Deep learning calibration model for PurpleAir PM2.5 measurements: Comprehensive Investigation of the PurpleAir network

Abstract

The limited number of PM_2.5 monitoring stations from the Environmental Protection Agency (EPA) across the Contiguous United States (CONUS) restricts PM_2.5 monitoring and associated policymaking efforts. Low-cost PM_2.5 stations, such as those from the PurpleAir network, offer a vital alternative to expand coverage in regions not monitored by the EPA. However, the accuracy of PurpleAir measurements has been questioned. This study introduces a deep learning (DL) approach, specifically a deep convolutional neural network (DeepCNN), to align hourly PM_2.5 data from PurpleAir with EPA PM_2.5 observations across the CONUS for the year 2021. Utilizing over nine million samples from 1595 PurpleAir stations located within 5 km of EPA stations, the DeepCNN demonstrates significant improvements in the agreement between PurpleAir and EPA observations. It increases the correlation coefficient (R) with EPA observations from 0.58 to 0.85 and reduces the mean absolute bias (MAB) from 4.99 to 2.98 μg/m³, achieving a 40% reduction in bias. The state-wise cross-validation also underscores the model's generalizability, with an average 11% improvement in R values and a 13% reduction in bias between PurpleAir and EPA PM_2.5 measurements in various states. Comparative analysis reveals that the accuracy of our DL-enhanced PurpleAir PM_2.5 (PM-DL) data significantly surpasses that of five previously established PurpleAir correction models, which show low R values of 0.55–0.58 and MABs ranging from 4.21 to 6.43 μg/m³ when validated against EPA data. This study underscores the need for more sophisticated models to better align PurpleAir PM_2.5 measurements to EPA standards. The PM-DL data can substantially mitigate the scarcity of reliable institutional PM_2.5 observations across the CONUS. By aligning PurpleAir PM_2.5 data with EPA observations, our model has the potential to augment the existing network with over ten thousand accurate monitoring stations, significantly expanding upon the nearly one thousand EPA stations currently in operation.