The Automated Seismologist: An Application of Deep Clustering to Detect Events and Low Energy Non -Volcanic Tremor in Japan

Abstract

The deep learning model, Deep Embedded Clustering (DEC), has been applied to the Japanese HiNet network to test its ability to identify and cluster different features in a seismic signal. The main focus of this work is to test whether DEC could (1) detect unknown singular patterns that may arise specifically before large or intermediate earthquakes; (2) identify low energy non-volcanic tremor (NVT) more efficiently than direct observation or statistical analysis of the seismic signal (e.g. root mean square or RMS) and (3) produce coherent clusters corresponding to other sources (anthropic activity, weather conditions). Training and testing on high passed data was successful at recognising day night perturbations globally across stations. For the NVT dataset, two clusters identified NVT: one could be verified visually, and the other correlated strongly with the RMS. Building on a previous application of DEC aimed at detecting and classifying anthropic activities in local array data, this thesis has broken new ground in adapting the method to the large scale HiNet network, and in testing DEC potential for seismic event detection and NVT discovery.