Influence of transboundary wildfire smoke on fine particulate matter concentrations at South Texas border schools

Abstract

Exposure to air pollutants in indoor and outdoor school environments poses significant health risks to children due to their heightened vulnerability. Increasing wildfire frequency has intensified transboundary smoke transport, with border communities among the first to experience associated air quality impacts. While the impacts of wildfires on air quality are increasingly recognized through satellite observations and modeling, ground-based observations remain scarce in under-monitored regions. Low-cost air sensors were deployed across schools in the Roma and Rio Grande City areas of Starr County, along the U.S.–Mexico border, for 146-days from November 2023 to April 2024. The raw sensor measurements were corrected using a collocation-based calibration with a research-grade instrument and analyzed alongside meteorological data, back trajectories, and satellite-derived smoke and fire products. Daily average fine particulate matter concentrations were significantly higher during smoke-impacted days compared with non-smoke periods (p < 0.001), with site-level increases ranging from 8.61% to 54.9%. Trajectory analysis integrated with these measurements indicate that the wildfire emissions from Mexico significantly influence fine particulate pollution in South Texas, during the wildfire seasons. These findings highlight the need for targeted mitigation strategies to decrease the exposure amongst border communities, especially that of children.

Introduction

Air pollution remains one of the leading environmental threats to human health worldwide. Exposure to elevated levels during early life is of particular concern, as it is associated with increased risk of both acute and chronic health effects. Children are more susceptible to the harmful effects of air pollution than adults due to several factors. These include a faster breathing rate, leading to greater air intake per body weight¹, less efficient nasal filtering², and an immature and developing immune system³. Asthma is the most common chronic disease in children, significantly impacting respiratory health. Exposure to air pollutants, particularly fine particulate matter (PM_2.5), has been shown to increase the risk of asthmatic symptoms such as wheezing, airway inflammation, and reduced lung function growth^4,5. Short-term exposure to ambient air pollutants was also associated with hospitalizations of children due to pneumonia⁶ and lower respiratory tract infections⁷. Several large-scale studies also indicated that exposure to air pollutants is associated with decreased academic performance in children^8–10 and with school absenteeism¹¹.

The school environment accounts for a major portion of children’s daily time-use activities. On average, children aged 5 to 18 years spend approximately 30.96 h per week at school, equivalent to ~6.2 h per weekday, making it the second largest daily time-use activity after sleep, which accounts for 67.20 h a week, equivalent to ~9.6 h per day¹². Hence, it is paramount to minimize the exposure levels at schools to reduce the detrimental impact on children’s short- and long-term health and their corresponding academic performances. Air pollution levels at schools are influenced by both indoor and outdoor sources, with outdoor emissions often playing the dominant role in determining the indoor levels. A study by Nwokorie et al.¹³ revealed that outdoor sources accounted for more than 95% of indoor PM_2.5 levels in schools in Port Harcourt, Nigeria. The school’s location also plays a key role in determining overall exposure burden, due to various factors such as proximity to roads or highways as well as other point or area sources of air pollution¹⁴. In the United States (U.S.), schools in low socio-economic status (SES) areas were associated with higher concentrations of PM_2.5 and NO₂ than schools with lower levels of poverty¹⁵. Similar disparities were also recorded by Cheeseman et al.¹⁶, which found disproportionate risk of air neurotoxicants in U.S. Public schools with a higher percentage of Hispanic, black, and Asian students. Exposure to PM_2.5 from wildfire smoke also negatively impacts primary school children more than secondary school children, leading to a 0.122% decrease in test scores¹⁷. These findings highlight the importance of implementing mitigation strategies to reduce exposure in school environments, particularly for younger students and those attending schools in low-SES areas.

A dearth of air quality data is a major setback for developing mitigation strategies and supporting data-driven decisions. Ambient air quality monitors operated by state and federal agencies are designed to represent regional conditions but often fail to capture finer-scale spatial variability at the neighborhood scale. This limitation is compounded by the fact that only one-third of U.S. counties have an ambient air quality monitoring station¹⁸. Low-cost air sensors are hence increasingly being used to complement traditional monitoring methods, enabling high-resolution spatial monitoring that improves understanding of local air quality trends at the neighborhood level, for example, at school locations^19–21. The U.S.–Mexico border stretches over ~3200 km, with an estimated ~8 million people living on the U.S. side across Arizona, California, New Mexico, and Texas²². Communities in this region experience limited access to health care, lower per capita income, and a large number of school children living in poverty and are from uninsured households²³. Approximately 57% of the population in these communities identify as Hispanics/Latinos²². These factors make the communities along the U.S.–Mexico border, especially children, more vulnerable to the adverse health impacts of air pollution.

This study utilizes measurements from low-cost air sensors deployed at schools in communities of Starr County located north of the U.S.–Mexico border. Starr County is located in the Lower Rio Grande Valley (RGV) region of South Texas. The county shares borders with Mexico to the south and west, with Zapata, Jim Hogg and Brooks counties to the north, and with Hidalgo County to the east. As of 2024, the county is home to approximately 65,000 residents²⁴. The study was conducted at schools in the Roma Independent School District (RISD). Figure 1 shows the location of Starr County and the Lower RGV region in Texas along with the locations of all schools of RISD. RISD hosts nine schools with a total of 5919 students in the 2023–2024 academic year²⁵, of whom 99.9% are Hispanic/Latino. This study aims to address multiple research gaps in understanding air quality in this region. There has been no ground-based ambient PM_2.5 data available in this county to date, making the data presented in this study the first such measurements. The 146-day monitoring period also captured wildfire smoke originating in Mexico, providing new evidence of how transboundary smoke influences air quality in border communities. This endeavor also helps characterize outdoor PM_2.5 concentrations at schools, as these measurements were conducted in outdoor microenvironments at schools and could guide local school and administrative authorities in developing mitigation strategies to reduce children’s outdoor exposure to wildfire smoke.

Fig. 1. Map showing the locations of Rio Grande Valley, and Starr County in Texas along with the locations of Roma Independent School District (RISD) schools which were the sampling sites (R1-R9).

a Location of Rio Grande Valley in Texas. b Location of Starr County in Rio Grande Valley. The green cross represents the nearest continuous ambient monitoring station from the study region. c Locations of sampling sites in Starr County are represented by red dots.

Results

Hourly averaged ambient PM_2.5 concentrations

Hourly ambient PM_2.5 concentrations were calculated to understand the short-term variations at the schools during the study period. Figure 2 shows the time series of 1-h averaged ambient PM_2.5 at different schools. Detailed descriptive statistics for the hourly averaged PM_2.5 are presented in Table S1. The mean ± standard deviation (SD) levels ranged from 12.12 ± 5.34 at R6 to 20.99 ± 13.40 µg/m³ at R9. Across all schools, median values are consistently lower than mean values, and 75th and 90th percentile values are higher than medians, indicating a positively skewed PM_2.5 distribution, as shown in Fig. S1. The highest PM_2.5 concentration of 152.34 µg/m³ was recorded at R5 on 30th December 2023 at 04:00, but no similar episodic increase was observed at other schools during this period. However, there were also periods when PM_2.5 concentrations at all sites ranged from 26.54 to 70.17 µg/m³, and mean concentrations across all sites ranged from 39.65 to 55.67 µg/m³ from 23:00 on 26th February to 10:00 on 27th February 2024. These results indicate that PM_2.5 was influenced by both localized sources affecting individual sites and also by broader-scale sources affecting all sites simultaneously.

Fig. 2. Time series of hourly-averaged PM_2.5 concentrations across different schools.

Different colors indicate PM_2.5 levels at the respective schools, as shown in the color bar legend. The black dashed line indicates the mean PM_2.5 across all sites at a given time.

24-hour averaged ambient PM_2.5 concentrations

The 24-h averaged ambient PM_2.5 concentrations were calculated to smooth out short-term spikes observed in hourly-averaged concentrations and to facilitate comparison with regulatory standards. Figure 3 shows the time series of 24-h averaged ambient PM_2.5 at different sites, with the dashed black line showing the mean PM_2.5 concentration across sites and the dashed red line indicating the regulatory standard. The United States Environmental Protection Agency’s (U.S. EPA) National Ambient Air Quality Standard (NAAQS) for 24-h averaged PM_2.5, i.e., 35 µg/m³ was used in this study for the comparison. Table S3 presents detailed descriptive statistics for the 24-h averaged PM_2.5 concentrations observed. The mean ± SD levels ranged from 12.09 ± 4.43 at R6 to 20.97 ± 11.05 µg/m³ at R9. The maximum daily average concentrations reached up to 54.27 µg/m³. On multiple days, measured PM_2.5 concentrations exceeded NAAQS. Out of the 146 monitored days, there were 10 days when at least two of the nine sampled sites exceeded the NAAQS concentration of 35 µg/m³. The mean PM_2.5 across all sites exceeded NAAQS on 5 days, with a maximum value of 41.69 µg/m³ on 27th February 2024. It is important to note that all days of exceedance occurred between 27th February and 8th April 2024. During the monitoring period from 15th November 2023 to 26th February 2024, the average PM_2.5 across sites was 15.31 µg/m³, whereas it increased to 20.18 µg/m³ between 27th February and 08th April 2024.

Fig. 3. Time series of 24-h averaged PM_2.5 concentrations across different schools.

Different colors indicate PM_2.5 levels at the respective schools, as shown in the color bar legend. The black dashed line indicates the mean PM_2.5 across all sites at a given time. The red dashed line indicates the National Ambient Air Quality Standard for PM_2.5.

Directional analysis for source attribution of PM_2.5 using wind speed and wind direction

Bivariate and polar plots combine wind direction and wind speed with the pollutant concentration and have been successfully used in studies to identify source directions^26,27. However, these approaches provide an aggregated representation of source direction, which misses the probable temporal variation in sources. To address this limitation, we employed a calendar plot (Fig. 4) and analyzed the PM_2.5 contribution weight as a function of wind direction (Fig. 5) and wind speed (Fig. 6), separately. This result aided our understanding of the influence of wind on daily PM_2.5 concentrations.

Fig. 4. Calendar plot of PM_2.5 against wind direction during the study period, with the legend indicating PM_2.5 in µg/m³.

The arrow indicates the direction the wind coming from, and the color of the box indicates PM_2.5 in µg/m³, corresponding to the color bar.

Fig. 5. Time series of daily ambient PM_2.5 concentrations as a function of wind direction.

Different colors as shown in the legend indicate the direction of the wind and its contribution to PM_2.5 over the day.

Fig. 6. Time series of daily ambient PM_2.5 concentrations as a function of wind speed.

Different colors as shown in the legend indicate wind speed and its contribution to PM_2.5 over the day.

As noted in Fig. 3, a higher mean of all measured PM_2.5 concentrations was observed between 27th February 2024 and 08th April 2024, when compared to the days before that in the study period. Figure 4 presents a calendar plot of daily PM_2.5 concentrations, with red color indicating higher PM_2.5 and green indicating lower PM_2.5. The arrows indicate the direction of incoming wind. Mean daily PM_2.5 concentration across sites exceeded the NAAQS of 35 µg/m³ on 7 days, as indicated in red in the plot. Notably, on all exceeding days, winds originated predominantly from the southeast (SE). The mean daily PM_2.5 concentrations ranged from 30 to 35 µg/m³ on 2 days, both of which were associated with SE winds. In contrast, the majority of days with lower PM_2.5 concentrations were generally observed under north (N) or north-west (NW) wind conditions. These patterns suggest that elevated pollution levels are associated with air masses arriving from the SE sector.

The contributions of different wind directions (Fig. 5) and wind speeds (Fig. 6) to PM_2.5 concentrations in the region further support this observation. Figure 5 shows that winds from the SE (purple) and S (orange) sectors contributed to higher PM_2.5 concentrations, with exceedances above 30 µg/m³ occurring primarily during these conditions, in the latter half of February 2024. Conversely, winds from N sector (red) were associated with comparatively low concentrations, typically below ~10 µg/m³. While these directional patterns provide insight into the source’s origin, they cannot definitively distinguish between local emissions and long-range transport. Figure 6 illustrating PM_2.5 concentrations as a function of wind speed, shows elevated concentrations were associated with higher wind speeds, particularly above 5.52 m/s (yellow). This also coincided with periods when the wind was predominantly from the S and SE direction. Increased pollutant concentrations at higher wind speeds may indicate long-range transport of pollutants rather than local sources^28–30.

Overall, the combined directional and wind-speed analyzes indicate that elevated PM_2.5 concentrations are associated with air masses from S and SE directions, particularly under higher wind speeds. To further understand the probable location of these sources, Potential Source Contribution Function (PSCF) and Concentration-weighted trajectory (CWT) analyzes are presented in the following section.

Trajectory-based source location identification

Back-trajectory-based PSCF and CWT analyzes were conducted for the study period as a whole and at monthly intervals to characterize spatial variations in probable sources. Figures 7 and 8 present the monthly PSCF and CWT maps, respectively, for the study period. PSCF and CWT maps for November 2023 and April 2024 do not represent the full calendar month as ambient PM_2.5 measurements were available only from November 15th 2023 through April 08th 2024. Hence, the November results reflect only the second half of the month, whereas the April results are based on data from only the first 8 days. The monthly maps reveal clear seasonal variability in potential source regions influencing ambient PM_2.5 concentrations in the study area. During November, December, and January (7a–c), PSCF values were generally below 0.3, while some grid cells in northern Mexico showed elevated probabilities. These patterns suggest that elevated ambient PM_2.5 concentrations are driven by episodic events during these months, likely originating in these regions of Mexico. However, the overall PSCF values indicate a comparatively limited contribution from long-range transport. In contrast, February, March, and April (7d–f) exhibit higher and more intense PSCF values extending farther south and southeast into Mexico and occasionally towards the Gulf region. March shows a relatively increased number of grid cells with PSCF values greater than 0.4, located in Mexico. This suggests that the increased ambient PM_2.5 concentrations observed during March and April 2024 were not likely caused by local pollution events but were instead driven by long-distance transboundary pollution.

Fig. 7. Potential Source Contribution Factor (PSCF) maps for PM_2.5 for various months.

The color of a pixel represents the factor intensity as shown in the color bar. a November 2023, b December 2023, c January 2024, d February 2024, e March 2024, f April 2024. The image was generated using the ground-based data collected and HYSPLIT trajectories, in the openAir R package.

Fig. 8. Concentration-weighted trajectory (CWT) maps for PM_2.5 for various months.

The color of a pixel represents the CWT value in μg/m³ as shown in the color bar. a November 2023, b December 2023, c January 2024, d February 2024, e March 2024, f April 2024. The image was generated using the ground-based data collected and HYSPLIT trajectories, in the openAir R package.

Similar patterns were observed in the CWT analysis, with low-concentration weighted values observed during November, December, and January (8a–c). In contrast, the latter 3 months of February, March, and April (8d–f) exhibited high values, mainly located in Mexico. The agreement in results between PSCF and CWT enhances confidence in the identified spatial pattern and supports the interpretation of increased long-range transport influence during the latter part of the study period. Further supporting these results, Figure S2 shows that concentrations exceeding the 90th percentile were more frequently associated with trajectories originating from the south toward the study area, mainly in Mexico, compared to the overall trajectory distribution. Figures S3 and S4 show the combined PSCF and CWT maps for the study region, which resonate with the same pattern of likely origin of high PM_2.5 concentrations from the Mexican region to the south, rather than local sources.

Wildfires in Mexico and their impact on ambient PM_2.5 concentrations

PSCF and CWT results primarily indicated long-range transported pollution, which impacted ambient PM_2.5 concentrations in the region and caused episodic increases, particularly in February, March, and April during the study period. Mexican wildfires, which begin around February and peak in mid-March³¹, are a major source of pollution during these months. Wang et al.³² showed in their study that fires in Mexico and Central America affect PM_2.5 and ozone concentrations in US Gulf Coast cities. On fire-impacted days, increases of 4.7, 3.9, 2.1, 3.6, and 2.8 µg/m³ were observed in Corpus Christi, Houston-Galveston-Brazoria, New Orleans, Mobile, and Pensacola areas, respectively. Pinakana et al.³³ also reported an increase in black carbon concentrations during the wildfire season of March and April 2023 in the Brownsville region of South Texas. Hence, to assess the impact of these wildfires on PM_2.5 concentrations, we used National Oceanic and Atmospheric Administration (NOAA)’s Hazard Mapping System (HMS) fire and smoke products.

The daily average ambient PM_2.5 concentrations were classified into two groups based on smoke coverage at the specific location. Figure 9 compares PM_2.5 concentrations across schools based on smoke coverage status. While group “0” represents PM_2.5 concentrations when there was no smoke coverage at the school, group “1” shows the concentrations when there was smoke coverage. Table S3 presents descriptive statistics for PM_2.5 across monitoring sites, with and without smoke. Across all sites, mean PM_2.5 concentrations during smoke episodes were consistently higher than on days without smoke. The Mann-Whitney U test showed that PM_2.5 concentrations were significantly higher on smoke-impacted days than on non-smoke-covered days when all school sites were analyzed together (p < 0.001). Site-specific analyzes indicated increases at multiple sites, although statistical significance varied across schools. To assess if this variability was caused by spatial heterogeneity, the coefficient of divergence (COD) was calculated. A COD value of less than 0.2 indicates high similarity, and the value closer to unity represents a difference in absolute concentrations. As shown in Table S4, all the values except for R3 to R6 were below 0.20, while the COD for R3 to R6 was 0.21. As this value is very close to 0.20, it can be interpreted that all sites exhibit similar trends throughout the study period, with minimal spatial heterogeneity. These results suggest that differences in statistical significance among sites were likely influenced by differences in sample sizes rather than by spatial heterogeneity, as sites without statistically significant differences generally had fewer than 100 non-smoke daily PM_2.5 data points.

Fig. 9. Box plots of daily-averaged PM_2.5 concentrations across monitored locations on days with different smoke coverage statuses.

Group “0” represents days without smoke, and group “1” represents days with smoke coverage. Boxes cover 25th to 75th percentile data, while the whiskers cover 5th to 95th percentile data. The circles in the box show the mean value, which is also labeled. The text above the line indicates the Mann-Whitney U test for PM_2.5 across different smoke coverage statuses, with significance codes: ^**p < 0.01, ^*p < 0.05, NS – not significant. The N values above each box indicate the number of days in each group.

Concentrations at R3 showed a significant difference (p < 0.01), with a substantial increase from a mean of 18.89 µg/m³ during non-smoke periods to 29.27 µg/m³ during smoke periods, a 54.9% increase. R8 showed a similar trend (p < 0.05), with a mean increase from 14.93 µg/m³ to 23.12 µg/m³, a 54.8% increase relative to no-smoke-coverage days. R1, R5, R6, R7, and R9 showed increases in mean concentrations of 12.51%, 28.13%, 8.61%, 32.25%, and 22.90%, respectively, during smoke-coverage days. Overall, the increases indicate that smoke presence significantly affects the variability of ambient PM_2.5 concentrations across monitoring sites. This variability in smoke’s impact on PM_2.5 can be explained by local topography and natural barriers, such as tree cover, at the sampling locations^34,35. To assess the impact of near-site-level natural barriers on the varying increases in PM_2.5 concentrations at the sites during smoke days, we used the Tree Canopy Cover data from the National Land Cover Database: NLCD TCC v2023.5³⁶. A buffer zone of 500 m was created around each site, and since the smoke from Mexican wildfires was emanating from the south, the mean canopy cover in the southern direction of each site was calculated. These values were then compared with the percentage increase in ambient PM_2.5 at each site on smoke-impacted days. A lower percentage increase in ambient PM_2.5 was observed at schools with higher mean canopy cover in the southern direction, and conversely, a higher increase was observed at schools with lower canopy cover in the same direction (Fig. S5). This indicates that canopy cover could act as an effective natural barrier in reducing the impact of smoke-related PM_2.5. Figure S6 shows the fire anomalies detected by the HMS fire product in the region of South Texas, Mexico, and Central America. The number of anomalies significantly increased in Fig. S6d–f, representing February 2024, March 2024, and April 01st to 08th 2024, respectively, when compared to Fig. S6a–c representing November 15th to 30th 2023, December 2023, and January 2024, respectively.

Discussion

This study provides the first multi-site, school-based ambient PM_2.5 assessment in Roma and Rio Grande City, situated in Starr County of South Texas, and, to the best of our knowledge, the first set of measurements collected in this county. The results indicate that ambient PM_2.5 concentrations were elevated between late February and early April 2024. Directional analysis showed elevated concentrations when air masses arrived from the S and SE, particularly at higher wind speeds, suggesting long-range transport as the likely source rather than local emissions. Back-trajectory analyzes using PSCF and CWT further supported this interpretation, with the source probability values being low in November through January and higher PSCF probabilities and elevated CWT values in February through April extending into northern and central Mexico. Trends in wildfire activity in Mexico, peaking toward mid-March, coincided with the period of PM_2.5 exceedances observed. The convergence of wind-sector dependence trajectory modeling and satellite products of smoke and fire indicates that the transboundary wildfire smoke was indeed a dominant driver of episodic air pollution events in the studied region.

Although this study’s results offer new insights into ambient PM_2.5 concentrations in Starr County and their sources, there are some caveats that should be taken into consideration while interpreting the findings. The low-cost air sensors used in this study were calibrated against a research-grade instrument. The optical particle counters are sensitive to aerosol composition, and in the absence of an Federal Equivalent Method (FEM) or Federal Reference Method (FRM) in the county, the ability to validate sensor measurements against regulatory-grade instrumentation was not possible. The NOAA’s HMS smoke product used in this study doesn’t necessarily confirm ground-level smoke presence. However, in the absence of other smoke or fire data, this product was instrumental in our understanding of wildfire impacts on PM_2.5. The data were also not collected over the year; a year-long measurement in future studies would provide additional insights into temporal trends in concentrations, the role of seasons, and sources. The study is also limited to PM_2.5 mass concentrations only. Chemical speciation, tracer-based source apportionment, or vertically resolved aerosol profile measurements could have provided more quantitative attribution to wildfire smoke. Future studies can incorporate these measurements to help further investigate the relative contributions of wildfire smoke and other regional emission sources in this region. Indoor PM_2.5 measurements were not conducted as part of this study; they would have quantified the impact of these wildfire-related smoke exposures in the schools’ indoor environments.

Despite these limitations, we believe that the current work contributes to the ever-growing field of air quality literature, especially in an understudied region such as Starr County. This study was conducted in the outdoor microenvironment of school campuses. Based on our findings, the school authorities can implement wildfire response plans by modifying outdoor activities during affected periods to reduce children’s exposure. Pollutant concentrations at the intra-urban level are impacted by the local topography, natural barriers, vegetation, tree cover, built environment, and point and area sources of air pollution³⁷. School authorities have incorporated greening of areas surrounding the schools to minimize the pollution exposure burden as reported in many research studies^38–40. However, plant species for greening initiatives must be chosen carefully, as those that emit high levels of volatile organic compounds (VOCs) and release large amounts of pollen may inadvertently deteriorate air quality and aggravate symptoms of wheezing and asthma in school-going children^39,41.

Methods

Instrumentation and sampling design

A low-cost air sensor TSI BlueSky^TM (Shoreview, MN, USA), model 8143, was used in this study to measure ambient PM_2.5 concentrations. These sensors are primarily operated with Sensirion SPS30 PM to provide PM data and have been used in multiple studies to understand PM_2.5 concentrations in various areas^42–45. The instruments were deployed at all schools on November 15th 2023, and collected data till April 08th 2024, for a total of 146 days. The sensors were installed simultaneously at all the study locations during the study period. Figure S7 shows pictures of the installations at each school.

Data preprocessing for quality assurance

The data collected from the TSI 8143 sensors was preprocessed to remove values outside the operational range. As per the TSI BlueSky sensor specifications, all the values with temperatures less than 40 °C, relative humidity less than zero and PM_2.5 concentration greater than 1000 µg/m³ were removed. Multiple studies reported that PM measurements by optical-based air sensors are affected by factors such as temperature, relative humidity (RH) and PM size fractions^46–49. To ensure the precision and accuracy of the data collected by these sensors, collocation amongst the sensors and with a reference method instrument is a widely adopted practice⁵⁰. A relationship between the meteorological variables and the PM measurements is then established, aiding in the development of a correction equation. In the absence of a FEM or FRM monitor for collocations, research-grade instruments such as TSI DustTrak II 8533⁵¹ and TSI DustTrak^TM DRX Aerosol Monitor 8534²⁷ are used. As the DustTrak is also dependent on an aerosol scattering method, Both et al.⁵² and Cheng et al.⁵³ reported that the measurements were 2.77 and 2.2 times higher than the gravimetric method measurements, respectively. In this study, the following Eq. 1 developed by Jaffe et al.⁵⁴ was used to correct the research-grade instrument measurements.

$${\mathrm{Corrected}\mathrm{PM}}_{2.5}={\mathrm{DustTrak}\mathrm{PM}}_{2.5}\left(\mu \mathrm{g}/{\mathrm{m}}^3\right)\mathrm{x}0.4193+4.414$$ 1

The corrected research-grade measurements were then compared with PM_2.5 measurements from a TSI 8143 instrument during an 11-day collocation period in an outdoor environment. All the sensors were collocated against the research-grade instrument, and the corresponding data were used to develop a correction factor via multiple linear regression, which is the most common correction method⁵⁵. The empirical correction equation using multiple linear regression had the PM_2.5 value, temperature, and relative humidity as predictor variables. Because of the limited collocation duration, the full collocation dataset was used to derive the correction equation. Prior to the correction, the low-cost air sensor data had a Root Mean Square Error (RMSE) of 13.19 µg/m³ and an R² of 0.69. After the correction model was applied, RMSE of 6.57 µg/m³ and R² of 0.746 were obtained. Because the full collocation dataset was used to derive the correction equation, these model statistics reflect a model fit rather than independent validation metrics. The following Eq. 2 was used to correct PM_2.5 measurements by TSI 8143 during the study period:

$${\mathrm{Corrected}\mathrm{PM}}_{2.5}=\left(0.452{\mathrm{x}\mathrm{LCS}\mathrm{PM}}_{2.5}\right)+\left(0.3044{\mathrm{x}\mathrm{T}}_{\mathrm{LCS}}\right)+\left(0.3010{\mathrm{x}\mathrm{RH}}_{\mathrm{LCS}}\right)\text{-}11.64$$ 2

where T_LCS, RH_LCS and LCS PM_2.5 are the temperature, relative humidity and PM_2.5 value reported by TSI 8143. The descriptive statistics of the predictor variables during the collocation period are presented in Table S5.

Due to technical issues, some sensors did not collect data on certain days, so they were replaced with fully functioning units to continue the sampling procedure. Hourly and daily mean PM_2.5 concentrations were calculated when at least 75% of data within each respective hour and day were available. Following the daily aggregation, data completeness over the monitored 146-day period was calculated. Figure S8 shows the daily PM_2.5 data completion by sampling location. The total completeness ranged from 29.3% at R2 to 99.8% at R5. As these sensors lacked the ability to connect to public Wi-Fi networks, the data were primarily collected and verified in person for operational status. Apart from R2 and R4, all sites collected at least 60% of the data during the study period. Although all the available data were presented in the time series analysis to illustrate temporal variability, the other analyzes excluded schools with data completeness below 60% of the monitoring period; fewer than 87 monitoring days.

Meteorological data

Meteorological data such as wind speed (WS) and wind direction (WD) play a key role in identifying the sources of air pollution. Due to the unavailability of ground-based meteorological stations, both at the sampling points and in the region, wind components U and V were extracted from the ERA5, which is the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis. The components were extracted using the Google Earth Engine platform, a cloud-based geo-big data platform that hosts a large catalogue of datasets⁵⁶. The wind speed and wind direction were calculated using the following Eqs. 3 and 4, respectively.

$$\mathrm{WS}=\sqrt{{U10}^2+{V10}^2}$$ 3

$$\mathrm{WD}=\mathit{atan}2\left(-U10,-V10\right)\times \left(180/\pi \right)$$ 4

Trajectory-based analysis and statistical analysis

Conditional Probability Function, Nonparametric Wind Regression, and PSCF are being increasingly used to understand the source attributions ^27,57–59. Trajectory-based analysis was conducted in this study to assess the impact of long- and near-range sources on pollutant concentrations. The Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model was used to generate backward air-mass trajectories for the sites, daily from November 15, 2023, to April 08 2024, at 6-h intervals. The trajectories were generated using the R package “splitr”⁶⁰, with the function hysplit_trajectory(·), at an arrival height of 500 m above ground level. The trajectories were then combined with ambient PM_2.5 measurements to identify the contribution of high-concentration back trajectories and to conduct PSCF and CWT analyzes. PSCF is a trajectory transport pathway-based analysis and was used to identify sources in multiple studies ^61–63. It calculates the probability of a source located at latitude i and longitude j, using the following Eq. 5:

$$\mathrm{PSCF}=\frac{m_{\mathit{ij}}}{n_{\mathit{ij}}}$$ 5

where n_ij represents the total number of trajectories passing through cell (i,j), and m_ij is the number of times the source concentration is higher than the 90th percentile passing through the cell.

The CWT method is similar to PSCF but provides greater insight into source strength, particularly between moderate and strong sources ^64,65. The following Eq. 6 is used for the analysis:

$$In\left({\underset{\_}{C}}_{ij}\right)=\frac{1}{{\sum }_{k=1}^N{\tau }_{ijk}}\sum _{k=1}^{N}In\left(C_k\right){\tau }_{ijk}$$ 6

where i and j are the grid indices, k is the trajectory index, N is the total number of trajectories, c_k is the measured PM_2.5 concentration upon arrival of k, τ_ijk is the residence time of trajectory, and higher ${\underset{\_}{C}}_{ij}$ values indicate those grid cells caused high concentrations at the observation site.

Spatial heterogeneity tests between sampling sites were performed by calculating the COD using the Eq. 7 below,

$$\mathrm{COD}=\sqrt{\frac{1}{p}{\sum }_{i=1}^p{\left(\frac{x_{\mathit{ij}}-x_{\mathit{ik}}}{x_{\mathit{ij}}+x_{\mathit{ik}}}\right)}^2}$$ 7

where x_ij and x_ik are the PM_2.5 concentrations at sites j and k respectively, and p is the total number of observations.

The analysis and plots of spatio-temporal variations were conducted in Origin Pro (Origin Lab Corporation, Northampton, MA, USA; Version 2025), and the source attribution analyzes were conducted in R, using the OpenAir package ⁶⁶.

Smoke and fire products

The National Oceanic and Atmospheric Administration (NOAA)’s widely used Hazard Mapping System (HMS) was used in this study. These data are produced from multiple polar and geostationary satellite observations ⁶⁷ and are available at a daily resolution at https://www.ospo.noaa.gov/products/land/hms.html#data. There are several limitations to these products, including their inability to detect small agricultural fires, the lack of information on plume sources, and dense cloud cover that impacts the product ⁶⁸. The smoke product also does not essentially represent the smoke cover at ground level all the time. Nevertheless, the products provide a broad range of information on fire and smoke, aiding in understanding the pollutant concentration trends in fire-impacted regions ^33,68–70.

Author contributions

S.D.P led the data analysis, result visualization, interpretation, and manuscript writing. K.B.S. led data collection, curation and assisted with data analysis. O.T, D.W., and K.S. assisted in funding acquisition, resources, and reviewed the paper. J.L.G. and Md.S.R validated the results, reviewed and edited the paper. A.U.R led the funding acquisition process, designed the study, supervised the data collection efforts, edited the first draft, and validated the research conclusions. The final version of the paper was reviewed and edited by all the authors.

Data availability

The ERA5 data is available at the Google Earth Engine data catalogue https://developers.google.com/earth-engine/datasets/catalog/ECMWF_ERA5_LAND_HOURLY. NOAA’s Hazard Mapping System Fire and Smoke products are publicly available at https://www.ospo.noaa.gov/products/land/hms.html#data. The HYSPLIT model was run using openair, whose documentation is available at https://openair-project.github.io/book/. The NLCD TCC v2023.5 is available at https://www.mrlc.gov/data; precisely at https://data.fs.usda.gov/geodata/rastergateway/treecanopycover/docs/v2023-5/nlcd_tcc_CONUS_2023_v2023-5_wgs84.zip.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44407-026-00096-z.