Calibrating networks of low-cost air quality sensors

Ambient fine particulate matter (PM2.5) pollution is a major health risk. Networks of low-cost sensors (LCS) are increasingly being used to understand local-scale air pollution variation. However, measurements from LCS have uncertainties that can act as a potential barrier to effective decision making. LCS data thus need adequate calibration to obtain good quality PM2.5 estimates. In order to develop calibration factors, one or more LCS are typically co-located with reference monitors for short or long periods of time. A calibration model is then developed that characterizes the relationships between the raw output of the LCS and measurements from the reference monitors. This calibration model is then typically transferred from the co-located sensors to other sensors in the network. Calibration models tend to be evaluated based on their performance only at co-location sites. It is often implicitly assumed that the conditions at the relatively sparse co-location sites are representative of the LCS network overall and that the calibration model developed is not overfitted to the co-location sites. Little work has explicitly evaluated how transferable calibration models developed at co-location sites are to the rest of an LCS network, even after appropriate cross-validation. Further, few studies have evaluated the sensitivity of key LCS use cases, such as hotspot detection, to the calibration model applied. Finally, there has been a dearth of research on how the duration of co-location (short-term or long-term) can impact these results. This paper attempts to fill these gaps using data from a dense network of LCS monitors in Denver deployed through the city's "Love My Air" program. It offers a series of transferability metrics for calibration models that can be used in other LCS networks and some suggestions as to which calibration model would be most useful for achieving different end goals.