A classification model for signals of seismic events based on vision transformer

Abstract

Accurate seismic event classifications offer crucial information to earthquake alarms, and precise classifications in time can effectively reduce casualties and losses in seismic events. This paper proposes a classification model using a Vision Transformer network trained and tested using 8667 time-series records intercepted from seismic observatories. Different forms of input, including Gramian angular field and short-time Fourier transform, are used as the image input, and the regular time-series input is also tested. The precisions of classifications of explosions, collapses, and natural earthquakes are 93.12%, 92.49%, and 92.63%, respectively. The overall accuracy of classification is 92.74%, with short-time frequency maps of multiple channels of seismic signals as input, showing the ability of the proposed network model to extract and identify features of different types of seismic events.