Using artificial neural networks in the temperature and humidity sounding of the atmosphere

The application of the radiative data inversion technique based on artificial neural networks (ANN) for the meteorological satellite sounding of the atmosphere is described. To increase the efficiency of solving inverse problems, the principal component method is used for the temperature and humidity profiles, as well as for IR radiation spectra, which allows the problem dimensionalities to be reduced substantially. Based on numerical experiments, errors of the temperature and humidity sounding are analyzed from the spectra of outgoing IR radiation (that were measured by the IKFS-2 instrument onboard the Meteor Russian satellite) using the iterative physical-mathematical (IPM) algorithm, multiple linear regression (MLR), and ANN-based methods. Appreciable advantages of the ANN-based method are revealed as compared to the MLR method. Therefore, in temperature sounding, the MLR method has a markedly large error at heights of 1–12 km (a difference of up to 1 K), while the IPM algorithm has almost the same error as the ANN method. The humidity determination error is about 10% when the ANN method is used at heights of 0–12 km. The IPM approach yields approximately the same error in the lower troposphere, but as the height increases the advantages of the ANN method grow.