Ob. 1
A new seismic database was constructed with data from selected volcanoes the represent different unrest and eruption scenarios. This allowed the organization of data and metadata in standardised formats. These data formed the basis for Ob.2 and Ob.3 of the project. They were used to produce catalogues of earthquakes with classification labels by applying Machine Learning methods. Seismic data were transformed to standard formats using open-source software. The size and variety of data formats represented a challenge. Results: In order to manage such a large set of data and metadata, in collaboration with the host institution group, I participated to design a mySQL relational database to archive and efficiently access seismic data and the related metadata. I contributed in writing a scientific paper focussed on the implementation and management of this database. Currently, this manuscript is under UGR internal review prior to submission to Computers & Geosciences, an ELSEVIER, an international peer reviewed journal.
Ob. 2:
I manually classified seismic signals for each case study volcano to increment the initial master DB of labelled events (that is, a catalogue of labelled signals extracted from continuous seismic data streams) created in Ob1. Events were separated on the basis of both their time and frequency domain characteristics to obtain a large dataset of more 10000 reliably classified seismic events. To achieve this objective, I overseen the implementation of the new Python Interface for the Classification of Seismic Signals (PICOSS). The following step was to create, with the support of the host research team, a new algorithm for supervised classification that allowed clustering of catalogued events in different families. This algorithm was based on the quantification of metrics linked to changes in amplitude and frequency within each event. Results: (i) I obtained a new labelled dataset and the corresponding parameters (ime and frequency domain) catalogues; (ii) I designed, in collaboration with the host research group a modular open source software, called PICOSS (doi:https://doi.org/10.1016/j.cageo.2020.104531) with a graphical user interface designed for detection, segmentation and classification, focusing on exportability and standardization of data formats. The users can select automatic or manual workflows to select and label seismic data from a comprehensive suite of tools, including deep neural networks; (iii) In addition, I have participated to build a new platform (VINEDA-doi: 10.3389/feart.2019.00335) for automatic identification of acoustic volcanic signals. I joined an international team (a joint collaboration between Italy, UK and Spain) focussing the analysis of infrasound signals in volcanic environments. Volcano infrasound used jointly with seismic data can provide valuable information on magma and eruption at active volcanoes.
Ob. 3:
Finally, I compared the automatically labelled seismic signals with the ones classified manually in order to benchmark the new classification algorithms via unsupervised ML procedures. This part of the project focused on the recognition of patterns and regularities in data. I established which set of features and architectures were most appropriate for signal characterization and to identify changes in seismic timeseries as potential precursor of every eruptive process. To reach this objective, I have used Deep Learning architectures implemented in the popular TensorFlow Python libraries. Results: (i) I analysed more than 55000 earthquakes using advanced ML algorithms in order to perform an automatic analysis aimed at characterizing the temporal evolution of clusters of similar events, that are well correlated with the occurrence of eruptive activity, suggesting that they could be a common precursor to eruptive events; (ii) I contributed to design, in collaboration with the host research group, a new method to detect subtle changes in continuous seismic data potentially associated with the evolution of unrest and impending eruptions, for which specialised architectures are required. We implemented a new framework to monitor changes in continuous seismic time series based on Bayesian Deep Learning theory. I participated to writing a scientific paper focussed on the implementation and management of this ML architecture, showing the results obtained. Currently, this manuscript is under review in IEEE-TGRS, an international peer reviewed journal.
In addition, other 12 scientific reports are open access on EGU, AGU and INGV conference repositories.
