Volcano-Seismic Event Detection and Clustering

Abstract

This study looks into unsupervised and supervised methods for detecting events in volcano-seismic time series data, segmenting the data, and clustering the segments where there is activity. This two-stage pipeline allows for the analysis of the signals without requiring the type of event to be identified at the offset and reduces the manpower required to analyze new data. Due to the resource intensive labeling process required to understand volcano-seismic signals it is important to explore unsupervised analysis techniques in this domain. The unsupervised methods are evaluated using supervised metrics including completeness, homogeneity, and V-measure scores. Alongside the unsupervised investigation, the use of intersection-based metrics that offer a clearer performance evaluation of the event segmentation task is motivated and the potential of gradient boosted trees for event detection is tested.