Earthquake detection in time-series of laser strainmeter measurements as a first step towards automatic signal classification.

Abstract

<p>Geophysical observatories around the world collect data on various natural phenomena within the Earth and on its surface. Many of these measurements are made automatically, sometimes at high sampling rates, so that enormous amounts of data accumulate over the years. Continuous analysis is important to classify current phenomena and decide which data are important and which can be downsampled later.</p><p>At Moxa Geodynamic Observatory, located in central Germany, several laser strainmeters have been installed in subsurface galleries in order to measure strain of the Earth's crust. These instruments run in north-south, east-west, and northwest-southeast directions. Nano-strain rates are determined with a sampling rate of 0.1 Hz almost continuously over distances of 26 and 38 m, respectively, since summer 2011.</p><p>Signals of tectonically induced crustal deformation are superimposed by other signals of greater amplitude, e.g., tides, changes in atmospheric pressure, hydrologic events such as heavy rainfall, and earthquakes. Classification of these events is important to better associate jumps in the temporal vicinity and to distinguish anomalies from instrument failures. To avoid time-consuming pattern recognition by hand, algorithms are required to do most of the work automatically. Due to recent advances in the field of artificial intelligence, it is possible to implement time series algorithms that are capable of unifying and automating many steps of data analysis. Although artificial intelligence applications are increasingly used to support data analysis, their use for time series of geophysical origin so far is not widespread outside of seismology.</p><p>In this contribution, an approach to automatically detect earthquakes in the strain data using 1D Convolutional Neural Networks is presented, including the generation of artificial training data with time series data augmentation. Also the training process and generation of new training data, based on classification by hand and false predictions of the trained model is described. The 1D Convolutional Neural Networks are able to identify almost all earthquakes in the strain data and have F1 values > 0.99, showing that their application has the potential to significantly reduce the time required in signal classification of observatory time series data.</p>