A Capsule Network Model for Aerosol Retrieval from DPC/Gaofen-5(02) Satellite Multi-Angle Polarimetric Observation

摘要

Satellite multi-angle polarimetric (MAP) observations provide crucial insights into the microphysical and optical properties of atmospheric aerosols. Recent advancements in multi-angle, multispectral, and polarized satellite observations have increased data content and complexity. While traditional methods like look-up tables and optimal estimation face challenges in fully utilizing these advanced datasets, deep learning approaches offer substantial advantages. However, deep learning models also have limitations, particularly regarding physical interpretability and the efficiency of processing high-dimensional observational data. To address these challenges, we propose MAP_CapsNet, a deep learning algorithm based on Capsule Networks (CapsNets) for aerosol multi-parameter retrieval. This algorithm combines the multi-dimensional modeling capabilities of CapsNets with vector radiative transfer models to retrieve aerosol optical and microphysical parameters. We applied it to MAP measurements from the Directional Polarimetric Camera (DPC) onboard the Gaofen-5(02) satellite to retrieve different aerosol parameters over China in 2022. The results were validated against Aerosol Robotic Network and Sun/sky-radiometer Observation Network data. The correlation coefficients ( R ) for aerosol optical depth and fine mode fraction exceed 0.935 and 0.782, respectively. The single scattering albedo also showed a moderate correlation ( R = 0.691). Compared with Moderate Resolution Imaging Spectroradiometer and Visible Infrared Imaging Radiometer Suite products, the DPC exhibited good spatial consistency and an enhanced ability to characterize aerosol properties due to higher spatial resolution and MAP capability. These findings highlight the DPC instrument’s potential for high-resolution, real-time monitoring of dust and haze pollution events.