Project Description
Supervisors
Dr Jessica Johnson (School of Environmental Sciences, University of East Anglia) – Contact me
Dr Ben Milner (School of Computing Sciences, University of East Anglia)
Dr Jason Lines (School of Computing Sciences, University of East Anglia)
Dr Zoë Mildon (School of Geography, Earth and Environmental Sciences, University of Plymouth)
Dr Irene Manzella (School of Geography, Earth and Environmental Sciences, University of Plymouth)
Gill Jolly (Natural Hazards and Risk Theme Leader, GNS Science, New Zealand)
Project Background
Geophysical hazards, such as volcanic eruptions, large earthquakes and landslides, threaten millions of people around the world. Many of these hazards are unpredictable, but there are often precursors that may be used to forecast damaging events. These precursors could be in the form of monitoring observations such as seismicity, deformation and gas emissions (1), or from data processed after the event has occurred, recognised in hindsight (2).
In typical hazard forecasting, observations of pre-cursors are studied and researchers use their prior knowledge to estimate the probability of an event occurring (3). However, Machine Learning (ML) tools are increasingly being applied to geosciences and hazard monitoring due to ML’s strength in identifying patterns (4,5). If patterns can be identified in hindsight, then they could be used as a precursor to forecast future events.
Research methodology
This project will use precursor sequences with ML techniques to investigate the relationship between datasets, answering questions like:
- Which data are the most important for forecasting?
- What is the minimum monitoring network required?
- Which data would be needed to decrease uncertainty?
- How does ML compare to expert elicitation?
The student will use geophysical, geological and geochemical data relating to specific geophysical hazards in New Zealand. Data will be obtained from online sources, published literature, and expert accounts. The student will train state-of-the-art time series classification algorithms on the precursor data, and devise performance measures for predictions. ML results will be quantitatively compared to existing methodologies using expert elicitation.
Training
Training will be given in geoscientific methods, machine learning techniques, and geophysical hazard processes (including earthquakes, volcanoes and landslides). The student will spend time in New Zealand, learning about the hazards and datasets.
Person specification
Applicants must hold, or expect to receive, a degree in a relevant geoscience, computing or physical sciences discipline. The student should be numerically literate and experience of using unix based operating systems and/or Python is desirable but not essential.