Using neural networks for GOME ozone retrieval - an overview

Martin Mueller

Atmospheric Chemistry & Dynamics Branch
NASA Goddard Space Flight Center

Neural network provide an interesting alternative to classical retrieval methods, because they are fast and accurate, and thoroughly exploit nonlinear correlations in the data. In a way they combine 4-dimensional statistical and physical information in a similar way assimilation does. Their application will be described for the case of ozone profile retrieval from the Global Ozone Monitoring Experiment (GOME), a UV spectrometer flying on ERS-2. The Neural Network Ozone Retrieval System (NNORSY) was trained on a data set of GOME radiances collocated with ozone measurements from ozonesondes, Halogen Occultation Experiment, Stratospheric Aerosol and Gas Experiment II, and Polar Ozone and Aerosol Measurement III. I will discuss results obtained from analyzing data from 1996 to 2001, error estimates gained from the neural network, sensitivity studies and possible combination with classical methods.