Development of a Network of Accurate Ozone Sensing Nodes for Parallel Monitoring in a Site Relocation Study

Recent technological advances in both air sensing technology and Internet of Things (IoT) connectivity have enabled the development and deployment of remote monitoring networks of air quality sensors. The compact size and low power requirements of both sensors and IoT data loggers allow for the development of remote sensing nodes with power and connectivity versatility. With these technological advancements, sensor networks can be developed and deployed for various ambient air monitoring applications. This paper describes the development and deployment of a monitoring network of accurate ozone (O3) sensor nodes to provide parallel monitoring in an air monitoring site relocation study. The reference O3 analyzer at the station along with a network of three O3 sensing nodes was used to evaluate the spatial and temporal variability of O3 across four Southern California communities in the San Bernardino Mountains which are currently represented by a single reference station in Crestline, CA. The motivation for developing and deploying the sensor network in the region was that the single reference station potentially needed to be relocated due to uncertainty that the lease agreement would be renewed. With the implication of siting a new reference station that is also a high O3 site, the project required the development of an accurate and precise sensing node for establishing a parallel monitoring network at potential relocation sites. The deployment methodology included a pre-deployment co-location calibration to the reference analyzer at the air monitoring station with post-deployment co-location results indicating a mean absolute error (MAE) < 2 ppb for 1-h mean O3 concentrations. Ordinary least squares regression statistics between reference and sensor nodes during post-deployment co-location testing indicate that the nodes are accurate and highly correlated to reference instrumentation with R2 values > 0.98, slope offsets < 0.02, and intercept offsets < 0.6 for hourly O3 concentrations with a mean concentration value of 39.7 ± 16.5 ppb and a maximum 1-h value of 94 ppb. Spatial variability for diurnal O3 trends was found between locations within 5 km of each other with spatial variability between sites more pronounced during nighttime hours. The parallel monitoring was successful in providing the data to develop a relocation strategy with only one relocation site providing a 95% confidence that concentrations would be higher there than at the current site.


Calculations
Mean Bias Error (MBE) and Mean Absolute Error (MAE) calculate the measurement error by examining the hourly differences between the 2B POM and the Crestline AMS Thermo 49i O3measurement. The MBE between the sensor and the reference O3 instrument provides a metric that indicates the tendency of the sensor to either under-or over-estimate the reference O3 concentrations during the pre-and post-deployment collocation periods. The units of both MBE and MAE are calculated in ppb, which is identical to the units of measurement. Care must be taken with the MBE statistic, as over-estimated errors will cancel out under-estimated errors in the calculation of the bias error. The MAE provides a better metric for actual measurement error between sensor and reference. The equations for MBE and MAE are found in equation S1 and S2, respectively.

Equation S2
Mean Absolute Error where, Xi is the 1-hr average measurement provided by the low-cost sensor Xt is the 1-hr average measurement provided by the Crestline Thermo 49i n is the number of 1-hr time-matched data pairs Mean Bias Deviation (MBD) and Mean Absolute Deviation (MAD) calculate the deviation in O3 concentrations between deployment locations by examining the hourly differences between the POM unit and the Crestline AMS Thermo 49i O3measurements. The MBD between the POM and Crestline provides a metric that indicates the tendency of a relocation site to either under-or over-estimate O3concentrations when compared to the Crestline AMS. The equations for MBD and MAD are found in equation S3 and S4, respectively.

Equation S4
Mean Absolute Deviation where, Xi is the 1-hr average measurement provided by the POM Xt is the 1-hr average measurement provided by the Crestline Thermo 49i n is the number of 1-hr time-matched data pairs

Hampel Filter
The Hampel Filter function applies a filter along a rolling sample of an input vector. The median and standard deviation of the median are computed for the rolling sample window. Points that exceed a set threshold for the standard deviation of the median of the rolling sample window are characterized as outliers and replaced with the rolling median value. The sample window consists of ten data points which would constitute 10 min of data. If a sample value differed from the rolling median by more than six standard deviations, the sample value was replaced with the median value for the rolling window.

% Confidence Interval Calculations
These calculations are adapted from the California Air Resources Board's (CARB's) Air Monitoring Technical Advisory Committee (AMTAC) document that provides guidelines for site relocation and Parallel Monitoring.

Equation S6
Upper Limit of 95% CI (U) = ̅ + ( * √ ), where, d ̅ = Mean Bias Deviation (Eq. S3) s = standard deviation of the MBD n = number of matching pairs df = degree of freedom = n-1 t = T score found in from at T-score distribution table, specific to degree of freedom and confidence interval level of certainty