Deep Network Radiative Transfer: A Revolution in Imaging Spectrometer Atmospheric Correction, Year 2

Atmospheric Radiative Transfer Models (RTMs) are computationally-complex simulations of photon transport in the atmosphere which are critical for estimating spectroscopic surface properties remotely from air and space platforms. They will be important for imaging spectrometer missions slated to fly in the early 2020s (e.g. NASA’s EMIT mission) and investigations planned for later in the decade (e.g. the Surface Biology and Geology Decadal Observable). However, such instruments produce data volumes that are not currently possible to analyze using our most accurate RTMs due to the extreme computational expense. This project uses machine learning strategies to provide multiple orders-of-magnitude improvements in RTM speed at negligible loss in accuracy while retaining the physical interpretability of the underlying physics-based model. We trained a neural network to emulate the physics-based solution over a relevant range of atmospheric conditions and observation geometries. This enables enhanced retrievals of surface properties under challenging atmospheric distortions.