Single Station Discrimination Between Tectonic Tremors and Regional Earthquakes: Case Studies in Taiwan and Parkfield

Abstract:

Non-volcanic tremor occurs not only in subduction zones of Pacific Plate (Obara and Kato, 2016) but in transform faults like San Andreas Fault (Nadeau and Dolenc, 2005). In Taiwan, ambient tremor occurs underneath a mountain in southern Central Range (Chuang et al., 2014). To identify tremor, the similarity and time lapse of the arrival of tremor bursts from multiple stations is demanded. Manual checks of multi-stations waveforms are oftentimes needed, to exclude loud noise and swarms of regional earthquakes. In this study we pay special attention to swarms of regional earthquakes (epicenter distance of 30-200 km) that contaminated the tremor identification in Taiwan. We want to know whether the advances of machine learning techniques enable an automatic search for patterns to discriminate tremor from regional events.

Here we chose k-Nearest Neighbor (kNN) by Cover and Hart, 1967 as classifier for its low requirement of hardware and fast computational speed. During the study period of 2016/Jan/1-Dec/31, 6,069 tremor and 7,779 M ≥ 3.0 regional earthquakes events at 3 CWBSN stations (ELD, STY, and WTP) are labeled and extracted features including maximum amplitude, energy of 2-8-Hz-band-pass-filtered waveform, etc. After building kNN model for each station we successfully differentiated the 2 classes with high accuracy and recall of 99.4 % and the 99.6 %, respectively. Applying Fisher’s class separability criterion, the optimal features are found to vary between stations (station ELD and WTP reach 1.4 but only 0.5 for STY).

Same technique is also applied in data in Parkfield (84,777 tremor and 2,612 M ≥3.0 regional events at 12 stations) that reach the high classification of accuracy at 91.0 % and tremor recall of 99.9 %. Feature with highest Fisher score (NPksDFTMax, with Fisher score of 0.357) however, is different from that of Taiwan. We also noticed the model accuracy dropped 1-2 % during leave-one-station-out experiment. The classification performance and optimal features in Taiwan reveals strong station dependency but not Parkfield. We concluded that the single-station discrimination between tremor and regional earthquakes can be successfully achieved through machine learning technique, the classification however, should be treated carefully with different features at different stations in Taiwan.