Tracking diurnal aerosol dynamics and smoke transport over East Asia using multi-geostationary satellite fusion (GK-2B, Himawari-9, FY-4B)

Continuous, high-frequency observations of aerosol optical depth (AOD) are crucial for understanding aerosol dynamics and regional air quality. However, individual geostationary (GEO) AOD products suffer from distinct spatiotemporal gaps and uncertainties driven by cloud screening, surface-reflectance assumptions, and strict algorithmic quality controls. To address these limitations, we propose a deep-learning fusion framework synergizing the spectral, temporal, and geometric complementarities of three GEO satellites (GK-2B, Himawari-9, and FY-4B) over East Asia. Central to our framework is a TabNet-based dynamic feature selection mechanism that adaptively weights heterogeneous multi-sensor inputs to suppress local retrieval errors. To improve spatial generalization, we implement a two-stage transfer learning strategy: pre-training with spatially extensive MODIS data to establish large-scale spatial priors, followed by fine-tuning with ground-based AERONET observations. Compared to single-satellite baselines, the fused AOD product from May 2024 to April 2025 improves accuracy and spatiotemporal coverage. Specifically, it yields an overall cross-validation correlation coefficient (R) of 0.94, reduces root-mean-square error (RMSE) by 47 %, and increases valid hourly matchups by 33.5 %. Independent-site validation shows consistent generalizability (R = 0.88). Furthermore, evaluations of extreme events, including severe biomass-burning and dust transport episodes, demonstrate that the fused AOD product can reconstruct hourly plume gradients during rapid transport. By improving the temporal characterization of complex aerosol environments, this deep learning-driven fusion approach provides a useful foundational dataset for future regional air-quality analyses and chemical transport model assimilation.