Automatic GNSS Coordinate Time Series Analysis for Very Large Networks
4 affiliations
- 1NOAA‐National Ocean Service National Geodetic Survey Silver Spring MD USA
- 2Now at Forterra Clarksburg MD USA
- 3NOAA‐National Ocean Service National Geodetic Survey Kansas City MO USA
- 4NOAA‐National Ocean Service National Geodetic Survey Seattle WA USA
Abstract
We present a new algorithm, AutoCF, for automatically constructing coordinate functions (CFs) for continuously operating GNSS reference stations (CORSs). The iterative algorithm alternates between application of LASSO and Ridge Regression methods, with hyperparameters successively tightened with each iteration to more judiciously scrutinize candidate breakpoints. We developed the algorithm using data from 60 CORSs from two independent analysis centers. These CORSs experienced displacements associated with four large magnitude earthquakes. Some stations are located in the near‐field of two of these earthquakes, while the other two earthquakes occurred at intermediate and far‐field distances from the nearest station. We show that automatically determined CFs track the observed site motion with weighted RMS (WRMS) residual scatter less than 2–3 mm and 5–7 mm for horizontal and vertical components, respectively. We tuned the algorithm such that estimates for site velocity are as close to the estimates generated by the independent processing centers as those centers' results are to one another. We then assessed the algorithm's performance on a new, unseen data set comprising 3226 coordinate time series generated independently at NOAA's National Geodetic Survey. The median WRMS misfits of the automatic fits to the time series are 1 mm for horizontal and 4 mm for vertical components. 587 stations (18%) were flagged for human assessment. A velocity solution computed with station trajectories parameterized according to the algorithm‐assisted analysis agrees with similar large‐scale solutions produced by three independent analysis centers at the level of 0.2 mm/yr in horizontal components.
Versions
Related
- Assessing Spatial and Temporal Variability of Vertical Land Motion Along Coastal Massachusetts Using Sentinel-1 InSAR and GNSS Time Series Analysis
- Convolutional LSTM and Self-Attention Networks in PyTorch for Decade-Ahead Forecasting of Land-Cover Time Series
- Large-scale flood monitoring based on time-series Sentinel-1 images and Z -score
- Multi-Band, Multi-GNSS Polarimetric GNSS-R Observations From SCOMAG Airborne Campaign
- Assessment of NAPGD2022's Performance in Oregon's Willamette Valley Using Astronomical Leveling and GNSS/Leveling
BibTeX
@article{bennett2026automatic,
title = {{Automatic GNSS Coordinate Time Series Analysis for Very Large Networks}},
author = {Richard A. Bennett and Phillip K. McFarland and Amy Whetter and Jordan Krcmaric},
journal = {Earth and Space Science},
year = {2026},
month = oct,
doi = {10.1029/2025ea004741},
url = {https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025EA004741},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Automatic GNSS Coordinate Time Series Analysis for Very Large Networks
AU - Richard A. Bennett
AU - Phillip K. McFarland
AU - Amy Whetter
AU - Jordan Krcmaric
PY - 2026
DA - 2026/10/10
JO - Earth and Space Science
DO - 10.1029/2025ea004741
UR - https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025EA004741
AB - We present a new algorithm, AutoCF, for automatically constructing coordinate functions (CFs) for continuously operating GNSS reference stations (CORSs). The iterative algorithm alternates between application of LASSO and Ridge Regression methods, with hyperparameters successively tightened with each iteration to more judiciously scrutinize candidate breakpoints. We developed the algorithm using data from 60 CORSs from two independent analysis centers. These CORSs experienced displacements associated with four large magnitude earthquakes. Some stations are located in the near‐field of two of these earthquakes, while the other two earthquakes occurred at intermediate and far‐field distances from the nearest station. We show that automatically determined CFs track the observed site motion with weighted RMS (WRMS) residual scatter less than 2–3 mm and 5–7 mm for horizontal and vertical components, respectively. We tuned the algorithm such that estimates for site velocity are as close to the estimates generated by the independent processing centers as those centers' results are to one another. We then assessed the algorithm's performance on a new, unseen data set comprising 3226 coordinate time series generated independently at NOAA's National Geodetic Survey. The median WRMS misfits of the automatic fits to the time series are 1 mm for horizontal and 4 mm for vertical components. 587 stations (18%) were flagged for human assessment. A velocity solution computed with station trajectories parameterized according to the algorithm‐assisted analysis agrees with similar large‐scale solutions produced by three independent analysis centers at the level of 0.2 mm/yr in horizontal components.
SN - 2333-5084
SN - 2333-5084
ER -