Enhancing Hourly AOD Retrieval from MSG-1/SEVIRI Imagery Integrating Deep and Transfer Learning

Geostationary Earth Orbit (GEO) satellites offer unique capabilities for capturing diurnal variations and providing valuable insights into aerosol cycles. However, publicly available hourly aerosol products with sufficient accuracy remain scarce across Europe, Africa, and West Asia, primarily due to the lack of shorter-wavelength (< 0.6 µm) channels on the Meteosat Second Generation (MSG) satellite series. Therefore, we developed a novel deep learning framework to retrieve hourly aerosol optical depth (AOD) at 550 nm over land in 2021 from MSG-1/SEVIRI imagery, which offers wider spatial coverage through Indian Ocean Data Coverage (IODC). This framework integrates an advanced time-sequence Transformer architecture with transfer learning, utilizing pre-training and fine-tuning techniques. The eXplainable Artificial Intelligence (XAI) analysis revealed that satellite observations across multiple wavelengths contribute 38% to the AOD retrieval, followed by viewing geometry (34%). In comparison with ground-based AOD measurements, our model achieves high accuracy, with an average ten-fold cross-validation (CV) R² of 0.88 and a root mean square error (RMSE) of 0.079. Additionally, our model maintains strong predictive performance in areas and periods lacking ground-based measurements, as evidenced by strong spatial- and temporal-based CV-R² values ranging from 0.71 to 0.86. The model performance is significantly improved when withholding each continent, showing marked increases in R (0.71–0.78) compared to models trained without transfer learning (0.23–0.47). Using the generated reliable 3-km-resolution AOD datasets, we capture pronounced diurnal aerosol variations, characterized by a gradual increase after sunrise, a peak around 10:00 UTC, and a decline by late afternoon, with average magnitude changes of approximately 26% ± 15% relative to the daily mean level (0.22 ± 0.14) on an annual scale, especially during the Northern Hemisphere summer, reaching 30% ± 19%. More importantly, we successfully tracked the rapid dispersion of aerosols and their transport process throughout the day during highly polluted events, driven by both natural and anthropogenic emissions, including dust storms, wildfires, and urban haze. These findings emphasize the unique value of our study for advancing aerosol research over under-monitored regions, particularly focusing on diurnal variations during extreme events.