GNSS-FM: A Self-Supervised Foundation Model for Daily GNSS Displacement Time Series

2026-06-05 · Source: Machine Learning · Field: Science & Research — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences, Environmental Science & Earth Systems · Depth: Expert, quick

Summary

GNSS-FM is a novel self-supervised foundation model designed for analyzing daily Global Navigation Satellite Systems (GNSS) displacement time series. Addressing the scarcity of labeled data in traditional supervised machine learning approaches for GNSS applications, GNSS-FM utilizes a dual-stream input that combines displacement and velocity-like increments. The model is pretrained using a masked latent prediction objective, adapted from wav2vec 2.0 with specific modifications for geodetic data, on an extensive dataset from over 17,000 globally distributed GNSS stations. Analysis of its learned codebook reveals that the model effectively captures key signal types, including seismic offsets, tectonic drift, and seasonal patterns. When fine-tuned on downstream tasks such as 90-day displacement forecasting and seismic step localization, GNSS-FM consistently outperforms existing task-specific baselines, demonstrating the significant potential of self-supervised pretraining for GNSS time series analysis.

Key takeaway

For research scientists developing machine learning solutions for geodetic applications, GNSS-FM demonstrates a compelling shift from supervised models. You should consider adopting self-supervised pretraining approaches, especially when labeled GNSS displacement data is scarce. This method significantly improves performance in tasks like 90-day displacement forecasting and seismic step localization. It offers a robust way to leverage vast amounts of unlabeled GNSS data for enhanced accuracy and broader applicability.

Key insights

GNSS-FM introduces self-supervised pretraining to overcome labeled data scarcity in GNSS time series analysis, outperforming supervised baselines.

Principles

• Self-supervised pretraining addresses labeled data scarcity.
• Dual-stream input improves time series representation.
• Pretraining captures diverse geodetic signal types.

Method

GNSS-FM uses a dual-stream input (displacement and velocity-like increments) and is pretrained via masked latent prediction with vector-quantized targets, adapted from wav2vec 2.0 for geodetic data.

In practice

• Forecast 90-day GNSS displacement.
• Localize seismic steps accurately.

Topics

• GNSS Displacement Time Series
• Self-Supervised Learning
• Foundation Models
• Geodetic Data Analysis
• Earthquake Cycle Monitoring
• Wav2vec 2.0 Adaptation

Best for: AI Scientist, Research Scientist

Related on AIssential

• From Short Histories to Long Futures: Horizon-Aware Graph Neural Networks for Long Horizon Forecasting — A multi-horizon GNN emulator improves long-range geophysical forecasting by learning state transitions and predicting increments.
• Physics-Guided Spatiotemporal Learning for Coastal Wave Peak Period Estimation from Video — Physics-guided deep spatiotemporal learning from video can accurately and consistently estimate coastal wave peak periods, overcoming traditional monitoring limitations.
• Self-Supervised Learning with Gaussian Processes — GPSSL uses Gaussian processes to learn data representations, offering inherent similarity grouping and uncertainty quantification.
• Time-Series Foundation Model Embeddings for Remaining Useful Life Estimation — Frozen time-series foundation models like Chronos-2 enable data-efficient RUL estimation with a lightweight regression head.
• MP3: Multi-Period Pattern Pre-training forSpatio-Temporal Forecasting — MP3 pre-trains multi-period patterns from long time series to resolve "temporal mirage" in spatio-temporal forecasting.
• LatentWave: JEPA Pretraining for Wireless Foundation Models — JEPA pretraining for wireless foundation models yields transferable representations by predicting latent space masks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.