Machine-learning-based detection of volcano seismicity using the spatial pattern of amplitudes

We propose a new algorithm, focusing on spatial amplitude patterns, to automatically detect volcano seismic events from continuous waveforms. Candidate seismic events arc detected based on signal-to-noise ratios. The algorithm then utilizes supervised machine learning to classify the existing candidate events into true and false categories. The input learning data are the ratios of the number of time samples with amplitudes greater than the background noise level at 1 s intervals (large amplitude ratios) given at every station site, and a manual classification table in which 'true' or 'false' flags are assigned to candidate events. A two-step approach is implemented in our procedure. First, using the large amplitude ratios at all stations, a neural network model representing a continuous spatial distribution of large amplitude probabilities is investigated at 1 s intervals. Second, several features are extracted from these spatial distributions, and a relation between the features and classification to true and false events is learned by a support vector machine. This two-step approach is essential to account for temporal loss of data, or station installation, movement, or removal. We evaluated the algorithm using data from Mt. Ontake, Japan, during the first ten days o f a dense observation trial in the summit region (2017 November 1-10). Results showed a classification accuracy of more than 97 per cent.