14:30 〜 14:45
[S22-10] Improving the symmetry of ambient seismic field correlation functions with machine learning
Seismic interferometry is a well-established method to retrieve the seismic wave propagation between a pair of seismic stations. Under the assumption that the ambient seismic wavefield is equipartitioned, the correlation function between the two sensors should yield the inter-station Green's function. However, such a condition is rarely fulfilled on Earth, as the ambient seismic field is generated by uneven distributions of sources, for example from ocean waves at low frequencies (<1 Hz). This uneven source distribution generally leads to non-symmetric correlation functions which can also be corrupted with non-physical wave arrivals. We focus on one year of continuous data recorded by Hi-net stations located in the Kii peninsula, Japan, and compute correlation functions between station pairs using short 15-min time-windows and stack them over 30 minutes. As the source of the ambient seismic field varies through the year, the raw stack of correlation functions over one year is strongly asymmetric for most station pairs. We propose to use machine learning techniques (e.g., Principal Component Analysis (PCA) and autoencoders) to reduce the dimensionality of the correlation function dataset for each station pair. By selecting correlation functions from the latent/low-dimension space of both methods, we show that the symmetry between the acausal and causal parts of the correlation functions can be improved. This additional processing step could help us to retrieve a better approximation of inter-station Green's functions, and therefore be useful for imaging purposes.