Harmonizing low-cost and regulatory air quality monitoring networks with interpretable semi-supervised learning: Reducing exposure misclassification in underrepresented communities.

Low-cost sensors (LCSs) can address gaps in regulatory air quality monitoring station (AQMS) distribution, but they face data quality issues and spatial misalignment challenges when calibrating large-scale LCS networks against AQMS networks. This study proposed a semi-supervised learning model that uses data augmentation via chained imputation (CI-DA) to address the spatial misalignment problem by synthesizing pseudo-LCS data, thereby enhancing the use of LCS in PM mapping. Tangshan, an industrial city in northern China, was selected as the case study area. The CI-DA model improved data quality for 82 % of LCSs post-calibration, reducing the Root Mean Square Deviation (RMSD) between LCS and AQMS data by 10.8 %. The CI-DA model improved predictive generalizability by harmonizing the LCS and AQMS networks, which increased the spatial validation R from 0.68 to 0.76 compared to the AQMS-only model. Moreover, it reduced exposure misclassification in industrial areas by approximately 20 % compared to the model using uncalibrated LCS data. By using model interpretation methods, we elucidated the mechanism by which CI-DA harmonizes LCS and AQMS data to improve PM prediction accuracy. The CI-DA model can reduce maintenance costs for LCS networks while enhancing exposure assessment accuracy in underrepresented communities, thereby promoting environmental justice.