Volcanic tremor is key to our understanding of active magmatic systems but, due to its complexity, there is still a debate concerning its origins and how it can be used to characterize eruptive dynamics. In this study we leverage machine learning (ML) techniques using 6 years of continuous seismic data from the Piton de la Fournaise volcano (La R\'eunion island) to describe specific patterns of seismic signals recorded during eruptions. These results unveil what we interpret as signals associated with various eruptive dynamics of the volcano, including the effusion of a large volume of lava during the August-October 2015 eruption, as well as the closing of the eruptive vent during the September-November 2018 eruption. The ML workflow we describe can easily be applied to other active volcanoes, potentially leading to an enhanced understanding of the temporal and spatial evolution of volcanic eruptiions
Deep Dive into Machine Learning Reveals the Seismic Signature of Eruptive Behavior at Piton de la Fournaise Volcano.
Volcanic tremor is key to our understanding of active magmatic systems but, due to its complexity, there is still a debate concerning its origins and how it can be used to characterize eruptive dynamics. In this study we leverage machine learning (ML) techniques using 6 years of continuous seismic data from the Piton de la Fournaise volcano (La R'eunion island) to describe specific patterns of seismic signals recorded during eruptions. These results unveil what we interpret as signals associated with various eruptive dynamics of the volcano, including the effusion of a large volume of lava during the August-October 2015 eruption, as well as the closing of the eruptive vent during the September-November 2018 eruption. The ML workflow we describe can easily be applied to other active volcanoes, potentially leading to an enhanced understanding of the temporal and spatial evolution of volcanic eruptiions
Long lasting seismic signals known as volcanic tremors are almost ubiquitously present in eruptive episodes at volcanoes (Jellinek & Bercovici, 2011). These seismic signals are thought to be critical in characterizing the magma migration pathways in the internal plumbing system of volcanoes, as they are typically linked with magma propagation (B. A. Chouet, 1996) With regards to the applications of ML to study and characterization of volcanoes, the primary applications thus far have been in the classification of volcano-seismic signals (Hibert et al., 2017;Malfante et al., 2018;Titos et al., 2019). In this work, we describe how statistical features derived from the continuous seismic signal recorded at Piton de la Fournaise volcano can be utilized to build machine learning models which reveal the characteristic eruptive tremor and eruptive dynamics of volcanic eruptions.
Piton de la Fournaise is an active volcano situated on La Réunion, a hot-spot basaltic island in
For this study we leverage 6 years of continuous seismic data recorded across the Observatoire Volcanologique du Piton de la Fournaise (OVPF) network shown in Figure 1. The network consists of stations equipped with short-period seismometers and broadband seismometers, recording at a sampling rate of 100Hz. The data shown in this manuscript are primarily from the seismic station of the Cratere Bory (BOR) site although the analysis described in sections 2.2 and 2.3 has been performed for most of the seismic stations situated in the the Enclos Fouqué belonging to the OVPF network caldera we find that data recorded at the BOR site provides the best performance for our eruptive state classifier. We note that in the period under study at BOR seismic data for the April 4 th eruption is missing, and thus this eruption does not appear in our dataset We also use data from the Enclos Sery Sud (CSS), Chateau Fort (FOR) and the Faujas
In order to reduce the continuous seismic signal recorded at various stations at Piton de la Fournaise to a set of tabular features, we first correct the seismic signal for each day and each station to remove the instrument response.
A filter bank is then applied in between the spectral range of 0.5 -26Hz with an initial spacing of 0.5 -2Hz and then a spacing of 1Hz, in a similar approach to that described by Rouet-Leduc, Hulbert, & Johnson, (2018). This results in 25 frequency bands. This frequency range is chosen as it encompasses the typical frequencies of volcanic tremor and volcano-tectonic events reported at Piton de la Fournaise and real Earth (Rouet-Leduc, Hulbert, & Johnson, 2018). In this case, these features consist of a range of percentiles, the range of the data, as well as normalized and non-normalized higher order moments (see Supplementary Information for table describing features). For the Bory Crater seismic station (BOR), this results in 4618 non-overlapping 1 hour time windows of eruption data with 990 features generated for each window point, or 39 features per spectral band, and 38021 windows in total over the 2013-2019 period.
We first use a supervised learning approach to determine the characteristics of eruptive tremor detected at the Piton de la Fournaise. In this case, the target label is the eruptive state of the volcano (erupting or dormant). For this binary classification task, we utilize a machine learning (ML) algorithm known known as gradient boosted decision trees (GBDT) (Friedman, 2002) with a cross-entropy loss function. We used the XGBoost implementation of GBDT to perform the modelling described in this paper, which enables parallel and distributed computation of the model (Chen & Guestrin, 2016).
The XGBoost model takes the features derived from a sliding time window over the seismic signal recorded at a station as inputs, and outputs a prediction on the eruptive state of the volcano. The training data in this case constitutes approximately 30% of the total data available for the BOR station, or 12000 time windows. The trained model is blind tested on the remaining 70% of the data, outputting an estimate for the eruptive state of the volcano during this period, without ever having actually seen this data. This allows us to estimate the consistency of the characteristics of the eruptive tremor throughout this period: if the spectral characteristics of the tremor are different between the training and the testing set, the model will be unable to effectively predict the eruptive state in the testing set.
Figure 2a shows the ability of our classifier to predict whether Piton de la Fournaise is undergoing an eruption or dormant based on features generated from a single time window of the continuous seismic signal recorded at the BOR station. The classifier was trained on the grey portion of the dataset (2013)(2014)(2015), and tested on the violet section (2015)(2016)(2017)(2018)(2019). The model performs well on the test set, with a precision of 0.99, accuracy of 0.97 and a recall scor
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
