ChangePointCNN‐GNSS: An AI Model for Assessing Change Points and Optimizing Site Velocity Estimation From Global GNSS Data

Abstract Estimating long‐term site velocities from Global Navigation Satellite System (GNSS)‐derived daily displacement time series is vital for studying secular tectonic motions and establishing regional and global geodetic reference frames. However, this estimation is complicated by displacements caused by earthquakes, equipment changes, hydraulic head changes, and other sources, which introduce change points in GNSS time series. This study introduces a two‐stage hybrid framework for automated change‐point detection in GNSS time series. The framework integrates (a) analytical methods, including a sliding‐window algorithm for instant change‐point detection and a cubic polynomial fit for transitional change‐point detection and (b) an artificial intelligence (AI) model, ChangePointCNN‐GNSS, which evaluates the suitability of candidate change points for site velocity estimation and iteratively optimizes analytical parameters. Unlike prior data‐driven approaches, our framework leverages an image‐driven method, employing a convolutional neural network (CNN) to visually assess and select the most suitable change‐point configuration for reliable site velocity estimation. Site velocities are computed from the longest change‐point‐free segment (minimum 4 years), processed independently for each station and direction. This integrated approach ensures robust site velocity estimation across large GNSS networks. The CNN is trained using approximately 6,000 time series plots with marked change points. Each plot is labeled as “good” if the detected change points are suitable for reliable site velocity estimation or “bad” if unsuitable. This study delivers long‐term site velocities (IGS20) for approximately 14,600 permanent GNSS stations worldwide, with a 95% confidence interval below 1 mm/year, offering a foundational data set for researchers in geodesy, tectonophysics, and hazard mitigation.