Leveraging machine learning algorithms to advance low-cost air sensor calibration in stationary and mobile settings

Abstract

Low-cost air sensing is changing the paradigm of ambient air quality management research and practices. However, consensus on a structured low-cost sensor calibration and performance evaluation framework is lacking. Our study aims to devise a standardized low-cost sensor calibration protocol and evaluate the performance of various calibration algorithms. Extensive collocation data were collected in stationary and mobile settings in two American cities, New York and Boston. We trained the calibration models using stationary data aggregated at various intervals to examine the performance of several commonly used calibration algorithms described in the literature. Linear models provide consistently satisfactory calibration results, indicating linear responses from the low-cost sensors in our stationary test environment. Its simplicity is recommended for citizen science and education usages. Models that can account for non-linear relationships, especially random forest, perform well and transfer between sensors better than generalized linear regression models for PM2.5 calibration, which should be adopted for regulatory and scientific purposes. Data collected in a mobile validation campaign in Boston were passed through the best-performing calibration models to assess their transferability. The results indicate that models trained with data from a different urban environment and season in the stationary setting did not transfer well to a mobile setting. It is recommended that low-cost sensors should be calibrated more often than suggested in Environmental Protection Agency's air sensor performance evaluation guidelines and used in an environment that is as similar as possible to the calibration environment.