Investigating large-scale change in volcanic time series data using machine learning analysis
Volcanic eruptions are characterised by large-scale transitions in behaviour, which include the transitions governing repose to unrest, unrest to eruption, and eruption to quiescence. Identification of these transitions is a fundamental goal in volcanology. This thesis aims to explore the use of mac...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2021
|
| Subjects: |
| _version_ | 1817932904765849600 |
|---|---|
| author | Manley, GF |
| author2 | Mather, T |
| author_facet | Mather, T Manley, GF |
| author_sort | Manley, GF |
| collection | OXFORD |
| description | Volcanic eruptions are characterised by large-scale transitions in behaviour, which include the transitions governing repose to unrest, unrest to eruption, and eruption to quiescence. Identification of these transitions is a fundamental goal in volcanology. This thesis aims to explore the use of machine learning approaches which are established for the monitoring of critical systems, such as healthcare or jet engine monitoring, to better understand the timing of large-scale transitions in volcanic activity. The first research chapter is an application of machine learning approaches to better understand the timing of eruption end, which has no systematic definition. Multi-class machine learning methods were applied to time series derived from volcano-seismic data from two volcanoes with contrasting eruption styles: Nevado del Ruiz, Colombia and Telica, Nicaragua. The resulting eruption end-dates agreed with previous, generic definitions of eruption end and provide an independent constraint on the timescale of eruption cessation processes. The second research chapter is an extension of this methodology to multi-parameter datasets from the 2005 – 2006 unrest and eruption of Augustine, Alaska. Both multi-class and novelty detection approaches, in which models are trained only on non- eruptive data, are applied. The resulting transition from non-eruptive to eruptive identified from the modelling agrees with previous estimates of the timing of dike initiation prior to eruption. Additionally, I explored the use of different types of data and different seismic catalogues for distinguishing between eruptive and non-eruptive activity. The final research chapter is an application of deep active learning to automated volcano-seismic event classification. Active learning is the process by which a machine learning model actively selects the training data which it learns from. This approach has the potential to reduce the quantity of labelled data required to train a successful model. The deep active learning approach is applied to two datasets from Nevado del Ruiz, Colombia and Llaima, Chile. Active learning models performed better on unseen testing data for both datasets and achieved better performance during training for the Nevado del Ruiz dataset. |
| first_indexed | 2024-03-07T07:42:14Z |
| format | Thesis |
| id | oxford-uuid:4165b523-01de-4323-aaa3-5216aeac3451 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-12-09T03:45:20Z |
| publishDate | 2021 |
| record_format | dspace |
| spelling | oxford-uuid:4165b523-01de-4323-aaa3-5216aeac34512024-12-07T17:04:35ZInvestigating large-scale change in volcanic time series data using machine learning analysisThesishttp://purl.org/coar/resource_type/c_db06uuid:4165b523-01de-4323-aaa3-5216aeac3451VolcanologySupervised learning (Machine learning)GeologyMachine learningEarth sciencesVolcanological researchEnglishHyrax Deposit2021Manley, GFMather, TPyle, DClifton, DRodgers, MKendall, JPoland, MVolcanic eruptions are characterised by large-scale transitions in behaviour, which include the transitions governing repose to unrest, unrest to eruption, and eruption to quiescence. Identification of these transitions is a fundamental goal in volcanology. This thesis aims to explore the use of machine learning approaches which are established for the monitoring of critical systems, such as healthcare or jet engine monitoring, to better understand the timing of large-scale transitions in volcanic activity. The first research chapter is an application of machine learning approaches to better understand the timing of eruption end, which has no systematic definition. Multi-class machine learning methods were applied to time series derived from volcano-seismic data from two volcanoes with contrasting eruption styles: Nevado del Ruiz, Colombia and Telica, Nicaragua. The resulting eruption end-dates agreed with previous, generic definitions of eruption end and provide an independent constraint on the timescale of eruption cessation processes. The second research chapter is an extension of this methodology to multi-parameter datasets from the 2005 – 2006 unrest and eruption of Augustine, Alaska. Both multi-class and novelty detection approaches, in which models are trained only on non- eruptive data, are applied. The resulting transition from non-eruptive to eruptive identified from the modelling agrees with previous estimates of the timing of dike initiation prior to eruption. Additionally, I explored the use of different types of data and different seismic catalogues for distinguishing between eruptive and non-eruptive activity. The final research chapter is an application of deep active learning to automated volcano-seismic event classification. Active learning is the process by which a machine learning model actively selects the training data which it learns from. This approach has the potential to reduce the quantity of labelled data required to train a successful model. The deep active learning approach is applied to two datasets from Nevado del Ruiz, Colombia and Llaima, Chile. Active learning models performed better on unseen testing data for both datasets and achieved better performance during training for the Nevado del Ruiz dataset. |
| spellingShingle | Volcanology Supervised learning (Machine learning) Geology Machine learning Earth sciences Volcanological research Manley, GF Investigating large-scale change in volcanic time series data using machine learning analysis |
| title | Investigating large-scale change in volcanic time series data using machine learning analysis |
| title_full | Investigating large-scale change in volcanic time series data using machine learning analysis |
| title_fullStr | Investigating large-scale change in volcanic time series data using machine learning analysis |
| title_full_unstemmed | Investigating large-scale change in volcanic time series data using machine learning analysis |
| title_short | Investigating large-scale change in volcanic time series data using machine learning analysis |
| title_sort | investigating large scale change in volcanic time series data using machine learning analysis |
| topic | Volcanology Supervised learning (Machine learning) Geology Machine learning Earth sciences Volcanological research |
| work_keys_str_mv | AT manleygf investigatinglargescalechangeinvolcanictimeseriesdatausingmachinelearninganalysis |