How can seismo-volcanic catalogues be improved or created using robust neural networks through weakly supervised approaches?

Abstract. Real-time monitoring of volcano-seismic signals is complex. Typically, automatic systems are built by learning from large seismic catalogues, where each instance has a label indicating its source mechanism. However, building complete catalogues is difficult owing to the high cost of data-labelling. Current machine learning techniques have achieved great success in constructing predictive monitoring tools; however, catalogue-based learning can introduce bias into the system. Here, we show that while monitoring systems recognize almost 90 % of events annotated in seismic catalogues, other information describing volcanic behavior is not considered. We found that weakly supervised learning approaches have the remarkable capability of simultaneously identifying unannotated seismic traces in the catalogue and correcting mis-annotated seismic traces. When a system trained with a master dataset and catalogue is used as a pseudo-labeller within the framework of weakly supervised learning, information related to volcanic dynamics can be revealed and updated. Our results offer the potential for developing more sophisticated semi-supervised models to increase the reliability of monitoring tools. For example, the use of more sophisticated pseudo-labelling techniques involving data from several catalogues could be tested. Ultimately, there is potential to develop universal monitoring tools able to consider unforeseen temporal changes in monitored signals at any volcano.