Global Positioning System (GPS) radio occultation is considered as a global sounding technique for providing atmospheric profiles useful for numerical weather prediction and climate change studies. The radio occultation system employs GPS receivers placed on a Low-Earth Orbit (LEO) satellite to sound the Earth's troposphere and ionosphere evaluating the additional delay affecting a radio signal when passing through the atmosphere due to the refractivity index. Several investigations have demonstrated that the accuracies about the retrieval of refractivity, temperature, pressure and humidity in the troposphere are comparable with traditional atmospheric remote sensing techniques.

The accuracy of tropospheric profile estimation is affected by the use of proper boundary conditions and by the presence of water vapor particularly in the lower troposphere, that complicates the interpretation of refractivity. Even though the atmospheric refractivity profiling by radio occultation is a well-defined problem, care must be taken to analyze factors affecting the interpretation of refractivity (N) such as the superrefraction or ducting that is commonly caused by big lapse of humidity on top of moist planetary boundary layer. The main lapse in refractivity, or large N-gradient, occurs in the interfacial layer on the top of planetary boundary layer that is characterized by significant decrease of partial pressure of water vapor and increase of temperature. Tropospheric profiles determination in wet conditions requires temperature profile availability at each GPS occultation from independent observations (i.e. radiosoundings or ECMWF data) to compute humidity profiles from the refractivity profile.

The purpose of this work is to create a method based on neural networks to solve tropospheric profiling in wet conditions overcoming the constraint of temperature profile availability at each GPS occultation from independent observations. Neural networks are trained with refractivity profiles as input, computed by Abel inversion of the optimized bending angle profiles obtained from the COSMIC (Constellation Observing System for Meteorology Ionosphere and Climate) Microsat Constellation satellites and provided by the COSMIC Data Analysis and Archive Center (CDAAC) of Boulder (Colorado). The outputs for the neural networks training are the dry and wet refractivity components obtained using geopotential, temperature, specific humidity and logarithmic surface pressure from the corresponding ECMWF 91 model levels analysis data. The choice to use directly dry and wet refractivity ECMWF profiles as network outputs instead of the dry and wet refractivity percentage from ECMWF profiles is under investigation. Also, the choice to estimate the partial pressure of dry air together with the dry and wet refractivity components from the network, instead of solving the ideal gas and hydrostatic equilibrium laws in dry conditions is under investigation as well. Preliminary results have shown that the error introduced by the neural networks is significantly lower with respect to the one exhibited after the integration of the hydrostatic equilibrium law. For a fast approach the Principal Components analysis is applied in the neural network training phase, expanding the high tropospheric levels on a basis of empirical orthogonal functions. The Principal Components technique permits a reduction of the number of descriptive profile parameters by exploiting the correlation among values at different altitudes.

The neural network training and the following independent test are performed over the entire ocean area between Tropics by using data on summer 2006. The output decomposition of the refractivity components together with the estimation of the dry pressure allows to retrieve temperature and pressure of water vapor from atmospheric refractivity formula at microwave wavelength, without the knowledge of the temperature profile as an independent source of information. Preliminary results regarding the tropical land area in the same temporal coverage have shown good performances of the neural networks using the Principal Component analysis for a fast approach, and a fair accuracy for wet profile estimation. Furthermore, such results can be improved with the availability of a larger data set for the neural network training covering a wider variety of atmospheric conditions.