Improving the Retrieval of Offshore-Onshore Correlation Functions With Machine Learning

The retrieval of reliable offshore-onshore correlation functions is critical to improve our ability to predict long-period ground motions from megathrust earthquakes. However, localized ambient seismic field sources between offshore and onshore stations can bias correlation functions and generate nonphysical arrivals. We present a two-step method based on unsupervised learning to improve the quality of correlation functions calculated with the deconvolution technique (e.g., deconvolution functions, DFs). For a DF data set calculated between two stations over a long time period, we first reduce the data set dimensions using the principal component analysis and cluster the features of the low-dimensional space with a Gaussian mixture model. We then stack the DFs belonging to each cluster together and select the best stacked DF. We apply our technique to DFs calculated every 30min between an offshore station located on top of the Nankai Trough, Japan, and 78 onshore receivers. Our method removes spurious arrivals and improves the signal-to-noise ratio of DFs. Most 30-min DFs selected by our clustering method are generated during extreme meteorological events such as typhoons. To demonstrate that the DFs obtained with our method contain reliable phases and amplitudes, we use them to simulate the long-period ground motions from a M-w 5.8 earthquake, which occurred near the offshore station. Results show that the earthquake long-period ground motions are accurately simulated. Our method can easily be used as an additional processing step when calculating offshore-onshore DFs and offers a new way to improve the prediction of long-period ground motions from potential megathrust earthquakes. Plain Language Summary Seismic waves from subduction earthquakes are generally characterized by a strong and elongated long-period component due to their propagation through complex velocity structures such as accretionary wedges. Seismic interferometry, which consists of cross-correlating continuous ambient seismic field signals at two seismic stations, can be used to retrieve the wave propagation between the two sensor's locations. However, the retrieval of clear wave propagation between offshore and onshore stations is difficult due to the characteristics of the ambient seismic field. We develop a method based on unsupervised learning to improve the quality of correlation functions between offshore and onshore sites. We apply our method to correlation functions calculated between an offshore station on top of the Nankai Trough, Japan, and surrounding onshore stations. The correlation functions retrieved with our method can be used to better simulate the ground motions from a M-w 5.8 earthquake, which occurred along the Nankai Trough. Improving our ability to retrieve accurate wave propagation between offshore and onshore stations is critical to better predict the long-period ground motion from potential megathrust earthquakes, which are likely to happen along subduction zones worldwide in the near future.