Detecting dynamical regimes by Self-Organizing Map (SOM) analysis: an example from the March 2006 phreatic eruption at Raoul Island, New Zealand Kermadec Arc

We propose a technique to improve the analysis of volcanic seismic data and highlight possible dynamical or precursory regimes, by using an efficient class of artificial neural network, the Self-Organizing Maps (SOMs). SOMs allow an automatic pattern recognition, as independent as possible from any a priori knowledge. In the training phase, volcanic tremor spectra are randomly presented to the network in a competitive iterative process. Spectra are then projected, ordered by time, onto the map. Every spectrum will take up a node on the map and their time evolution on the map can highlight the existence of different regimes and the transitions between them. We show a practical application on data recorded at Raoul Island during the period around the March 2006 phreatic eruption which reveals both a diurnal anthropogenic signal and the post-eruption system excitation.