Retrieving aerosol single scattering albedo from FY-3D observations combining machine learning with radiative transfer model

Abstract

This study proposed a new method to retrieve aerosol single scattering albedo (SSA) over land for the Medium Resolution Spectral Imager-II (MERSI-II) onboard the Fengyun-3D (FY-3D). Considering both accuracy and retrieval efficiency, the method combines machine learning with an aerosol optical model constructed from mixed aerosol components. A sample dataset, containing 4 bands of apparent reflectance simulated by the radiative transfer model and corresponding geometric conditions, aerosol and land surface information, is constructed for training and validating machine learning models. Three Back Propagation Neural Network (BPNN) SSA retrieval models are built based on the theoretical basis of SSA retrieval, and the sensitivity of SSA retrieval accuracy to input parameter errors is analyzed. The results show that BPNN-based SSA retrieval models can replace the iterative optimal solution process to a certain extent, achieving quick retrieval of satellite SSA. The BPNN SSA retrieval models are applied to FY-3D MERSI-II observations and validated using AERONET SSA products. The results indicate that the BPNN SSA retrieval model, which uses solar zenith angle, satellite zenith angle, relative azimuth angle, aerosol optical depth (AOD), surface altitude, bi-directional reflectance distribution function (BRDF) parameters (bands 1–2), and apparent reflectance (bands 1–4) as inputs, performs better than others. The retrievals show good consistency with AERONET SSA products with a correlation coefficient of approximately 0.5 and a root mean square error (RMSE) of 0.045 (0.034) at 470 nm (550 nm). In addition, more than 66 % of the SSA retrievals are within the expected error of ±0.05.