Abstract
<jats:p>Multi-angle polarimetry provides enhanced sensitivity to aerosol absorption, but retrieving reliable single scattering albedo (SSA) over land remains difficult when instrument noise, calibration drift, and forward-model simplifications induce persistent simulation–observation discrepancies. Here we develop a domain-adaptive neural retrieval framework for multi-angle polarimetric (MAP) observations that is trained on radiative transfer simulations and explicitly adapts to real measurements using paired feature alignment, domain-adversarial training, and limited observation fine-tuning. We use the Gaofen-5 Directional Polarimetric Camera (DPC)—a particularly challenging MAP dataset with post-launch drift and elevated polarimetric uncertainty—as an application case to test robustness under realistic errors. Direct deployment of a simulation-trained model to DPC observations does not transfer reliably because collocated radiances exhibit band-dependent offsets and large degree of linear polarization (DoLP) scatter, confirming substantial domain shift. After adaptation, out-of-fold validation against AERONET yields mean AOD (R ≈ 0.81), mean derived white-sky albedo (R ≈ 0.94), and SSA(443 nm) (R ≈0.53) for AOD(443 nm) > 0.4 (n = 140). Controlled noise-injection experiments further quantify the observation-error regimes under which a simulation-only retrieval would remain viable, indicating that random scatter is a stronger limiter than plausible systematic bias for SSA retrieval. Together, these results show that simulation-trained MAP retrieval can remain effective on noisy sensors when explicit simulation and observation(sim-obs) adaptation is incorporated, and they provide quantitative observation-accuracy targets relevant to next-generation polarimeters.</jats:p>