Scene-Aware Correction of Scalar-to-Vector Reflectance in the Visible NO2 Retrieval Window

Polarization affects forward modeling in the 420–455 nm visible NO2 retrieval window, yet operational retrievals require computational efficiency. We generated 205,200 paired scalar DISORT and vector VLIDORT simulations for a clear atmosphere and four pure-aerosol scenes (sulfate, sea salt, soot, and mineral dust), covering four NO2 vertical column densities, 19 wavelengths, viewing geometries, and surface albedos. Vector Stokes-I reflectance was the reference. We compared scene-specific inverse-distance-weighted look-up tables (IDW-LUTs) with unified and scene-specific backpropagation (BP), histogram-based gradient boosting (HGB), and Transformer models. Test samples were held out by geometry–surface state. The scalar baseline yielded an RRMSE of 3.0620%. IDW-LUT, unified BP, and unified HGB reduced this error by 44.90%, 79.48%, and 82.27%, respectively. The unified Transformer performed best (MARE = 0.1133%; RRMSE = 0.2402%; P95 = 0.3008%; 92.15% RRMSE reduction), at 0.4636 ms per 19-wavelength spectrum and 0.1159 MiB storage. The unified Transformer offers a strong accuracy–storage balance, while unified BP was the fastest among the unified machine-learning models. Conclusions are limited to clear-sky, pure, known-aerosol scenes; mixed aerosols, clouds, profile variability, instrument effects, and end-to-end NO2 retrievals remain to be tested.