Our analyses aimed to assess data and modeling performance from multi-city low-cost sensors and existing regulatory monitoring networks. We selected core-based statistical areas (CBSAs) with at least seven EPA regulatory monitors and seven PurpleAir sensors between 1 January 2018, and 31 December 2021 (Table 1). The choice of requiring at least seven sensors from each type (EPA monitors and PurpleAir sensors) in the study is based on our goal to ensure a sufficient spatial coverage for LUR models. The specific number, seven, is chosen in reference to Lu et al. (2022), which deployed seven sensors of each type in six cities across the United States, and this configuration showed promising results in assessing data and modeling performance for low-cost sensors in different urban environments. Following this approach, we aimed to ensure a similar level of representativeness in our analyses.

Hourly PM_2.5 measurements are downloaded from the EPA Air Quality System (AQS) database for Federal Reference Method (FRM) monitors and Federal Equivalent Method (FEM) monitors. In addition, we downloaded publicly available outdoor PM_2.5 measurements from the PurpleAir website via an open-source R package: AirSensor, developed by the South Coast Air Quality Management District (South Coast AQMD) and Mazama Science. The raw data was aggregated into hourly averages using the quality control function provided by AirSensor. The function creates a PM_2.5 time series by averaging the A and B channels and removing the invalidate date when 1) the measurement count is lower than 20, 2) the hourly difference between A and B channels is higher than 5, and 3) the hourly percent difference between A and B channels is higher than 70%. Samples with PM_2.5 measurements greater than 1,000 μg/m³ and with missing or abnormal temperature and humidity readings were removed (humidity readings should be within 0%–100%; temperature readings should be within 20°F to 140°F). Approximately 15% of the raw PurpleAir measurements were removed due to these quality issues. We applied the correction to the PurpleAir PM_2.5 measurements based on the U.S.-wide calibration proposed by Barkjohn et al. (2021).

Comparing the monthly hourly PM_2.5 concentrations between EPA and PurpleAir in the selected study regions (Figure 1), measurements from the PurpleAir sensors generally have a lower median value than the EPA monitors. There could be several reasons why EPA air quality measurements are consistently higher than PurpleAir measurements. One reason could be that the EPA uses more sophisticated and expensive air quality monitoring equipment, which is subject to strict quality control and calibration procedures to ensure the accuracy of the measurements. In contrast, PurpleAir monitors are designed for personal use and are not subject to the same level of quality control and calibration, which can lead to a more significant margin of error in the readings. Another reason could be the placement of the monitors. The EPA often selects monitoring sites based on strict criteria to ensure that they represent the air quality in a given area (Raffuse et al., 2007). In contrast, PurpleAir monitors are typically installed by individuals in their own homes or businesses, and the placement of the monitors can vary widely (Ardon-Dryer et al., 2020). It is also worth noting that the EPA and PurpleAir use different methods to measure air quality. The EPA primarily measures PM_2.5 using gravimetric analysis, while PurpleAir uses a laser-based sensor technology. While both methods are considered reliable, they can produce slightly different measurements depending on factors such as humidity and temperature.

Traditional LUR models are statistical methods developed to estimate PM_2.5 concentrations based on geographical features. The models can create an empirical relationship between PM_2.5 concentrations and land use variables such as traffic volume, distance to major roads, population density, and presence of emission sources. A typical LUR model can be formulated as: C = β 0 + β 1 X 1 + β 2 X 2 + … + β n X n + ε (1) where C is the PM_2.5 concentration, β 0 is the intercept term, X 1 , X 2 , …, X n are the predictor variables, i.e., different land use, traffic, and geographical features, β 1 , β 2 , …, β n are the coefficients for these variables that indicate the strength and direction of the relationships between each predictor variable and PM_2.5 concentration, and ε is the error term that captures the variation in PM_2.5 concentration not explained by the model. The β coefficients are estimated using regression techniques, most commonly multiple linear regression. The goal is to find the set of β values that minimizes the sum of the squared differences between the observed and predicted PM_2.5 concentrations.

In this study, we used automated machine learning (detailed in Section 2.3.3) to construct more sophisticated versions of LUR models. Traditional LUR models, illustrated in Eq. 1, use multiple linear regression, which assumes a linear relationship between predictor variables and the target outcome. However, real-world relationships between land-use variables and PM_2.5 concentrations may be nonlinear or involve complex interactions, which can be better captured with machine learning techniques.

We analyzed 4 years of available PM_2.5 concentration data for LUR modeling. Hourly measurements of PM_2.5 concentration were used to compute monthly averages for each hour (e.g., January 2021 will have 24 h averages; monthly-hourly thereafter). These monthly-hourly averages were used to train and test the LUR models. To test the usefulness of low-cost sensor networks for developing AutoML-LUR models, we developed three types of models 1) using the EPA measurements, 2) using the PurpleAir measurements, and 3) incorporating PM_2.5 measurements from EPA monitors and PurpleAir sensors for the eight CBSAs. For each monitoring location, we prepare the geographic covariates following the guidelines from the Multi-Ethnic Study of Atherosclerosis and Air Pollution (Keller et al., 2015; Kirwa et al., 2021).

Quantifying human exposure to air pollution depends on two factors: 1) the population living within the area and 2) the air pollution concentration to which they are exposed. Combining the two factors, existing studies utilized population-weighted annual average concentration as a score to estimate population exposures, thus giving greater weight to the air pollution exposure where most people live (Reis et al., 2018). The population-weighted exposure is generally defined as E = P × C , where E is the annual mean population exposure for a certain area, P is the respective population number in this area, and C is the annual mean concentration of the pollutant for this area.

To account for human mobility in the estimation of PM_2.5 exposure, we used the publicly available human mobility dataset SafeGraph to calculate population-weighted exposure. Specifically, we used the SafeGraph Patterns data product, which contains mobility patterns from approximately 47 million (around 10% of) mobile devices in the United States (Squire, 2019; Coston et al., 2021). The monthly Patterns data product used in this study provides anonymized counts of how many people visit commercial points of interest (POIs) each month, which can be divided by the visitor’s home census block group (CBG). Note that the term “visitors” is not used in the conventional sense to distinguish between residents and non-residents. Rather, it refers to individuals visiting Points of Interest (POIs) in a specific area. This can be helpful for understanding how mobility patterns may impact exposure to air pollution, as people who spend more time in areas with higher pollution concentrations may be at greater risk of exposure. Since the data is aggregated into monthly patterns, home CBG information is not directly linked to a specific device, differentiating SafeGraph mobility data from other individual mobility tracking datasets, such as Call Detailed Records (CDR) or mobility survey datasets. Therefore, patterns are aggregated and approximated as community-level mobility instead of individual-based. For the visitor population, we used the ‘popularity by hour’ field from SafeGraph data for each point of interest (POI) and aggregated the values into each CBG.

Based on the estimated population, a mobility-based exposure (MBE) is calculated by matching visit locations with PM_2.5 concentrations estimated from the AutoML-LUR model. Specifically, MBE is calculated as follows: M B E c b g , m h = P M H × C (2) where P M H is the percentage of mobility-based visitor number estimated from SafeGraph for a particular census block group over all visitors in a particular month and hour, and C is the PM_2.5 concentration derived for the census block group for a particular month and hour from the 1 km-resolution AutoML-LUR results.

The model prediction performance varied temporally and spatially (Table 3). Replacing EPA data with PurpleAir data to train AutoML-LUR models resulted in significant decreases in the root mean squared error (RMSE), with a decrease of 0.51 μg/m³ (SD = 0.22 μg/m³) on average. In addition, using EPA + PurpleAir data to train AutoML-LUR models also resulted in significant decreases in RMSE compared to using EPA data only, with a decrease of 0.32 μg/m³ (SD = 0.26 μg/m³) on average. However, diverging patterns of RMSE and R²values are observed across different regions. For example, in regions such as Chicago and Houston, lower RMSE and higher R ² values were observed when only PurpleAir sensor data were used. However, in other regions such as Los Angeles, models trained on combined EPA and PurpleAir data outperformed those using only PurpleAir data. Several underlying factors may explain these observed variations. Firstly, the geographical placement of sensors is a significant determinant. PurpleAir sensors, predominantly purchased by private residents, are often situated within residential areas. In contrast, EPA monitors are strategically positioned near pollution emission sources or within areas characterized by particular land use types. Consequently, the heightened performance of models in regions using solely PurpleAir data (such as Chicago) could be attributed to their greater sensitivity in capturing the air quality variations within residential environments. Secondly, the spatial distribution and density of sensors is another potential influencer. In the regions where combined sensor data yielded higher performance (such as Los Angeles), the heterogeneous sensor locations and increased sensor density could be contributing factors. The amalgamation of data from both types of sensors provides a broader representation of air quality variations across distinct land uses and proximities to emission sources. Furthermore, the level of congruity between readings from different sensors may impact model performance. As supported by Figure 1, PurpleAir sensors demonstrate greater consistency among themselves compared to their consistency with EPA monitors, thus resulting in enhanced performance when solely PurpleAir data is employed.

The trained models allow the mapping of pollutant concentrations in the eight CBSAs. Local variations of PM_2.5 concentrations also vary by the input and can be significantly different among models (Figures 3A–C). AutoML-LUR models trained with EPA data tend to produce a higher concentration of PM_2.5 than models trained with PurpleAir data. Models trained with EPA + PurpleAir data showed higher spatial variance than the other two models, with high PM_2.5 concentrations varying spatially according to the observation used to train the AutoML-LUR model. Examples of predicting PM_2.5 for different hours in January 2021 in Los Angeles are shown in Figure 3D–G).

Figure 4 demonstrates the monthly trends of PM_2.5 predictions in the target CBSAs. Most CBSAs show a seasonal trend of PM_2.5 concentrations, but the levels of PM_2.5 concentrations vary significantly depending on location. For example, the levels of PM_2.5 concentrations tend to be the highest during the winter months in Los Angeles, primarily because of the increased emission and weather patterns in winter months, such as temperature inversions which can trap pollutants near the ground leading to higher PM_2.5 level (Wallace et al., 2010). On the other hand, PM_2.5 concentrations increased during the summer months in San Francisco, primarily because of natural sources such as wildfires. Some other CBSAs have higher PM_2.5 concentrations during summer months due to a variety of factors, such as increased heat and sunlight, which can lead to the formation of ground-level ozone, a major component of smog. Other factors that may contribute to the high PM_2.5 levels during summer include increased use of air conditioning and wildfires, which can release particles into the air.

Permutation feature importance was calculated to represent the increase in the prediction error of the model after we permuted feature’s values, which breaks the relationship between the feature and the true outcome (Fisher et al., 2019). Analyses of feature importance revealed the key factors that were spatially correlated with PM_2.5 concentrations (Figure 5). Spatially aggregated feature importance scores showed that satellite AOD, temporal indicators, and meteorological variables were highly correlated with PM_2.5 concentrations (Figure 5A). City-wise feature important scores reflect city-wide features most relevant to air pollutant concentrations. In most CBSAs, satellite AOD, meteorological variables, temporal indicators, NDVI, and land use remain highly correlated with PM_2.5 concentrations. Figure 5B–I provides a decomposed analysis for each CBSA. Spatial heterogeneity is observed regarding the CBSA’s most relevant sources of PM_2.5 other than AOD and meteorological conditions (e.g., distance to large airports for Chicago; imperviousness for Houston; nearby primary road length for Las Vegas; emission for Los Angeles, New York and Riverside; distance to coastline for New York, Riverside, and San Francisco; elevation for Phoenix and San Francisco).

Mobility-based exposure (MWE) for each census block group was calculated to analyze visitor exposure to air pollution, which was shown to vary significantly in terms of spatial distribution and temporal variation. Spatial patterns of MWE reveal hotspots of high visitor number weighted exposure to PM_2.5 and places where visitors are less exposed to this pollutant. For Los Angeles, the highest MWE values are found in areas near major transportation hubs and business districts, such as the Los Angeles International Airport, the Port of Los Angeles and Port of Long Beach, and Hollywood (Figure 6A). These values can vary over the time of the day and month of the year, as illustrated in the enlarged areas (Figures 6B–F). The months chosen to illustrate the temporal variation in Figures 6B–F were selected to represent different seasons throughout the year, to capture the potential seasonal differences in PM_2.5 exposure and visitor patterns in various areas of Los Angeles. These months also provide a comprehensive picture of visitor exposure to PM_2.5 throughout the year. Most of the CBGs demonstrate a relatively consistent pattern of MWE throughout the day, which might result from the levels of PM_2.5 or a relatively low number of visitors throughout the day. CBGs that are significantly deviating from the general temporal patterns show clear diurnal trends. One typical trend is that MWE increases during the day, peaking in the late morning to early afternoon as people arrive for work or other activities. MWE declines in the late afternoon and evening as people finish their activities and return home or to their lodging. Another typical trend is that MWE remains high at night and decreases during the day, which might result from high PM_2.5 concentrations or high visitor numbers at night.

MWE to PM_2.5 is a measure of the average exposure of visitors to a particular area, considering both the concentration of PM_2.5 in the air and the number of visitors present in the area. Supposing either of these factors is exceptionally high, the MWE value may appear as a hotspot within the region, indicating a potential area of concern for air quality and public health. However, it is important to note that a high MWE value on its own does not necessarily indicate significant exposure risks without examining both factors separately. For example, a high MWE value may be due to a high number of visitors in an area with relatively low PM_2.5 concentrations (Figure 7A. Airport), or it could be due to a relatively low number of visitors in an area with very high PM_2.5 concentrations (Figure 7B. Long Beach). In the first case, the risk of exposure to PM_2.5 for visitors may be relatively low, while in the second case, the risk may be much higher. Therefore, examining the visitor numbers and PM_2.5 concentrations separately is important to get a complete picture of potential exposure risks. This can help policymakers and public health officials develop targeted strategies to reduce exposure to PM_2.5 and improve air quality in areas with high levels of visitor activity.