A Hybrid Machine Learning Approach for Modeling Tropospheric Zenith Wet Delay with Enhanced Generalization Performance.

Tropospheric zenith wet delay (ZWD) is one of the major error sources for space geodetic techniques and plays a vital role in meteorological research.  Accurate prior estimates for ZWD can significantly improve the performance of geodetic applications, such as precise kinematic positioning. Current single machine learning ZWD models have limitations in modeling the high spatiotemporal variations of moisture in the lower atmosphere and in their generalization capabilities. To mitigate these limitations, this work introduces a hybrid learning framework that combines multiple machine learning models. The proposed model offers stronger generalization capabilities, improving the ZWD modeling and forecasting accuracy.

When comparing the RMSE, the proposed model outperforms existing machine and deep learning-based ZWD models, the empirical GPT-3 model, and the traditional models such as the Saastamoinen and Askne & Nordius models. In the blind case, when surface meteorological data are not used, the RMSE is reduced by 25.76% compared to the GPT-3 model. When using surface meteorological parameters, the proposed model achieves improvements of 47.05% and 34.24% compared to Saastamoinen and Askne & Nordius, respectively.

The generalization capabilities of the models were evaluated at non-modeled sites. The proposed model demonstrates improvements in overall external performance, with a particularly significant increase of 26.14% in the blind case compared to GPT-3. When sites access meteorological data, the model shows improvements of 45.23% and 34.31% compared to Saastamoinen and Askne & Nordius, respectively.

The spatiotemporal analysis shows the improved stability and precision of the proposed model over the other models evaluated in this work, indicating promising prospects for it in real-time and rapid geodetic applications.