Spatial-Temporal Analysis of PM2.5 and NO2 Concentrations Collected Using Low-Cost Sensors in Peñuelas, Puerto Rico

The U.S. Environmental Protection Agency (EPA) is involved in the discovery, evaluation, and application of low-cost air quality (AQ) sensors to support citizen scientists by directly engaging with them in the pursuit of community-based interests. The emergence of low-cost (<$2500) sensors have allowed a wide range of stakeholders to better understand local AQ conditions. Here we present results from the deployment of the EPA developed Citizen Science Air Monitor (CSAM) used to conduct approximately five months (October 2016–February 2017) of intensive AQ monitoring in an area of Puerto Rico (Tallaboa-Encarnación, Peñuelas) with little historical data on pollutant spatial variability. The CSAMs were constructed by combining low-cost particulate matter size fraction 2.5 micron (PM2.5) and nitrogen dioxide (NO2) sensors and distributed across eight locations with four collocated weather stations to measure local meteorological parameters. During this deployment 1 h average concentrations of PM2.5 and NO2 ranged between 0.3 to 33.6 µg/m3 and 1.3 to 50.6 ppb, respectively. Peak concentrations were observed for both PM2.5 and NO2 when conditions were dominated by coastal-originated winds. These results advanced the community’s understanding of pollutant concentrations and trends while improving our understanding of the limitations and necessary procedures to properly interpret measurements produced by low-cost sensors.


Introduction
The recent development of low-cost (<$2500) air quality (AQ) sensors has created new opportunities for individuals with various expertise to engage in AQ monitoring [1]. Traditionally, AQ monitoring has been limited to expensive regulatory monitors that require technical expertise

Instrumentation
EPA developed a new version of the Citizen Science Air Monitor (CSAM) based on lessons learned from a previous deployment [2]. Briefly, the CSAMs consisted of low-cost (<$2500) OEM sensors, an Arduino Uno microprocessor for data acquisition, and an 8 GB secure digital (SD) card for data storage. For this deployment each CSAM was designed to collect real-time measurements at 5 min intervals of the particulate matter size fraction 2.5 micron (PM 2.5 ), total volatile organic compounds (tVOCs), temperature ( • C), and relative humidity (RH; % at • C). The CSAMs measured PM 2.5 using an OPC-N2 Particle Monitor (Alphasense, Essex, United Kingdom) [32] and tVOCs with a Baseline Mocon piD-TECH sensor (MOCON Inc, Lyons, NY, United States) [33]. The OPC-N2 uses light particle counting to measure the concentration of suspended particles in the air sampled via an internal pump and has a detection limit of 0.1 µg/m 3 . The Baseline Mocon piD-TECH sensor uses a photoionization detector to measure a sum of all responding chemicals with a detection limit of 0.5 ppb isobutylene. Additionally, four CSAMs were equipped with a CairPol CairClip (Cairpol, Poissy, France) to measure real-time concentrations of nitrogen dioxide (NO 2 ) [34]. The CairClip used a gas-specific inlet filter combined with dynamic air sampling in an integrated system to measure real-time NO 2 concentrations with a detection limit of 1 ppb. During the deployment, four Vantage Vue weather stations (Davis Instruments, Hayward, CA, United States) were collocated with four CSAMs (CSAM 301, 304, 305, and 353/355) to provide meteorological parameters (rainfall, temperature, humidity, wind speed, and direction) every 30 min. Measurements collected from the tVOC sensors are not discussed in this manuscript pending additional investigation regarding sensor performance.

Deployment Area
A citizen science led effort conducted between October 2016-February 2017 deployed nine CSAMs across eight ambient locations in a southern area of Puerto Rico (Tallaboa-Encarnación, Peñuelas) with little historical data available on pollutant spatial variability. These locations, shown in Figure 1, were within close proximity to residential communities, a major highway (Hwy 2), and various potential industrial sources. During the deployment two CSAMs were collocated to assess the long-term performance of the low-cost sensors. CSAMs were operated on approximately a weekly schedule where citizen scientists retrieved data and maintained all operations. Citizen scientists also recorded any visually observed pollution episodes or perceived smells during these routine site visits.

Instrumentation
EPA developed a new version of the Citizen Science Air Monitor (CSAM) based on lessons learned from a previous deployment [2]. Briefly, the CSAMs consisted of low-cost (<$2500) OEM sensors, an Arduino Uno microprocessor for data acquisition, and an 8 GB secure digital (SD) card for data storage. For this deployment each CSAM was designed to collect real-time measurements at 5 min intervals of the particulate matter size fraction 2.5 micron (PM2.5), total volatile organic compounds (tVOCs), temperature (°C), and relative humidity (RH; % at °C). The CSAMs measured PM2.5 using an OPC-N2 Particle Monitor (Alphasense, Essex, United Kingdom) [32] and tVOCs with a Baseline Mocon piD-TECH sensor (MOCON Inc, Lyons, NY, United States) [33]. The OPC-N2 uses light particle counting to measure the concentration of suspended particles in the air sampled via an internal pump and has a detection limit of 0.1 µg/m 3 . The Baseline Mocon piD-TECH sensor uses a photoionization detector to measure a sum of all responding chemicals with a detection limit of 0.5 ppb isobutylene. Additionally, four CSAMs were equipped with a CairPol CairClip (Cairpol, Poissy, France) to measure real-time concentrations of nitrogen dioxide (NO2) [34]. The CairClip used a gas-specific inlet filter combined with dynamic air sampling in an integrated system to measure real-time NO2 concentrations with a detection limit of 1 ppb. During the deployment, four Vantage Vue weather stations (Davis Instruments, Hayward, CA, United States) were collocated with four CSAMs (CSAM 301, 304, 305, and 353/355) to provide meteorological parameters (rainfall, temperature, humidity, wind speed, and direction) every 30 min. Measurements collected from the tVOC sensors are not discussed in this manuscript pending additional investigation regarding sensor performance.

Deployment Area
A citizen science led effort conducted between October 2016-February 2017 deployed nine CSAMs across eight ambient locations in a southern area of Puerto Rico (Tallaboa-Encarnación, Peñuelas) with little historical data available on pollutant spatial variability. These locations, shown in Figure 1, were within close proximity to residential communities, a major highway (Hwy 2), and various potential industrial sources. During the deployment two CSAMs were collocated to assess the long-term performance of the low-cost sensors. CSAMs were operated on approximately a weekly schedule where citizen scientists retrieved data and maintained all operations. Citizen scientists also recorded any visually observed pollution episodes or perceived smells during these routine site visits.   For clarity, the locations of the CSAMs and Vantage Vue weather stations are discussed in terms of the west, north, and southeast regions of the study area. The west region includes CSAMs 305, 352, and the West weather station. The north region includes CSAMs 304, 351, and the North weather station. The southeast region includes CSAMs 301, 302, 303, the collocated CSAMs 353/355 and the South and East weather stations.

Collocation Period
A 1-week collocation of CSAMs and weather stations was performed at the Environmental Quality Board (EQB) regulatory site in Ponce, Puerto Rico. While a period of sensor collocation versus regulatory monitoring during the deployment was a part of the study design, ultimately the local regulatory agency was unable to provide these data. The collocation period however identified improperly functioning sensors and established a median response per pollutant as a reference signal. Linear regressions were used to normalize each low-cost sensor and weather station to the reference signal to allow for spatial comparisons. Sensors that correlated poorly (R 2 < 0.50) with the median reference signal were not included in the final analysis. This resulted in the exclusion of measurements from one PM 2.5 sensor in the west (CSAM 305). Prior to deployment a multipoint calibration was performed on the NO 2 sensors. A baseline check of each PM 2.5 sensor was performed in a particle clean room as well as laboratory calibration of the RH, temperature, and tVOC sensors prior to deployment. Results of the collocation period are described in more detail in an earlier publication [35].

Data Analysis and Quality Assurance Procedures
To address quality assurance concerns about low-cost sensor data, detailed validation procedures were followed. The impact of these data validation procedures is discussed elsewhere [35]. Briefly, data were flagged for exclusion based on four criteria in the following order: human interference, non-responsive signal, influence of environmental conditions, and exceedance of the limit of detection. Data from the PM 2.5 sensors were flagged when RH exceeded 90%. The removal of data collected under elevated RH conditions improved the precision between the PM 2.5 sensors by 10.7% and accounted for between 2.5-21.8% of the dataset across CSAMs.
An EPA developed Microsoft Excel Macro Analysis Tool (MAT) was used for initial data analysis efforts [36]. The time interval, sampling rate, limit of detection (LOD), averaging period, and data completeness were required to be specified. The MAT data completeness constraint specified the percentage of data between the averaging period required to produce an average value. Once initialized, the MAT automatically time aligned the provided data and output averaged data with a visual representation. The validated 5 min data collected by the CSAMs were averaged with the MAT to 1 h and 24 h values using a data completeness of 80%.
The relative spatial variably between CSAM locations was explored using the Pearson coefficient (r) and coefficient of divergence (COD). The Pearson coefficient describes how well correlated two pollutant distributions are but not how similar. The COD describes the degree of homogeneity between two pollutant distributions on a scale of 0-1 defined as: where σ represents the standard deviation between CSAM 353 and 355 and µ represents the mean concentration between CSAM 353 and 355. The CV describes the precision between two sensors using the ratio of the standard deviation and mean expressed as a percentage, with zero indicating perfect precision.

Meteorological Conditions
Meteorological conditions were measured by four Vantage Vue weather stations at 30 min intervals. The weather stations were deployed alongside the CSAMs in a wind rose pattern to ensure results were representative of the entire deployment area. A summary of the monthly average temperature ( • C), average RH (%), average wind speed (mph), and median wind direction ( • ) are reported in Table 1. Values for these parameters in November are not shown for the West weather station due to a delayed deployment. Temperatures were generally stable during the deployment period while RH decreased from November to February as Puerto Rico's traditional rainy period concluded. Similarly, monthly average wind speed (WS) increased during the deployment at each weather station with winds originating predominantly from the east to south-southeast direction. Active winds were defined as ≥ 2 mph. The East and South weather stations were the most active with 1 h average winds occurring 42% and 51.5% of the deployment respectively. At the East weather station 78.6% of the active winds approached from the east to southeast direction. The south weather station had the most active winds dominated by two directions. Winds between 2-3 mph arrived from the north-northeast to northeast direction (29.7% of active winds), while stronger winds between 3-6 mph originated from between the east and southeast directions (59.6% of active winds). The North weather station observed the least active winds (17.3% of deployment), typically approaching from the southeast to south-southeast direction at speeds between 2-4 mph. The West weather station observed active winds 27% of the deployment with 95.8% approaching from the southeast direction at WS predominantly greater than 6 mph. The strong winds observed by the East and West weather stations relative to the North and South weather stations were likely due to their proximity to the shoreline.
Similarly, the reduced wind activity at the North weather station was likely due to being located more inland. The relationship between the local meteorological conditions and pollutant concentrations at these locations was explored here to better understand the micro-environmental conditions.

Long-Term CSAM Collocation
During the deployment, two CSAMs (353 and 355) in the southeast region were collocated to assess the change in correlation and precision over time between replicates of the PM 2.5 and NO 2 sensors. The correlation and precision were calculated on a weekly basis using normalized 5 min data. The weekly correlation between the replicate PM 2.5 and NO 2 sensors were also calculated using normalized 1 h average data to demonstrate the effect of increased averaging time on data quality. Previous evaluations of low-cost sensors have demonstrated that over time a sensor's response can change due to a shift in baseline and degradation in the sensing mechanism of both light-scattering and gas-phase based sensors [20,37].
Change in correlation and precision over the duration of the deployment for the PM 2.5 (black dots) and NO 2 (purple dots) sensors are shown in Figure 2A,B. At the start of the deployment the PM 2.5 sensors were initially better correlated (r: 0.98 > 0.93) and more precise (CV: 5% < 23%) compared to the NO 2 sensors. Other studies have also observed greater inter-variability among gas-phase sensors compared to light-scattering sensors [15,17,38]. Figure 2A compares the correlation between 5 min (dashed line) and 1 h (solid line) average data from the PM 2.5 and NO 2 sensors. Averaging improved the correlation between the PM 2.5 sensors from 0.75 to 0.92 and the NO 2 sensors from 0.87 to 0.95. Figure 2A,B indicate that during the first 8 weeks the 5 min correlation between the PM 2.5 sensors ranged between 0.93 and 0.99 and the precision between 3% and 7%. After week 8 the correlation (r = 0.07-0.77) and precision (CV = 25-122%) between the PM 2.5 sensors were suddenly reduced. Further inspection revealed CSAM 353 began reporting a non-responsive PM 2.5 signal during week 9 until being restarted to begin week 12. After week 12 CSAM 353 was non-responsive 96.2% of the remaining deployment compared to 0.9% of the deployment for CSAM 355. This intermittent PM 2.5 response is highlighted in Figure 2A   likely due to being located more inland. The relationship between the local meteorological conditions and pollutant concentrations at these locations was explored here to better understand the micro-environmental conditions.

Long-Term CSAM Collocation
During the deployment, two CSAMs (353 and 355) in the southeast region were collocated to assess the change in correlation and precision over time between replicates of the PM2.5 and NO2 sensors. The correlation and precision were calculated on a weekly basis using normalized 5 min data. The weekly correlation between the replicate PM2.5 and NO2 sensors were also calculated using normalized 1 h average data to demonstrate the effect of increased averaging time on data quality. Previous evaluations of low-cost sensors have demonstrated that over time a sensor's response can change due to a shift in baseline and degradation in the sensing mechanism of both light-scattering and gas-phase based sensors [20,37].
Change in correlation and precision over the duration of the deployment for the PM2.5 (black dots) and NO2 (purple dots) sensors are shown in Figure 2A,B. At the start of the deployment the PM2.5 sensors were initially better correlated (r: 0.98 > 0.93) and more precise (CV: 5% < 23%) compared to the NO2 sensors. Other studies have also observed greater inter-variability among gas-phase sensors compared to light-scattering sensors [15,17,38]. Figure 2A compares the correlation between 5 min (dashed line) and 1 h (solid line) average data from the PM2.5 and NO2 sensors. Averaging improved the correlation between the PM2.5 sensors from 0.75 to 0.92 and the NO2 sensors from 0.87 to 0.95. Figure 2A,B indicate that during the first 8 weeks the 5 min correlation between the PM2.5 sensors ranged between 0.93 and 0.99 and the precision between 3% and 7%. After week 8 the correlation (r = 0.07-0.77) and precision (CV = 25-122%) between the PM2.5 sensors were suddenly reduced. Further inspection revealed CSAM 353 began reporting a non-responsive PM2.5 signal during week 9 until being restarted to begin week 12. After week 12 CSAM 353 was non-responsive 96.2% of the remaining deployment compared to 0.9% of the deployment for CSAM 355. This intermittent PM2.5 response is highlighted in Figure 2A    During the first 6 weeks the 5 min correlation between NO 2 sensors ranged between 0.86 and 0.95 while the precision improved from 23% to 10%. This initial improvement in precision was potentially due to the electrochemical sensors stabilizing to new environmental conditions. During weeks 8 to 11 the 5 min correlation (r = 0.88-0.89) and precision (CV = 10-13%) were stable. Figure 2A,B show the NO 2 sensors maintained correlation (r = 0.89) during week 12, but for unexplained reasons a decrease in precision was observed (CV = 66%). During weeks 13 to 16 the correlation (r = 0.87-0.54) and precision (CV = 17-35%) between NO 2 sensors gradually decreased. This gradual increase in variability between collocated sensor replicates was likely due to potential degradation of the sensor transducer [20,37]. The median value of CSAM 353 and 355, referred to as CSAM353/355, was used for all following temporal and spatial pollutant analysis reported here. During periods when only one of the two collocated CSAMs were reporting, then the value of the functioning CSAM was used for comparisons.

1 h Average Pollutant Concentrations
The normalized 5 min PM 2.5 and NO 2 measurements were averaged on an hourly basis using the MAT [36] with a data completeness of 80%. Distributions of 1 h average PM 2.5 and NO 2 concentrations are shown in Figure 3A,B. The box represents the interquartile range of 25th and 75th percentile and the whiskers indicate the 5th and 95th percentile. The horizontal line in each box is the median concentration. The number of 1 h average data points measured at each CSAM location is displayed along the x-axis. The variation in data points collected between CSAM locations was the result of sensor failure and/or interference of environmental conditions. This resulted in a data completeness between 8.6% and 80.0% for the PM 2.5 sensors and between 97.4% and 99.7% for the NO 2 sensors. Sensor performance and quality assurance procedures are discussed in more detail in a previous publication [35]. During the first 6 weeks the 5 min correlation between NO2 sensors ranged between 0.86 and 0.95 while the precision improved from 23% to 10%. This initial improvement in precision was potentially due to the electrochemical sensors stabilizing to new environmental conditions. During weeks 8 to 11 the 5 min correlation (r = 0.88-0.89) and precision (CV = 10-13%) were stable. Figure 2A,B show the NO2 sensors maintained correlation (r = 0.89) during week 12, but for unexplained reasons a decrease in precision was observed (CV = 66%). During weeks 13 to 16 the correlation (r = 0.87-0.54) and precision (CV = 17-35%) between NO2 sensors gradually decreased. This gradual increase in variability between collocated sensor replicates was likely due to potential degradation of the sensor transducer [20,37]. The median value of CSAM 353 and 355, referred to as CSAM353/355, was used for all following temporal and spatial pollutant analysis reported here. During periods when only one of the two collocated CSAMs were reporting, then the value of the functioning CSAM was used for comparisons.

1 h Average Pollutant Concentrations
The normalized 5 min PM2.5 and NO2 measurements were averaged on an hourly basis using the MAT [36] with a data completeness of 80%. Distributions of 1 h average PM2.5 and NO2 concentrations are shown in Figure 3A,B. The box represents the interquartile range of 25th and 75th percentile and the whiskers indicate the 5th and 95th percentile. The horizontal line in each box is the median concentration. The number of 1 h average data points measured at each CSAM location is displayed along the x-axis. The variation in data points collected between CSAM locations was the result of sensor failure and/or interference of environmental conditions. This resulted in a data completeness between 8.6% and 80.0% for the PM2.5 sensors and between 97.4% and 99.7% for the NO2 sensors. Sensor performance and quality assurance procedures are discussed in more detail in a previous publication [35].   [39]. An additional study operated a MET One Neighborhood monitor at the University of Puerto Rico's Rio Piedras campus from 22 November 2017 to 22 December 2017 and observed an average 1 h PM 2.5 concentration of 3.5 ± 2.3 µg/m 3 [40]. The MET One is a low-cost PM sensor that uses light-scattering detection similar to the PM sensor deployed in this study. We reference these studies here only to provide a general review of the localize air quality conditions in other areas of Puerto Rico. The mean PM 2.5 concentrations (5.03-6.32 µg/m 3 ) reported by the MET One in urban areas of Puerto Rico were only slightly greater than the median 1 h average PM 2.5 concentrations observed in this study [39].
The maximum 1 h average median PM 2.5 concentration (33.6 µg/m 3 ) observed during the deployment was at the South weather station (CSAM 353/355). This 1 h average concentration was examined beyond the required quality assurance procedures to ensure the validity of this data point. Unfortunately, during this period CSAM 353 was non-responsive, preventing the comparison of the collocated CSAMs. Evaluation of normalized 5 min data during this period on 23 December 2016 reveals a bimodal PM 2.5 event. The first event starts at approximately 8:00 AM and peaks at 141.6 µg/m 3 by 8:25 AM. A second smaller event begins at 8:40 AM and peaks at 12.0 µg/m 3 by 9:00 AM. Further inspection of the normalized 5 min data during this period identified a similar bimodal PM 2.5 event of lesser magnitude at a nearby location across from Hwy 2 (CSAM 302). At this location, PM 2.5 concentrations suddenly increased from 5.8 to 26.1 µg/m 3 at 9:00 AM, followed by a second event at 9:20 AM that peaked at 13.3 µg/m 3 at 9:30 AM. The response of the PM 2.5 sensor to this event (peak followed by gradual decline) in addition to the observation of a similar event of a lesser magnitude at a nearby location indicates this data point was suggestive of a true PM 2.5 event. The observed variation in PM 2.5 concentrations between CSAMs was indicative of micro-environmental conditions that were then explored temporally and spatially.
The 1 h average median NO 2 concentrations shown in Figure 3B ranged between 5.9 and 8.7 ppb. The lowest 1 h average median NO 2 concentration was observed at the west location (CSAM 352), similar to the PM 2.5 concentrations. The west location was positioned the farthest from Hwy 2 followed by the north (~100 m) and southeast (~15 m) locations. The 1 h average median NO 2 concentrations observed at these locations decreased as a function of the distance from Hwy 2. The north location (CSAM 351) consistently observed higher 1 h average NO 2 concentrations despite being located farther from Hwy 2 and having a lower median NO 2 concentration than the southeast location (CSAM 353/355). This is shown in Figure 3B by the wider distribution skewed towards higher NO 2 concentrations at the north location. The maximum 1 h NO 2 concentration reported by EQB monitoring stations during our deployment period was 159 ppb (Caguas, PR) and 195 ppb (Guaynabo, PR) in 2016 and 104 ppb (Caguas, PR) and 28 ppb (Guaynabo, PR) in 2017. The maximum 1 h NO 2 concentration of 50.6 ppb observed by the north location was still lower than the NAAQS 1 h NO 2 concentration of 100 ppb.

Spatial Analysis of 1 h Average PM 2.5 and NO 2 Concentrations
The spatial comparison of 1 h average PM 2.5  The spatial comparison between PM2.5 distributions across CSAM locations is shown in Figure 4. The correlations and COD values shown in Figure 4 were also recalculated while excluding the previously discussed elevated concentration observed at the South weather station to determine the impact on analysis. The removal of this data point improved the correlation between CSAMs 351 and 353/355 (r = 0.64-0.92) but did not significantly affect any other correlations or COD values. This elevated concentration was therefore excluded. The PM2.5 distributions in the southeast, excluding CSAM 302, were well correlated (r = 0.92-0.99). Comparisons between the southeast and north regions were slightly less correlated (r = 0.81-0.97) but indicated similar responses. The correlation between CSAM locations in the north (r = 0.81) was less than in the southeast, possibly due to differences in wind activity. As previously mentioned, the wind activity in the southeast region on the coast was more active (42-51.5%) than in the mainland to the north (17.3%).  The spatial comparison between PM 2.5 distributions across CSAM locations is shown in Figure 4. The correlations and COD values shown in Figure 4 were also recalculated while excluding the previously discussed elevated concentration observed at the South weather station to determine the impact on analysis. The removal of this data point improved the correlation between CSAMs 351 and 353/355 (r = 0.64-0.92) but did not significantly affect any other correlations or COD values. This elevated concentration was therefore excluded. The PM 2.5 distributions in the southeast, excluding CSAM 302, were well correlated (r = 0.92-0.99). Comparisons between the southeast and north regions were slightly less correlated (r = 0.81-0.97) but indicated similar responses. The correlation between CSAM locations in the north (r = 0.81) was less than in the southeast, possibly due to differences in wind activity. As previously mentioned, the wind activity in the southeast region on the coast was more active (42-51.5%) than in the mainland to the north (17.3%).
The 1 h average PM 2.5 concentrations reported by CSAM 351 in the north were similar (COD = 0.08-0.18) and correlated (r = 0.81-0.96) to all locations. This agreement with the other locations was likely due to either CSAM 351's relatively central location or low data completeness (10.1%). The western most location and the elevated location in the southeast (CSAM 302) were uncorrelated with all locations except CSAM 351. Figure 4 shows the relatively low PM 2.5 distribution in the west was spatially unique (COD = 0.24-0.32) when compared to the 2 locations with the highest median 1 h average PM 2.5 concentrations (CSAMs 302 and 304). Similarly, CSAM 302 was the most spatially unique location with the majority of comparisons resulting in COD values > 0.20, despite other CSAMs located nearby in the southeast region. The lack of correlation and spatial homogeneity between CSAM 302 and the other southeast locations in conjunction with the relatively higher median 1 h average PM 2.5 concentration ( Figure 3A) was suggestive of a local source of PM 2.5 unique to this location. Alternatively, the lack of correlation and low PM 2.5 concentrations reported by the west location compared to other CSAMs was suggestive this location was more representative of background conditions. To maximize the data available for comparisons, the three CSAM locations that collected the most 1 h average PM 2.5 measurements (CSAMs 302, 304, and 353/355) were used for additional spatial and temporal analysis [35].
It was previously shown the median 1 h average NO 2 concentrations decreased across regions (southeast-west), as a function of distance from Hwy 2. Spatial analysis of the 1 h average NO 2 concentrations in Figure 5 shows the southeast and west locations were equivalently correlated (r = 0.70) and spatially homogeneous (COD = 0.17) as the collocated NO 2 sensors. The similar response from the southeast and west locations likely indicated a similar near-road source with the distance from the road influencing the difference in median 1 h average concentrations ( Figure 3B). The lower correlation (r = 0.59-0.64) and dissimilarity (COD = 0.22-0.23) of the NO 2 concentrations observed at the north location, in conjunction with higher maximum concentrations ( Figure 3B), indicated a unique response potentially due to an additional source of NO 2 . To further explore these spatial relationships, wind speed and direction were examined to isolate conditions when distribution of concentrations were correlated or similar.  Figure 4 shows the relatively low PM2.5 distribution in the west was spatially unique (COD = 0.24-0.32) when compared to the 2 locations with the highest median 1 h average PM2.5 concentrations (CSAMs 302 and 304). Similarly, CSAM 302 was the most spatially unique location with the majority of comparisons resulting in COD values > 0.20, despite other CSAMs located nearby in the southeast region. The lack of correlation and spatial homogeneity between CSAM 302 and the other southeast locations in conjunction with the relatively higher median 1 h average PM2.5 concentration ( Figure 3A) was suggestive of a local source of PM2.5 unique to this location. Alternatively, the lack of correlation and low PM2.5 concentrations reported by the west location compared to other CSAMs was suggestive this location was more representative of background conditions. To maximize the data available for comparisons, the three CSAM locations that collected the most 1 h average PM2.5 measurements (CSAMs 302, 304, and 353/355) were used for additional spatial and temporal analysis [35].
It was previously shown the median 1 h average NO2 concentrations decreased across regions (southeast-west), as a function of distance from Hwy 2. Spatial analysis of the 1 h average NO2 concentrations in Figure 5 shows the southeast and west locations were equivalently correlated (r = 0.70) and spatially homogeneous (COD = 0.17) as the collocated NO2 sensors. The similar response from the southeast and west locations likely indicated a similar near-road source with the distance from the road influencing the difference in median 1 h average concentrations ( Figure 3B). The lower correlation (r = 0.59-0.64) and dissimilarity (COD = 0.22-0.23) of the NO2 concentrations observed at the north location, in conjunction with higher maximum concentrations ( Figure 3B), indicated a unique response potentially due to an additional source of NO2. To further explore these spatial relationships, wind speed and direction were examined to isolate conditions when distribution of concentrations were correlated or similar.

Spatial Analysis of PM 2.5 and NO 2 Concentrations as a Function of Wind Conditions
The spatial relationship of pollutants between CSAM locations was further explored by examining the effect of changing wind conditions. Using the South weather station as a reference, 1 h average meteorological data were used to identify the following 3 wind conditions: inactive winds ( Figures 6A  and 7A), costal-originated winds ( Figures 6B and 7B), and mainland-originated winds (Figures 6C  and 7C). Periods of inactive winds occurred 48.5% of the deployment at the South weather station. Figures 6A and 7A shows during periods of inactive winds at the South weather station the North and East weather stations similarly observed weak (<2 mph) winds from the SW and E directions, respectfully. Figures 6B and 7B shows the ocean-originated winds (30.7% of deployment) were characterized by strong WS approaching from the east-southeast to south-southeast direction that weaken the farther inland. Figures 6C and 7C shows as winds transitioned to a mainland-originated wind approaching from the north-northeast (15.3% of deployment), the winds in the north diminish and the East weather station shifted back to a weak easterly direction, similar to the period of inactive winds (Figures 6A and 7A). This demonstrated the coastal and mainland winds were only active during the period of ocean-originated winds ( Figures 6B and 7B). In Figure 6A-C, the three CSAM locations compared were located at the North weather station (CSAM 304), South weather station (CSAM 353/355), and in a southern residential community (CSAM 302). The South weather station was a near-road location (~15 m from Hwy 2) located at a local public school. The southern residential community (CSAM 302) was located approximately 0.5 km adjacent to the South weather station on the opposite side of Hwy 2 at a 30 m elevation. The North weather station was located approximately~2.3 km north of the South weather station in a residential community 2 km west of the Peñuelas Valley Landfill.
The 1 h average PM 2.5 concentrations were spatially homogeneous (COD = 0.16-0.20) and similarly correlated (r = 0.60-0.73) between all locations during the period of inactive winds shown in Figure 6A. As coastal-originated winds began to approach, PM 2.5 concentrations peaked across all 3 locations and maintained spatial homogeneity (COD = 0.14-0.18). Although PM 2.5 distributions were similar across all 3 locations, only the north and southern near-road location were correlated (r = 0.92). As winds transitioned from coastal to mainland-originated, the north and southern near-road locations remain correlated (r = 0.80). During all three wind conditions ( Figure 6A-C) the PM 2.5 concentrations were greater in the north relative to the south, similar to the PM 2.5 distributions shown in Figure 3A. Figure 6A-C indicate similar responses between the north and southern near-road locations as correlations improved with increased wind activity. The southern residential location was only similarly correlated during periods of inactive winds and relatively low PM 2.5 concentrations. A possible explanation could be the southern residential location was on the opposite side of Hwy 2, upwind of potential near-road PM 2.5 sources. This location was therefore possibly more representative of local background concentrations. Other near-road studies have utilized similar site selections with respect to wind direction to differentiate between background and on-road pollution [41,42].  In Figure 7A-C 3 near-road CSAM locations in the southeast (CSAM 353/355), north (CSAM 351), and west (CSAM 352) were compared during the previously defined wind conditions. The 1 h average NO 2 concentrations shown in Figure 7A were more variable (COD = 0.20-0.21) and uncorrelated (r = 0.46-0.57) during inactive winds. Figure 7A shows the farthest near-road location in the west observed lower NO 2 concentrations. Similar to the PM 2.5 concentrations, NO 2 concentrations peaked at every location during coastal-originated wind conditions. During coastal-originated wind conditions, shown in Figure 7B, the west and southeast near-road locations were similar (COD = 0.13) and correlated (r = 0.69) indicating the influence of a similar near-road source (Hwy 2). The north near-road location was dissimilar (COD = 0.24) and uncorrelated (r = 0.39-0.46) compared to the west and southeast locations. The north near-road location also observed relatively higher NO 2 concentrations than the southeast location despite a farther distance from Hwy 2. This spatially unique location with relatively higher concentrations could be indicative of an additional source to the south-southeast of the north location shown in Figure 7B. When wind conditions changed to being mainland-originated, minimum NO 2 concentrations were observed at all 3 locations probably due to the sites no longer being downwind of the near-road sources. Comparisons between distributions of NO 2 concentrations indicated homogeneity (COD = 0.13-0.15) and correlations varied as a function of distance between the compared locations as shown in the bar plot inset in Figure 7C. The wind condition shown in Figure 7C was likely representative of local background NO 2 concentrations. The analysis of Figure 7A-C demonstrated data from the study being useful to to spatially compare pollutant concentrations as a function of wind conditions and identify isolated locations and/or periods with elevated concentrations. In Figure 7A-C 3 near-road CSAM locations in the southeast (CSAM 353/355), north (CSAM 351), and west (CSAM 352) were compared during the previously defined wind conditions. The 1 h average NO2 concentrations shown in Figure 7A were more variable (COD = 0.20-0.21) and uncorrelated (r = 0.46-0.57) during inactive winds. Figure 7A shows the farthest near-road location in the west observed lower NO2 concentrations. Similar to the PM2.5 concentrations, NO2 concentrations peaked at every location during coastal-originated wind conditions. During coastal-originated wind conditions, shown in Figure 7B, the west and southeast near-road locations were similar (COD = 0.13) and correlated (r = 0.69) indicating the influence of a similar near-road source (Hwy 2). The north near-road location was dissimilar (COD = 0.24) and uncorrelated (r = 0.39-0.46) compared to the west and southeast locations. The north near-road location also observed relatively higher NO2 concentrations than the southeast location despite a farther distance from Hwy 2. This spatially unique location with relatively higher concentrations could be indicative of an additional source to the south-southeast of the north location shown in Figure 7B. When wind conditions changed to being mainland-originated, minimum NO2 concentrations were observed at all 3 locations probably due to the sites no longer being downwind of the near-road sources. Comparisons between distributions of NO2 concentrations indicated homogeneity (COD = 0.13-0.15) and correlations varied as a function of distance between the compared locations as shown in the bar plot inset in Figure 7C. The wind condition shown in Figure 7C was likely representative of local background NO2 concentrations. The analysis of Figure 7A-C demonstrated data from the study being useful to to spatially compare pollutant concentrations as a function of wind conditions and identify isolated locations and/or periods with elevated concentrations.

Temporal Comparisons
To explore the weekly variation between CSAM locations, 1 h average PM2.5 and NO2 distributions were time-aligned and binned hourly. Any data gaps in Figure 8 were the result of one or more non-functioning sensors during the period. The spatial variation between CSAM locations was compared using daily average r and COD values calculated in a pairwise fashion. Temporal analysis allowed concentrations to be compared on an hourly basis to identify diurnal trends and pollution events independent of wind conditions. The weekly variation in PM2.5 concentrations between CSAMs 302, 304, and 353/355 is shown in Figure 8A. Across all three locations, PM2.5 concentrations were the lowest on the weekend (3.6-4.7 µg/m 3 ) and continuously increased during the week until peaking on Thursday (5.4-6.5 µg/m 3 ). Similar to earlier observations, concentrations in the north (CSAM 304) were relatively higher compared to the southeastern locations. The southeastern near-road location (CSAM 353/355) observed the largest difference between average PM2.5 concentrations on the weekend (3.9 µg/m 3 ) compared to during the week (5.3 µg/m 3 ). The adjacent location across Hwy 2 observed the least difference between weekend (4.3 µg/m 3 ) and weekday (4.8 µg/m 3 ) PM2.5 concentrations. A possible explanation for the relatively lower weekend PM2.5 concentrations was reduced on-road vehicles. This would also explain why the largest change in PM2.5 concentrations were observed at the near-road location. During the week the north and southeastern near-road locations were correlated everyday (r = 0.84-0.95), with exception to Friday (r = 0.64) when none of the locations were correlated. This maintained correlation between locations was likely due PM2.5 concentrations being dominated by on-road vehicles during the week. It was previously shown by Figure 6A-C that the relationship between the north and southeastern near-road locations improved when sites were downwind of Hwy 2. Similar to the spatial analysis of PM2.5 concentrations above, the southeastern residential community was only correlated (r = 0.91-0.95) and similar (COD = 0.14-0.18) to the adjacent near-road site during low PM2.5 concentrations on the weekend.

Temporal Comparisons
To explore the weekly variation between CSAM locations, 1 h average PM 2.5 and NO 2 distributions were time-aligned and binned hourly. Any data gaps in Figure 8 were the result of one or more non-functioning sensors during the period. The spatial variation between CSAM locations was compared using daily average r and COD values calculated in a pairwise fashion. Temporal analysis allowed concentrations to be compared on an hourly basis to identify diurnal trends and pollution events independent of wind conditions. The weekly variation in PM 2.5 concentrations between CSAMs 302, 304, and 353/355 is shown in Figure 8A. Across all three locations, PM 2.5 concentrations were the lowest on the weekend (3.6-4.7 µg/m 3 ) and continuously increased during the week until peaking on Thursday (5.4-6.5 µg/m 3 ). Similar to earlier observations, concentrations in the north (CSAM 304) were relatively higher compared to the southeastern locations. The southeastern near-road location (CSAM 353/355) observed the largest difference between average PM 2.5 concentrations on the weekend (3.9 µg/m 3 ) compared to during the week (5.3 µg/m 3 ). The adjacent location across Hwy 2 observed the least difference between weekend (4.3 µg/m 3 ) and weekday (4.8 µg/m 3 ) PM 2.5 concentrations. A possible explanation for the relatively lower weekend PM 2.5 concentrations was reduced on-road vehicles. This would also explain why the largest change in PM 2.5 concentrations were observed at the near-road location. During the week the north and southeastern near-road locations were correlated everyday (r = 0.84-0.95), with exception to Friday (r = 0.64) when none of the locations were correlated. This maintained correlation between locations was likely due PM 2.5 concentrations being dominated by on-road vehicles during the week. It was previously shown by Figure 6A-C that the relationship between the north and southeastern near-road locations improved when sites were downwind of Hwy 2. Similar to the spatial analysis of PM 2.5 concentrations above, the southeastern residential community was only correlated (r = 0.91-0.95) and similar (COD = 0.14-0.18) to the adjacent near-road site during low PM 2.5 concentrations on the weekend. This further supported the belief that the north and southeastern near-road locations were dominated by on-road PM 2.5 sources not detected by the location upwind of Hwy 2 (CSAM 302).
This further supported the belief that the north and southeastern near-road locations were dominated by on-road PM2.5 sources not detected by the location upwind of Hwy 2 (CSAM 302). Temporal analysis of NO2 concentrations, shown in Figure 8B, revealed a bimodal distribution of NO2 concentrations at each near-road location occurring on a daily basis. NO2 concentrations peaked at each location at approximately 6:00 AM and 6:00 PM. Other near-road studies have observed similar trends in NO2 concentrations indicating this response was likely a function of morning and evening on-road commuters. Daily average NO2 concentrations were consistent on a weekly basis, with higher concentrations observed at the north near-road location (10.6 ± 0.7 ppb) compared to the west (7.9 ± 0.2 ppb) and southeast (8.8 ± 0.5 ppb) locations. Spatial analysis demonstrated the west (CSAM 352) and southeast (CSAM 353/355) near-road locations were similar (COD = 0.15-0.18) every day and correlated (r = 0.63-0.76) every day, except Sunday. The relationship between the west and southeast locations, despite distance, indicated NO2 concentrations were likely being dominated by a common source such as Hwy 2. Figure 7A-C revealed the correlation and similarity between the southeast and west locations improved when active winds were blowing downwind or upwind of Hwy 2. The north location on average was equivalently similar to both the west (COD = 0.20 ± 0.03) and southeast (COD = 0.20 ± 0.01) locations as was shown in Figure 7A-C. However, the north location on average was more correlated with the west (r = 0.70 ± 0.08) than the southeast (r = 0.63 ± 0.07). This same relationship between locations was observed during non-coastal-originated source winds ( Figure 7A,C). This was expected as daily wind activity was dominated by non-coastal-originated winds. Although the southeast location was closer to the north than the west location, correlations were lower and NO2 distributions were more dissimilar, indicating a potential additional source of NO2 unique to the north near-road location. This belief was supported by the spatial analysis that demonstrated the north near-road location observed the largest delta change in NO2 concentrations (7.8 ppb) and became uncorrelated (r = 0.39-0.46) as coastal-originated winds approached from upwind of the near-road source.

Discussion
In this study citizen scientists deployed and maintained low-cost AQ sensors for an extended period which provided the basis for extensive data analyses on the spatial and temporal variability Temporal analysis of NO 2 concentrations, shown in Figure 8B, revealed a bimodal distribution of NO 2 concentrations at each near-road location occurring on a daily basis. NO 2 concentrations peaked at each location at approximately 6:00 AM and 6:00 PM. Other near-road studies have observed similar trends in NO 2 concentrations indicating this response was likely a function of morning and evening on-road commuters. Daily average NO 2 concentrations were consistent on a weekly basis, with higher concentrations observed at the north near-road location (10.6 ± 0.7 ppb) compared to the west (7.9 ± 0.2 ppb) and southeast (8.8 ± 0.5 ppb) locations. Spatial analysis demonstrated the west (CSAM 352) and southeast (CSAM 353/355) near-road locations were similar (COD = 0.15-0.18) every day and correlated (r = 0.63-0.76) every day, except Sunday. The relationship between the west and southeast locations, despite distance, indicated NO 2 concentrations were likely being dominated by a common source such as Hwy 2. Figure 7A-C revealed the correlation and similarity between the southeast and west locations improved when active winds were blowing downwind or upwind of Hwy 2. The north location on average was equivalently similar to both the west (COD = 0.20 ± 0.03) and southeast (COD = 0.20 ± 0.01) locations as was shown in Figure 7A-C. However, the north location on average was more correlated with the west (r = 0.70 ± 0.08) than the southeast (r = 0.63 ± 0.07). This same relationship between locations was observed during non-coastal-originated source winds ( Figure 7A,C). This was expected as daily wind activity was dominated by non-coastal-originated winds. Although the southeast location was closer to the north than the west location, correlations were lower and NO 2 distributions were more dissimilar, indicating a potential additional source of NO 2 unique to the north near-road location. This belief was supported by the spatial analysis that demonstrated the north near-road location observed the largest delta change in NO 2 concentrations (7.8 ppb) and became uncorrelated (r = 0.39-0.46) as coastal-originated winds approached from upwind of the near-road source.

Discussion
In this study citizen scientists deployed and maintained low-cost AQ sensors for an extended period which provided the basis for extensive data analyses on the spatial and temporal variability of select air pollutants. Two types of low-cost sensors were collocated for an extended period to evaluate their performance in a harsh costal environment. The extended collocation provided observations that a sudden change in response between some PM 2.5 sensors had occurred that would have been otherwise undetected. This demonstrates the need to incorporate routine QA checks during deployment periods to verify measurements. Experimental designs that include collocation at the beginning and end of a deployment are able to account for natural degradation in sensor performance but are unable to identify real-time sensor failure in situations where a sensor reports values unreflective of reality. Additionally, it has been shown that sensor performance is a function of localized environmental conditions and collocations performed outside the deployment area are typically not useful as part of the development of a correction algorithm [18]. Studies have attempted to address this issue by deploying AQ instruments with higher precision and accuracy that have been calibrated against a reference instruments alongside the low-cost sensors and periodically rotating them about the low-cost network [43].
An advantage of real-time measurements was shown to be the ability to establish diurnal trends across pollutants. This allowed specific days and hours when concentration peaked to be identified to contribute to communities understanding of pollutants. Although these low-cost sensors were not free of issues (low data completeness, environmental influences, poor agreement) rigorous QA guidelines allowed for the validation of an extensive data set. A pending manuscript details the QA process for verifying these data and examines the success and failures of the deployment from a citizen science perspective. These low-cost sensors demonstrated the ability to advance a community's understanding of pollutant concentrations and trends while improving our understanding of the limitations and necessary procedures to properly interpret measurements produced by low-cost sensors.

Conclusions
1 h average concentrations of PM 2.5 (0.3-33.6 µg/m3) and NO 2 (1.3-50.6 ppb) were relatively low across the deployment area. The median 1 h average PM 2.5 concentrations were greater in the north region and lower in the west region. The median 1 h average NO 2 concentrations varied as a function of distance from Hwy 2. Despite the low range of observed PM 2.5 and NO 2 concentrations, relationship between CSAM locations identified uncorrelated and heterogeneous pollutant distributions. The relationships established between CSAM locations were further explored as a function of wind conditions using the Vantage Vue weather stations. Wind conditions across the island were predominantly calm (WS < 2 mph), with the majority of active winds originating from the coastal direction. PM 2.5 concentrations peaked when conditions were dominated by coastal-originated winds. Correlations between PM 2.5 concentrations in the north (CSAM 304) and southeastern residential community (CSAM 353/355) improved with increased wind activity indicating a similar response. PM 2.5 concentrations at the neighborhood location (CSAM 302) upwind of Hwy 2 became less correlated with increasing wind activity and increasing PM 2.5 concentrations. Similar to PM 2.5 concentrations, NO 2 concentrations peaked as coastal-originated winds approached with NO 2 sensors located downwind of Hwy 2. The lowest NO 2 concentrations were observed at all locations as winds shifted to mainland-originated winds and the NO 2 sensors were upwind of the near-road source. During the mainland-originated regime, correlations between NO 2 concentrations varied as a function of distance between CSAM locations. When the NO 2 sensors were downwind of Hwy 2 the north location observed the largest increase in concentration and became less similar and uncorrelated, indicating a potential additional local source of NO 2 . Pollutant distributions were then analyzed temporally to explore how concentrations vary on a weekly basis. PM 2.5 concentrations were lowest on the weekends and gradually increased during the week until peaking on Thursday. PM 2.5 concentrations in the north (CSAM 304) and southeastern residential community (CSAM 353/355) were correlated every day of the week, with exception to Friday, while the neighborhood location (CSAM 302) was only correlated on the weekend when concentrations were lower. NO 2 concentrations consistently peaked at 6:00 AM and PM daily likely indicating a response to morning and evening on-road commuters. Similar to the trends observed during the spatial analysis, correlations were lower and NO 2 distributions were more dissimilar at the north near-road location, indicating a potential additional source of NO 2 .
Author Contributions: R.W., M.C., E.H., M.O., P.S., D.S. conceived and planned the experiments. D.S. carried out the field deployments and citizen scientists' leadership efforts. S.R. and R.W. took the lead in authoring the manuscript and summarizing key findings. All authors provided critical feedback and helped shaped the research, analysis, and manuscript. The dataset used in support of this article will be available upon publication at the EPA environmental dataset gateway (https://edg.epa.gov) (EPA 2018). The dataset can be retrieved by searching for the name of the senior EPA author (R.W.).
Funding: The U.S. Environmental Protection Agency through its Office of Research and Development funded and conducted the research described here in association with Region 2 through a Regional Applied Research Effort (RARE). It has been reviewed by the U.S. EPA and approved for publication. The effort was supported through a contract with Jacobs Technology Inc. (EP-C-15-008). Mention of trade names or commercial products does not constitute endorsement or recommendation for use.