Linear regression models and neural networks for the fast emulation of a molecular absorption code.

The background scene generator MATISSE, whose main functionality is to generate natural background radiance images, makes use of the so-called Correlated K (CK) model. It necessitates either loading or computing thousands of CK coefficients for each atmospheric profile. When the CK coefficients cannot be loaded, the computation time becomes prohibitive. The idea developed in this paper is to substitute fast approximate models for the exact CK generator; using the latter, a representative set of numerical examples is built and used to train linear or nonlinear regression models. The resulting models enable an accurate CK coefficient computation for all the profiles of an image in a reasonable time.