HIGH-RESOLUTION MACHINE LEARNING-BASED CALIBRATION OF A LOW-COST PARTICULATE MATTER SENSOR VIA INCORPORATION OF ENVIRONMENTAL PARAMETERS

Abstract

Low-cost sensors for air quality monitoring are gaining widespread attention due to their ease of operation, affordability, and ability to provide high-resolution data across both spatial and temporal scales. However, their performance remains sub-par compared to conventional monitoring methods and is vulnerable to environmental and meteorological parameters such as temperature (T) and relative humidity (RH). In this study, PM2.5 measurements from an optical low-cost particulate matter sensor are calibrated using a higher-grade optical PM2.5 sensor as a secondary reference. Three calibration models—Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Multilinear Regression (MLR)—were deployed, and their performance was compared, including the influence of environmental parameters T and RH. Results show that ANN models performed as well as or better than the MLR model. In general, ANN outperformed the base linear regression model, suggesting that with additional data, further model refinement, and hyperparameter optimization, deep learning methods could potentially enhance the performance of low-cost environmental sensors.