Evaluating uncertainty in sensor networks for urban air pollution insights

Abstract

Ambient air pollution poses a major global public health risk. Lower-cost air quality sensors (LCSs) are increasingly being explored as a tool to understand local air pollution problems and develop effective solutions. A barrier to LCS adoption is potentially larger measurement uncertainty compared to reference measurement technology. The technical performance of various LCSs has been tested in laboratory and field environments, and a growing body of literature on uses of LCSs primarily focuses on proof-of-concept deployments. However, few studies have demonstrated the implications of LCS measurement uncertainties on a sensor network's ability to assess spatiotemporal patterns of local air pollution. Here, we present results from a 2-year deployment of 100 stationary electrochemical nitrogen dioxide (NO2) LCSs across Greater London as part of the Breathe London pilot project (BL). We evaluated sensor performance using collocations with reference instruments, estimating ∼ 35 % average uncertainty (root mean square error) in the calibrated LCSs, and identified infrequent, multi-week periods of poorer performance and high bias during summer months. We analyzed BL data to generate insights about London's air pollution, including long-term concentration trends, diurnal and day-of-week patterns, and profiles of elevated concentrations during regional pollution episodes. These findings were validated against measurements from an extensive reference network, demonstrating the BL network's ability to generate robust information about London's air pollution. In cases where the BL network did not effectively capture features that the reference network measured, ongoing collocations of representative sensors often provided evidence of irregularities in sensor performance, demonstrating how, in the absence of an extensive reference network, project-long collocations could enable characterization and mitigation of network-wide sensor uncertainties. The conclusions are restricted to the specific sensors used for this study, but the results give direction to LCS users by demonstrating the kinds of air pollution insights possible from LCS networks and provide a blueprint for future LCS projects to manage and evaluate uncertainties when collecting, analyzing, and interpreting data.