Estimation of High-Temporal-Resolution PM2.5 Concentration from 2019 to 2023 Using an Interpretable Deep Learning Model

The FY-4A satellite represents a new generation of geostationary platforms, providing high-temporal-resolution observations over China. However, challenges remain in effectively leveraging the FY-4A satellite data for high-temporal-resolution PM2.5 concentration estimation, particularly regarding the unclear key parameters required for accurate estimation and the limited interpretability of models. This study utilizes an interpretable deep learning framework that integrates FY-4A Top-of-Atmosphere (TOA) reflectance data, meteorological variables, and auxiliary data to estimate surface high-temporal-resolution PM2.5 concentrations from 2019 to 2023. A multicollinearity test was applied to optimize feature selection, while the SHapley Additive exPlanations (SHAP) method was used to enhance model interpretability. The results indicate that parameters such as TOA02, TOA03, TOA04, and boundary layer height (BLH) significantly influence model performance across years. The model demonstrates strong predictive ability in the Beijing–Tianjin–Hebei (BTH) region, achieving an average R2 of 0.83. Root mean square error (RMSE) values remained below 15 µg/m3, aligning well with ground-based monitoring data. These findings demonstrate that combining high temporal satellite data with interpretable deep learning provides a reliable approach for long-term, high-temporal-resolution PM2.5 monitoring in regions.