Quantifying and Classifying Volcano Video and Infrasound Datasets via Computer Vision and Machine Learning Algorithms

Abstract

Volcanoes are dangerous and complex with processes coupled to both the subsurface and atmosphere. Effective monitoring of volcanic behavior during and in between periods of crisis requires a diverse suite of instruments and processing routines. Acoustic microphones and video cameras are typical in long-term deployments and provide important constraints on surficial and observational activity yet are underutilized relative to their seismic counterpart. This dissertation increases the utility of infrasound and video datasets through novel applications of computer vision and machine learning algorithms, which help constrain source dynamics and track shifts in activity. Data analyzed come from infrasound and camera installations at Stromboli Volcano, Italy and Villarrica Volcano, Chile and are diverse in terms of the recorded activity. At Villarrica, a computer vision algorithm quantifies video data into a set of characteristic features that are used in a multiparametric analysis with seismic and infrasound data to constrain activity during a period of crisis in 2015. Video features are also input into a machine learning algorithm that classifies data into five modes of activity, which helps track behavior over weekly and monthly time scales. At Stromboli, infrasound signals radiating from the multiple active vents are synthesized into characteristic features and then clustered via an unsupervised learning algorithm. Time histories of cluster activity at each vent reveal concurrent shifts in behavior that suggest a linked plumbing system between the vents. The algorithms presented are general and modular and can be implemented at monitoring agencies that already collect acoustic and video data.