A robust and flexible satellite aerosol retrieval algorithm for multi-angle polarimetric measurements with physics-informed deep learning method

Abstract

The multi-angle polarimetric (MAP) satellite measurements provide abundant information concerning aerosol optical/microphysical properties. In this study, we present a robust and flexible aerosol retrieval algorithm for MAP measurements based on physics-informed deep learning (PDL) method. Different from optimized inversion that needs iterative Radiative Transfer (RT) calculations of all the unknowns, the PDL method can model the whole MAP observations with each retrieved aerosol parameter separately with the pre-training of RT simulations. Furthermore, the training of PDL can make full use of the prior information from ground-based aerosol inversions and satellite surface products, and provides an effective constraint to avoid unphysical values. To examine performance of PDL algorithm, we retrieve aerosols over eastern China from POLDER-3 measurements during 2007–2009. Comparison with AERONET products shows high correlations (R > 0.91) for both POLDER-3 PDL Aerosol Optical Depth (AOD) and fine AOD. Despite lower correlations caused by a small portion of poor retrievals, PDL coarse AOD and Single Scattering Albdeo (SSA) is very consistent with AERONET results. Also, PDL retrievals perform well as the best estimates of optimized methods such as GRASP (Generalized Retrieval of Aerosol and Surface Properties). With an outstanding performance in accuracy and efficiency, the flexible PDL algorithm exhibits great potential for operational retrieval of MAP satellite measurements.