Estimating Retrieval Errors From Neural Network Inversion Schemes—Application to the Retrieval of Temperature Profiles From IASI

Abstract

Neural networks (NNs) are becoming increasingly more popular to interpret remote sensing observations in many contexts. Their flexibility and accuracy have been a true advantage for the retrieval of many geophysical variables in the atmosphere, land, and ocean, over the last three decades. Uncertainty of the retrieved products is important to assess the quality of the retrievals but also to combine products <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a posteriori</i> , combine them in a complex way to study for instance the water cycle, or assimilate them in numerical weather prediction (NWP) centers. However, no easy solution has been proposed so far to estimate the retrieval uncertainties of NN inversion schemes. A simple, pragmatic, and easy-to-implement scheme based on the input space clustering is proposed here to perform this task. Tests are conducted using an application aiming at retrieving atmospheric profiles based on improved atmospheric sounding in the infrared (IASI) high-spectral resolution observations in the infrared.