Aerosol satellite retrieval can provide detailed aerosol information on a large scale, which becomes one of the main ways of global aerosol research. However, rapid and accrue aerosol retrieval by satellite is challenging, typically requiring radiation transfer models (RTMs) and surface reflectance (SR). An aerosol retrieval algorithm (SEMARA) combining the simplified aerosol retrieval algorithm and simplified and robust surface reflectance estimation can obtain local high-precision aerosol optical depth (AOD) without RTMs and SR datasets, while the method cannot perform large-scale and long-term aerosol retrieval. Hereby, a machine learning (ML) method based on the fully connected neural network (FCNN) and SEMARA was proposed. The new method optimizes the traditional sample construction of the ML and can achieve aerosol retrieval at a larger spatial and temporal scale. Moderate resolution imaging spectroradiometer data were applied to AOD retrieval on four typical regions globally. The AOD retrievals were validated using aerosol robotic network measurements in comparison to MOD04_3K AOD and the SEMARA. The accuracy validation indicators of the new method, in which the root-mean-square error (RMSE) was 0.109, mean absolute error (MAE) was 0.072, Pearson correlation coefficient (R) was 0.8983, and approximately 79.69% of the retrievals fell within the expected error (EE), performed better than MOD04_3K (RMSE = 0.1972 MAE = 0.1403, R = 0.7692 and Within EE = 55.24%) and the SEMARA method (RMSE = 0.2465 MAE = 0.1106, R = 0.0.5968 and Within EE = 72.85%) in all study regions, and the AOD retrievals can better reflect the spatial variation of AOD with better spatial continuity and coverage.
