Single station seismic observations enable high resolution localization of tectonic tremor sources using a vision transformer

Abstract

Abstract Precisely locating tectonic tremor events is essential for understanding subduction zone dynamics. However, the lack of impulsive phase arrivals typically necessitates multi-station observations, posing a fundamental challenge for high-resolution monitoring using limited seismic records. Here, we show that single-station three-component wavefields can encode sufficient information to constrain tremor source locations using TremorViT, a vision transformer-based framework designed to estimate three-dimensional coordinates together with aleatoric uncertainties. Applied to the Nankai subduction zone, TremorViT achieves a mean epicentral error of 4.8 km using a single station, which improves to 2.5 km when independent station-level estimates are integrated. Using this approach, we localized 68,101 tremor events between January and September 2016, corresponding to approximately 1135.0 h of cumulative tremor activity. These metrics are substantially larger than those reported in existing network-based catalogs. This discrepancy likely reflects differences in detection sensitivity, event definition, and temporal windowing, even as the model preserves previously documented spatiotemporal patterns. This framework enables high-resolution tremor detection and offers a more granular view of tectonic activity, including potential signals associated with geodetically undetected slow slip events.