Lowcost sensors are widely used for air quality monitoring due to their affordability, portability and easy maintenance. However, the performance of such sensors, such as PurpleAir Sensors (PAS), is often affected by changes in environmental (e.g., temperature and humidity) or emission conditions, and hence the resulting measurements require corrections to ensure accuracy and validity. Traditional correction methods, like those developed by the USEPA, have limitations, particularly for applications to geographically diverse settings and sensors with no collocated referenced monitoring stations available. This study introduces BaySurcls, a Bayesianoptimised surrogate model integrating deep learning (DL) algorithms to improve the PurpleAir sensor PM2.5 (PAS2.5) measurement accuracy. The framework incorporates environmental variables such as humidity and temperature alongside aerosol characteristics, to refine sensor readings. The BaySurcls model corrects the PAS2.5 data for both collocated and noncollocated monitoring scenarios. In a case study across multiple locations in New South Wales, Australia, BaySurcls demonstrated significant improvements over traditional correction methods, including the USEPA model. BaySurcls reduced root mean square error (RMSE) by an average of 20% in collocated scenarios, with reductions of up to 25% in highvariation sites. Additionally, BaySurcls achieved Nash–Sutcliffe Efficiency (NSE) scores as high as 0.88 in collocated cases, compared to scores below 0.4 for the USEPA method. In noncollocated scenarios, BaySurcls maintained NSE values between 0.60 and 0.78, outperforming standalone models. This improvement is evident across multiple locations in New South Wales, Australia, demonstrating the model’s adaptability. The findings confirm BaySurcls as a promising solution for improving the reliability of lowcost sensor data, thus facilitating its valid use in air quality research, impact assessment, and environmental management. © 2024 by the authors.
