New insights into volcanic ash: database creation and exploitation with machine learning

• Main Author: Benet, Damià
• Other Authors: Susanna Jenkins
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Geology::Volcanoes and earthquakes
• Online Access: https://hdl.handle.net/10356/166396
• Institution: Nanyang Technological University

A central challenge in volcanology is to anticipate the most likely outcome of a restless volcano. The main approach to address this problem has been monitoring the geophysical and geochemical signals that may warn of an impending eruption. However, interpretation of the processes driving the activity may not be straightforward, and data can be limited. In this thesis, I investigate the use of volcanic ash particles for monitoring volcanoes in a robust and reproducible manner. First, I study a time series of nine ash samples from Nevados de Chillán Volcanic Complex (Chile, 2016–2018). I find that the proportions of the main particle types (juvenile, lithic, altered material, and free-crystal) change over time, and when combined with seismic data, allow to propose the processes that drove the phreatic eruptions and up to dome extrusion. Second, I developed a standardized methodology for particle image acquisition and processing, analysis of 33 features related to the shape, texture and color, and classification into main particle types. I apply this methodology to create a web-based and publicly available volcanic ash database (VolcashDB) with 6,300 particles from 12 samples from 8 volcanoes of various compositions and eruptive styles. The web interface allows users to browse, obtain visual summaries and download the database contents. Third, I explore the use of machine learning techniques for particle classification (supervised learning). I find that the Extreme Gradient Boosting model, which classifies particle feature data, has great interpretability, and can help establishing which are the most diagnostic observations for classification of the main particle types. Classification of particle images using the Vision Transformer model yields very accurate results, but with significant variations depending on the eruptive style, and thus its application to samples not represented in the database is unclear. I believe that this thesis represents an advance towards a more robust and reproducible use of volcanic ash to monitor activity, and thus to help decide the most likely outcome during a volcanic crisis and mitigate eruption hazards.