A Multi-Wavelength Deep Neural Network Framework for Synergistic Retrieval of AOD, FMF, and AAOD from TROPOMI

Aerosol optical depth (AOD), fine-mode fraction (FMF), and absorption aerosol optical depth (AAOD) are essential for quantifying aerosol extinction and related climate and air-quality effects. Yet, most satellite retrievals target a single wavelength or parameter. In this study, a deep neural network (DNN) framework was developed to synergistically retrieve AOD, FMF, and AAOD from Sentinel-5P/TROPOMI at seven wavelengths across 380–772 nm. Parameter-specific feature engineering was designed by incorporating physical linkages among aerosol optical properties. Bayesian optimization was employed to tune hyperparameters, and SHAP (Shapley additive explanations) was used to interpret feature contributions. The proposed model demonstrated high accuracy and robustness on an independent test set. The retrieved AOD showed excellent agreement with AERONET (R = 0.960, MAE = 0.034, RMSE = 0.070), and similarly strong performance was achieved for FMF (R = 0.955, MAE = 0.027, RMSE = 0.039). For AAOD, an overall correlation of 0.86 was obtained (MAE = 0.005, RMSE = 0.008). Comparisons with existing satellite products indicated globally consistent spatial patterns and improved spatial continuity under high aerosol loading. Overall, the proposed data-driven approach enhances the efficiency and coverage of multi-parameter aerosol retrieval while maintaining high accuracy, supporting absorbing aerosol monitoring, aerosol-type discrimination, and climate-effect assessment.