the MLR calibration model is expressed as follows:

(1)

where the q₁(x)−q₄(x) are simple arithmetic functions of x and a₁ − a₄ are fitting coefficients.

LCS measurements show a high sensitivity to seasonal variations in ambient T and RH. This necessitates explicitly accounting for seasonal trends during model calibration [25]. Following the methodology established in our previous study [26], the dataset for each cluster was partitioned into two distinct subsets, one corresponding to dry periods and the other to humid periods, to account for the influence of these meteorological factors on sensor performance. The “dry” data subset encompasses the months between 1. The scientific literature presents two primary approaches for incorporating humidity effects into calibration frameworks for PM_2.5 measurements. The first approach is based on κ-Köhler theory, which provides a theoretical basis for understanding how ambient water vapor condenses onto atmospheric aerosol particles. This condensation process is governed by the hygroscopic properties of the aerosol population and the prevailing ambient conditions. According to κ-Köhler theory, the empirical relationship between dry and wet aerosol mass concentrations can be expressed as [27]:

(2)

where κ is a parameter that characterizes the aerosol hygroscopicity and is determined by aerosol composition. If the aerosol composition is known, a tabled κ value can be used. The value κ = 0.62 is recommended for polluted environments containing a mixture of organic and inorganic aerosols [28].

The second approach for modelling the RH impact on PM_2.5 is through a linear correction of the form [12]:

(3)

Where the fitting coefficients r₁ and r₂ are determined dynamically during model fitting. This is done by applying a simple linear fit to each cluster, assuming and .

In the first calibration step, fitting was performed separately for each location on the two seasonal data subsets. An example of data from a collocated InfoAer-RNMCA pair is shown in Fig 2A, illustrating the generic behavior of LCS and reference PM_2.5 during the dry and humid periods, and the improvements to the raw LCS data introduced by calibration. The best fit calibration model is also shown. The average R² across all clusters is listed in Table 1. The best models, based on R^2, are highlighted. More detailed results of model fitting, including several statistical indicators per each individual cluster, are available as Supporting Information in the S3 Data archive.

Download:

Table 1. Performance statistics (mean ± standard deviation of R²) for calibration models fitted independently to each sensor cluster. The humidity correction functions f(RH) and k(RH) are implemented according to Eqs. (2) and (3), respectively. Predictor variables raised to the 0th power indicate exclusion from the regression model. The highest-performing model is marked in bold, while the second and third-ranked models are shown in italics.

https://doi.org/10.1371/journal.pclm.0000780.t001

Download:

Fig 2. Calibration model performance evaluation.

(a) Comparison of calibrated (red and green lines) and uncalibrated (light blue line) LCS measurements versus reference data (black line) for a representative cluster (the B23 cluster is shown), demonstrating the behavior of the LCS and reference PM2.5 during dry and humid periods, and the improvement over the raw LCS achieved through calibration. Note that the calibrated LCS PM2.5 fully overlaps the reference PM2.5. (b) Box plot analysis of the accuracy of the LCS PM2.5 calibrated with the three models (M1, M2, and M3) across all nine clusters. R2 distribution statistics are shown separately for the dry and humid seasonal periods. On the boxplots, the thick line represents the median, the box lower and upper limits are the Q1 (25%) and Q3 (75%) quartiles, while the whiskers extend to 1.5 times the Q1 − Q3 interquartile range.

https://doi.org/10.1371/journal.pclm.0000780.g002

Several key conclusions can be drawn from the performance analysis presented in Table 1. The calibration model incorporating NO₂ and RH as predictor variables (M3) demonstrates the most consistent performance across all clusters and achieves the highest mean R² values. However, given the limited availability of NO₂ measurements in many monitoring applications, alternative model formulations that exclude this parameter warrant consideration. Among these models, the formulation incorporating RH and T as predictors (M2) performs best. It should be noted that the most parsimonious model (M1), which relies solely on the raw LCS PM_2.5 measurements and no additional environmental parameters, still achieves reasonable calibration accuracy. This simple linear-scaling model serves as a slope-correction algorithm and provides empirical evidence for the LCS measurement system’s fundamentally linear response relative to reference-grade instrumentation.

Boxplots of the R² values for the three calibration models (M1, M2, and M3) across all clusters are shown in Fig 2B for the humid and dry data subsets. All models demonstrate substantially superior performance during humid periods, whereas they show reduced accuracy and increased inter-cluster variability during dry conditions. Models M2 and M3 perform exceptionally well during humid months, with R² values approaching ~0.9. In contrast, performance during dry periods ranges from 0.55 to 0.8 (Fig 2B). This seasonal dichotomy is particularly noteworthy, given that a substantial portion of the aerosol sensor calibration literature focuses on correcting for hygroscopic effects on particle measurements. The hygroscopic nature of atmospheric aerosols typically results in an enhanced light-scattering response from nephelometer-based sensors at high relative humidity. This happens because the presence of bound water increases the particle size and refractive index. However, for the Clarity Node-S sensors deployed at this study site, even the simplest linear correction model (M1) achieves R² > 0.7 relative to reference measurements during humid periods. This demonstrates the inherently linear response characteristics of these instruments under high-humidity conditions. This performance level is consistent with the manufacturer’s specifications for typical operating conditions (Section 2.2). In contrast, the loss of linearity becomes more pronounced during dry periods. This is demonstrated by the significantly degraded performance of the simple slope correction model (M1) and by the generally reduced accuracy of all calibration approaches relative to humid-period performance. In addition, the mean performance statistics for models M2 and M3 are adversely affected by one or two extreme outlier clusters, which will be addressed in subsequent analysis.

In the second calibration stage, cluster B23 is designated as the reference calibration site. The choice was made because the cluster had the closest proximity between the LCS sensor and the reference station among all available clusters. This cluster also demonstrates superior measurement concordance, characterized by the lowest root-mean-square errors and the highest correlation coefficients relative to reference measurements across the entire cluster ensemble. The three calibration models are then fitted to the B23 reference dataset and the resulting parameterized model equations are presented in Eqs. (4) to (6). To rigorously assess model transferability and spatial generalizability, these models, with fixed coefficients derived from the B23 reference cluster, are validated on independent datasets from the remaining clusters (i.e., sites not used during model training), as detailed in the subsequent validation analysis.

(4)

(5)

(6)

Several noteworthy observations emerge from the parameterization of models M2 and M3. First, the coefficients for RH and T are positive, which is inconsistent with established principles of aerosol hygroscopicity. In theory, water vapor uptake by hygroscopic aerosol particles increases the measured particle mass relative to dry conditions. This suggests that the humidity correction coefficients should be negative to compensate for this hygroscopic growth effect. However, the LCS measurements consistently underestimate reference PM_2.5 concentrations across all seasons (as shown by the systematic offset between the cyan and black curves in Fig 2A). This underestimation is unusual for the LCS, which typically shows a positive bias relative to reference instrumentation due to hygroscopic enhancement and other factors, especially during winter. We hypothesize that this underestimation stems from aggressive manufacturer pre-calibration algorithms applied to the sensor firmware. An alternative explanation involves potential systematic errors in the sensor’s size-selective separation of the PM_2.5 and PM₁₀ fractions. The latter hypothesis is supported by the fact that corresponding PM₁₀ measurements do not show similar underestimation patterns. Regardless of the underlying mechanism, this systematic underestimation results in artificially suppressed PM_2.5 concentrations throughout the measurement period. The most severe manifestation occurs during the summer months when measurements occasionally register spurious zero values despite the presence of measurable atmospheric PM.

The second notable characteristic is the negative coefficient of NO₂ in model M3. As a key precursor species for secondary aerosol formation, NO₂ participates in atmospheric chemical processes that generate nitrate-containing PM. These secondary nitrate aerosols begin to form in the ultrafine size range (nucleation mode, < 10 nm) and subsequently undergo coagulation and condensational growth. They then transition through the Aitken mode (10 − 100 nm) before reaching the accumulation mode (100 − 1000 nm), where they contribute significantly to PM_2.5 mass concentrations. This formation pathway contrasts considerably with that of primary aerosol sources, such as mineral dust, which are emitted into the coarse mode (≤ 2.5 μm) without undergoing atmospheric processing. High NO₂ concentrations, therefore, serve as indicators of high ultrafine and fine particle populations within the aerosol size distribution. However, the limitations of LCS in detection efficiency result in poor sensitivity to ultrafine particles (those in the nucleation and smaller Aitken modes). As a result, during periods of elevated NO₂-driven secondary aerosol formation, LCS measurements systematically underestimate total PM_2.5 mass. This phenomenon is the inverse of hygroscopic effects: while humidity enhances measured particle mass through water uptake, high NO₂ conditions reveal measurement deficits due to limitations in sensor size selectivity. Therefore, the divergent contributions of NO₂ and RH in calibration models indicate different physical mechanisms that affect sensor response characteristics.

(b) Validation of calibration model

In the second calibration stage, the three models are rigorously validated against independent datasets from all remaining clusters that were not used during model training. The validation results are summarized in Table 2. The performance evaluation shows that all three models achieve an R² value greater than 0.7 at most validation sites. The European Commission’s Joint Research Center considers R² _> 0.9 comparable to reference systems, whereas R² _> 0.75 is considered good performance for supplementary monitoring [29]. Thus, our results indicate good overall accuracy and confirm that the models can be successfully transferred across the study domain. Models M2 and M3 show comparable performance, with only minor differences in accuracy across validation locations, suggesting that both formulations effectively capture the essential sensor-reference relationships. While model M1 shows slightly lower accuracy than the more complex models, it still performs competitively, with only modest degradation. This confirms the fundamental linear relationship between LCS and reference measurements across the network. Minimal working code to reproduce the validation results is given in the Supporting Information as S1 Code.

Download:

Table 2. Statistical indicators of accuracy for calibration models over the identified sensor clusters. 95% confidence intervals are shown in paranthesis. CI is estimated by resampling 1000 times the calibrated daily PM2.5 values for each station for the whole year, then calculating the statistical indicators for each sample and extracting the 2.5% and 97.5% quantile values from the obtained distribution.

https://doi.org/10.1371/journal.pclm.0000780.t002

Several validation clusters show anomalous performance characteristics. At these sites, multiple calibration models produce substantially degraded accuracy that cannot be attributed to random measurement uncertainty alone. First, the RNMCA B9 station shows systematic deviations from the expected sensor-reference relationships. The PM_2.5 concentrations at this reference station are consistently and significantly lower than those measured at all other RNMCA locations across the network. A comparative analysis shows that the co-located InfoAer sensor (uncalibrated) has a stronger correlation with the second-nearest RNMCA B23 station (r = 0.92) than with its designated reference station B9 (r = 0.87). This suggests anomalous behavior. The B9 station is located in a school courtyard, characterized by dense vegetation. This may produce localized meteorological conditions and enhanced dry deposition processes, systematically reducing PM concentrations relative to urban background levels. On the other hand, the observed anomalies may result from unidentified instrumental malfunctions or systematic measurement biases affecting the B9 reference station.

Two additional locations (B16 and B20) exhibit abnormal calibration performance and require further investigation. A systematic analysis indicates that these stations record exceptionally high PM_2.5 concentrations during the summer months relative to all other network sites, despite similar meteorological conditions across the study domain. Both stations are located at major traffic intersections with high vehicular density. This proximity could lead to steep spatial concentration gradients from localized emission sources. However, other LCS devices and RNMCA stations deployed along the same roadway do not exhibit similar anomalies. To investigate whether the anomalous PM concentrations at B16 and B20 result from suboptimal spatial co-location or excessive sensor separation distances within clusters, we have conducted a systematic analysis of calibration model performance as a function of inter-sensor distance and station classification (Figs 3A and 3B). Linear regression analysis across all validation clusters (excluding the reference calibration cluster B23) yields the counterintuitive result that model accuracy appears to improve with increasing sensor separation distance. However, this paradoxical relationship represents a statistical artifact arising from the limited cluster sample size rather than a genuine physical phenomenon.

Download:

Fig 3. Spatial dependency analysis of calibration model performance.

(a) Scatter plot showing the relationship between the accuracy of the calibration models (R2 values for models M1 − M3) and the inter-sensor separation distance for each cluster. Distance is measured between the corresponding InfoAer sensors and the RNMCA reference stations. (b) Box plot comparison of model M2 calibration performance (R2 distribution) across sensor clusters stratified by RNMCA station classification categories.

https://doi.org/10.1371/journal.pclm.0000780.g003

Within the 1.25 km co-location radius, we conclude that sensor separation distance does not affect the calibration model’s performance. This assessment does not account for additional meteorological variables, such as wind speed, wind direction, and atmospheric stability, that could influence spatial representativeness. Regarding station classification, deployment in high-traffic environments improves model accuracy. However, these categorical comparisons are not statistically significant: the urban background category contains only two clusters, B23 (the reference station with the highest accuracy) and B9 (the extreme outlier with the lowest accuracy), whereas the industrial category comprises only cluster B21. Given these analytical limitations, we cannot definitively identify the underlying mechanisms responsible for the anomalous behavior at these sites. We attribute the degraded performance at outlier stations to unidentified site-specific factors that are not captured by the current model formulations.

The averages of the statistical performance indicators for the M2 model across all eight validation clusters are as follows: MBE = −1.2 ±± 3.9 μg m⁻³, RMSE = 5.4 ±± 2.2 μg m⁻³, and R² = 0.69 ±± 0.20. The corresponding indicators for the M3 model are MBE = −3.0 ± 4.0 μg m⁻³, RMSE = 5.8 ± 2.9 μg m⁻³, and R² = 0.63 ± 0.32. The performance differences between the two models are of the same order of magnitude as the standard deviations of their respective statistics, suggesting that the difference is not statistically significant. Consequently, either model could be considered acceptable for deployment. Here, we selected M2 because its average performance metrics were marginally superior to those of M3. Therefore, M2 was implemented to calibrate the entire InfoAer sensor network. The calibrated InfoAer PM_2.5 measurements obtained through this methodology are subsequently used in the following section to evaluate Bucharest Municipality’s compliance with EU air quality standards and guidelines.

(c) Compliance with the EU air quality framework

To evaluate Bucharest’s compliance with EU and WHO air quality standards, the M2 calibration model was applied to data from 42 InfoAer network devices that had successfully passed the pre-processing quality control procedures. The calibrated measurements consistently exhibit higher exceedance frequencies than the uncalibrated data, indicating a systematic underestimation of PM_2.5 concentrations in raw InfoAer sensor outputs. These disparities are particularly evident when evaluated against EU air quality guidelines.

Fig 4 shows the frequency of threshold exceedances for each sensor under the EU and WHO regulatory frameworks. The uncalibrated data exceed the daily PM_2.5 threshold at only 8 of 42 sensors (19%), whereas calibrated data show threshold violations at all 42 sensors (100%). The mean number of exceedances of the EU and WHO daily PM_2.5 thresholds across the InfoAer network are 71.9 ± 15.9, and 135.0 ± 21.4 days, respectively, for calibrated data, compared to 12.5 ± 8.8 and 70.2 ± 16.0 days for uncalibrated data. For comparison, RNMCA reference stations record average exceedance durations of 64.6 and 36.4 days for the EU threshold, and 138.2 and 60.4 days for the WHO threshold.

Download:

Fig 4. Annual exceedances of daily PM_2.5 thresholds recorded by InfoAer sensors in 2024:

(a) EU and (b) WHO. The frequency of exceedances for calibrated and uncalibrated data is represented by circles and diamonds, respectively. Connecting lines between paired data points show the magnitude of the difference attributable to sensor calibration..

https://doi.org/10.1371/journal.pclm.0000780.g004

These results underscore the importance of rigorous calibration procedures for LCS data in quantitative air quality studies. Although the calibration framework employed here does not meet the requirements for reference equivalence, the optimized LCS is sufficient to indicate that urban air quality can be substantially underestimated by public authorities if uncalibrated LCS data are used for any air quality assessment.

(d) Spatial and seasonal variations in PM levels

To analyze the spatial patterns in PM_2.5 distribution, we calculate the deviation of each sensor from the network-averaged number of EU daily threshold exceedances across the InfoAer monitoring array. The results are presented in Fig 5A. Several key patterns emerge from the spatial analysis. The city center shows average pollution levels, with no noticeable spatial clustering of extreme values. In general, sensors located in peripheral areas of the city record high PM_2.5 concentrations, with two distinct clusters of high concentrations identified: (1) the boundary region between Districts 1 and 6, spanning the Giulești-Sârbi and Bucureștii Noi neighborhoods, and (2) the outer section of District 5, including the Ghencea, Rahova, Ferentari, and Giurgiului neighborhoods.

Download:

Fig 5. Spatial and temporal variability of PM_2.5 concentrations across the InfoAer and RNMCA monitoring networks.

(a) Deviation from the network-mean number of EU daily PM_2.5 threshold exceedances for stations from both monitoring networks. σ represents the standard deviation of annual exceedance frequencies across all InfoAer stations in 2024. δ denotes the deviation of individual InfoAer and RNMCA site exceedance counts from the network-wide mean. The administrative boundaries of Bucharest’s six districts are delineated by purple lines. The base map data was obtained from OpenStreetMap as described in the caption of Fig 1. (b) Monthly-averaged PM_2.5 concentrations across the combined dataset. The shaded regions represent the ± 0.5 (dark orange) and ±1.5 (light orange) standard deviation confidence bands around the monthly means.

https://doi.org/10.1371/journal.pclm.0000780.g005

The monthly-averaged PM_2.5 concentrations across all InfoAer stations are shown in Fig 5B. A pronounced seasonal pattern is evident, with significantly higher PM_2.5 concentrations during the winter months. This seasonal variation is attributed to two primary meteorological and anthropogenic factors. First, reduced atmospheric temperatures result in lower mixing-layer heights in the troposphere, thereby increasing surface-level aerosol concentrations due to limited vertical dispersion. Second, residential heating during the winter months constitutes an additional source of PM emissions relative to the summer months. While reduced atmospheric mixing contributes relatively uniformly to PM_2.5 enhancement across the urban area, residential heating is a source of emissions that varies spatially. As shown in Fig 5B, the inter-station standard deviation of monthly PM_2.5 concentrations is substantially higher during winter. This increased spatial variability suggests that individual monitoring sites make different contributions during this season, supporting the hypothesis that residential heating activities concentrated around the city periphery significantly contribute to high air pollution levels in these areas.

Spatial trends indicate that vehicular emissions and regional aerosol transport are the dominant sources of pollution, with peripheral heating activities intensifying winter concentrations. However, definitive correlations between PM_2.5 concentrations and specific emission sources cannot be established without robust source apportionment analysis, as empirical associations may not reflect actual causal relationships. The next step for this research program is to develop a comprehensive land-use regression model that incorporates spatially resolved data on traffic density, industrial emissions, residential heating patterns, and other anthropogenic activities.

4. Conclusion

This study presents the calibration of an LCS network deployed throughout Bucharest, Romania. The “InfoAer” network comprises 44 Clarity Node-S devices, which are validated against reference measurements from the Romanian National Environmental Protection Agency. The InfoAer sensors exhibit a substantial underestimation of PM_2.5 concentrations, particularly during the summer months. This bias is likely due to manufacturer pre-calibration algorithms being optimized for different environmental conditions. Sensors positioned within a 1.25 km radius of regulatory monitoring stations are selected for calibration model development. Multiple calibration approaches are evaluated using ambient temperature, relative humidity, and NO₂ concentrations as predictor variables. A parsimonious linear model that incorporates terms for temperature and relative humidity significantly improves sensor accuracy. The calibrated InfoAer measurements exhibit a strong correlation with the reference data. Then, this calibration model is applied to the entire InfoAer monitoring network.

Analysis reveals that all Bucharest sensors exceeded the EU daily PM_2.5 threshold, with the network averaging more than 60 violations in 2024. In contrast, uncalibrated data indicate threshold violations at only 8 stations, resulting in a systematic underestimation of air quality degradation. This disparity demonstrates the importance of rigorous sensor calibration for accurate air quality assessment. Seasonal analysis demonstrates high average PM_2.5 concentrations during winter months, accompanied by increased inter-sensor variability. This enhanced spatial heterogeneity suggests a non-uniform distribution of PM sources across the urban area. The spatial distribution of annual-averaged PM_2.5 concentrations in 2024 confirms that the highest pollution levels occur in peripheral urban areas. Overall, these results characterize Bucharest as a city experiencing chronic deterioration in air quality, with EU daily threshold violations occurring approximately every 5 days, and this pattern is further exacerbated during the winter months by peripheral emission sources.

In addition to developing comprehensive air pollution mitigation strategies, regulatory authorities should prioritize the following two components: (1) the establishment and expansion of robust air quality monitoring networks, and (2) the implementation of systematic network maintenance and sensor replacement protocols. The maintenance imperative is essential given that most LCSs have an operational lifespan of only 2 − 3 years, after which measurement drift increases exponentially, compromising data quality and reliability. Multi-annual datasets are essential for evaluating long-term air quality trends and assessing whether pollution-reduction policies are achieving measurable environmental improvements. Network expansion is significant in areas with high spatiotemporal variability in pollutant concentrations. In Bucharest, PM_2.5 concentrations exhibit the greatest seasonal variation in peripheral urban areas, particularly in zones with the lowest density of monitoring stations in the current network. Strategic expansion of the InfoAer LCS network and the RNMCA regulatory monitoring system in outer municipal areas represents an essential investment for comprehensive air quality management. Such expansion would enable: (1) improved spatial resolution of pollution mapping, (2) enhanced detection of localized emission sources, (3) more accurate assessment of population exposure levels, and (4) evidence-based evaluation of targeted intervention strategies.

LCS network calibration is the first step in establishing high-accuracy air quality monitoring systems at regional and national scales. Several essential research issues require systematic investigation to advance the field toward operational deployment. First, the spatial transferability of calibration models across different urban environments is a significantly understudied research area. Inter-city variations in sensor hardware (different manufacturers and models), local meteorological conditions, and aerosol physicochemical properties present substantial difficulties for model generalization. Different sensor architectures show distinct response characteristics under identical ambient conditions. This means that robust, manufacturer-agnostic calibration frameworks must be developed that demonstrate performance across diverse sensor platforms. Second, multi-annual longitudinal studies are essential for characterizing sensor drift patterns and developing time-dependent correction algorithms. Such investigations should prioritize regions with established long-term monitoring infrastructure to enable the creation of comprehensive validation datasets. Third, several promising research directions warrant systematic exploration, including ensemble machine learning models, the integration of satellite-derived aerosol optical depth, the incorporation of aerosol chemical composition data, and the development of hybrid physics-statistical models.

Supporting information

S1 Text. Description of datasets.

https://doi.org/10.1371/journal.pclm.0000780.s001

(DOCX)

S1 Data. InfoAer LCS dataset.

https://doi.org/10.1371/journal.pclm.0000780.s002

(CSV)

S2 Data. RNMCA reference dataset.

https://doi.org/10.1371/journal.pclm.0000780.s003

(CSV)

S1 Code. Calibration model R code.

https://doi.org/10.1371/journal.pclm.0000780.s004

(R)

S3 Data. Results of calibration study.

https://doi.org/10.1371/journal.pclm.0000780.s005

(ZIP)

Acknowledgments

The authors would like to thank Karunya Institute of Technology and Sciences and West University of Timisoara, Timisoara, Romania for providing all required facilities throughout the research work.

References

1. Greenstone M. Air Quality Life Index. Annual update. 2024. accessed April 2025. https://aqli.epic.uchicago.edu/reports/
2. EU Air Quality Standards. Accessed April 16, 2025. https://environment.ec.europa.eu/topics/air/air-quality/eu-air-quality-standards_en
3. WHO global air quality guidelines: particulate matter (PM2.5 and PM10), ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide. Accessed April 16, 2025. https://www.who.int/publications/i/item/9789240034228
4. Correia AW, Pope III CA, Dockery DW, Wang Y, Ezzati M, Dominici F. Effect of air pollution control on life expectancy in the United States: an analysis of 545 US counties for the period from 2000 to 2007. Epidemiology. 2013;24(1):23–31.
- View Article
- Google Scholar
5. Ebenstein A, Fan M, Greenstone M, He G, Zhou M. New evidence on the impact of sustained exposure to air pollution on life expectancy from China’s Huai River Policy. Proc Natl Acad Sci U S A. 2017;114(39):10384–9. pmid:28893980
- View Article
- PubMed/NCBI
- Google Scholar
6. van Donkelaar A, Hammer MS, Bindle L, Brauer M, Brook JR, Garay MJ, et al. Monthly Global Estimates of Fine Particulate Matter and Their Uncertainty. Environ Sci Technol. 2021;55(22):15287–300. pmid:34724610
- View Article
- PubMed/NCBI
- Google Scholar
7. Raysoni AU, Pinakana SD, Mendez E, Wladyka D, Sepielak K, Temby O. A Review of Literature on the Usage of Low-Cost Sensors to Measure Particulate Matter. Earth. 2023;4(1):168–86.
- View Article
- Google Scholar
8. Liang L. Calibrating low-cost sensors for ambient air monitoring: Techniques, trends, and challenges. Environ Res. 2021;197:111163. pmid:33887275
- View Article
- PubMed/NCBI
- Google Scholar
9. Morawska L, Thai PK, Liu X, Asumadu-Sakyi A, Ayoko G, Bartonova A, et al. Applications of low-cost sensing technologies for air quality monitoring and exposure assessment: How far have they gone?. Environ Int. 2018;116:286–99. pmid:29704807
- View Article
- PubMed/NCBI
- Google Scholar
10. Directive 2008/50/EC of the European Parliament and of the European Council of 21 May 2008 on ambient air quality and cleaner air in Europe. Annex I – A. Data quality objectives. Accessed 16 April, 2025. https://eur-lex.europa.eu/eli/dir/2008/50/oj/eng
11. EC Working Group on Guidance for the Demonstration of Equivalence. Guide to the demonstration of equivalence of ambient air monitoring methods. 2010. Accessed 16 April, 2025. https://www.aces.su.se/reflab/wp-content/uploads/2016/11/Demonstration_of_Equivalence_of_Ambient_Air_Monitoring.pdf
12. Giordano MR, Malings C, Pandis SN, Presto AA, McNeill VF, Westervelt DM, et al. From low-cost sensors to high-quality data: A summary of challenges and best practices for effectively calibrating low-cost particulate matter mass sensors. Journal of Aerosol Science. 2021;158:105833.
- View Article
- Google Scholar
13. Hong G-H, Le T-C, Tu J-W, Wang C, Chang S-C, Yu J-Y, et al. Long-term evaluation and calibration of three types of low-cost PM2.5 sensors at different air quality monitoring stations. Journal of Aerosol Science. 2021;157:105829.
- View Article
- Google Scholar
14. Patel MY, Vannucci PF, Kim J, Berelson WM, Cohen RC. Towards a hygroscopic growth calibration for low-cost PM 2.5 sensors. Atmospheric Measurement Techniques. 2024;17(3):1051–60.
- View Article
- Google Scholar
15. Feng Z, Zheng L, Ren B, Liu D, Huang J, Xue N. Feasibility of low-cost particulate matter sensors for long-term environmental monitoring: Field evaluation and calibration. Sci Total Environ. 2024;945:174089. pmid:38897458
- View Article
- PubMed/NCBI
- Google Scholar
16. European Commission 2024. Key decisions of November 2023 infringement package. Accessed April 16, 2025. https://ec.europa.eu/commission/presscorner/detail/en/inf_23_5380
17. Action brought on 12 October 2018 — European Commission v Romania (Case C-638/18). Accessed April 16, 2025. https://eur-lex.europa.eu/legal-content/EN/TXT/?qid=1557176797270&uri=CELEX:62018CN0638
18. Clarity Node-S Air Sensor. Accessed 2025 April 16. https://www.clarity.io/products/clarity-node-s
19. Comde-Derenda Air Pollution Monitor APM-2. Accessed April 16, 2025. https://www.comde-derenda.com/produkte/apm-2-2/
20. Equivalency of the results obtained during 2023 with automated equipment with respect to the reference method for PM₁₀ and PM_2.5 measurements from RNMCA stations in Brasov, Bucuresti, Cluj, Dolj, Iasi. ANMP NEPA. report. 2023. Accessed April 16, 2025. https://www.anpm.ro/rapoarte-anuale-de-evaluare-a-calitatii-aerului/
21. Kumar V, Sahu M. Evaluation of nine machine learning regression algorithms for calibration of low-cost PM2.5 sensor. Journal of Aerosol Science. 2021;157:105809.
- View Article
- Google Scholar
22. Ma J, Yu Z, Qu Y, Xu J, Cao Y. Application of the XGBoost Machine Learning Method in PM2.5 Prediction: A Case Study of Shanghai. Aerosol Air Qual Res. 2020;20(1):128–38.
- View Article
- Google Scholar
23. Kang J, Choi K. Calibration Methods for Low-Cost Particulate Matter Sensors Considering Seasonal Variability. Sensors (Basel). 2024;24(10):3023. pmid:38793878
- View Article
- PubMed/NCBI
- Google Scholar
24. Blaga R, Gautam S. Improving PM10 sensor accuracy in urban areas through calibration in Timișoara. npj Clim Atmos Sci. 2024;7(1).
- View Article
- Google Scholar
25. Crilley LR, Shaw M, Pound R, Kramer LJ, Price R, Young S, et al. Evaluation of a low-cost optical particle counter (Alphasense OPC-N2) for ambient air monitoring. Atmos Meas Tech. 2018;11(2):709–20.
- View Article
- Google Scholar
26. Svenningsson B, Rissler J, Swietlicki E, Mircea M, Bilde M, Facchini MC, et al. Hygroscopic growth and critical supersaturations for mixed aerosol particles of inorganic and organic compounds of atmospheric relevance. Atmos Chem Phys. 2006;6(7):1937–52.
- View Article
- Google Scholar
27. Karagulian F, Barbiere M, Kotsev A, Spinelle L, Gerboles M, Lagler F, et al. Review of the Performance of Low-Cost Sensors for Air Quality Monitoring. Atmosphere. 2019;10(9):506.
- View Article
- Google Scholar
28. Aix M-L, Schmitz S, Bicout DJ. Calibration methodology of low-cost sensors for high-quality monitoring of fine particulate matter. Sci Total Environ. 2023;889:164063. pmid:37201842
- View Article
- PubMed/NCBI
- Google Scholar
29. Malyan V, Kumar V, Sahu M. Significance of sources and size distribution on calibration of low-cost particle sensors: Evidence from a field sampling campaign. Journal of Aerosol Science. 2023;168:106114.
- View Article
- Google Scholar

[ref1] 1. Greenstone M. Air Quality Life Index. Annual update. 2024. accessed April 2025. https://aqli.epic.uchicago.edu/reports/

[ref2] 2. EU Air Quality Standards. Accessed April 16, 2025. https://environment.ec.europa.eu/topics/air/air-quality/eu-air-quality-standards_en

[ref3] 3. WHO global air quality guidelines: particulate matter (PM2.5 and PM10), ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide. Accessed April 16, 2025. https://www.who.int/publications/i/item/9789240034228

[ref4] 4. Correia AW, Pope III CA, Dockery DW, Wang Y, Ezzati M, Dominici F. Effect of air pollution control on life expectancy in the United States: an analysis of 545 US counties for the period from 2000 to 2007. Epidemiology. 2013;24(1):23–31.
View Article
Google Scholar

[5] View Article

[6] Google Scholar

[ref5] 5. Ebenstein A, Fan M, Greenstone M, He G, Zhou M. New evidence on the impact of sustained exposure to air pollution on life expectancy from China’s Huai River Policy. Proc Natl Acad Sci U S A. 2017;114(39):10384–9. pmid:28893980
View Article
PubMed/NCBI
Google Scholar

[8] View Article

[9] PubMed/NCBI

[10] Google Scholar

[ref6] 6. van Donkelaar A, Hammer MS, Bindle L, Brauer M, Brook JR, Garay MJ, et al. Monthly Global Estimates of Fine Particulate Matter and Their Uncertainty. Environ Sci Technol. 2021;55(22):15287–300. pmid:34724610
View Article
PubMed/NCBI
Google Scholar

[12] View Article

[13] PubMed/NCBI

[14] Google Scholar

[ref7] 7. Raysoni AU, Pinakana SD, Mendez E, Wladyka D, Sepielak K, Temby O. A Review of Literature on the Usage of Low-Cost Sensors to Measure Particulate Matter. Earth. 2023;4(1):168–86.
View Article
Google Scholar

[16] View Article

[17] Google Scholar

[ref8] 8. Liang L. Calibrating low-cost sensors for ambient air monitoring: Techniques, trends, and challenges. Environ Res. 2021;197:111163. pmid:33887275
View Article
PubMed/NCBI
Google Scholar

[19] View Article

[20] PubMed/NCBI

[21] Google Scholar

[ref9] 9. Morawska L, Thai PK, Liu X, Asumadu-Sakyi A, Ayoko G, Bartonova A, et al. Applications of low-cost sensing technologies for air quality monitoring and exposure assessment: How far have they gone?. Environ Int. 2018;116:286–99. pmid:29704807
View Article
PubMed/NCBI
Google Scholar

[23] View Article

[24] PubMed/NCBI

[25] Google Scholar

[ref10] 10. Directive 2008/50/EC of the European Parliament and of the European Council of 21 May 2008 on ambient air quality and cleaner air in Europe. Annex I – A. Data quality objectives. Accessed 16 April, 2025. https://eur-lex.europa.eu/eli/dir/2008/50/oj/eng

[ref11] 11. EC Working Group on Guidance for the Demonstration of Equivalence. Guide to the demonstration of equivalence of ambient air monitoring methods. 2010. Accessed 16 April, 2025. https://www.aces.su.se/reflab/wp-content/uploads/2016/11/Demonstration_of_Equivalence_of_Ambient_Air_Monitoring.pdf

[ref12] 12. Giordano MR, Malings C, Pandis SN, Presto AA, McNeill VF, Westervelt DM, et al. From low-cost sensors to high-quality data: A summary of challenges and best practices for effectively calibrating low-cost particulate matter mass sensors. Journal of Aerosol Science. 2021;158:105833.
View Article
Google Scholar

[29] View Article

[30] Google Scholar

[ref13] 13. Hong G-H, Le T-C, Tu J-W, Wang C, Chang S-C, Yu J-Y, et al. Long-term evaluation and calibration of three types of low-cost PM2.5 sensors at different air quality monitoring stations. Journal of Aerosol Science. 2021;157:105829.
View Article
Google Scholar

[32] View Article

[33] Google Scholar

[ref14] 14. Patel MY, Vannucci PF, Kim J, Berelson WM, Cohen RC. Towards a hygroscopic growth calibration for low-cost PM 2.5 sensors. Atmospheric Measurement Techniques. 2024;17(3):1051–60.
View Article
Google Scholar

[35] View Article

[36] Google Scholar

[ref15] 15. Feng Z, Zheng L, Ren B, Liu D, Huang J, Xue N. Feasibility of low-cost particulate matter sensors for long-term environmental monitoring: Field evaluation and calibration. Sci Total Environ. 2024;945:174089. pmid:38897458
View Article
PubMed/NCBI
Google Scholar

[38] View Article

[39] PubMed/NCBI

[40] Google Scholar

[ref16] 16. European Commission 2024. Key decisions of November 2023 infringement package. Accessed April 16, 2025. https://ec.europa.eu/commission/presscorner/detail/en/inf_23_5380

[ref17] 17. Action brought on 12 October 2018 — European Commission v Romania (Case C-638/18). Accessed April 16, 2025. https://eur-lex.europa.eu/legal-content/EN/TXT/?qid=1557176797270&uri=CELEX:62018CN0638

[ref18] 18. Clarity Node-S Air Sensor. Accessed 2025 April 16. https://www.clarity.io/products/clarity-node-s

[ref19] 19. Comde-Derenda Air Pollution Monitor APM-2. Accessed April 16, 2025. https://www.comde-derenda.com/produkte/apm-2-2/

[ref20] 20. Equivalency of the results obtained during 2023 with automated equipment with respect to the reference method for PM₁₀ and PM_2.5 measurements from RNMCA stations in Brasov, Bucuresti, Cluj, Dolj, Iasi. ANMP NEPA. report. 2023. Accessed April 16, 2025. https://www.anpm.ro/rapoarte-anuale-de-evaluare-a-calitatii-aerului/

[ref21] 21. Kumar V, Sahu M. Evaluation of nine machine learning regression algorithms for calibration of low-cost PM2.5 sensor. Journal of Aerosol Science. 2021;157:105809.
View Article
Google Scholar

[47] View Article

[48] Google Scholar

[ref22] 22. Ma J, Yu Z, Qu Y, Xu J, Cao Y. Application of the XGBoost Machine Learning Method in PM2.5 Prediction: A Case Study of Shanghai. Aerosol Air Qual Res. 2020;20(1):128–38.
View Article
Google Scholar

[50] View Article

[51] Google Scholar

[ref23] 23. Kang J, Choi K. Calibration Methods for Low-Cost Particulate Matter Sensors Considering Seasonal Variability. Sensors (Basel). 2024;24(10):3023. pmid:38793878
View Article
PubMed/NCBI
Google Scholar

[53] View Article

[54] PubMed/NCBI

[55] Google Scholar

[ref24] 24. Blaga R, Gautam S. Improving PM10 sensor accuracy in urban areas through calibration in Timișoara. npj Clim Atmos Sci. 2024;7(1).
View Article
Google Scholar

[57] View Article

[58] Google Scholar

[ref25] 25. Crilley LR, Shaw M, Pound R, Kramer LJ, Price R, Young S, et al. Evaluation of a low-cost optical particle counter (Alphasense OPC-N2) for ambient air monitoring. Atmos Meas Tech. 2018;11(2):709–20.
View Article
Google Scholar

[60] View Article

[61] Google Scholar

[ref26] 26. Svenningsson B, Rissler J, Swietlicki E, Mircea M, Bilde M, Facchini MC, et al. Hygroscopic growth and critical supersaturations for mixed aerosol particles of inorganic and organic compounds of atmospheric relevance. Atmos Chem Phys. 2006;6(7):1937–52.
View Article
Google Scholar

[63] View Article

[64] Google Scholar

[ref27] 27. Karagulian F, Barbiere M, Kotsev A, Spinelle L, Gerboles M, Lagler F, et al. Review of the Performance of Low-Cost Sensors for Air Quality Monitoring. Atmosphere. 2019;10(9):506.
View Article
Google Scholar

[66] View Article

[67] Google Scholar

[ref28] 28. Aix M-L, Schmitz S, Bicout DJ. Calibration methodology of low-cost sensors for high-quality monitoring of fine particulate matter. Sci Total Environ. 2023;889:164063. pmid:37201842
View Article
PubMed/NCBI
Google Scholar

[69] View Article

[70] PubMed/NCBI

[71] Google Scholar

[ref29] 29. Malyan V, Kumar V, Sahu M. Significance of sources and size distribution on calibration of low-cost particle sensors: Evidence from a field sampling campaign. Journal of Aerosol Science. 2023;168:106114.
View Article
Google Scholar

[73] View Article

[74] Google Scholar

Figures

Abstract

1. Introduction

2. Data and method

(a) Study area

(b) Low-cost PM sensor (LCS) data

(c) Reference data

3. Air quality assessment

(a) Calibration of LCS data

(b) Validation of calibration model

(c) Compliance with the EU air quality framework

(d) Spatial and seasonal variations in PM levels

4. Conclusion

Supporting information

S1 Text. Description of datasets.

S1 Data. InfoAer LCS dataset.

S2 Data. RNMCA reference dataset.

S1 Code. Calibration model R code.

S3 Data. Results of calibration study.

Acknowledgments

References