Abstract: To address the limitations of the Saastamoinen model in calculating zenith hydrostatic delay(ZHD)due to its reliance on ground-measured atmospheric pressure data and the lack of meteorological instruments at most GNSS stations, this study proposes an improved method based on a general regression neural network(GRNN).By integrating radio occultation data and radiosonde station observations to construct a training dataset, a GRNN-ZHD prediction model was developed.Combined with ZTD derived from GNSS observations of the Crustal Movement Observation Network of China(CMONOC), a novel model for retrieving precipitable water vapor(PWV)was established.The results demonstrate that the GRNN model achieves an average RMSE of 15.23 mm for ZHD retrieval, showing a 46.8% improvement compared to the GPT3 model(28.64 mm).For PWV retrieval, the GRNN model achieves an average RMSE of 5.17 mm, outperforming the GPT3 model's 10.76 mm (51.9% accuracy improvement).Among the 20 validation stations, the GRNN model maintains PWV retrieval deviations below 7 mm at 15 stations, whereas the GPT3 model achieves this threshold at only 3 stations.

Key words: Saastamoinen model, zenith hydrostatic delay, general regression neural network, precipitable water vapor