Improving Seismic Tomography of Southern Taiwan and Eastern Offshore Regions from the SALUTE Amphibious Array and Machine Learning-Based Arrival Time Analysis

Seismic tomography is vital for understanding complex tectonic processes and assessing geological hazards in seismically active regions like Southern Taiwan, where the Eurasian continental subduction transitions into the Luzon Arc collision. To enhance the lateral and depth resolution of tomographic images, we expanded station coverage and addressed data gaps in mountainous areas by deploying the dense amphibious SALUTE array across this transition zone since October 2021. A machine learning (ML) approach was employed for precise and efficient P- and S-phase picking from three-component velocity waveform data recorded by totally 236 seismic stations, integrating the SALUTE array with existing permanent networks in southern Taiwan.

Using SeisBench and the pre-trained EQTransformer model for automatic phase picking, followed by phase association and event’s origin time and hypocenter determination with the Gaussian Mixture Model Associator (GaMMA) and event relocation using the hypo3D program with a 3D velocity model, we detected four times more seismic events than the CWA catalog for one-month test data. Moreover, P amd S arrival picks increased by about 1.7 and 1.8 times, respectively, demonstrating the potential of ML methods to improve earthquake catalogs and the following tomographic inversion.

After thoroughly validating all phase picks and removing false detections, we applied the LOTOS-12 tomography package to iteratively invert for 3-D P- and S-wave velocity models while simultaneously relocating earthquake sources. Preliminary results reveal a prominent velocity contrast of up to ±20% in the upper 20 km of the crust across the Chaochou Fault (CCF), with low velocities beneath the western sedimentary coastal plain and high velocities beneath the eastern metamorphic Central Range (CR). These onshore features are generally consistent with previous models. Besides, low-velocity crustal anomalies deepen eastward, and are underlain by a high-velocity structure, likely representing the eastward-subducting Eurasian slab. Incorporating offshore stations in our dataset has enabled us to resolve two isolated, shallow low-velocity anomalies beneath the Southern Longitudinal Trough and the Luzon Arc. Expanding station coverage along the CCF and in the CR and eastern offshore regions through the SALUTE array, we anticipate significant improvements in the tomographic images, with enhanced lateral and depth resolution and greater structural detail.