Abstract View
Performance of Correction Models for Accurate PM2.5 Estimation from Low-Cost Air Quality Sensor Data
DINUSHANI SENARATHNA, Vijay Kumar, Shantanu Sur, Suresh Dhaniyala, Supraja Gurajala, Sumona Mondal, Clarkson University, Potsdam
Abstract Number: 291
Working Group: Aerosol Standards
Abstract
Environmental epidemiology requires accurate estimates of pollutant levels in ambient air to evaluate the health effects from long-term exposure. Among these pollutants, particulate matter smaller than 2.5 µm in diameter (PM2.5) has been shown to be closely associated with various acute and chronic health problems. In the United States, Environmental Protection Agency (EPA) provides reliable measurements of PM2.5 but the impact is limited by a sparse distribution of monitoring sites. Air quality measurements using low-cost sensors such as PurpleAir (PA) are currently being used to overcome this limitation. However, measurements from these sensors could be noisy, and an appropriate correction model will be required to get an accurate estimate of PM2.5. In this work, we aim to evaluate and improve PA sensor-derived PM2.5 using EPA data as a gold standard and determine the optimal distance between EPA and PA sites from a calibration perspective. We collected hourly PM2.5 measurements from 14 PA sensors located in the Cook County of Chicago from August 2019 to July 2020 and developed a modeling approach that utilizes multiple regression techniques together with distance analysis. Interestingly, while considerable improvement was observed when multiple PA sensors were used in the model instead of a single PA sensor, the performance of the correction model did not depend on the relative distance of the PA sensors from the EPA sites. Additionally, our results suggest that temperature is a more influential (R2 = 0.60 ~ 0.80) factor for the correction models compared with relative humidity (R2 = 0.50 ~ 0.60). These correction models along with data pre-processing techniques implemented in this work could help extract accurate air quality prediction from the low-cost sensor networks.