Gaussian Process Filtering for Calibration of Low-Cost Air-Pollution Sensor Network Data

• Activity Number: 561 - Statistical Analyses for Environmental Monitoring
• Type: Contributed
• Date/Time: Thursday, August 11, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Claire Monique Heffernan* and Abhi Datta and Roger Peng and Kirsten Koehler and Drew R. Gentner
• Companies: Johns Hopkins University and Johns Hopkins University and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins University and Yale University
• Keywords: spatial statistics; Gaussian Process; Bayesian statistics; air pollution; low-cost sensors

Abstract

Low-cost air pollution sensors, offering hyper-local characterization of pollutant concentrations, are becoming increasingly prevalent in environmental and public health research. However, low-cost data can be noisy and biased, and often need to be field-calibrated by co-locating low-cost sensors with reference-grade instruments. We show that the common procedure of regression-based calibration using co-located data underestimates high air-pollution concentrations, and fails to utilize the spatial correlation in concentrations. We propose a novel spatial filtering approach to co-location-based calibration of low-cost networks that mitigates the underestimation issue by using an inverse regression and incorporates spatial correlation by second-stage modeling of the true concentrations using a conditional Gaussian Process. Our approach works with one or more co-located sites in the network and is dynamic, leveraging spatial correlation with the latest available reference data. Through extensive simulations and an application on a low-cost PM2.5 network in Baltimore, we demonstrate how spatial filtering improves estimation of pollutant concentrations, especially peak concentrations.