Use of neural network approach in determining total and tropospheric column ozone content based on measurements of outgoing thermal radiation

For monitoring ozone total column (OTC) and tropospheric ozone (TRO) in the vertical column of the atmosphere, only satellite remote sensing techniques can provide global continuous measurements can be provided only by satellite remote sensing methods. Among them, methods based on measurements of the spectra of outgoing thermal infrared radiation allow us to obtain information on OTC and TRO in the absence of solar radiation, including during the polar night period.
We consider techniques for obtaining OTC and TRO on the basis of such spectra measured by the IKFS-2 instrument from Meteor M satellite No. 2. The techniques are based on artificial neural networks (ANN) and principal component methods. The approximation error of OTC when trained with ANN is on average close to 3 %, in the tropical region it is less than 2 %, and in the polar regions in winter-spring period it increases up to 6-8 %. The methodology for determining the OTC has been described many times before, so we analyse in detail in the report the ANNs only for the determination of TRO. For this purpose, we considered different training datasets and optimised the structure of the ANN. The approximation error of the TRO when training the optimal ANN is about 3.4 e.d.. 
The methods were applied to the processing of spectra measured from Meteor M satellite No. 2 in 2015-2022. The obtained OTC values were compared with data from the TROPOMI instrument on the S5P satellite and ground-based observations (Dobson and Brewer instruments, direct measurements of the Sun). The standard deviations of the differences are about 2.7% for both comparisons. The results of the TRO determination were compared with data from stations in the IRWG-NDACC network. On average, the standard deviations of the differences between satellite and ground-based TRO observations are about 3 DU, which is ~15% of the TRO derived from FTIR measurements. The spatial and temporal variability of OTC and TRO is analysed, and examples of measurement results are presented. In this paper, we have shown that, although the regression-based approach is not considered optimal for solving inverse remote sensing problems, the adequacy and completeness of the training dataset allows obtaining valuable results. We would also like to draw attention to the fact that the use of the simplest ANN - a perceptron with one hidden layer - allows obtaining quite satisfactory results. 
The research was supported by the St. Petersburg University (project NO. 124032000025-1).