Improving Aerosol Retrieval From MISR With a Physics-Informed Deep Learning Method

Abstract

The Multi-angle Imaging SpectroRadiometer (MISR) measurement with a large range of scattering angles provides valuable information about aerosol microphysical properties. The current MISR algorithm utilizes pre-defined aerosol mixtures in lookup tables (LUT) to infer aerosol types and microphysical parameters, which performs well globally but remains subject to considerable uncertainties in regional scales. To make efficient use of MISR measurement, we developed a physics-informed Deep Learning (PDL) method to retrieve aerosol optical/microphysical parameters over land in eastern China. By combining the physical constraint of radiative transfer simulation and modeling ability of DL methods, each aerosol parameter can be modeled with the whole used MISR measurements separately with high computational efficiency. PDL Aerosol Optical Depth (AOD) and fine AOD(FAOD) have high correlation coefficients (R>0.95) with Aerosol Robotic Network (AERONET) observations, with 89% and 81% values falling into expected error (EE) envelope of ± (0.05+20%AODAERONET) respectively. Despite only a slightly higher accuracy than recent MISR Version 23 products, PDL retrievals have solved the underestimation problem of AOD and FAOD at moderate-high values (>0.4). Besides better constraint of abnormal values in coarse AOD(CAOD), PDL algorithm significantly improves retrieval accuracy of MISR Single Scattering Albedo (SSA). With reliable and robust performance, PDL algorithm provides a flexible and efficient aerosol retrieval framework for emerging multi-angle polarimetric measurements.