Zenith Wet Delay Fusion Based on Generalized Regression Neural Network

Objectives Accurate modeling of zenith wet delay (ZWD) is crucial for enhancing the precision of global navigation satellite system (GNSS) positioning and meteorological applications. While multiple data sources exist for ZWD modeling, their comprehensive utilization is challenged by multi-source heterogeneity, uneven precision, and significant systematic biases.

Methods This paper employs a generalized regression neural network which is known for its strong nonlinear approximation capability, to correct the relatively lower-quality ZWD products from global/regional assimilation and prediction enhanced system-mesoscale (GRAPES) model and the the fifth generation ECMWF reanalysis (ERA5). The correction uses high-quality radiosonde (RS) ZWD as a reference. The modified GRAPES ZWD and ERA5 ZWD are then fused with RS ZWD to achieve a bias-free multi-source ZWD product, resulting in 1 032 fused ZWD grid sets.

Results Validation against RS ZWD in 2016—2017 reveal systematic biases in the original GRAPES ZWD and ERA5 ZWD. After correction, the overall bias of both datasets approaches zero, and root mean square (RMS) is reduced by 22.6% and 16.0%,respectively. The spatiotemporal variations of bias, standard deviation, and RMS are also significantly attenuated.

Conclusions The fused ZWD product demonstrates high accuracy and stability, offering a reliable data source for advanced GNSS applications.