Cleaning volcano-seismic event catalogues: a machine learning application for robust systems and potential crises in volcano observatories

Complete and precise volcano-seismic event catalogues are important not only for the statistical value that they possess for describing past volcanic activity, but also because they constitute the input for automated systems that help monitor volcanic activity in real time. Computer systems are valuable assets in the task of volcano-seismic event classification because in theory they can have improved performance compared to humans due to speed, consistency, and unbiasedness. However, such systems are trained with data from previously created catalogues of events, and as such, if catalogues have noise, the systems will learn incorrectly. In this work, we propose the implementation of a methodology that is relatively easy and fast to apply for the identification of potentially mislabeled events in a seismic event catalogue. We compare the results of applying the procedure to two open catalogues from Cotopaxi and Llaima volcanoes. The first catalogue is believed to have an unknown but potentially significant level of noise, while the other is assumed to be clean. We further validate our results for one of the datasets with volcano observatory scientists in a blind-review fashion to demonstrate some of the hypotheses that can arise in a catalogue with a presumably important level of noise. We conclude that the methodology is valid for identifying potentially mislabeled seismic events and can help in assessing the quality of a given catalogue.