In urban environments, air quality has complex spatial and temporal patterns. Diverse emission sources are present with large variations in emission rate and source type on scales of hundreds of meters. In addition, dispersion of pollutants into the urban environment is affected by the topography of the urban landscape and the associated wind flows, which also vary on length scales of ∼ 100 m (Vardoulakis et al., 2003; Lateb et al., 2016). Conventional approaches to air quality monitoring rely on a limited number of relatively high-cost instruments that lack the spatial resolution needed to characterize these variations, opting instead to target spatial averages. This averaging hampers our attempts at source attribution and understanding of mixing, chemistry, and human exposure in cities where emissions vary on spatial scales that are small compared to typical observations or models.

One approach to obtaining higher spatial resolution observations is passive sampling, which has been implemented using inexpensive sampling devices that can be later analyzed in bulk. Passive samplers do not require electrical power to function properly and are collected and analyzed 1 to 2 weeks after deployment. Such protocols provide high spatial resolution but also have significant drawbacks. Spatial resolution is gained at the expense of temporal resolution, and analysis after collection of the samplers is time consuming; thus passive sampling has typically been used only in short-duration experiments (e.g., Krupa and Legge, 2000; Cox, 2003). Furthermore, as a result of boundary layer dynamics, passive sampling in urban areas is likely dominated by the high concentrations found at night and relatively insensitive to daytime variability.

Recent developments in low-cost sensors for trace gases and particulate matter, as well as advances in software and hardware enabling low-cost data communication, have made high-density, high-time-resolution air quality monitoring networks possible. Devices and networks of devices are emerging that are low cost, report at a time resolution of seconds, and are capable of long-term deployment, providing potential for improvement over the two major weaknesses of passive sampling. Examples include metal oxide sensors used to measure O3, CO, NO2, and total volatile organic compounds (e.g., Williams et al., 2013; Bart et al., 2014; Piedrahita et al., 2014; Moltchanov et al., 2015; Sadighi et al., 2017), and electrochemical sensors used to measure CO, NO, NO2, O3, and SO2 (e.g., Mead et al., 2013; Sun et al., 2015; Jiao et al., 2016; Hagan et al., 2017; Jerrett et al., 2017; Mueller et al., 2017). These different low-cost sensor systems have been evaluated and compared (Borrego et al., 2016; Papapostolou et al., 2017). While these studies found low-cost trace gas sensors to be successful at qualitatively characterizing the variability of air quality in an urban area, challenges related to selectivity and stability remain, hindering more quantitative interpretation of the data.

The current generation of low-cost sensors is not as easily tied to a gravimetric calibration standard as many of the passive samplers. Calibration is known to vary with sensor age, temperature, and in some cases humidity. In addition, many of the sensors have responses to gases other than the target analyte (Mead et al., 2013; Spinelle et al., 2015, 2017; Cross et al., 2017; Mueller et al., 2017; Mijling et al., 2018; Zimmerman et al., 2018). One approach to addressing this challenge is to combine periodic re-calibration and co-location with regulatory reference instruments in the lab or the field (Williams et al., 2013; Moltchanov et al., 2015; Jiao et al., 2016; Mijling et al., 2018). Field calibration is preferred as in-lab performance is often a poor approximation of sensor behavior under ambient conditions (Piedrahita et al., 2014; Masson et al., 2015). However, either method requires considerable time investment by trained personnel, especially as the number of sensors increases. The requirement of time-consuming and labor-intensive calibration then offsets the low-cost advantage of the sensors.

In this paper, we explore an automated, in situ strategy for the calibration of individual sensors embedded in an air quality sensor network that includes both low-cost sensors and anchor points of higher-grade, well-calibrated instrumentation. The BErkeley Atmospheric CO2 Observation Network (BEACO2N) is a low-cost, high-density greenhouse gas (CO2) and air quality (CO, NO, NO2, O3, and particulate matter) monitoring network located in San Francisco Bay Area, California (see Fig. 1 and Shusterman et al., 2016). As of this writing, BEACO2N consists of approximately 50 sensor “nodes”, deployed with approximately 2 km horizontal spacing. Most of the nodes are mounted on the roofs of schools and museums. In previous work, we described an approach to CO2 sensing and calibration (Shusterman et al., 2016). Here, we focus on CO, NO, NO2, and O3.

We begin by describing laboratory experiments and in-field comparisons to co-located reference instruments that give an initial characterization of the sensors and provide insight into the effects of temperature, humidity, and cross-sensitivity to non-target analytes. Then we describe an in situ calibration procedure that accounts for these variables without requiring co-location with a reference instrument. The calibration procedure is finally verified against regulatory quality measurements not used in the procedure itself.

Details of the node design and deployment are described in Shusterman et al. (2016). Briefly, each BEACO2N node contains a Vaisala CarboCap GMP343 non-dispersive infrared sensor for CO2, a Shinyei PPD42NS nephelometric particulate matter sensor, and a suite of Alphasense electrochemical sensors: CO-B4, NO-B4, either NO2-B42F or NO2-B43F, and either Ox-B421 or Ox-B431. All sensors are assembled into compact, weatherproof enclosures as shown in Fig. 2. Two 30 mm fans are located on either side of the enclosure to facilitate airflow through the node. A Raspberry Pi microprocessor collects data via a serial-to-USB converter for CO2 and an Adafruit Metro Mini microcontroller for all other sensors. Then, data collected every 5 or 10 s are transmitted to a central server using a direct on-site Ethernet connection or a local Wi-Fi network.

The Alphasense B4 electrochemical gas sensing series that we use employs a four-electrode approach. The electrodes are embedded in an electrolyte solution separated from the atmosphere by a semi-permeable membrane. The gas of interest diffuses through the membrane into the electrolyte, where it contacts a “working” electrode and is either oxidized (in the case of NO and CO) or reduced (NO2 and O3). The potential at the working electrode is maintained at a constant value with respect to a “reference” electrode. Electric charge produced at the working electrode is balanced by the complementary redox reaction at a “counter” electrode, generating an electric current. The sensor also contains an “auxiliary” electrode, which shares the working electrode's catalyst structure, but is isolated from the ambient environment, accounting for fluctuations in the background current associated with other processes at the electrode and electrolyte. Subtracting the auxiliary current from the working current gives a corrected current dependent on the gas concentration.

The working and auxiliary currents detected by the sensors are converted to working and auxiliary voltages using amplifiers in the individual sensor boards (ISBs) provided by Alphasense. Over the mixing ratio range of interest, the sensors' responses to the gases of interest are approximately linear. We derive mixing ratios from the observed voltages by subtracting an offset and then scaling by a constant (Eqs. 1–4): COambient=(VCO-zeroCO)/kCO,NOambient=(VNO-zeroNO)/kNO, NO2ambient=(VNO2-zeroNO2)/kNO2-(rNO-NO2×NOambient),NO2ambient=(VNO2-zeroNO2)/kNO2+(rCO2-NO2×CO2ambient), O3ambient=(VO3-zeroO3)/kO3-(rNO2-O3×NO2ambient). Here, CO, NO, NO2, and O3 with the subscript “ambient” refer to the gas mixing ratios (ppb) in air; VCO, VNO, VNO2 and VO3 are the signals (mV) measured by each sensor, which is the voltage of the auxiliary electrode subtracted from the voltage of the working electrode; zeroCO, zeroNO, zeroNO2 and zeroO3 indicates the voltage measured in the absence of analyte; and kCO, kNO, kNO2 and kO3 represent the linear sensitivity factor that converts mV to ppb. Additional terms corresponding to the cross-sensitivities of the NO2 and O3 sensors appear in Eqs. (3a), (3b), and (4), where rNO-NO2 is the cross-sensitivity of the NO2-B42F sensor to NO gas, rCO2-NO2 is the cross-sensitivity of the NO2-B43F sensor to CO2 gas, and rNO2-O3 is the cross-sensitivity of both the O3-B421 and O3-B431 sensors to NO2 gas.

There are a total of eight sensitivities and zero offsets, as well as two cross-sensitivity terms. All of these may also vary with time, temperature, and humidity. Thus we need a calibration strategy that constrains 10 parameters in a single instant as well as the variation of those 10 parameters in response to the environmental variables and time. We begin by characterizing the sensors in both laboratory and outdoor environments.

We evaluate BEACO2N in terms of four factors: drift, noise, cross-sensitivity, and temperature dependence. The humidity dependence is included in the temperature dependence, as there is no evidence for independent humidity dependence and relative humidity exhibits an anti-correlation with temperature in the field. In the laboratory, a range of mixing ratios of target gases were delivered to a chamber containing the full suite of four Alphasense B4 sensors: CO, NO, NO2, and O3. Zero air was supplied by a Sabio 1001 compressed zero-air source and blended with calibration gases using a Thermo Scientific model 146i multi-gas calibrator.

Noise. Alphasense reports 2σ noise of ±4, ±15, ±12, and ±15 ppb for CO, NO, NO2, and O3, respectively over concentrations from 0 to 200 ppb at time resolution of a second. In our laboratory, noise (±2σ) was measured for ambient ppb levels with 10 s time resolution and was seen to be ±10 ppb for CO, ±3 ppb for NO, ±6 ppb for NO2 (NO2-B42F and NO2-B43F), and ±12 ppb for O3 (O3-B421 and O3-B431).

Cross-sensitivity. We measured the cross-sensitivity of all four of the trace gas sensors to the non-target gases. The NO2 sensors and O3 sensors were the only ones to exhibit sensitivity to other species. The O3 sensor (O3-B421 and O3-B431) demonstrated 100 % sensitivity to NO2. This sensor is now being marketed by Alphasense as an odd-oxygen (Ox≡O3+NO2) sensor. In addition, the NO2-B42F sensor was found to possess a significant NO sensitivity (130 %) that exceeds the cross-sensitivity specified in the Alphasense documentation (< 5 %). The NO2-B43F sensor was found to have 0.002 % sensitivity to CO2 gas, which is in the range of the cross-sensitivity specified in the Alphasense documentation (< 0.1 %). However, given that typical ambient CO2 concentrations are 4 orders of magnitude larger than NO2 concentrations, this relatively small cross-sensitivity to CO2 gas manifests as a significant interference in the NO2 sensors. These cross-sensitivities are represented in Eqs. (3) and (4).

Temperature dependence. Electrochemical sensors are known to have temperature-dependent sensitivities and zero offsets. Alphasense reports sensitivities and zero offsets for a temperature range between -30 and 50 ∘C. The sensitivities in their data sheets vary with temperature by +0.1 to +0.3 % K-1 (referenced to sensitivity at 20 ∘C) and the zero offsets are indicated to vary little except at high temperatures. We observed similar but slightly larger variations via in situ comparison to co-located reference instruments. We observed temperature dependence in the sensitivities of +0.3 to +5 % K-1 and no variation in the zero offset of the CO, NO2, and O3 sensors from 10 to 24 ∘C (Fig. 3). However, the zero offset of the NO sensor exhibited a strong temperature dependence of 0.34 mV K-1.

Drift. Two laboratory calibrations were performed roughly 10 weeks apart and the zero offsets and sensitivities are shown in Table 1. Over the 10-week interval, zero drift was equivalent to -15.9, -2.3, +15.8, and -12.7 ppb for CO, NO, NO2, and O3, respectively. Alphasense reports the stability over time for the zero offset to be < ±100, 0 to 50, 0 to 20, and 0 to 20 ppb yr-1 for these sensors, respectively; over this 10-week interval, the observed zero drift was within the range of these specifications. However, it is a large fraction of the annual drift specification and further experiments would be warranted to test whether the zero measured is stable over a full year within the specified tolerances. The drift in the sensitivity (in % of kX) was -15.9, -17.7, -20.6, and -53.2 %. Alphasense reports < 10, 0 to -20, -20 to -40, and < -20 to -40 % yr-1 for CO, NO, NO2, and O3 calibration factors, respectively. We find that drift for the CO and O3 sensitivities exceeded the manufacturer specifications, but that the NO and NO2 sensitivity drifts were within the specified tolerances.

Here, we propose a model for field calibration that leverages (1) useful cross-sensitivities, (2) chemical conservation equations, (3) knowledge of the global and/or regional background of pollutants, and (4) assumptions based on well-known characteristics of urban air quality and local emissions. The result is a calibration procedure for the drift and temperature dependencies of the 10 calibration parameters that does not require co-location with a reference instrument or prior laboratory experiments for each sensor. The first constraint we apply is the O3 sensors' cross-sensitivity to NO2. Laboratory measurements indicate that this cross-sensitivity is 100 % and we fix it at that value.

We evaluate the efficacy of our calibration method using a BEACO2N node co-located with reference instruments at the Laney College monitoring site maintained by the Bay Area Air Quality Management District (BAAQMD). Here we consider data collected from February to April 2016, calibrate them according to the procedure described above (following Sect. 3.1 to 3.5), and compare it against the BAAQMD data. Reference data are collected by a TECO 48i CO analyzer and a TECO 42i NOx analyzer. Ozone data from the “Oakland West” location, the closest ozone-monitoring site maintained by BAAQMD, were used for multiple linear regression of Eq. (5). The zero offset for CO was calculated using BAAQMD data from the Bodega Bay background site (see Fig. 1; Guha et al., 2016) as local “supersite” data were unavailable during this period. A background site closer to the network would likely improve our ability to constrain the CO zero offset; a reference instrument for that purpose was installed in summer 2017.

In our calibration procedure, the cross-sensitivities and temperature dependence are corrected for better accuracy. Table 3 shows the reduction in mean absolute error (MAE) that results when cross-sensitivity and temperature dependence issues are considered during multiple linear regression of Eq. (5). Here, MAE is calculated after conducting the sensitivity correction explained in Sect. 3.1, but before the offset correction in Sect. 3.3. Fully calibrated, hourly averaged BEACO2N sensor data are compared to reference data in Fig. 6. For NO, NO2, O3, and CO the mixing ratio measured agrees reasonably well with the reference instrument with correlation coefficients of 0.88, 0.61, 0.69, and 0.74 and MAE of 3.63, 4.12, 5.04, and 54.93 ppb, respectively. The noise (±2σ) in the differences between the calibrated hourly BEACO2N data and reference data is 9.74 ppb for NO, 9.97 ppb for NO2, 13.04 ppb for O3, and 116.23 ppb for CO. These noise values are dominated by the Alphasense noise except in the case of CO, where noise is evenly split between the low-cost electrochemical sensors and the reference instruments.

Figure 7 shows a week-long time series of fully calibrated air quality data from four BEACO2N sites in 2017 (see Fig. 1). BEACO2N nodes capture the short-term variability associated with local emissions, superimposed on the diurnal variation caused by mixing and changes in the height of the boundary layer. Large mixing ratios of NO, NO2, and O3 are observed at the Hercules and Ohlone sites, likely representing strong NOx emissions from an oil refinery nearby. The spatial variability of trace gases observed at these four BEACO2N sites provides a more diverse perspective on emissions compared to that provided by the one regulatory monitoring site in the vicinity.

The emission ratios of CO and NOx were also investigated using the BEACO2N data from sample locations. Figure 8 shows ratios observed at the Laney College site. The slope of CO / NOx varies from 4.43 to 12.99 across five BEACO2N sites, reflecting spatial variations in local sources. Sites near roads with more diesel vehicles, such as Laney College, show lower CO / NOx ratios, as expected given diesel vehicles' higher NOx emissions. The range of observed CO / NOx emission ratios is similar to the values reported by McDonald et al. (2013).

Calibration of low-cost sensors is necessary for quantitative analysis. In this paper, we have described a truly low cost, routine in-field calibration method and the evaluation of a fully calibrated low-cost, high-density air quality sensor network. The Alphasense B4 electrochemical gas sensors are able to detect typical diurnal cycles in gas concentrations as well as short-term changes corresponding to chemical reactions and local emissions. These capabilities of the sensors are utilized for a field calibration protocol that does not require co-location with reference instrumentation, but does require reference instruments to be sited within the network domain. The calibrated dataset demonstrates the accuracy required to resolve information relevant to urban emission sources, such as CO / NOx emission ratios. Through this work, we can realize the promise of low-cost, high-density sensor networks as a viable approach for atmospheric monitoring.