Machine Learning-Based Detection and Localization of Tectonic Tremors in the Japan Trench

Abstract

Shallow tectonic tremors near trenches have been detected due to the advancement of offshore observation networks. Traditionally, tremors were identified by cross-correlating envelope waveforms between seismic stations. However, this method has struggled to differentiate tremor signals from earthquakes and sometimes missed tremors during active tremor episodes. Addressing these challenges is crucial for monitoring tremors in seismically active regions, such as the Japan Trench. We developed a machine learning-based tremor monitoring system using a dense network of cable-type ocean-bottom seismometers (S-net) in the Japan Trench. The system analyzed continuous waveforms recorded from August 2016 to August 2024. Our analysis detected seven times more tremors than the previous study using envelope cross-correlation. The newly identified tremors expanded the known spatial distribution of tremor activity, both along the strike and dip, revealing a complementary spatial relationship between tremors and earthquakes. Additionally, our catalog improved temporal resolution, uncovering spatiotemporal patterns of tremors synchronized with slow slip events. Seismic energy rates of tremors were calculated, showing spatial variations along the strike and dip, with higher rates near asperities of large earthquakes. A positive correlation between energy rates and recurrence intervals was found, suggesting that these spatial variations reflect frictional heterogeneities on plate boundaries. The enhanced spatiotemporal resolution of our tremor catalog provides valuable insights into the relationship between slow and fast earthquakes.