Improving hourly PM_2.5 estimation from geostationary satellite TOA reflectance through physically constrained multi-task learning

Geostationary satellite-based PM_2.5 mapping increasingly relies on data-driven models using top-of-atmosphere (TOA) reflectance. However, most existing approaches train a single estimator on temporally aggregated samples, implicitly assuming a stationary relationship between TOA radiative signals and near-surface PM_2.5. This assumption overlooks the pronounced diurnal variability captured by geostationary satellites and limits the representation of hour-dependent aerosol–radiation interactions. To address this limitation, we developed a physically constrained multi-task learning framework, termed the Multi-Temporal Decoupled Attention Network (MTDAN), for hourly PM_2.5 estimation from geostationary satellite TOA reflectance. The framework integrates three components: (1) a physically guided surface–atmosphere decoupling module that suppresses surface-background interference and enhances aerosol-sensitive information; (2) a multi-task attention architecture that jointly learns shared and hour-specific relationships between PM_2.5 and explanatory variables; and (3) a temporal-constrained loss function that exploits diurnal continuity to improve the consistency of sequential predictions. Experiments over eastern China using year-long Himawari-8 observations and ground-based PM_2.5 measurements demonstrated the effectiveness of the proposed framework. MTDAN achieved an overall site-based validation R² of 0.89 with an RMSE of 9.27 μg m⁻³, outperforming representative machine-learning and deep-learning models by reducing estimation errors by 21.7 %–27.8 %. The model successfully reproduced hourly pollution evolution, captured localized hotspots, and reconstructed season-dependent diurnal PM_2.5 dynamics. Ablation analyses further confirmed the complementary contributions of physical decoupling, adaptive feature weighting, and temporal-consistency constraints. These results demonstrate that explicitly modeling hour-dependent aerosol–radiation relationships can substantially improve geostationary satellite-based PM_2.5 estimation. The proposed framework provides an interpretable and robust solution for high-frequency air-quality monitoring and offers a general strategy for integrating physical constraints and temporal context into Earth observation applications.