Abstract

Ground displacements offer crucial insights into Earth’s deformation, yet their analysis is challenged by the diverse processes influencing the signals. We highlight here two distinct, data-driven approaches aimed at enhancing our understanding and monitoring of geophysical processes using daily to bi-weekly geodetic datasets assessing both short-term and long-term seasonal deformation processes.  The first project introduces a Temporal Convolution Network (TCN) designed to assess non-tidal loading signatures in Global Navigation Satellite Systems (GNSS) displacement time series. By leveraging a continuous global dataset from nearly 12,000 GNSS stations covering January 2002 to June 2024, the TCN model evaluates the non-tidal influences that are often masked by seasonal oscillations and other artefacts. The project successfully improves upon traditional numerical loading products by reducing the root mean square error (RMSE) by approximately 4%, thereby offering a robust tool for validating and enhancing existing loading models.  The second project presents a proof-of-concept Convolutional Neural Network (CNN) aimed at generating strain maps directly from stacks of wrapped Interferometric Synthetic Aperture Radar (InSAR) interferograms. Although this approach is still a work in progress, early results demonstrate its potential to extract subtle deformation signals and produce high-resolution strain estimates, particularly in regions where GNSS coverage is sparse. Together, these projects illustrate the capability of machine learning methodologies to advance geodetic signal analysis, with the completed TCN model providing immediate benefits to nontidal loading assessments, and the evolving CNN approach paving the way for future developments in InSAR-based strain mapping.