Deep Learning and Machine Learning Applied to the Detection and Classification of Volcano-Seismic Events at Piton de la Fournaise Volcano

Abstract

The Piton de la Fournaise volcano (PdlF) on the island of La Réunion is one of the most active and best monitored volcanoes in the world. Its frequent eruptions make it a natural laboratory for developing new methods and evaluating their performance over multiple eruption sequences. In this work, we present a Deep Learning (DL) model for volcanic earthquake detection and two models for classification based on DL and Machine Learning (ML) algorithms. The detection model is based on encoder–decoder layers that extract high-order features in the time domain that are hidden in the seismograms. The first classification model consists of a simple convolutional neural network that uses the short-time Fourier transform of the signals as input data. A second classifier is based on ML approach and uses hand-crafted features. We show that our detection model, trained on ~ 7 000 volcano-seismic events recorded at PdlF between 2014 and 2021, outperforms previous DL-based models in detecting volcano-seismic events, achieving an accuracy of 98.15% on the testing dataset. Seven classes of signals are considered for classification models: volcano-tectonic (VT) events, rockfall, long-period events, volcanic tremors, tectonic events, anthropogenic noise and environmental noise. Both tested classification models achieve an accuracy of 96.55% in the testing dataset. By applying these models to the continuous data recorded at PdlF in 2019, we are able to detect and classify 1.5 times more VT events than the catalog provided by the Observatory. The detection model takes 28 s to process 24 h seismograms and from a few to a maximum of 70 s for classification.