GNSS-FM Exposes How Everyday Satellite Positioning Infrastructure Can Be Rebuilt as a Global Ground-Motion Observatory

The arXiv preprint GNSS-FM introduces a self-supervised foundation model pretrained on daily displacement time series from more than 17,000 globally distributed GNSS stations, using a dual-stream architecture that ingests both raw displacements and velocity-like increments before applying a wav2vec 2.0-style masked latent prediction objective with vector-quantized targets adapted for geodetic signals. This methodology sidesteps the labeled-data bottleneck that has constrained supervised ML applications in geodesy, where seismic offsets, tectonic drift, and seasonal loading dominate the signal space. Unlike earlier task-specific models that required hand-labeled earthquake catalogs, GNSS-FM learns a compact codebook whose entries align with physically meaningful regimes, then transfers to 90-day displacement forecasting and seismic-step detection with measurable gains over strong baselines. The work remains a preprint and has not undergone peer review; its 17k-station corpus, while large, still under-samples tectonically active regions with sparse receiver networks and inherits the well-known multipath and reference-frame artifacts common to public GNSS archives. What the original coverage underplays is the infrastructure-level inversion underway: the same continuous GNSS signals already collected for surveying, autonomous navigation, and telecommunications timing are being silently repurposed into a planetary-scale deformation sensor without new hardware. This pattern echoes the earlier migration of weather models from physics-only to foundation-model hybrids and suggests GNSS data may soon feed real-time strain-rate fields for operational earthquake and volcano early-warning systems. Limitations include the daily sampling rate, which precludes detection of sub-daily transients, and the absence of uncertainty quantification in the current codebook representations. Two related works underscore the trajectory: the 2023 FourCastNet paper demonstrated that self-supervised pretraining on global atmospheric reanalysis yields skillful medium-range forecasts, while a 2024 study on self-supervised seismic waveform embeddings showed analogous transfer gains for event detection. GNSS-FM therefore sits at the intersection of these trends, indicating that ubiquitous location infrastructure is quietly becoming the next large-scale unlabeled corpus for geophysical foundation models.