Improving estimation of surface PM _2.5 by including satellite observations of gases, aerosols, and radiation in tandem

Abstract Shanxi Province, a major hub of coal production and industrial activity in China, faces severe PM 2.5 pollution and complex co-emission sources that challenge air quality monitoring. This study developed multiple machine learning models to estimate daily PM 2.5 concentrations in 2022 by integrating ground-based PM 2.5 observations, Himawari-8 measurements, CO, NO 2 and ultraviolet aerosol index (UVAI) derived from the TROPOspheric Monitoring Instrument (TROPOMI), and ERA5. Ten-fold cross-validation showed that incorporating TROPOMI-derived CO, NO 2 and UVAI produced the highest predictive performance, reducing root mean square error (RMSE) and mean absolute error (MAE) by 14% and 16% respectively. Meanwhile, SHapley Additive exPlanations (SHAP) analysis and partial dependence analysis identified CO as the most important species driving PM 2.5 prediction, with all three variables showing positive contributions to PM 2.5 that increased with higher values. This emphasis on the importance of CO and NO 2 is physically realistic given that CO is co-emitted with primary aerosols, and NO 2 is a secondary aerosol precursor. However, such work has not been previously considered by the community focusing on surface PM 2.5 prediction using AI models and satellite observations. The model revealed pronounced spatiotemporal gradients in PM 2.5 across Shanxi, peaking in the final 5 weeks before Chinese New Year (62.0 μ g m −3 ), and minimizing in summer (21.0 μ g m −3 ), while remaining persistently high in southern and eastern industrial regions. The steady PM 2.5 rise in the final 5 weeks before Chinese New Year, suggesting that pre-holiday production activities may contribute to short-term emission intensification, while demonstrating that traditional season-based analysis approaches may miss these extremes. These findings demonstrate the effectiveness of integrating multi-sensor satellite observations with interpretable machine learning for characterizing pollution dynamics in complex industrial regions.