Zenith Wet Delay Fusion Based on A Generalized Regression Neural Network

  Objectives:   Accurate modeling of zenith wet delay (ZWD) is beneficial to improving the accuracy of GNSS navigation and positioning and GNSS meteorological applications. There are many data sources for ZWD modeling. However, there are problems such as multi-source heterogeneity, unequal precision, and even severe system deviation, making it challenging to utilize multi-source ZWD data comprehensively.   Methods:   To address this issue, a generalized regression neural network (GRNN) with strong nonlinear approximation capability and the high-quality radiosonde (RS) ZWD is used to modify the relatively low-quality GRAPES_MESO (GRAPES) ZWD and ERA5 ZWD. And then, the modified GRAPES/ERA5 ZWD and RS ZWD are combined to realize the unbiased fusion of multi-source ZWD, and 1032 ZWD fusion images are obtained.   Results:   Validated by the RS ZWD, there are different degrees of systematic biases between GRAPES/ERA5 ZWD and RS ZWD in 2016-2017. After modification, the overall Bias of GRAPES and ERA5 ZWD are close to 0, and the RMS is decreased by 22.6% and 16.0%, respectively. The spatiotemporal variations of GRAPES/ERA5 ZWD Bias, STD, and RMS are better attenuated by modification.   Conclusions:   The fused ZWD has high accuracy and stability and can provide a reliable ZWD data source for GNSS applications.