GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 - POSTER GNGTS 2023 CFM: a Convolutional network for First Motion polarity classification of earthquake waveforms. G. Messuti 1 , S. Scarpetta 1 , O. Amoroso 1 , F. Napolitano 1 , P. Capuano 1 1 Dipartimento di Fisica “E.R. Caianiello”, Università degli Studi di Salerno, Fisciano (SA), Italy Introduction The determination of reliable first-motion polarities of P-waves plays a key role in computing earthquake focal mechanisms. Manual procedures are time-consuming and affected by human error, mainly when dealing with small magnitude events. Automatic procedures can avoid these drawbacks. Polarity identification is not a classification task easily expressed in terms of mathematical procedures. For this reason, the use of machine learning approaches results necessary to accomplish this task. The goal of the present work is the development of the Convolutional First Motion (CFM) network, a Deep Convolutional Neural Network (Goodfellow et al., 2016) used to classify seismic traces based on the first motion polarity of P-waves. The network has been trained using approximately 150 ˙ 000 waveforms related to the INSTANCE catalog (Michelini et al., 2021), specifically designed for the application of machine learning techniques. The first step of the analysis has been the selection of suitable seismic traces, carried out following a thorough investigation using unsupervised learning techniques such as Principal Component Analysis (PCA, Bishop et al., 2006) and Self-Organising Maps (SOM, Kohonen, 2013). To this end, we devised a procedure consisting of two rounds of analysis, through which a clustering process individuated the groups of traces to be discarded. A second dataset, consisting of the manually picked waveforms relative to the seismic sequence that occurred between 2010 and 2014. at Mt. Pollino area in Italy (Napolitano et al., 2021), was used to test the network on totally unknown traces. Various attempts at training the CFM network brought out the best architecture and the best performing configurations. An accuracy of about 99% was achieved by using the initial test set (derived using part of the INSTANCE catalog), and an accuracy of about 96% by using the dataset of Napolitano et al., 2021. Dataset and waveforms pre-processing The waveforms collected from the INSTANCE catalog (Michelini et al., 2021, Fig. 1a) were used as the starting dataset to train the network and evaluate its performance. The dataset includes events with magnitude between 0.0 and 6.5. All traces have a sampling rate of 100 Hz. For our purposes, we selected the vertical component of velocimeter waveforms with assigned P-wave polarity. The selection allowed the use of 161 ˙ 199 waveforms. Such a large dataset inevitably contains some