Neural Network Ozone Profile Retrieval System for GOME Spectra (NNORSY-GOME)

A new approach for retrieving ozone profiles from ERS2-GOME spectral data has been developed, which relies on feed-forward neural networks to perform the data inversion. By using GOME spectral data from selected wavelength regions, instrument (e.g. scan angle), geolocation and UKMO temperature profile data as input, neural networks have been trained to determine the ozone profile from 1–60 km geopotential height in a one-step inversion scheme. In order to train neural networks, an extensive database of collocated GOME and ozone profile measurements is necessary. Ozone profiles from sondes collected by the World Ozone and Ultraviolet Data Centre (WOUDC), as well as HALOE, SAGE II and POAM III limb measurements have been utilised for this purpose, constituting about 70000 training collocations. While training takes a certain amount of time, it is only needed once, the actual retrieval process is by a factor of 103 to 105 faster than classical methods.The NNORSY retrieval results indicate that the neural network successfully extracts information on the tropospheric ozone distribution from the GOME spectra. Overall, there is good agreement with collocated ozone-sondes and a reasonable agreement of large scale ozone field structures in comparison with a tropospheric chemistry and transport model. However, due to the structure of the training data set, some inaccuracies remain at extreme SZAs and over the oceans, where there is little sonde data available.The method is thus not yet optimised for tropospheric ozone retrieval. In the future, we plan to improve error checking and cloud treatment, dataset distribution and possibly training procedures, i.e. by using specialised networks. Adaptation to upcoming sensors is also envisioned, and should be possible with relatively little effort.