A Novel Physics-Informed Transformer Algorithm (PITA) for Near–Real-Time Synergistic Multi-Parameter Aerosol Retrieval From FY-4B/AGRI Geostationary Satellite
Аннотация
Accurate retrieval of aerosol properties from geostationary satellites is critical for capturing rapid atmospheric variability, yet remains difficult due to uncertain aerosol-model assumptions, clear-pixel misclassification over bright surfaces, and limited physical interpretability of purely data-driven approaches. We present a Physics-Informed Transformer Algorithm (PITA) for aerosol retrieval that couples radiative transfer (RT) physics with a Transformer architecture to enable near – real-time synergistic retrieval of aerosol optical depth (AOD) at 550 nm and the Ångström exponent (AE; 440–675 nm) from Fengyun-4B (FY-4B) Advanced Geosynchronous Radiation Imager (AGRI) observations. PITA integrates: 1) six regionally optimized seasonal aerosol models derived via deep clustering of 11-year ground-based aerosol observations, reducing the mean absolute error (MAE) of forward-simulated apparent reflectance by 57.6% relative to MODIS Collection 6 models; 2) a critical-reflectance clear-pixel identification scheme to improve screening over bright surfaces; and 3) an RT-constrained Transformer network with RT-derived features, with SHAP attribution indicating these physics features contribute > 70% of model importance. Sample-based ten-fold cross-validation yields AOD performance of R = 0.94 and MAE = 0.07 (76% within the expected error, EE) and AE performance of R = 0.77 and MAE = 0.16 (93% within EE). Compared with the operational FY-4B land aerosol product, PITA improves AOD correlation by 52.5% and increases AE correlation from − 0.12 to 0.73. Sensitivity analyses show negligible bias at low aerosol loading (AOD < 0.23) and weak dependence on particle size. Without external reanalysis inputs, PITA offers a physically interpretable framework for near – real-time, high-accuracy geostationary aerosol monitoring.