Retrieving Hourly Aerosol Absorption From Space Using Physical Informed Deep Learning

Abstract Retrieving aerosol absorption properties, such as single scattering albedo (SSA) and absorption aerosol optical depth (AAOD) from single‐view satellite observations remains a significant challenge. This study introduces the Physics‐Informed Neural Network for Aerosols (PINA) framework to retrieve hourly aerosol absorption properties from Himawari‐8/9 data. PINA integrates an inversion network with a high‐efficiency surrogate radiative transfer model, providing a differentiable path for physical consistency. By employing an adaptive weighting strategy that transitions from data‐driven learning under clean conditions to physics‐dominant learning under polluted conditions, PINA achieves high precision for aerosol optical depth (AOD) ( R = 0.96, Root Mean Squared Error [RMSE] = 0.10), AAOD ( R = 0.86, RMSE = 0.02), and SSA ( R = 0.63, RMSE = 0.04), outperforming traditional operational satellite products. PINA also demonstrates robust spatiotemporal generalization, successfully capturing the hourly dynamic evolution of absorption signatures during extreme wildfire smoke, dust, and complex urban haze events.