A Machine-Learning-Based Missing Data Interpolation Method for GNSS Time Series

The long-term GNSS time series contain abundant geodynamic and geophysical information. These time series are inevitably subject to random or continuous missing data caused by environmental or human activities. However, many mathematic methods cannot be directly applied to analyse these time series when missing data arise. Though some efforts have been made to interpolate the random missing data, the long gaps still cannot be interpolated properly. This paper seeks to apply machine learning (ML) models in missing data interpolation. These ML models are trained by using 12 site-motion-related physical variables, including the Sun’s and Moon’s coordinates, temperature, atmospheric pressure, and hydrology. Then the missing data are generated from these trained ML models. This method is tested on seven GNSS stations, and the results show that the interpolation precision can averagely reach 5.5 mm for 2-year gaps, and 4.9 mm for 1-year gaps. This new missing data interpolation method shows good performance in restoring incomplete time series, which will be helpful for further analysis of GNSS time series.