Physics-Guided Neural Network Speeds Up Atmospheric Correction for Remote Sensing

Researchers have developed a physics-aware multi-fidelity surrogate framework, pKANrtm, to emulate atmospheric correction coefficients for optical remote sensing. The method uses paired simulations from 6S and libRadtran radiative transfer models, sampled via Latin Hypercube Sampling, for Sentinel-2 bands. The high-fidelity targets include path reflectance, total transmittance, and spherical albedo. The Kolmogorov-Arnold Network predicts the residual between low-fidelity 6S and high-fidelity libRadtran outputs, reducing computational cost for dense look-up-table generation, sensitivity analysis, and operational preprocessing. The study is published on arXiv.