Neural Network Emulation of VSWIR Atmospheric Radiative Transfer Models(NPO-50675-1)

Conventional Visible Shortwave Infrared (VSWIR) imaging spectrometer atmospheric correction evolved from multi-band approaches and generally does not exploit the full spectral measurement. We hypothesize that a pure spectroscopic approach can improve atmospheric inversion accuracy to minimize regional biases in global-scale investigations. Such techniques are pervasive in atmospheric remote sounding disciplines, where Optimal Estimation (OE) retrieval theory (Rodgers et al., 2000) inverts a radiance spectrum to recover a consistent physical model incorporating, for example, aerosol and H2O, the interactions between scattering, absorption, and the coupling between surface reflectance and atmosphere. This enables a statistically rigorous treatment of uncertainty with the potential to recover information on spectrally-broad signals. Our software demonstrates a proof of concept that overcomes the primary computational roadblock of these methods: fast execution of line-by-line Radiative Transfer (RT) models. Neural Network (NN) emulation can replicate the results of the MODTRAN 6.0 line-by-line A-band model to high accuracy and a five order of magnitude improvement in execution time. This demonstrates potential for OE to provide significant advances in the accuracy and modeling power of VSWIR atmospheric correction.