Aerosol Optical Depth (AOD) and Fine Mode Fraction (FMF) are important information for air quality research. Both are mainly obtained from satellite data based on a radiative transfer model, which requires heavy computation and has uncertainties. We proposed machine learning-based models to estimate AOD and FMF directly from Geostationary Ocean Color Imager (GOCI) reflectances over East Asia. Hourly AOD and FMF were estimated for 00–07 UTC at a spatial resolution of 6 km using the GOCI reflectances, their channel differences (with 30-day minimum reflectance), solar and satellite viewing geometry, meteorological data, geographical information, and the Day Of the Year (DOY) as input features. Light Gradient Boosting Machine (LightGBM) and Random Forest (RF) machine learning approaches were applied and evaluated using random, spatial, and temporal 10-fold cross-validation with ground-based observation data. LightGBM (R2 = 0.89–0.93 and RMSE = 0.071–0.091 for AOD and R2 = 0.67–0.81 and RMSE = 0.079–0.105 for FMF) and RF (R2 = 0.88–0.92 and RMSE = 0.080–0.095 for AOD and R2 = 0.59–0.76 and RMSE = 0.092–0.118 for FMF) agreed well with the in-situ data. The machine learning models showed much smaller errors when compared to GOCI-based Yonsei aerosol retrieval and the Moderate Resolution Imaging Spectroradiometer Dark Target and Deep Blue algorithms. The Shapley Additive exPlanations values (SHAP)-based feature importance result revealed that the 412 nm band (i.e., ch01) contributed most in both AOD and FMF retrievals. Relative humidity and air temperature were also identified as important factors especially for FMF, which suggests that considering meteorological conditions helps improve AOD and FMF estimation. Besides, spatial distribution of AOD and FMF showed that using the channel difference features to indirectly consider surface reflectance was very helpful for AOD retrieval on bright surfaces.
