A Weighted Mean Temperature Forecast Model Based on Fused Data and Generalized Regression Neural Network and Its Impact on GNSS-based Precipitable Water Vapor Estimation

The weighted mean temperature (Tm) is a crucial variable in mapping zenith wet delays from the global navigation satellite system (GNSS) to precipitable water vapor (PWV). Existing empirical models for Tm estimation mainly utilize a single data source, either radiosonde (RS) data or reanalysis data. And these models assume that the Tm follows a predefined linear periodic function, failing to capture the detailed nonlinear Tm variations, and yielding low accuracy. This study developed a Tm forecast model (GRNN-F) using the generalized regression neural network (GRNN) and fused Tm data. The performance of GRNN-F was evaluated using Tm from the RS sites not involved in modeling. The results demonstrate strong agreement between GRNN-F and RS, with a bias of 0 K, a standard deviation (STD) of 3.26 K, and a root mean square (RMS) error of 3.26 K. Compared with the traditional predefined function-based model (GPT3) and the linear model (GTm), GRNN-F exhibits a 28.35% and 34.67% reduction in STD and a 30.93% and 36.58% reduction in RMS. Compared to the single data source-based models, GRNN-F demonstrates a significant advantage in Tm forecast, especially at moments with more and larger sudden Tm variations. Moreover, GRNN-F outperforms two comparative models based on random forest and backpropagation neural network on the same fused data across different cases. The theoretical mean PWV relative error derived from GRNN-F is only 1.15%, with less than 1% of the sites accounting for 19%, whereas other models fail to achieve this level of accuracy at any site.