Machine Learning-Based Retrieval of Aerosol and Surface Properties Over Land From the Gaofen-5 Directional Polarimetric Camera Measurements

Aerosol properties, including aerosol optical depth (AOD), aerosol absorption optical depth (AAOD), single scattering albedo (SSA), and fine mode fraction (FMF), are essential in studying aerosol climate effects. Spectral multiangle polarimetry (MAP) has been recognized as a promising technique for comprehensive retrievals of global aerosol optical properties from space. As one of the very few MAP sensors in space, the Directional Polarimetric Camera (DPC) onboard the GaoFen (GF)-5 satellite has great potential to provide these critical aerosol parameters. However, retrievals of aerosol parameters from DPC, especially SSA and AAOD, still remain limited. This study introduces a machine-learning algorithm using the eXtreme gradient boosting (XGBoost) model to retrieve AOD, AAOD, SSA, FMF, as well as surface albedo (expressed as the directional hemispherical reflectance, DHR) over land from DPC multiangle reflectances and degree of linear polarization (DOLP), using AERONET aerosol measurements and Moderate Resolution Imaging Spectroradiometer (MODIS) DHR data as the training target. Cross-validation indicates high retrieval accuracy, with correlations exceeding 0.75 for all parameters under sufficient aerosol loading. Notably, the accuracy of SSA retrieval is comparable to that of the Polarization and Directionality of the Earth’s Reflectance (POLDER) products, with 73% of the independently retrieved 670-nm SSA falling within the ±0.03 error envelope (EE) when 670-nm AOD is above 0.30. Gridded products also effectively capture the spatial and seasonal variability of aerosol properties worldwide, such as in regions dominated by biomass burning and dust. This study confirms the capability of DPC for aerosol property retrievals, which could serve as an important technique and data source for global aerosol and climate monitoring.