Fig. 1: Schematic architecture of SSEgenerator and SSEdetector.
From: Multi-station deep learning on geodetic time series detects slow slip events in Cascadia
a Overview of the synthetic data generation (SSEgenerator). In the matrix, each row represents the detrended GNSS position time series for a given station, color-coded by the value of the position. The 135 GNSS stations considered in this study are here shown sorted by latitude. The synthetic static displacement model (cf. b panel), due to a Mw6.5 event, at each station is convolved to a sigmoid to model the SSE transient, and is added to the ultra-realistic artificial noise to build synthetic GNSS time series. b Location of the GNSS stations of MAGNET network used in this study (red triangles). An example of synthetic dislocation is represented by the black rectangle, with arrows showing the modeled static displacement field. The heatmap indicates the locations of the synthetic ruptures considered in this study, color-coded by the slab depth. The dashed black contour represents the tremor locations from the PNSN catalog. c High-level representation of the architecture of SSEdetector. The input GNSS time series are first convolved in the time domain. Then, the Transformer computes similarities between samples of each station, learning self-attention weights to discriminate between the relevant parts of the signals (here, slow slip transients) and the rest (e.g., background noise), and a probability value is provided depending on whether slow deformation has been found in the data.
