The classification of volcano-seismic signals is a key task in volcano observatories worldwide. Historically, however, automatic classification routines have typically encountered difficulties when applied to such environments. The evolving nature of an active volcanoes subsurface affects the attenuation structure, resulting in non-stationary time series containing signals of varying frequency content, amplitude and duration. In addition, dense seismic networks around volcanoes are recording exponentially more data, rendering traditional visual-manual classification obsolete.
Lately, machine learning routines have emerged as promising set of methods to exploit the data contained in the latest seismic catalogues, proven to perform classification to a standard similar to a human expert, with unparalleled efficiency.
We compare a range of machine learning methods for supervised classification of volcano-seismic signals, trained on data from Santiaguito Volcano over the period 2018-2020. The training dataset is compiled from manually labelled signals and contains 4 class types: explosions, volcano-tectonic earthquakes, tremors and noise. Spectrograms of the time series are used as the training inputs. With over 1000 examples for each signal type, a large dataset is available for training the algorithm.
The results from this work will help to set up a routine for the automatic classification of seismic signals recorded at Santiaguito. Such tools to aid the workflow of analysts are crucial for volcanic hazard mitigation. Furthermore, the results provide insight into the general performance of deep learning-based classification routines when applied to volcanic settings.