Loic Viens, Chris Van Houtte

Seismic interferomestry is an established method for monitoring the temporal evolution of the Earths physical properties. We introduce a new technique to improve the precision and temporal resolution of seismic monitoring studies based on deep learning. Our method uses a convolutional denoising autoencoder, called ConvDeNoise, to denoise ambient seismic field correlation functions. The technique can be applied to traditional two-station cross-correlation functions but this study focuses on single-station cross-correlation (SC) functions. SC functions are computed by cross correlating the different components of a single seismic station and can be used to monitor the temporal evolution of the Earths near surface. We train and apply our algorithm to SC functions computed with a time resolution of 20 minutes at seismic stations in the Tokyo metropolitan area, Japan. We show that the relative seismic velocity change (dv/v(t)) computed from SC functions denoised with ConvDeNoise has less variability than that calculated from raw SC functions. Compared to other denoising methods such as the SVD-based Wiener Filter method developed by Moreau et al. (2017), the dv/v results obtained after using our algorithm have similar precision. The advantage of our technique is that once the algorithm is trained, it can be apply to denoise near-real-time SC functions. The near-real-time aspect of our denoising algorithm may be useful for operational hazard forecasting models, for example when applying seismic interferometry at an active volcano.