Field calibration of fine particulate matter low-cost sensors in a highly industrialized semi-arid conurbation

Abstract

Low-cost sensors (LCS) for suspended particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5) have attracted worldwide attention for crowdsourcing air quality data. Here, we analyze one year’s worth of PM2.5 data from light-scattering LCS deployed in Monterrey, Mexico, one of the most polluted conurbations of Latin America. We also tested the Extreme Gradient Boosting (XGBoost) algorithm for classification and field calibration of the PM2.5 data derived from the LCS. Regression model performance increased from a low baseline (compared to other studies) of R2 ≈ 0.3 to R2 ≈ 0.5, with XGBoost outperforming the other machine learning algorithms tested. Differences in local climate and emission conditions emphasize the significance of considering regional distinctions when interpreting and comparing LCS responses and field calibration efforts. When using rank-level confusion matrices, True Positive air quality classification of predicted PM2.5 levels by XGBoost rated between 71% and 88%.