Converting Smaller Seismic Arrays to Larger-N Synthetic-Aperture Arrays Using Deep Learning

Abstract:

The recent development of low-cost nodal-style seismic instruments has led to the deployment of large-N seismic arrays for many seismological applications. Nonetheless, for broad-band sensors, the numbers of stations in an array can still be limited. Also, for future seismic experiments ever denser seismic arrays will be needed. In this study, convolutional neural networks (CNNs) with a U-net architecture are used to train smaller seismic arrays to have responses more like larger seismic arrays. We first investigate synthetic examples with linear seismic arrays where a small number of irregularly spaced stations with as little as 10% of the number of sensors are used to reconstruct a much denser seismic array. Also, the sparse irregularly spaced stations do not need to be at the same locations as the denser reconstructed array stations. By incorporating different levels of Gaussian and correlated noise in the training and validation process, de-noising can also be accomplished with noise levels up to several times larger than the desired signals. We then investigate 2D sparsely sampled small-N seismic arrays for densifying to larger-N arrays and also denoising the data. We study sparse spiral arrays which have moderately good beam response patterns, and convert these using deeper learning to denser 2D array deployments. Also, as in the linear array designs, denoising of the sparsely sampled data can also be accomplished when noise is incorporated into the training and validation process. These results are currently being applied to the spiral array being used in the 1-year array deployment for the AutoCorr Experiment to create a larger-N synthetic-aperture array to study seismic ambient noise and coda waves from earthquakes.