Neural network scheme for the retrieval of total ozone from Global Ozone Monitoring Experiment data

A novel approach to retrieving total ozone columns from the ERS2 GOME (Global Ozone Monitoring Experiment) spectral data has been developed. With selected GOME wavelength regions, from clear and cloudy pixels alike plus orbital and instrument data as input, a feed-forward neural network was trained to determine total ozone in a one-step inverse retrieval procedure. To achieve this training, ground-based total ozone measurements from the World Ozone and Ultraviolet Data Center (WOUDC) for the years 1996-2000, supplemented with Dobson-corrected Total Ozone Mapping Spectrometer (TOMS) data to provide global coverage, were collocated with GOME ground pixels into a training data set. Validation of the neural-network-retrieved ozone values relative to independent ground stations yielded a rms error of better than 11 Dobson units. Comparisons performed on the basis of operationally available TOMS and GOME level-3 maps exhibit good agreement in general, with a latitude-dependent offset.