Detection and location of aftershocks of the 2020 Western Idaho earthquake using neural networks

Abstract:

Aftershocks are small earthquakes that follow the mainshock. Typically, hundreds to thousands of aftershocks occur following large earthquakes, but most of them have relatively small magnitudes, making them difficult to be detected. Traditional detection approaches include STA/LTA, waveform similarity and template matching, which are based on the amplitudes of seismograms. Recently, machine learning algorithms have been widely utilized to detect and locate earthquakes. In this study, we transfer events and noises (training datasets) into 2D Time-Frequency feature maps, which are used to train our convolutional neural network (CNN). We select a detection region (42° N-48° N, 119° W-110° W) that centered around the 2020 Western Idaho earthquake, there are 52 stations within the region and we select one-month long continuous records after the mainshock as our testing datasets. After detection, we utilize another trained CNN model to automatically pick P wave arrival times. We then use a nonlinear location program to locate events around the mainshock area. The 31 March 2020 Mw 6.5 Western Idaho earthquake is the largest earthquake occurred at Idaho in the past 30 years, and there are only 670 aftershocks in the USGS catalog. Based on our method, we build a more completed catalog that includes 2-3 times of events more than the current catalog. We observe a clear aftershocks distribution pattern, which is consistent with earthquake energy back projection results. This study helps us to better understand the rupture processes of the 2020 Western Idaho event.