Seismic ambient noise tomography of central Italy using a deep learning algorithm

This work aims at testing a convolutional neural network (CNN), developed by Zhang et al. (2020), to measure group velocity of Rayleigh waves, extracted from records of background seismic noise, and at studying the crustal structure of Central Italy. I investigate an area in Central Italy by means of seismograms recorded by 73 seismic stations located approximately in the Central and Northern Apennines. The study is part of a larger project, named MUSE (Multiparametric and mUltiscale Study of Earthquake preparatory phase), that has the goal, among others, of detecting the spatial and temporal evolution of the velocity in the Earth’s crust. In this context, I have obtained the group velocity maps representative of the entire time span, from 01/01/2010 to 01/05/2021. In this study, I employ the technique of seismic ambient noise interferometry to extract Rayleigh wave measurements from the fully diffuse wavefields (Shapiro and Campillo, 2004). I then use CNN and other methods to determine the dispersion characteristics of the Rayleigh wave fundamental mode: a traditional method, requiring operator interaction; a classical automated procedure; and the recent neural network. I then invert each of the three dispersion data sets retrieved with the different approaches, to compute group velocity maps at different periods. I compare the performance of the different methods employed, and thus evaluate the performance of the CNN network, by comparing the maps among them, with geological observations, and also with a pre-existing model from the literature (Molinari et al., 2015). The CNN method shows excellent potential, but – at the current stage of development – it needs more accurate and specific training to reach the precision of manual picks. My maps image the crustal structure of the Northern Apennines area with unprecedented detail. This work poses the basis for further studies, to image the time variations of 3D structure (i.e., 4D tomography).