Two-stage approach for earthquake detection using multiple clustering-based classification

SUMMARY Deep learning (DL) approach has gained attention for earthquake (EQ) detection. To alleviate the problem of training data shortage, transfer learning (TL) provides a useful framework to adapt pre-trained models, typically through tuning of model parameters. Nonetheless, the current practice still requires considerable data, which hinders its application where only a small number of data is available. Instead of TL, we propose a novel two-stage of model correction as a solution to this important and ubiquitous problem in EQ detection. In the proposed approach, a pre-trained DL model is directly applied to waveform data in the target domain (first stage), and the cases that are classified as an earthquake signal (i.e. positive cases) are further classified as positives and negatives using a non-DL classification method (second stage). Our classification method for the second stage is based on multiple clustering, which characterizes local waveform patterns in terms of amplitude scale in specific time segments that are inferred in a data-driven manner. This characterization captures complex high-dimensional waveform patterns in a low-dimensional space, which leads to the effective classification of true and false positives. Furthermore, the proposed method is useful when only true positive waveforms are labelled (PU classification). Both synthetic and real data analysis clearly demonstrated effectiveness of unsupervised waveform characterization of the proposed method.