This document contains instructions for running application P and S Waves Detection Tool within the EPISODES Platform. The application is a tool for automatic detection of the first arrival time of the P and S waves. The input seismogram is scanned to find potential events. It uses Deep Learning technique based on a convolutional neural network to process the seismograms and detect the first arrival times. The model is based on Ross, Meier, Hauksson and Heaton (2018). The implementation is based on SeisBench a toolbox for machine learning in seismology (Woollam, Münchmeyer, Tilmann et al. 2022).

To obtain more general information about working with applications within the Platform, see Applications Quick Start Guide.

CATEGORY Source and Shaking Parameters Estimation

KEYWORDS Waveform viewing, Picking on waveform

CITATION If you use the results or visualizations retrieved from this application in a publication, then you must cite the data source as follows:

Ross, Meier, Hauksson and Heaton (2018). Generalized Seismic Phase Detection with Deep Learning. doi: 10.1785/0120180080.

Woollam, Münchmeyer, Tilmann et al. (2022). SeisBench — A Toolbox for Machine Learning in Seismology. doi: 10.1785/0220210324.

Zhu, W., & Beroza, G. C. (2019). PhaseNet: a deep-neural-network-based seismic arrival-time picking method. doi: 10.1093/gji/ggy423

Orlecka-Sikora, B., Lasocki, S., Kocot, J. et al. (2020) An open data infrastructure for the study of anthropogenic hazards linked to georesource exploitation., Sci Data 7, 89, doi: 10.1038/s41597-020-0429-3.

Input file specification

The application requires a MiniSEED Waveform (or any subtype) containing seismograms on which the wave arrival time should be detected. The waveform has to contain E, N and Z channels for each station.

Figure 1. Application input files specification

Filling form values

The application form consists of four parameters - see Figure 2. The first parameter allows you to choose weights obtained by training the Machine Learning model, where the Original option are the original weights published by the authors of the models ([1] for the GPD model and [2] for the PhaseNet model), and all the others are weights obtained for public data described in [3]. The Stride parameter sets the stride in samples for point prediction models, and the P phase threshold and the S phase threshold set the threshold of probability for which the phase is detected.

Figure 2. Application form with values filled

[1] Ross, Meier, Hauksson and Heaton (2018). Generalized Seismic Phase Detection with Deep Learning. doi: 10.1785/0120180080.

[2] Zhu, W., & Beroza, G. C. (2019). PhaseNet: a deep-neural-network-based seismic arrival-time picking method. doi: 10.1093/gji/ggy423

[3] Woollam, Münchmeyer, Tilmann et al. (2022). SeisBench — A Toolbox for Machine Learning in Seismology. doi: 10.1785/0220210324.

Produced output

The main output of the application is a QuakeML file containing picks marking the phase detections. The file is named after the input MiniSEED file with -picks.xml suffix added. The picks are displayed on the original input file within the application output view (Figure 3).

Figure 3. Application output - MiniSEED waveform with resulting picks displayed

As an additional information, the application provides also the detection probability values, saved to the file named after the input MiniSEED with -annotations.mseed suffix added. The probability can be examined (see Figure 4 and Figure 5) after expanding the Wave detection probability option (bottom of Figure 3).