A Sequence-to-Sequence Transformer Model for Satellite Retrieval of Aerosol Optical and Microphysical Parameters from Space

Aerosol optical and microphysical properties determine their radiative capabilities, climatic impacts, and health effects. Satellite remote sensing is a crucial tool for obtaining aerosol parameters on a global scale. However, traditional physical and statistical retrieval methods face bottlenecks in data mining capacity as the volume of satellite observation information increases rapidly. Artificial intelligence methods are increasingly applied to aerosol parameter retrieval, yet most current approaches focus on end-to-end single-parameter retrieval without considering the inherent relationships among multiple aerosol properties. In this study, we propose a sequence-to-sequence aerosol parameter joint retrieval algorithm based on the transformer model S2STM. Unlike conventional end-to-end single-parameter retrieval methods, this algorithm leverages the encoding-decoding capabilities of the transformer model, coupling multi-source data such as polarized satellite, meteorological, model, and surface characteristics, and incorporates a physically coherent consistency loss function. This approach transforms traditional single-parameter numerical regression into a sequence-to-sequence relationship mapping. We applied this algorithm to global observations from the Chinese polarimetric satellite (the Particulate Observing Scanning Polarimeter, POSP) and simultaneously retrieved multiple key aerosol optical and microphysical parameters. Event analyses, including dust and pollution episodes, demonstrate the method's responsiveness in hotspot regions and events. The retrieval results show good agreement with ground-based observation products. This method is also adaptable to satellite instruments with various configurations (e.g., multi-wavelength, multi-angle, and multi-dimensional polarization) and can further improve its spatiotemporal generalization performance by enhancing the spatial balance of ground station training datasets.