An assessment of support vector machines for volcanic infrasound signal classification on local scales

Abstract:

Acoustic signals produced by volcanoes are varied, complex, and sometimes temporally overlapping. Machine learning has shown promise for parsing out and classifying signals from geophysical waveform data. While global catalogs of labeled infrasound-producing events exist, the application of supervised machine learning to local (< 10 km) volcanic infrasound signals has to date been limited by a lack of robust training datasets. To circumvent this, previous studies have applied unsupervised learning methods to cluster similar signals while locating them to various vents. Here, we apply a supervised learning approach to classify explosive volcanic signals recorded by a local infrasound network at Yasur volcano, Vanuatu. Yasur is an ideal volcano for this goal due to 1) its near-continuous Strombolian activity consisting of hundreds of explosions per day and 2) the presence of two distinct vents in its crater. We use data from six sensors deployed around the crater, all less than 500 m from either vent. We generate a large training dataset by locating explosions using a backprojection code that integrates finite-difference time-domain propagation modeling to calculate realistic infrasound travel times over topography. We then use the support vector machine method (SVM) to classify signals as belonging to either vent. We use metrics from both the time and frequency domains of the recorded signals to construct the feature space required for SVM. Previous studies have applied SVM to classify global infrasound signals, but the large influence of path effects at these source-receiver distances dominates the signal feature space. This work focuses on the local scale, where differences in waveforms more directly relate to source characteristics. This emphasis allows us to compare SVM to other, more traditional techniques such as waveform cross-correlation. Our aim is to examine whether the vent of origin for a given explosion can be determined from signal characteristics alone, as well as identify temporal changes in activity. We also seek insight into how machine learning models can be used to extract physical information from infrasound data.