Revealing Hidden <i>T</i> -Wave Sources in the South China Sea from an Island Seismometer Using Deep Metric Learning

Abstract T waves are hydroacoustic signals generated when seismic waves interact with complex seafloor topography, with energy converted into oceanic acoustic waves that propagate through the Sound Fixing and Ranging channel at acoustic speeds (∼1.5 km/s). In semienclosed basins like the South China Sea, complex bathymetry creates challenging detection conditions due to variable signal patterns. We developed a deep metric learning framework for T-wave identification using land-based seismic data from the South China Sea (2012–2024). Our semisupervised approach categorized earthquakes M ≥ 5.0 into three classes: no T phase, single T phase, and multiple T phase. The deep learning model achieved superior performance with a macro F1-score of 94.61%, particularly excelling in detecting complex multiple T-wave patterns. Application to smaller earthquakes (M &amp;lt; 5.0) revealed 116 previously undetected T-wave events. Spatial analysis showed that T-wave events are concentrated along the western edge of the Philippine plate and the Manila trench, with multiple T-phase events predominantly located east of the Taiwan Island, likely due to complex bathymetry in that region. In addition, a subset of nearshore earthquakes along the Vietnamese coast generates detectable T waves, suggesting that shallow coastal waters may enable seismic-to-acoustic energy coupling. These findings advance understanding of T-wave generation mechanisms in semienclosed basins and provide new insights into the spatial patterns of seismic-to-acoustic conversion. The expanded T-wave catalog and automated detection framework offer valuable tools for marine seismic monitoring and contribute to improved understanding of T-wave propagation in complex bathymetric settings. Because the framework is developed from a single island seismometer, its applicability to other stations or regions remains to be assessed in future studies.