Abstract

One of the most challenging tasks in volcanic data analysis is the classification of seismic events. By knowing them, it is possible to take decisions in advance, providing benefits for the neighboring societies as, for instance, how such events may impact users' life and cropland areas. Although there are several approaches to perform such task, Deep Neural Networks (DNN) have been barely considered to deal with seismic signals, as discussed in our related work. In this sense, we started our research with a wide set of experiments to analyze the DNN performance while discriminating seismic activities through two common network architectures: a 2D Convolutional Neural Networks (CNN) and the Long Short-Term Memory (LSTM) network. In attempt to draw a parallel of DNN and classical supervised learning strategies, we extended our study by comparing those two architectures against the Multilayer Perceptron (MLP), which would be the simplest and most common baseline to take into account. Our MLP network was implemented after combining a basic architecture with elements from DNN models as dropout, flatten, and batch normalization layers. From this first analysis, we confirmed the need of additional neural network layers to obtain good classification results for seismic events, given the MEP required more DNN-based pipeline operations to improve its overall performance, making it comparable to the previous results obtained by our research group. As a natural extension, we designed a new DNN architecture by assessing additional network layers specifically devoted to extract features from seismic signals, thus improving the overall classification of Volcano activities involving the following events: Volcano-Tectonic (VT), Long Period (LP), Tremor (TR) and Tectonic (TC). Such design motivated the study of current DNN architectures tackling similar problems to classify raw signals from which SoundNet was taken as the most prominent candidate, thus leading to a new CNN architecture here proposed and referred to as SeismicNet. As our final contribution, SeismicNet provided classification results among the best in the literature without demanding explicit signal pre-processing steps though. (C) 2020 Elsevier B.V. All rights reserved.