In order to retrieve the atmospheric ozone distribution from the UV-visible satellite spectrometer GOME (Global Ozone Monitoring Experiment), we have modelled inverse radiative transfer directly, using a multilayer perceptron (MLP) type neural network. This network was trained on a data set of measured GOME radiances as input, and collocated ozone profile measurements from ozonesondes, SAGE II, HALOE and POAM III as target values. A partial training method has been devised for dealing with incomplete target data, because neither occultation instruments nor ozone sondes cover the full retrieval height range (1-60 km). Around 70000 collocations from 1996 to 2001 were used for training, and another 12000 for cross-validation in a test data set. Network input consists of a combination of spectral, geolocation and climatological information (latitude and time), the latter making the use of external a priori ozone profiles unnecessary. We designate our method as Neural Network Ozone Retrieval System (NNORSY).

In the stratosphere, the method globally reduces standard deviation (StD) with respect to a well-tried ozone climatology by around 40%. Tropospheric ozone can also be retrieved in many cases, reducing the StD by 10-30% globally. The neural network was found capable of correcting for instrument degradation, pixel cloudiness and scan angle effects on its own, based on the input data provided. Remaining inhomogeneities in the geographical distribution of training data, combined with differing ozone field variability causes retrieval quality to vary with latitude and season.

Since retrieval only requires one forward propagation through the network, NNORSY is about 10³-10⁵ times faster than classical, local retrieval techniques like, for instance, Optimal Estimation. Therefore the method is well suited for real-time application and fast data reprocessing. An operational near-real-time prototype of the system is already running successfully at three GOME receiving stations.

In order to better characterize single output ozone profiles, a number of local error estimation methods has been investigated. Vertical resolution of the profiles was assessed empirically in comparisons with the high-resolution collocations, and seems to be in the order of 4-6 km.

Further developments of NNORSY could involve improvement of training data composition, input parameter optimization, more sophisticated network error functions and training methods, as well as adaptation to other sensors.