An Empirical Model for Atmospheric Temperature Estimation Using a Geographic Grid Neural Network

This study presents an empirical model for atmospheric temperature (T) estimation based on a Geographic Grid Neural Network (GGNN) framework. As a critical parameter in GNSS tropospheric tomography, T facilitates the conversion of wet refractivity (Nw) to water vapor density (ρv). The proposed GGNNTemp model integrates grid-specific neural networks across a 1°×1° geographic grid, where each node employs a multilayer perceptron with month and altitude as inputs to predict temperature. Comparative evaluations against traditional neural networks and the GPT2w model demonstrate the superior performance of the GGNNTemp model, achieving a BIAS value of 0.01 K and an RMSE of 1.60 K—significantly outperforming GPT2w’s 12.37 K RMSE. The model exhibits consistent accuracy across varying latitudes, altitudes (up to 10 km), and seasons, with ρv conversion errors maintained below 5‰. This work highlights the efficacy of GGNN technology in modeling geophysical parameters and advances its potential for broader applications in GNSS meteorology.