Real-time Total Ozone and Ozone Profiles Retrieved from Gome Data Using Neural Networks∗

A new approach for retrieving total ozone from ERS2-GOME spectral data has been developed, which relies on feed-forward neural networks to perform the data inversion. Using selected GOME wavelength regions, instrument and geolocation data as an input, networks have been trained to determine atmospheric ozone in a one-step procedure. In order to train a neural network, an extensive database of collocated GOME and ozone measurements is necessary. In the case of total ozone Dobson and Brewer measurements mainly from the World Ozone and Ultraviolet Data Center (WOUDC) have been utilized, while ozone profiles stem from sondes collected by WOUDC, as well as HALOE and SAGE II limb measurements. A special training method has been employed for dealing with different valid height ranges of these profiles. Results of the total ozone retrievals are being validated with ground stations and cross-compared against TOMS V7 Level 3 ozone data. The neural network is obviously capable of correcting for limb-viewing as well as for instrument degradation effects and clouds. This also holds for the neural network ozone profile retrieval, whose results compare well qualitatively with classical retrieval methods and independent sonde and satellite data. The methodology developed here is independent of any physical parameterizations apart from those used in data preprocessing, which is performed by standard software. It is fast enough to allow real-time profile retrievals even on PCs. Adaptation of the method to upcoming high-resolution sensors could be carried out easily.