Air Quality in Cities: What Do Sparse Sensor Networks Miss?

Abstract

Stationary air quality networks are increasingly being used to refine urban air pollution exposure assessments and inform local government decision-making, yet to date, a comprehensive assessment of the extent to which such measurements provide sufficient spatial resolution to yield meaningful insights in these contexts is lacking. In this study, the spatial representativeness of sparse air quality monitoring measurements was characterized by pairing measurements at a fixed station with simultaneous mobile measurements of PM_2.5 at a dense number of surrounding locations within a 250 m radius. Data collection was conducted at three contrasting areas across the Boston Metro area (MA, USA) during two seasons (Spring and Summer 2025). Under typical conditions, the mean difference between stationary and mobile measurements across all sites and seasons was 2.0 μg/m³ (median: 1.5 μg/m³, IQR: 0.6–2.8 μg/m³); however, certain dynamic events created differences exceeding 300% in relative terms and 800 μg/m³ in difference terms. In 10.4% of cases, with this proportion reaching as high as 21.8% at one site during spring, measurements were sufficiently different to result in different AQI classifications, indicating that local-scale concentration differences can translate into different public-facing air quality messages. Euclidean distance alone could not explain the differences. Higher relative humidity was associated with larger Center–Mobile discrepancies (r = −0.84), though this association should be interpreted cautiously given the limited number of sampling days. Two observed extreme events (a building smoke plume and a nearby idling truck) produced sharp, localized pollution spikes, providing an opportunity to quantify the extent to which short-lived sources and street-canyon effects can drive steep gradients at small spatial scales. This evidence suggests that while sparse sensor networks may be sufficient to characterize short-term air pollution exposures in relatively homogeneous settings, such an approach may not be sufficient in complex urban environments.

1. Introduction

Air pollution is a major global health risk, − with the World Health Organization attributing 6.7 million premature deaths annually to air pollution. Specifically, studies consistently find strong associations between exposure to air pollutants and respiratory illnesses, including asthma, chronic obstructive pulmonary disease, and pneumonia, , as well as elevated risk among children and older adults. , Working toward solutions to this epidemic requires both exposure estimation and the design of health-protective interventions, the success of which depends critically on an accurate understanding of the conditions individuals actually experience in their immediate surroundings.

Yet available air quality data tend to be spatially sparse. In the U.S., such data either come from federal regulatory networks (tens of stations per state) or from networks of emerging lower-cost air quality sensors marketed to the general public (e.g., PurpleAirg) exploration in the air pollution exposure literature of using such networks of stationary monitors to provide information on air quality, typically with some assumption about the radius of applicability or approach for interpolation or extrapolation across the study area. − However, a persistent challenge with using stationary monitoring networks lies in the uncertainty regarding the extent of spatial representativeness of each monitor. Air pollutant concentrations can be especially heterogeneous in urban environments as they vary based on spatial emission patterns, weather conditions (especially those driving chemical reactions and secondary pollutant formation), , and surrounding land use. , For instance, it has been observed that traffic signals that create congestion, as well as pedestrian zones with a high density of shops, restaurants, and cafes, often act as localized hot spots that generate elevated peaks of air pollution in their immediate surroundings. Given that such factors create increasingly complex environments in cities, this highlights the need to assess uncertainty introduced into estimates of localized air pollution exposure as derived from stationary monitoring networksthat is, to better understand whether these sparse networks are sufficient for approximating urban air quality or to quantify what these sparse networks might be missing.

Previous research has demonstrated that mobile air quality sensors can more accurately identify localized pollution hotspots and better characterize personal exposures, − with pollutant variability at scales of tens of meters. Typically, by aggregating repeated traverses conducted over days to months, mobile monitoring can produce fine-grained maps spanning multiple neighborhoods to entire cities. ,, Recent mobile monitoring studies have similarly shown that low-cost PM_2.5 sensors can identify fine-scale spatial variability that may not be captured by sparse fixed monitoring networks. , Previous work also reported increasing sensor uncertainty with increasing platform speed, , and therein, the sampling approach of this study helps to minimize speed-related uncertainty while maintaining high spatial resolution. In dense urban settings, prior studies have estimated that stationary monitors correlate with long-term average conditions within less than 1 km²; however, long-term averages may not capture the relevant information needed to assess health risk. As a result, it remains unclear how closely a nearby stationary monitor reflects pollutant conditions experienced on the pedestrian scale at a given moment in time. Addressing this uncertainty requires direct, simultaneous comparisons between stationary and nearby mobile measurements within a defined local radius.

This work therefore implements a targeted quantitative study to examine the extent to which stationary air quality monitors represent (or miss) conditions within their immediate urban surroundings. Specifically, taking PM_2.5 as a case study, measurements are made at ultrahigh spatial density (subblock resolution) within a 250 m radius of a fixed location to observe variability representing pedestrian experience. Measured differences are quantified in terms of their potential impact on exposure assessments and evaluated in the context of urban elements, weather patterns, and unique events. The results suggest the importance of reconsidering how sparse data are used for exposure assessment, as well as the importance of accounting for temporally short but potentially frequent urban pollution events.

2. Materials and Methods

The study design centered on a stationary fixed monitoring station, where continuous air quality data were recorded while measurements were simultaneously collected at high density within a radius of 250 m using a mobile approach. This was implemented at three locations in Greater Boston, Massachusetts, over two seasons (spring and summer). An overview of the study and subsequent analyses is illustrated in Figure and described in detail in the following sections.

1.

Overview of study setup and experimental methodology. Field equipment (a) is used to cover a set study area (b), resulting in comparable time-series data (c). (a) Mobile Sensor setup (left) and Center Station stationary tripod (right). (b) Example of a single Mobile-Sensor trajectory for a given experiment day, with corresponding tree-canopy coverage and building height. (c) Illustrative example comparing Mobile-Sensor and Center Station readings with 1σ error bands around 2 min averaged measurements.

2.1. Sensor Suite for Data Collection

PM_2.5 concentrations and meteorological variables were recorded at both a central location (“Center Station”, representing a typical location of a single sensor in an urban air quality sensor network) and surrounding locations using a mobile setup. The same sensors were used for all measurements; specifications are summarized in Table . The PurpleAir sensor logs readings on an SD card for later download, while the Kestrel DROP logs to internal memory for later download via Bluetooth. Wind data were recorded manually by the operators.

1. Specifications of Sensors Deployed in the Study for Measurement of Particulate Matter and Other Environmental Conditions.

Target	Sensor Model	Sampling Period (s)	Mechanism	Range	Accuracy
PM2.5	PurpleAir Classic	1	Light scattering	0–500 μg/m3	Only consistency reported: ±10 μg/m3 at <100 μg/m3, ±10% at 100–500 μg/m3 (ref )
T	Kestrel DROP 3	20	Thermistor	–10–55 °C	T ±0.5 °C
RH			Capacitive polymer	10–90% @ 25 °C	RH ±2%
Wind speed	Wintact WT8907	10	Rotor-based	0–45 m/s	±3% or ±0.1 m/s

At the Center Station, all sensors were mounted on a single tripod at 183 cm (6 ft), with height selected to reflect a measurement relevant to human exposure while avoiding interference from the study operators and passing pedestrians. For mobile monitoring, each “Mobile Sensor” set was mounted on a lightweight pole that could be easily carried by hand, with sensors also positioned at 183 cm (6 ft). The anemometer was installed at the top of each pole. Both systems are illustrated in Figure a.

2.2. Study Locations and Timing

Study locations were selected to represent areas that people frequently traverse and congregate in, such that results would have meaningful implications for real-world individual exposure. The study area was discretized using the Uber H3 geospatial indexing system at Resolution 8. At this resolution, the Greater Boston area is divided into cells with an average area of 0.73 km² and an edge length of approximately 0.46 km. The visit frequencies were calculated by spatially joining GPS-mobility tracks to grid cells using data from the BostonWalks study. From this visit frequency distribution, the ten most popular cells were selected as candidate locations. These candidate locations were further screened according to the following criteria: (1) urban form diversity: representation of distinct environments such as open green spaces, residential blocks, neighborhood commercial corridors, and downtown street canyons, and (2) field safety and operational practicality (e.g., secure areas for equipment storage, charging, and setup). This resulted in the selection of three final study areas, specifically the Christian Science Plaza (“CS”), Downtown Boston (“DT”), and the eastern side of MIT campus (“MIT”). Urban characteristics of each are shown visually in Figure (left column) and summarized in more detail in Figure S5.

2.

Difference maps of PM_2.5 for three study areas (rows) and two study seasons (columns). Each dot marks a measured location, with color indicating percent difference between Mobile and Center measurements, and marker size (area) indicating the difference. Bold marker edges denote locations where the Mobile Sensor and Center Station measurements fall into different AQI_PM categories. Note that the color bar and the dots are scaled uniformly throughout the manuscript.

Air quality in each location was characterized in two different seasons to capture seasonal variability, specifically in March and June 2025. Each seasonal measurement effort consisted of 4–5 days of sampling to complete data collection for all 3 areas, with measurements taken consistently from approximately 10 a.m. to 2 p.m. daily. Days of extreme weather (such as precipitation) were avoided. Average temperatures in the two seasonal data-collection windows were 9.9 °C and 28.8 °C, respectively, capturing reasonable late-winter and midsummer weather for this region.

2.3. Site Sampling Design

A set of evenly distributed measurement locations was generated across the area within a 250 m radius of Center Stations. Using the OpenStreetMap road network, street segments were interpolated at 25–60 m intervals, subject to practical constraints related to road length, land use, and sidewalk accessibility, to create a dense lattice of potential measurement locations. To remove anomalies and redundancy, nearby nodes were grouped using a DBSCAN algorithm with a 25 m distance threshold, and each cluster centroid was designated as a sampling location. Sampling locations were then projected back onto the nearest accessible street and verified to ensure feasibility (e.g., practical access). The final layout provided evenly spaced measurement locations with average nearest-neighbor distances of 53 m at CS, 56 m at MIT, and 47 m in DT, capturing spatial variation across the testing locations while maintaining reasonably uniform coverage density (see Figure S6) around each center station. This process yielded 77, 87, and 57 sampling points, respectively, in the three study areas designated CS, DT, and MIT. The total number of points differed due to differences in number of road and pedestrian pathway segments in each study area.

2.4. Data Collection

On each measurement day, data collection followed a standardized multistep protocol. The procedure included (1) sensor warm-up and consistency checks, (2) placement of the Center Station, (3) Mobile Sensor data collection, and (4) end-of-day consistency checks. Consistency checking consisted of collecting 10 min of collocated data collection, i.e., all Mobile Sensors located adjacent to the Center Station, which was analyzed to confirm no change in relative cross-calibration factors. For collection of mobile measurements, the sensors were continuously recording time-stamped data while the mobile operator carried the Mobile Sensor setup between locations and remained at each designated location for 10 min for the sensors to reach steady state after shifting locations. GPS location of the Mobile Sensor was recorded using operator cellular devices. At each location, operators also recorded environmental observations on standardized forms, including arrival/departure times, wind speed and direction at both start and end of the 10 min, and ratings of surrounding tree coverage, traffic volume, ambient odor, etc. (each scored on a scale of 1–5). Wind conditions across sites and seasons are summarized in Figure S8 and Table S2. To minimize variability in ratings between operators, all operators were trained on standardized examples. Three Mobile Sensors were simultaneously deployed on each experiment day, resulting in measurement of 40–70 locations per day.

2.5. Data Analysis

2.5.1. Sensor Cross-Calibration

Before beginning field data collection, cross-calibration of all PurpleAir PM_2.5 sensors (both Center and Mobile) was completed in a lab setting. The Mobile PM_2.5 readings were then adjusted using the linear regression relationships established using the results from cross-calibration to correct for sensor-to-sensor bias (see Figures S1–S3).

2.5.2. Data Alignment and Postprocessing

Timestamps on sensor readings were used to temporally align the data across all locations and sensor devices. Preliminary experiments quantifying sensor reading stabilization time frames under urban conditions seen in this study indicated that approximately 8 min were needed after moving the Mobile Sensors before data could be considered representative of the new location. This stabilization time was determined from a PurpleAir equilibrium test (Figure S4). Therefore, the first 8 min of data from each measurement location were discarded, with the final 2 min of PM_2.5 readings averaged to report conditions at that location. The same 2 min window of data from the Center Station was used to calculate the comparable reading.

2.5.3. Comparative Analyses

Both the differences and relative differences between Center Station and Mobile Sensors were calculated for each mobile measurement location. The difference is defined as

$$\mathrm{\Delta }\mathrm{P}{\mathrm{M}}_{2.5}=\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{Center}}-\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{Mobile}}$$ 1

and relative difference as

$$\mathrm{\Delta }\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{\%}}=\frac{\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{Center}}-\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{Mobile}}}{\mathrm{P}{\mathrm{M}}_{2.5,\mathrm{Center}}}\times 100\mathrm{\%}$$ 2

where PM _2.5,Center and PM_2.5,Mobile represent the 2 min averages extracted in Section .

To provide insight into the meaning of these numbers for human health, the Air Quality Index for PM (AQI_PM) was derived for both the Center Station and Mobile Sensors based on their cross-calibrated PM_2.5 concentrations, following the U.S. Environmental Protection Agency (EPA) breakpoint method. The resulting AQI_PM values were categorized according to standard AQI thresholds using the regulatory schedule in effect at the time of the experiments.

2.5.4. Exposure Implications

To assess how well stationary monitors represented local exposure under varying urban, meteorological, and temporal conditions, the Mobile Sensors operators were considered as proxies for mobile exposure subjects. Because at each sampling point the operator remained stationary for approximately 10 min while carrying a pole-mounted Mobile Sensor, their study pathway yielded paired measurements from the Mobile Sensors and the Center Station over the same time window (Figure b,c). For each visit, the differences and percent differences between the cross-calibrated Center and Mobile readings were computed.

Temperature and relative humidity were binned at 4 °C and 10% intervals, respectively, for visualization purposes; in the ordinary least-squares (OLS) regression described in Supporting Information Section S3, these variables were treated as continuous predictors for the formal inference. Between-group differences in discrepancy across temperature campaigns were additionally tested using a two-sample t test. How these discrepancies varied with temperature, humidity, time of day, distance from the Center Station, and study location was then examined using Spearman rank correlations for continuous predictors and Kruskal–Wallis tests for categorical groupings (location, season). The trends should be interpreted as exploratory given the small number of sampling days, which introduces temporal confounding between meteorological conditions and unobserved day-specific factors. To characterize typical daily conditions and limit the influence of extreme outliers, the analysis was restricted to the lower 95% of observed pollution readings.

3. Results and Discussion

3.1. Spatial Heterogeneity

The heterogeneity in PM_2.5 measured across the three sites and two studied seasons is visualized in Figure , where each measurement location is represented as a dot for which the color indicates the percentage difference with respect to the Center station (eq ) and the marker size (the area of the marker) reflects the difference with respect to the Center station (value of eq ). Locations where the Center and Mobile data fell into different AQI_PM categories, providing an additional policy-relevant perspective on how concentration differences may be communicated to the public, have bold symbol outlines. It can be seen that air quality levels varied significantly within the study area of 250 m radius. Taking site DT as an example, during the spring measurements, 19 of 87 (21.8%) of Mobile sites recorded values in a different AQI_PM category than the Center Station, with the average difference in PM_2.5 concentration between the Center Station and surrounding points being 6.1 μg/m³, of notable magnitude with respect to the AQI category breakpoint at 9 μg/m³. However, the level of spatial discrepancy differed notably between the three areas. For example, in spring, the CS site had only 10.4% of Mobile sites in a different AQI category from the Center station but showed an average difference of 16.1 μg/m³. This elevated average was attributable to an atypical boiler ignition event (described in detail in Section ) that considerably increased apparent differences. MIT had the most homogeneous conditions in spring, with an AQI_PM difference rate of only 3.5% and an average concentration difference of 1.4 μg/m³, but even so, points within the study area varied from 85.9% lower concentration to 162.8% higher concentration as compared to the Center station. For reference, Table S1 provides a complete summary of statistics comparing Mobile and Center stations across all sites and seasons.

Given the similarity in bulk characteristics such as tree cover, building density, and building height, these results raise an interesting question as to the driving factors behind the recorded heterogeneity. Leveraging observation forms completed at each Mobile location, air quality patterns were compared to street layout, pedestrian density, traffic density, and perceived tree coverage; however, no significant relationships were found (a complementary regression using GIS-derived morphological covariates is presented in Section S3, which similarly leaves most variance unexplained). Although the complexity of urban dynamics relative to the amount of data collected in this study makes assigning causality difficult, these steep concentration gradients, often occurring over distances of only tens of meters, are consistent with the physics and spatial scales of localized features that are well described in the literature, such as street canyons, building wakes, and traffic congestion. ,

The significance of heterogeneity observed also differed between seasons. Generally, PM_2.5 had a larger difference between Mobile and Center in spring than in summer, especially when excluding exceptional events (see Section ). For example, in DT, the AQI_PM category difference rate dropped from 21% in spring to 6.9% in summer, while at CS, the AQI_PM difference rate dropped from 10.4% to 0%. This may reflect patterns reported in prior work, where increases in vertical mixing during summer reduce spatial heterogeneity, whereas cooler-season conditions favor pollutant accumulation and localized heterogeneity. However, at the MIT site, the opposite was observed: the rate of difference in AQI category actually increased from 3.5% in spring to 19.3% in summer even while the percentage differences in concentration dropped significantly. Closer analysis revealed this was due to summer PM_2.5 concentrations hovering near the regulatory AQI breakpoint of 9 μg/m³ such that even a 1 μ g/m³ difference (very near sensor precision levels) can trigger a category change. This highlights a real challenge associated with understanding and reporting risk: on one hand, measurement precision suggests that categorical aggregates are potentially more appropriate, while on the other hand, it is clear that air quality levels varying near thresholds can lead to results that are inconsistent with meaningful risk differences. Consequently, AQI category differences should be interpreted alongside the underlying concentration differences, as threshold-based classification systems can amplify perceived differences of relatively small concentration differences when observations occur near regulatory breakpoints.

3.2. Localized Pollution Events

Although unexpected, the documentation of two highly localized pollution events during this study provided an opportunity to illustrate real urban pollution scenarios and the associated exposure differences that these can create within urban environments. Both events were observed during the spring measurement campaign and are summarized in Figure .

3.

Highly localized pollution events captured at two locations. (a) Smoke plume event observed over the CS study area caused by a boiler ignition issue and (b) delivery truck idling adjacent to the Center Station in the DT area. The middle panel provides a comparison of Center Station and Mobile Sensor PM_2.5 readings, while the right panel localizes these measurements in space, overlaid on the site maps. Together, these events demonstrate that short-lived, localized emission sources can produce concentration gradients of hundreds of μg/m³ over distances of less than one block, creating hazardous exposure conditions that would almost certainly be missed by a sparse stationary monitoring network. Note that the color bar and the dots are scaled uniformly throughout the manuscript and that the y-axis range of the time-series plots is identical.

First, in the morning of March 19, 2025, a distinct smoke plume was observed over the CS area, later identified as a boiler ignition issue from an adjacent building. The plume remained visible for approximately 10 min (09:50–10:00 a.m.), though measurements captured event aftermath until approximately 10:40 a.m. During the event, the fixed Center Station at the CS area recorded PM_2.5 concentrations peaking near 800 μg/m³ while Mobile Sensors did not register a spike of comparable intensity, or in some cases registered no elevated pollution at all. Following the event, once the Mobile Sensors were relocated closer to the emission source (starting at 10:22 and 10:33), their readings became up to 11.9 μg/m³ higher than the Center Station. This event documented a localized yet representative urban scenario: pollution reached “Hazardous” AQI levels at specific locations, while nearby areas remained virtually unaffected. Such contrast illustrates an extreme case of spatial heterogeneity in urban air quality that poses particular risk to individuals with underlying health conditions.

The second event was documented at the DT location when a delivery truck parked and began idling adjacent to the Center Station (March 20, 2025, 09:55–10:05 a.m.). During this time, similar to the smoke plume event, the Center Station recorded sharp increases in PM_2.5 concentrations, reaching up to 400 μg/m³surpassing typical readings during wildfire events, again pushing into the “Hazardous” AQI category and reinforcing the need for anti-idling laws. In this case, although the Mobile Sensors were positioned less than one block away, the event measured at the Center Station did not significantly impact air quality at the Mobile Sensor locations. This extreme localized heterogeneity may have been driven by building density at this location, where buildings are especially tall relative to road width (see Figure S5); however, unlike the boiler fire, vehicle idling represents an unfortunately extremely common type of pollution event in urban settings. This study shows that these events are both significant in terms of potential health effects and extremely likely to be missed by any sensor network with density less than measuring on every block (economically intractable). This raises important questions about how we might capture these types of pollution events, both to better understand their impacts but also to reduce them, e.g., through better anti-idling enforcement programs.

Together, these incidents illustrate the pronounced spatial variability of extreme pollution events, even within small domains (250 m radius). Short-term, localized emissions can create steep concentration gradients and hazardous conditions for urban residents, yet they are unlikely to be captured by conventional monitoring networks that rely on sparse, widely spaced stations.

3.3. Implications for Exposure Assessment

Results of the proxy exposure analysis based on Mobile operator trajectories are provided in Figure . While the most obvious potential predictor, distance from the Center Station did not explain observed concentration differences (Figure b). However, results from the two seasonal data collection campaigns revealed significant differences. A between-group comparison confirmed that discrepancies differed significantly between the cool (<14 °C, mean = −0.77 μg/m³, n = 199) and warm (>14 °C, mean = −1.41 μg/m³, n = 177) campaigns (two-sample t test, t = 2.25, p = 0.025). This difference should nonetheless be interpreted cautiously: the two campaigns sampled a bimodal temperature distribution with no observations between approximately 13 and 21 °C, so the relationship cannot be characterized across the continuous range. Moreover, because sites were not revisited on multiple days within each campaign, temperature is confounded with day-specific conditions, and we treat this association as exploratory rather than evidence of a causal temperature mechanism. Humidity analyses showed a consistent trend in which higher humidity was correlated with greater differences between Center and Mobile monitors. While differences varied across testing locations, humidity readings binned at 10% intervals showed a strong negative correlation (r = −0.84) between humidity and Center–Mobile agreement, indicating that higher humidity reduces the accuracy of exposure assessment in ambient conditions. However, given the small number of sampling days, this correlation should be treated as exploratory. Humidity is necessarily confounded with unobserved day-specific conditions, and the association does not establish a causal meteorological mechanism. Analyses did not reveal any consistent relationship between monitoring representativeness and time of day.

4.

Measurement differences across seasons and locations. (a) Median difference (IQR shaded) between Center and Mobile measurements binned by temperature, relative humidity, and time of day, indicating conditions under which stationary monitors underestimate or overestimate exposure. (b) Averaged, cross-calibrated PM_2.5 difference between Center and Mobile Sensors as a function of Euclidean distance from the Center Station (IQR shaded). (c) Spatial variation in cross-calibrated PM_2.5 differences across the three study locations.

A comparison across the sites shows statistically significant and varied results in terms of measurement differences when basing results on the Center station versus actual Mobile measurements (see Figure c, where data are averaged across both seasons and restricted to the lower 95th percentile of observations to remove the effect of exceptional eventsFigure S7 provides results without this filtering). Among the three study areas, the smallest mean difference was observed at MIT (1.34 μg/m³, IQR: 0.45–1.81), followed by CS (1.8 μg/m³, IQR: 0.52–2.39) and DT (2.8 μg/m³, IQR: 0.89–3.58). Pooled across all sites and seasons, the mean difference was 2.0 μg/m³ (IQR: 0.6–2.8 μg/m³), based on the lower 95th percentile of observations to exclude exceptional pollution events. Notably, these measurable mean differences are overlain on highly dynamic conditions, e.g., in the downtown area, the Center Station both underestimated and overestimated exposure by up to 15 μg/m³ (Figure c). Overall, these findings indicate that while stationary monitors can approximate short-term pedestrian-scale exposure reasonably well under certain conditions (e.g., less complex built environments), they can both significantly under- and overestimate instantaneous exposure in highly heterogeneous urban areas or under specific meteorological conditions. It is important to note that this study characterizes short-term spatial disagreement at fixed moments in time, and understanding the true implications of this for human exposure will require further study assessing whether these observed spatial disagreements persist across longer time scales in a way that maintains the discrepancy even under temporal averaging.

3.4. Implications and Suggestions for Future Work

A formal regression analysis (Supporting Information, Section S3, Table S3) found that relative humidity was the only significant predictor of Center–Mobile discrepancy, while distance to the Center Station, tree coverage, and temperature were not. This suggests that no universally “useful” radius of representativeness can be defined for stationary air quality monitors. However, temperature and relative humidity in this model only reflect conditions at each measurement point. Spatial variation in micrometeorological conditions across mobile locations was also not examined; therefore, the relationship between local micrometeorological gradients and Center–Mobile PM_2.5 discrepancy remains untested in these results. Additional comparisons between wind speed and PM_2.5 differences are provided in Figures S9 and S10, which show no consistent statistically significant relationship across sites and seasons. Also, the model explained only a small fraction of the observed variance (adj. R ² = 0.16), indicating that the majority of short-term spatial variability is not captured by the morphological and meteorological covariates measured here.

For short-term instantaneous exposure assessment, fixed monitors may serve as reasonable proxies in relatively homogeneous environments but require supplemental mobile sampling and careful interpretation in complex urban environments. Exposure assessments based solely on fixed monitors may underestimate or overestimate short-term pedestrian exposure in precisely the environments where variability is highest and population densities mean a larger number of individuals are being affected. These findings carry important implications for how sparse air quality monitoring data are used in public health research. Across all areas and seasons, stationary monitors disagreed with nearby mobile measurements in AQI_PM category in 10.4% of paired observations, with discrepancies of up to 800 μg/m³, large enough to span multiple AQI categories and to bias exposure estimates. AQI category disagreements should be interpreted in the context of sensor uncertainty. At the EPA PM_2.5 breakpoint of 9 μg/m³, the 95% prediction intervals of the sensor cross-calibration models corresponded to an uncertainty of approximately ±1.5–2.0 μg/m³. Consequently, some AQI category differences observed near regulatory thresholds may reflect measurement uncertainty rather than meaningful differences in exposure, whereas larger discrepancies likely represent true spatial variability in air quality. These category differences should be interpreted cautiously because threshold-based classification systems can amplify small concentration differences when observations occur near regulatory breakpoints. Therefore, AQI results are presented here as a complementary communication-oriented metric alongside concentration differences. The two captured exceptional pollution events further illustrate that urban pollution sources can produce Hazardous-level conditions invisible to monitors less than a block away, and, especially in the case of freight delivery, such events are likely to recur frequently and at the same location, with meaningful health implications for nearby residents and workers. The difficulty in properly characterizing exposures compounds further when accounting for individual mobility, as people move continuously through microenvironments that no single monitor captures.

Since denser sensor networks remain too cost-prohibitive to generate data at the scales needed to resolve block-level variability, researchers may benefit from efforts at quantifying representativeness and uncertainty, flagging morphologically complex study areas as more uncertain, and, where possible, supplementing fixed data with mobile sampling. For example, studies can prioritize mobile monitoring campaigns in high-variance corridors or during high-risk weather regimes to improve exposure models or to guide the placement of new permanent monitors, thereby generating data sets more likely to achieve meaningful protections for people who live, work, or travel through areas where air quality may vary in highly dynamic ways.

To improve the specificity of these recommendations, it would be beneficial to expand future research to include similar studies across a wider range of climate zones, weather conditions, and urban contexts, e.g., to better understand the impacts of different transportation systems, heating modes, and industrial activities in urban settings. Integrating meteorological variables (wind, turbulence, and vertical mixing) and traffic sensors within dense stationary sensor networks could facilitate a finer-scale understanding of drivers of heterogeneous conditions and support the development of correction models for sparse network data. Moreover, expanding this framework to include indoor air quality measurements could yield a more comprehensive view of human exposure, since individuals spend the majority of their time indoors, where pollutant levels are influenced by ventilation, building materials, and occupant activities.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.6c03923.

• Extended description of sensor cross-calibration; additional information on spatial context of study areas; Regression Analysis of Center–Mobile Discrepancy Predictors; and DOI link for experimental data (PDF)

R.S. and N.B.: joint first authorship. Q.R.W. and A.V.M.: joint senior authorship.

The authors declare no competing financial interest.