Mapping pollution exposure and chemistry during an extreme air quality event (the 2018 Kīlauea eruption) using a low-cost sensor network

Significance Poor air quality is a global public health issue, contributing to millions of premature deaths per year worldwide. Low-cost air quality sensors are a promising tool to improve monitoring capabilities. In this study, we built and deployed a low-cost sensor network for emergency response during an extreme air quality event, the 2018 Kīlauea Lower East Rift Zone eruption. This network was used to estimate fine-scale population exposures to multiple pollutants, to measure the chemical transformation of volcanic emissions, and to provide real-time observations as part of emergency management efforts.


Supplementary Information Text Materials
A schematic of the multi-pollutant air quality sensor (MPAQS) used in this study is shown in Fig. S1; the design is based on previous sensor nodes from our group (1). Each MPAQS node contains an electrochemical cell for measuring SO2 (Alphasense SO2-B4, Alphasense Ltd), an optical particle counter (OPC) to measure PM (Alphasense OPC-N2), and one RH+TA sensor (SHT3x-ARP, Sensirion). The sensors and electronics are contained in a waterproof polycarbonate enclosure (218.19 x 167.39 x 129.79 mm, YH-080604, Polycase Inc). Each node is powered by photovoltaic panels (2 x 9 W) and rechargeable battery (44 Wh) (Voltaic Systems Inc.). Sensors, electronics, and communications are integrated by a custom printed circuit board and the node is controlled by custom software on a micro-controller (Electron, Particle Inc.).
In addition to the MPAQS nodes, there were six prototype SO2-only nodes (see 1 for full details) and five prototype PM-only nodes in the network. These SO2-only nodes were used in the initial deployment near the LERZ eruption. These additional nodes also measure TA and RH and use the same power and communications equipment as the MPAQs. The SO2 sensors in all nodes throughout the network are identical while the PM-only nodes use a different sensor (PMS5003, Plantower) than the MPAQS. Sensor nodes were deployed to 33 locations across the Island (Table S1).

Sulfur dioxide measurements and calibration
The SO2 sensors (Alphasense SO2-B4) are electrochemical devices in which a redox reaction generates a measurable electrical signal in proportion to the amount of SO2 present in the atmosphere. In the MPAQS node design, the sensors are exposed directly to ambient atmosphere and passively ventilated through apertures in the node enclosure. For the prototype SO2-only nodes, ambient air is actively drawn through a flow-tube and across sensors by a small fan (see 1 for full design details).
The electrochemical SO2 sensors' electrical output signal is sensitive to ambient air temperature (TA) in addition to SO2. Sensors are calibrated by algorithm based on sensorspecific i) linear sensitivity to SO2 concentration and ii) non-linear baseline response to TA: where ppb is SO2 mixing ratio in parts per billion by volume, V is measured voltage (mV) from the 'working electrode' of the sensor,  is the temperature-dependent baseline, and  is instrument sensitivity (mV ppb -1 ). The baseline TA exponential function parameters (a, b, c) for each sensor are determined using observed variations in sensor voltage with TA over a training period of several days (Fig. S2). Factory-determined  (~1 mV per 3 ppb) for each sensor was confirmed before deployment using a laboratory calibration chamber at constant temperature.
The calibration procedure was evaluated using co-located measurements with HDOH reference instruments during a 48-hour period at the Konawaena HDOH station (July 7-9, 2018, SO2 values range from 0 -50 ppb) and a 24-hour period at the Pahala HDOH station (May 27-28, 2018 SO2 values range from 0-800 ppb). The reference instrument at each site is a regularly-calibrated UV fluorescence monitor (Teledyne 100E) and reference data are available online as 5-minute averages. Overall, average uncertainty for all electrochemical sensors is 7.3 ppb (average mean absolute error, MAE, of 5-minute values during the calibration period) with individual sensors' MAE ranging from 2.3-13.6 ppb (Table S2). Based on earlier measurements, sensor drift is negligible over the span of several months (1). For quality control, data are flagged and withheld from analysis for three hours after the sensor is powered on to allow the sensor to warm-up. Additionally, a standard deviation filter (±5 times s.d.) is applied to the 1-minute time series to screen any unrealistic short-term spikes (0.6% of (1) 1 min values). The 1-minute SO2 dataset from July 15 -August 1 used for this analysis is shown in Fig. S3.

Particulate matter measurements and calibration
For PM, ambient air is continuously drawn to the sensor through an aperture in the node enclosure via a 7 cm length of conductive tubing. The MPAQS PM sensors are laser optical particle counters (OPCs) in which laser scattering by airborne particles provides information on particle number and size distributions. The sensor uses a 658 nm laser and provides particle counts in 16 size bins (0.38-17 m). Mass concentrations (PM1, PM2.5, and PM10) are calculated from binned counts using assumed particle shape (spherical), refractive index (1.5), and particle density (1.65 g mL -1 ). A different type of low-cost PM sensor (a nephelometer) was used in the five PM-only nodes. The smallest particle size measured by each sensor is 380 nm (OPC) and 300 nm (nephelometer), according to manufacturer specifications.
The PM measurements from both sensors are sensitive to relative humidity (though nephelometer and OPC RH-responses are not identical (2)), due to water uptake by the highly hygroscopic sulfate particles (Fig. S4). To account for this, an RH correction based on -Köhler theory (3) was applied to enable direct comparison with reference instruments. The reference instruments use a beta-ray attenuation technique (Beta-ray Attenuation Monitor, Met-One Instruments) and ambient air is dried to 30% RH prior to analysis (hourly data, available online). Hygroscopic growth of sulfuric acid particles can occur at lower RH levels, however, there was good agreement with reference measurements in this environment during relatively dry conditions (RH<50%) and application of the RH correction below this threshold resulted in poorer agreement and so was not applied.
The correction factor (f) assumes an average bulk particle density ( ) and hygroscopic growth factor (): In this environment,  and  values representative of sulfuric acid (=1.84, =1.19) were used (4). As a sensitivity test, ammonium sulfate values (=1.77, =0.53) were also applied and the resulting difference to mean PM2.5 at the Konawaena reference station is minimal (MAE values differ by 0.7 g m -3 , relative to reference).
Overall, there is good inter-instrument agreement between nephelometers (mean PM2.5 values of individual sensors are within 0.5 g m -3 ), while variation between individual OPCs is greater (differences up to 5 g m -3 ). To minimize inter-instrument differences, a secondary linear calibration factor is applied to the LCS measurements based on comparison to reference instruments. The average adjustment to mean PM2.5 from the linear calibration factor is 5.9 g m -3 and measurements from the two sensor types (OPC and nephelometers) are statistically indistinguishable after the RH correction and linear calibrations have been applied. Overall, the average uncertainty of hourly PM2.5 values for all sensors, relative to a reference instrument, is 4.5 g m -3 (Table S2). This magnitude of uncertainty is comparable to that of the reference instrument itself in ambient conditions (5).
Additionally, the one-minute PM timeseries are subject to a standard deviation quality control filter (±5 times s.d.) to screen any unrealistic short-term spikes (0.5% of 1-minute values). The 1-minute PM2.5 dataset from July 15 -August 1 used for this analysis is shown in Fig. S5.

Population data and analysis
Overall, approximately 61% of the Island's population (106,211 people) lived within 5 km of an AQ monitoring station in operation at some point during the eruption. Of this population, ~10,000 lived within 5 km of multiple stations and an additional 16,442 people live within 5 km of a PM monitor only, but not SO2. Of the 87,400 remaining people, a majority (56%) lived upwind of the eruption and were largely unexposed to vog during the 2018 eruption (Table S4).

(3)
The population near each network node's location is determined in a GIS using circular buffers of varying radii from 0.5-10 km ( Figure S6B, Tables S3-S4). To determine population coverage for the entire network, the individual buffer populations for each node are summed and any overlapping buffers are merged to avoid double-counting.
The average residents' distance (d) to an air quality monitoring station is the populationweighted straight-line minimum distance of each populated grid-cell (di) to the nearest AQ station: where pi is the individual grid-cell population and ptot is total population.
Additionally, population-weighted mean concentration [cp] are calculated as the populationweighted mean of SO2 (ppb) and PM2.5 (g m -3 ) across multiple nodes (Figs. S7-S8): An hourly timeseries of accumulated population exposure (persons x hourly concentration) from both the LCS and regulatory networks are shown in Fig. S9. On average, the accumulated regulatory network total is 61% of the LCS network for SO2 and 29% for PM2.5. The difference in pollutant distributions across the population between networks is also illustrated in Fig. S10. This figure shows pollutant exposure distributions based on regulatory network observations (5 stations, with ~28,000 people total living within 5 km of the stations) using the same approach as in Figure 3 of the main text. For this application, simulated air parcels (about one parcel per minute) were tracked backwards in time with the model for 60 hours or until they left the model domain. The HYSPLIT time-reversed runs were initialized hourly and output minutely to characterize the average path the air parcels had taken to arrive at each of the twelve stations ( Figure S12). The air parcels that could be traced backwards to the LERZ and summit eruption coordinates (the dominant source of SO2 during the eruption) were used to estimate the age of the volcanic plume. The plume ages are calculated by subtracting the start time from the time the particle (air parcel) arrives at the source. To simulate stochastic processes in the atmosphere, a random component is added to the advection component by the mean flow. During the study period, winds were predominantly from the northeast and the majority of parcels remained within the subtropical boundary layer (<~1800 m) by the trade wind inversion. During daytime, vog was carried onshore and upslope from localized sea-breeze circulations due to solar heating of the land surface.

Meteorological model
Additionally, downwind of the Island in the lee of the prevailing flow, there is a large persistent clockwise eddy circulation (7) that transports vog first northward and then eastward towards the western coast. This flow pattern results in similar plume ages (20.6-21.1 hours) at the five stations along the western coast as the dispersed plume was carried ashore ( Figure S12).    (Table S2) for July 15 -August 1, 2018.   (Table  S2) for July 15 -August 1, 2018.   . Each unique color represents a measurement location (see map insets) and color bands are stacked so that the y-axis is the cumulative total across each respective network. The difference between networks is due to different numbers of people within a 5 km radius of network stations. Population-weighted averages, which account for this difference, are shown in Fig. S8.

Fig S10.
Population exposure to volcanic pollutants, measured by the regulatory network (DOH) over the 15-day study period (equivalent to Figure 3, but for the regulatory network rather than the LCS network). For reference, the LCS network distribution from Figure 3 of the main text is also shown as gray bars. Panels (A-B): Mean pollutant distribution as a function of cumulative near-node population (residents living within 5 km of each node). Bar width is proportional to nearby population and bar height is the average pollutant concentration measured by each node. Sensor nodes are differentiated by color, as shown on the inset map. Stations are arranged from lowest to highest average concentration. Panels (C-D): Population distribution as a function of hourly exposure frequency to SO2 and PM2.5. Here, the distribution of hourly concentrations experienced by each sensor node is weighted by population within 5 km of the node and arranged by average concentration.   (Table S2). Plume age uncertainties (horizontal error bars) are the interquartile range of calculated parcel travel times between the LERZ and measurement location during each hour of the study period. To incorporate these uncertainties into confidence (dashed lines) and prediction (gray shading) intervals (95%), the model is fit to an array of random points uniformly sampled with the uncertainty bounds at each measurement point (n=10 points at each location, total n=130).  Table S3. Population (%) living within various radii of low-cost sensor (LCS, see Table S1) and Hawaii Department of Health (HDOH) networks at any point during the eruption. Spatial buffers are merged so that residents living near multiple stations are not counted twice.