CFM: a Convolutional network for First Motion polarity classification of earthquake waveforms.

<p>The knowledge of the crustal stress field is essential in the evaluation of the seismic hazard of an area.To this aim, it is necessary to derive reliable focal mechanisms mainly when small earthquakes have to be included in the computation. The first motion focal mechanism solution techniques are still widely used in modern softwares. The determination of P-wave polarities with manual procedures can lead to human errors and it is time-consuming. Automatic procedures can avoid these drawbacks. Polarity identification is not a classification task easily expressed in terms of mathematical procedures, in fact classical automatic procedures can lead to worse results than those obtained by human operators. For this reason, the use of machine learning approaches results necessary to accomplish this task.With low computational costs, real-time analysis capabilities, no need for complicated pre-processing procedures, and truly competitive results, properly designed convolutional networks can be the answer to various problems, including those related to seismology.&#160;</p> <p>In our work, we present the Convolutional First Motion (CFM) network, a Deep Convolutional Neural Network (DCNN) used to classify seismic traces based on first motion polarities of P-waves. We used waveforms contained in two datasets. We prepared the first dataset selecting approximatively 150&#729;000 waveforms contained in the Italian seismic catalogue INSTANCE, specifically designed for the application of machine learning techniques. To this end we devised an analysis procedure using Principal Component Analysis and Self-Organising Maps, through which a clustering process individuated groups of suitable traces. A second dataset, not specifically designed for machine learning techniques, is prepared manually picking approximatively 4&#729;000 waveforms of earthquakes occurred between 2010 and 2014 at Mt. Pollino area in Italy, avoiding possible overlapping of waveforms between the two datasets. The network, trained on ~130&#729;000 time windows centred on P-wave arrival times of waveforms in the INSTANCE catalogue, achieved accuracies of 95.7% and 98.9% on two <em>test sets</em>: the Mt. Pollino dataset and part of the INSTANCE catalogue. Further testing showed that if we give the network waveforms with uncertain arrival times, it acquires robustness to this type of noise, still showing high-level of performance.</p> <p>We infer that the CFM network would be suitable in succession to automatic techniques that derive P-wave arrival times, for example techniques in which deep learning is used, in order to cover the entire data processing phase with machine learning. Given the incredible ability of DCNNs to model and process large volumes of data and their remarkable performance, it is reasonable to assume that deep learning will soon become the norm even in the context of first-motion polarity determination.&#160;</p> <p>This work was partially supported by the PRIN-2017 MATISSE project (no. 20177EPPN2), funded by the Ministry of Education and Research.</p>