Detecting Low-frequency Earthquakes in Western Japan Using Convolutional Neural Networks

Low-frequency earthquakes (LFEs) occur largely during slow slip events at subduction interfaces. Their actions affect the stress state of the seismogenic zone and potentially link to larger ordinary earthquakes. Detecting and analyzing low-frequency earthquakes is crucial for a better understanding of the subduction process. As LFEs lack distinct, impulsive body waves, it is challenging to detect them using classical techniques. The most common method for identifying low-frequency earthquakes involves autocorrelation or template matching techniques that often have low computational efficiencies and therefore difficult to be applied to large dataset. Fortunately, a rapid transformation has occurred in the field of computer vision in recent years due to the emergence of convolutional neural networks (CNNs). These CNNs are powerful variants of supervised machine learning which significantly improve the computational detecting similar waves in seismology. In this paper, we attempt to detecting low-frequency earthquakes 2018 in Shikoku, Japan using a highly scalable CNN developed by Perol et.al, 2018. We adopt the continuous waveforms from April, 2004 to March, 2011 recorded by two Hi-net stations, KWBH and YNDH, in central Shikoku. Based on the LFE catalog from Ohta et.al, 2017, we obtain 1222 events for the training set and 195 events for the testing set. Our preliminary results show that we detect 3 folds of new tremors compared to the existing catalog. These newly detected LFEs are confirmed with visual inspections and cross-correlation tests with template events. We also manage to determine the locations of these tremors which agree with the previous studies of LEFs in this area as well. Our initial attempt demonstrate that CNN is effective in detecting LFEs. We plan to expand our effort to detect the LFEs across western Japan and Cascadia.