Machine Learning-Based Post-Hoc Calibration of Low-Cost Air Quality Sensors Using Reference Monitoring Stations

Sammanfattning

Air pollution is a major environmental and public health challenge worldwide. Fine particulate matter (PM2.5) is particularly harmful because it can penetrate deep into the lungs and bloodstream, contributing to respiratory and cardiovascular diseases. Although regulatory-grade reference monitoring stations provide accurate measurements, their high-cost limits spatial coverage. Low-cost sensors (LCS) offer a cost-effective alternative but require calibration to improve measurement reliability. This thesis presents a machine learning-based post-hoc calibration framework for low-cost PM2.5 sensors using reference monitoring station data. A Linear Tree Regressor was selected because it combines nonlinear decision-tree partitioning with interpretable local linear regression models. The study used the SensEURCity dataset collected in Antwerp, Belgium, comprising approximately 18.7 million minute-level observations from 34 low-cost sensors and 9 reference monitoring stations. PM2.5 concentration, relative humidity, temperature, and atmospheric pressure were used as model inputs. Three research questions were investigated: (1) whether machine learning improves calibration accuracy compared with the U.S. Environmental Protection Agency (EPA) correction method, (2) how many reference stations are required for effective calibration, and (3) how spatial distance affects calibration error. The Linear Tree Regressor achieved MAE = 2.31, RMSE = 3.91, and R² = 0.83, outperforming the EPA correction method (MAE = 3.55, RMSE = 5.29, R² = 0.69). The EPA performance was consistent with previous validation studies, while the proposed model further improved calibration accuracy. Results also showed that effective calibration could be achieved using a limited number of reference stations. Furthermore, the proposed model maintained lower prediction error than the EPA method across most practical urban deployment distances. These findings demonstrate the potential of machine learning to support accurate, scalable, and cost-effective urban air quality monitoring.