Unsupervised deep machine learning methods to discriminate icequakes in seismological data from Neumayer Station, Antarctica

Unsupervised machine learning methods are gaining attention in the seismological community as more and larger datasets of continuous waveforms are collected, since they tackle the chal- lenging task of learning without using hand-labelled data. In this thesis, we present some neural network approaches to the task of unsupervised learning, and apply them to discriminate and cluster different seismological events, including icequakes and earthquakes. We focus mainly on contrastive learning, which has recently been showing great success in the field of computer vision and other domains, and we transfer these methods to the domain of seismology. We implemented and tested data augmentation strategies for seismological data, and applied the contrastive learning method SimCLR as well as the deep clustering method DEC. For this pur- pose, we created and partially labelled different datasets containing various waveforms including many icequakes detected by an STA/LTA algorithm on continuous waveform recordings from the geophysical obervatory at Neumayer station, Antarctica. We demonstrate the effectiveness of our approach using quantitative evaluation on the labelled dataset as well as qualitative evaluation of the clustering on a larger unlabelled dataset.