A second generation of the neural network model for predicting weighted mean temperature

In global navigation satellite system (GNSS) meteorology, the weighted mean temperature (T-m) is a variable parameter in the conversion between zenith wet delay errors of GNSS and precipitable water vapor. The combined models of T-m, which are modeled with a combination of T-m seasonal variations and relationships between T-m and site meteorological measurements (mainly site measured temperature), have been proven to be of relatively higher accuracy. In this study, an improved combined model for T-m called the NN-II model was developed and is the second generation of the NN model. Similar to the NN model, NN-II is a combined model and is modeled by using the neural network model. The NN model was only designed for T-m estimates near the surface, while NN-II was designed for T-m estimates from the surface to almost the top of the troposphere. Compared with the NN model, the NN-II model shows some advanced features in terms of model design: modeled T-m data cover from the surface to almost the top of the troposphere, a more accurate seasonal T-m from the GTrop-T-m model is used, and the input variables are refined. Due to these refinements, the bias and RMSE of NN-II for global T-m from the surface to almost the top of the troposphere are 0.08 K and 3.34 K, respectively, and this new model shows 29.1% and 40.6% improved accuracies compared to those of the GTrop-T-m model and the NN model, respectively. The accuracy advantage is maintained over different heights of the troposphere on a global scale.