An efficient transformer-based aerosol retrieval algorithm for the polarization crossfire (PCF) satellite senor suite: development and validation

Multi-angle polarimetric (MAP) satellite observations provide information on aerosol optical and microphysical properties. In this study, we propose an effective transformer-based deep learning (DL) algorithm for aerosol retrieval using MAP observations, utilizing synergistic observations from the Directional Polarimetric Camera (DPC) and POSP (Particulate Observing Scanning Polarimeter) sensors of the polarization crossfire (PCF) sensor suite. The use of these two sensors overcomes limitations of traditional DL approaches that rely solely on sparse ground-based stations. The proposed algorithm involves two main steps: (1) a date base of over 41,000 highconfidence aerosol samples are is constructed using Aerosol Robotic Network (AERONET) data supplemented by selected high-quality-screening data retrieved from the Particulate Observing Scanning Polarimeter (POSP sensor) from using the Generalized Retrieval of Atmosphere and Surface Properties (GRASP) algorithm; (2) A transformer model is trained on DPC MAP measurements, to learn the nonlinear representation linking spaceborne measurements with targets for estimating aerosol optical depth (AOD), fine-mode AOD (FAOD), and coarse-mode AOD (CAOD) at 550 nm. Comparison of the aerosol parameters derived from application of the transformer model to DPC observational data to AERONET reference data observations, shows substantial improvement over models trained solely on AERONET data, with high Pearson correlation coefficients (R) of 0.855 (AOD), 0.812 (FAOD), and 0.793 (CAOD), and slopes of 0.837, 0.813, and 0.801. Comparison with POSP/ GRASP aerosol products shows improved spatial coverage, especially over bright surfaces and during extreme aerosol events such as dust and forest fires. This scalable and transferable framework integrates optimal estimation (OE) interpretability with DL efficiency, achieving 3–4 orders of magnitude speedup over GRASP methods and offering a promising solution for next-generation MAP satellite mission.