Global retrieval of stokes vector of water-leaving radiance via Unet-based framework

The polarization characteristics of water-leaving radiance ( L w ) exhibit marked sensitivity to the micro-optical properties of oceanic particles, making multi-angle polarimetric observations a crucial approach for solving nonlinear, ill-posed, and underdetermined radiative transfer (RT) equations in ocean color remote sensing. However, most existing statistically optimized search-based atmospheric correction (AC) models neglect the multi-angle Stokes vector contribution of L w in retrieving polarimetric aerosol properties. In this study, a polarization-based AC model for the multi-angle Stokes vector of L w was developed using Fully Connected U-net (FUCN-AC) and global RT simulations database. And, the global spatial distribution patterns of inversion errors for the L w Stokes vector, and the impact of instrument noise on the FUCN-AC with respect to observation geometry, were systematically investigated. The results showed that the global mean absolute percentage errors (MAPEs) for the L w Stokes parameters ( I , Q , and U ), remote sensing reflectance ( R rs ), aerosol optical thicknesses (AOTs), and degree of polarization (DOP) at 490 nm were 2.24 %, 5.86 %, 6.38 %, 2.2 %, 15.94 %, and 6.06 %, respectively. The maximum MAPE for each parameter in coastal waters were mainly concentrated along the coastlines of the Eurasian continent in the Northern Hemisphere and the coastal waters of Antarctica. Additionally, field-measured polarized radiances validated the FUCN-AC model with a MAPE of 39.84 % for DOP, which might be attributed to uncorrected instrumental polarization and aerosol model mismatches. Moreover, the FUCN-AC model demonstrated significantly larger MAPE for the Stokes parameter Q and AOT at 490 nm, exceeding those for Stokes parameter I by factors of approximately 4.12 and 4.95, respectively. Furthermore, instrument noise exerted the most pronounced impact on Stokes parameter I and R rs , with the MAPEs increasing by factors of 2.54 (3.80) and 2.61 (3.90) under 3 % (5 %) noise levels, respectively. Overall, the FUCN-AC achieves the nonlinear decoupling between atmospheric and ocean polarization signals, and provides technical support for the large-scale, high-frequency dynamic monitoring of the multi-angle Stokes vector of L w .