Seismic stations used by the Piton de la Fournaise volcano observatory (OVPF) to monitor seismicity on Piton de la Fournaise volcano and on Réunion Island. Triangle color indicates station type: blue for broadband stations, green for short‐period 1‐Hz three‐component stations, and red for short‐period 1‐Hz vertical analog stations.
Examples of seismic events recorded by stations CIL, FJS, BOR, and RVL (in that order). All traces are corrected for instrument response and shown as velocity seismograms. Scale bars are given to the left of the traces. All events are plotted at the same scale except for the first one.
Simplified schematic of the Random Forest (RF) instances training strategy for (a) a single station and (b) multiple stations. The best features of the single‐station RF classifiers are concatenated into the feature set for the multistation RF classifiers. Only the best features of the multistation RF are used for class prediction on new data.
Examples of feature histograms for the same four stations from Figure 2 showing the relative frequencies of feature values for each of the eight classes. Each panel shows the values of one feature at one station. The histograms are color coded according to class. (a) Duration at station BOR; (b) ratio of duration to maximum amplitude at CIL; (c) centroid frequency at FJS; and (d) median frequency of the signal spectrum at RVL.
A matrix indicating the 10 most discriminating features selected by each of the eight RF classifiers (the four single‐station classifiers and the four three‐station classifiers). The features and classifiers are plotted along the x and y axes, respectively. A black grid cell indicates that the feature was one of the 10 most discriminating ones for the given classifier.
Confusion matrixes for the four single‐station RF classifiers. Each row of a confusion matrix shows the distribution of event types predicted by the automated classifier for a given manually classified event type. The scores (out of 100) derived from the fivefold cross validation of each classifier are shown next to the station names, together with their 2σ uncertainties.
Confusion matrixes for the four multistation RF classifiers. The scores (out of 100) derived from the fivefold cross validation of each classifier are shown next to the station names, together with their 2σ uncertainties.
Example classification predictions plots shown to the OVPF operators. The star indicates the preferred class. The probability distribution over the classes gives some indication of the confidence that can be placed on the classification: strong for events (a) and (b) and weak for events (c) and (d).
Confusion matrixes and scores obtained with training sets generated using (a) sampling with replacement and (b) sampling without replacement. Confusion matrixes and standard accuracy scores for head‐to‐head classification experiments for (c) summit and deep volcano‐tectonic (VT) events and (d) summit VT events and rockfalls.
Description of terms: s(t) the windowed raw seismogram, e(t) its envelope, a(τ) its auto‐correlation function, si(t) the windowed seismograms filtered in the 0.1–1 Hz (i=1), 1–3 Hz (i=2), 3–10 Hz (i=3), 10–20 Hz (i=4), and 20–50 Hz (i=5) frequency bands, ei(t) their corresponding envelopes, ts and te the start and end times of the window, tmax the time of the peak amplitude, Kurt(X)=μ4(X)/σ4(X) the kurtosis of distribution X in which μ4(X) indicates the fourth moment of X and σ its standard deviation, Skew(X)=μ3(X)/σ3(X) the Skewness of distribution X in which μ3 indicates the third moment of X, S(ν) the fast Fourier transform of s(t), νmax the frequency at which |S(ν)| is maximum, and |S(ν)|i the ith quartile of |S(ν)|.
↵* Features not also used by Hibert et al. (unpublished manuscript, 2017; see Data and Resources).