Aerosol optical properties are critical for understanding climate change and assessing aerosol pollution risks. However, it remains challenging to simultaneously and accurately retrieve aerosol optical thickness (AOD), fine-mode fraction (FMF), and single scattering albedo (SSA) from satellite-based sensors offering only multispectral, single-angle measurements. Previous physical methods struggle to meet the rapidly increasing requirements for both accurate and real-time aerosol property retrievals needed in certain applications. LookUp-Table (LUT) approaches offer high computational efficiency, while Optimal-Estimation (OE) methods generally provide improved accuracy with considerably higher computational demands. To address these challenges, we developed ORION-DL (Operational Retrieval of Aerosols from Polarimetric Observations using Deep Learning), a deep learning framework that integrates pretraining and physical constraints to enhance aerosol retrieval from multispectral, single-angle polarization measurements. The ORION-DL model consists of two modules. The feature-extraction module employs multihead attention to separately process continuous and categorical features, and the parameter-regression module embeds physical constraints to ensure physically consistent outputs. Pretraining was performed with satellite products to enhance generalization, followed by fine-tuning using AERONET data to optimize accuracy. Independent validation shows that pretraining substantially improves retrieval performance, with ORION-DL outperforming tree-based models (XGBoost and LightGBM). Compared with NASA MODIS and VIIRS products, ORION-DL achieves more accurate simultaneous retrievals of AOD, FMF, and SSA. Application to the record-breaking 2023 North American wildfires further confirms the model’s capability to capture the spatiotemporal evolution of aerosol properties and to quantify aerosol radiative forcing directly from satellite retrievals.
Ji et al. (Mon,) studied this question.