An inversion algorithm using neural networks to retrieve atmospheric CO total columns from high‐resolution nadir radiances

Up to the first years of the next millennium, several observation programs of the troposphere are scheduled, including the Infrared Atmospheric Sounding Interferometer, which uses Fourier transform spectroscopy to record the radiance of the Earth-atmosphere system with a nadir-viewing geometry. The Interferometric Monitor for Greenhouse Gases (IMG), launched aboard the Advanced Earth Observing System in August 1996, was a precursor of these forthcoming missions. A new inversion algorithm based on neural network techniques is in development to retrieve trace gases from high-resolution nadir radiances. Neural networks offer a technical alternative to classical methods and allow efficient inversion calculations as required to treat the huge volume of data which will be provided by continuous observation of the atmosphere from space. To develop a network to retrieve the carbon monoxide total column, realistic simulations of the IMG measurements were obtained by coupling a three-dimensional chemical-transport model with a high-resolution line-by-line radiative transfer code adjusted to the instrumental features. The application of the algorithm on simulated data allowed the checking of its performance: for about 99% of the cases, the relative inversion error was less than 10%. This algorithm has been applied to the spectra recorded by the IMG instrument between June 16 and 19, 1997. Global-scale distributions of CO total columns were obtained for the first time by using a neural network, and this technique proved its ability to achieve real-time inversion of atmospheric CO.