Enhancing accuracy of air quality sensors with machine learning to augment large-scale monitoring networks

Low-cost sensors have revolutionized air quality monitoring, however, precision is questioned compared to reference instruments. Hence, the performance of two widely used PM2.5 Sensors, Purple Air (PA) and ATMOS, were evaluated over a 10-month period in the North Western-Indo Gangetic Plains (NW-IGP). In-field collocation with Beta Attenuation Monitor found low R2 values; 0.40 for ATMOS and 0.43 for PA. To calibrate and improve the accuracy of sensors, five Machine Learning (ML) models and an empirical relative humidity correction methodology were used separately for both sensors. Out of these, the Decision Tree outperformed others, and R2 values improved to 0.996 for ATMOS and 0.999 for PA. Root mean square error reduced from 34.6 µg/m3 to 0.731 µg/m3 for ATMOS and from 77.7 µg/m3 to 0.61 µg/m3 for PA, while using DT as a calibrating model. The study reveals the best-performing ML model for correcting PM2.5 sensor data, enhancing the accuracy of air quality monitoring systems.