Application of data analysis and machine learning techniques to improve baseline volcano and mountain hazards monitoring

In this work, several machine learning and data analysis tools were applied to various datasets consisting of seismo-acoustic recordings. Due to the complexities and uncertainties of each dataset, several criteria were proposed for the adequate application of these techniques for different tasks. The first was to assess the quality of two open-access volcano seismic datasets, one considered noisy (Cotopaxi volcano, Ecuador), and the other clean (Llaima volcano, Chile). By applying a catalogue cleaning procedure, metrics and benchmarks were defined to rapidly assess each dataset's quality. After analyses, the Cotopaxi dataset showed numerous mislabelled events, which was confirmed by performing a blind test from experts' assessment. In contrast, Llaima's catalogue yielded few mislabelled events, validating its cleaner status. A second task consisted of expanding a dataset of volcanic explosions from nearly 10 years (2006-2016) of continuous acoustic recordings from a seismo-acoustic network on the flanks of Tungurahua volcano, Ecuador. The original explosions were identified by human inspection using an amplitude threshold and omitted small to medium sized explosions. To expand this catalogue, a series of successive steps combining traditional and novel data analysis with (un)supervised machine learning tools was applied to continuous recordings from one station. This led to more than 29,000 new explosions being detected that were grouped and linked to changes in Tungurahua's activity. Finally, a third dataset consisting of several months (March-July 2021) of novel continuous acoustic recordings from the Squamish River (Mount Cayley, BC.), was explored. To use these recordings as a potential tool for natural hazards monitoring in the region, an acoustic activity baseline was established by harnessing the extreme weather conditions caused by the 2021 Western North American heat wave. Recordings of local weather parameters, especially water level gauges, were used to relate part of the acoustic measurements to the river's discharge rate using rapid data analysis tools. Anomalous signals deviating from this baseline are proposed as a means to identify future natural hazards. This work showcases the flexibility -yet care- with which machine learning and data analysis tools can be applied to monitor volcano and mountain hazards, and lays the groundwork for future developments.