Identifying P and S phases using supervised learning algorithms
Abstract
The last years have seen a massive increase in the application of machine learning (ML) algorithms to seismology. This is, however, not a new field for ML. In the Icelandic SIL seismic system an artificial neural network was used to identify P and S phases on a station specific basis already back in the mid 1990s (Bödvarsson et al., 1996). Here we revisit the problem for the Swedish National Seismic Network using two types of supervised learning algorithms; logistic regression and multilayer perceptron (MLP, an artificial neural network). As input data we use station specific parametric information on detected phases, referred to as phase-logs, which have been identified as P or S and associated to seismic events by manual inspection. The phase-logs include data such as onset time, duration, signal and noise amplitudes, three-component analysis (coherency and apparent velocity) and references to previous and next phases. We implement both algorithms using the Keras deep learning library in Python and evaluate the results in terms of accuracy, precision and recall. We find that both the logistic regression and MLP algorithms perform very well, with an accuracy of above 90% for 62 out of 64 stations, and that the MLP is consistently better than the logistic regression at 96% accuracy for 62 of 64 stations, with the two remaining at above 90%. Similarly the MLP perform better than logistic regression for precision and recall. Our data set includes varying numbers of labeled phases for each station, from more than 40,000 down to only about 2,000. In addition, on average more than 80% of the data are P phases. We therefore test the identification algorithms on under-sampled data sets with balanced numbers of P and S phases and find that accuracy numbers stay similar but that precision decreases slightly and recall increases, i.e. higher proportion of actual S phases become correctly identified (example of precision-recall trade-off).
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.S43D0686L
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS