This research introduces InSARDenoiser, a deep-learning framework designed to extract small-amplitude (millimeter-scale) ground deformation from noisy Interferometric Synthetic Aperture Radar (InSAR) time series. The architecture utilizes a spatiotemporal attentive convolutional U-Net, combining a deep spatial U-Net with a Transformer-based bottleneck to simultaneously track transient deformation in space and follow its temporal signature. To overcome the scarcity of labeled geodetic data, the model was trained using a mathematical formulation that generates realistic multi-scale atmospheric noise. Validated with real data from the North Anatolian Fault, the model successfully retrieved millimeter-scale deformation associated with a slow earthquake, with high correlation with independent local creepmeter measurements.

This study presents SSEdetector, an end-to-end deep-learning-based pipeline for the detection of Slow Slip Events (SSEs) in raw, non-post-processed multi-station GNSS time series. The model integrates a Convolutional Neural Network (CNN) for feature extraction and a Transformer with a self-attention mechanism to identify temporal transients buried in geodetic noise. Trained on realistic synthetic datasets derived from the geometry of the Cascadia subduction zone, the detector identified 78 SSEs between 2007 and 2022. We observe a ~2-day time lag between the initiation of slow deformation and the onset of tremor, suggesting that aseismic slip may trigger the rupture of nearby small-scale asperities.

This study addresses the limitations of single-station and 2D-CNN geodetic analysis by introducing SSEdenoiser, a Spatiotemporal Graph Neural Network (STGNN) that leverages spatiotemporal information. Unlike traditional methods, this architecture treats the GNSS network as a graph with learnable adjacency matrices, enabling the model to directly uncover hidden spatial relationships between GNSS stations from the data. By coupling graph recurrent units with spatiotemporal Transformers, the model extracts aseismic deformation transients with sub-millimeter precision. When applied to the Cascadia region, the denoised displacement rates showed a remarkable spatiotemporal correlation with independent tremor catalogs, effectively retrieving slip migration and recognizing events closely spaced in time and space.

Utilizing 15 years of high-resolution denoised GNSS data, we revisit the scaling laws and rupture mechanics of slow slip on the Cascadia megathrust. The study demonstrates that SSE scaling laws (moment-duration) are significantly influenced by detection thresholds; however, all events are constrained by a fundamental upper bound on moment rates. The data reveal an unbounded-to-bounded rupture transition at magnitude M≈6.2. Smaller events initiate as 2D expanding cracks (circular ruptures) within the transition zone, while larger events, upon reaching the geometric boundaries of this zone (~40 km width), transition to 1D pulse-like lateral propagation. This evidence supports a physical continuum of slow slip events of varying sizes, modulated by the plate interface geometry.

This research evaluates the efficacy of different geodetic data representations—1D time series, 2D differential images, and 3D image time series (ITS)—for the automatic inversion of seismic source parameters. The study identifies that ITS-based Transformer models significantly outperform other architectures by effectively leveraging the spatial coherency and temporal variability of GNSS positioning. Specifically, the ITS representation better constrains the wavelength of deformation, improving the characterization of offshore and deep earthquakes where the signal-to-noise ratio is typically low. The framework demonstrated its robustness on real datasets at the Japan Trench, effectively estimating magnitudes and centroid locations for both subduction and crustal thrust events.