Speaker
Description
There is a severe lack of air pollution data around the world. Low cost sensors (LCS) for measuring air pollution offer a possible path forward to remedy the lack of data, though they require careful calibration as the manufacturer-reported data often show large biases against reference monitors. Traditionally, calibration has occurred by co-locating LCS with local reference monitors and developing correction models using statistical techniques. To address this, we first present a series of traditional “colocation” corrections models and performance evaluations in a variety of global cities, including New York City, Kinshasa (DRC), Kampala (Uganda), Accra (Ghana), and New Delhi (India). We then present a globally-applicable Gaussian Mixture Regression (GMR) probabilistic model trained on co-locations from at least 5 global cities that span across wide temperature, relative humidity, and PM2.5 ranges. The model is tested on at least 20 independent co-location datasets that the GMR has not seen. We also compare the data corrected by the universal GMR to a more traditional, local correction factor, where available. GMR has proven successful for correcting LCS data: in Kinshasa, the GMR-corrected PurpleAir data resulted in R2 = 0.88 when compared to the MetOne BAM-1020, and in Accra, the GMR lowered the Mean Absolute Error of Clarity data from 7.51 𝜇g/m3 to 1.93 𝜇g/m3. The wide breadth of the universal GMR allows for correction of LCS data without the need for a local co-location which enables the correction of data from 10,000+ PurpleAir sensors around the world.
