Automatic classification of triggered tectonic tremor with deep learning

Studying instant dynamic triggering of tectonic tremor by surface waves from large distant earthquakes contributes to understanding background ambient tremor activity and faulting mechanisms in the tremor source zone. Detection, identification, and verification of the subtle signals associated with triggered tremor requires tedious interactive examination of thousands of seismograms per earthquake as well as investigations and eventual exclusion of glitches that look like triggered tremor. In this study, we apply a machine learning algorithm - a deep convolutional neural network trained with pairs of data points (a Siamese network) - to detect triggered tremor. To assemble training data, we choose 43 regions in which triggered tremor has been previously reported or was detected by our own work in progress. From all available seismic stations in these regions, we collect waveforms in windows containing surface waves from large earthquakes, which include plain surface waves as well as surface waves with triggered tremor signals superimposed. These regions include Japan (western and eastern Shikoku, southern and northern Kii, Tokai, Kanto, western and southeastern Kyushu, central Hokkaido, and southwestern Ryukyu Trench), Alaska (Aleutian Arc and central-south Alaska), northwest Canada (eastern Denali fault and Queen Charlotte Margin), Cascadia subduction zone (northern and southern Vancouver Island, Washington State and southern Oregon), California (Parkfield, San Jacinto fault, and Calaveras fault), Guerrero in Mexico, Cuba, Haiti, central Colombia, Ecuador, southern Chile, New Zealand (central and northern Hikurangi Margin, central Alpine Fault, Kaikoura, Southland, and Northland), Taiwan (central and northern Central Range and in the northeastern and southeastern offshore areas), Indonesia, and southern Italy. We are using pairs of waveforms from stations in these regions to train our convolutional neural network. We will report on the efficacy and efficiency of this machine learning approach and compare it to the performance of other approaches to this problem, such as detecting triggered tremor via a signal to noise ratio, or using apparent peak ground velocity as a likelihood predictor.