Summary Seismograms contain a variety of signals in addition to direct waves. The local later phase in the S coda (LPS; strongly reflected or scattered S-waves by subsurface heterogeneities when the hypocentre, station, and the origin of the later phase are located relatively close to each other) is useful for investigating fine-scale subsurface heterogeneities such as significant velocity contrasts and localized scatterers. With the development of dense seismic observation networks, accurate and rapid automatic processing techniques for large numbers of seismograms have become important. The introduction of deep learning techniques enables us to automate the processing of seismograms, such as phase detection and arrival time picking of direct waves with high quality and speed. To utilise the information on LPSs contained in large volumes of seismic waveform data, we developed the LPS-detector, a convolutional neural network-based automatic LPS detection model. We trained the LPS-detector using single-station data from the Moriyoshi volcanic area in northeastern Japan, which is known for observing distinct LPSs. We create an original training dataset with 5 000 earthquakes in this area. The training dataset was separated into the training (3 200), validation (800), and test (1 000) subsets. The LPS-detector yielded a 0.910 area under the curve (AUC) of the receiver operating characteristic (ROC) curve and 0.966 AUC of the precision-recall (PR) curve. These scores indicate good performance for automatic LPS detection, which is comparable to that of manual detection. Additionally, we confirmed that the LPS-detector could detect LPSs at other stations if the waveform characteristics were similar to those of the training subset. The large volume of the LPS catalogue obtained by the LPS-detector provides a new insight into the LPS origin in the Moriyoshi volcanic area. The LPS-detector detected 3 951 new LPSs outside of the training period. Combined with manual detection, 7 599 LPSs were detected in this area. Based on the comprehensive LPS catalogue, we found that the origin of LPSs in this area lies just beneath the earthquake swarm and is inclined at approximately 12° with a south-westward dip. These results indicate that the LPS-detector enables us to extract LPS information from seismogram big data and explore fine-scale subsurface structures.
Amezawa et al. (Thu,) studied this question.