GroningenNet: Deep Learning for Low-Magnitude Earthquake Detection on a Multi-Level Sensor Network

Abstract

Automatic detection of low-magnitude earthquakes has become an increasingly important research topic in recent years due to a sharp increase in induced seismicity around the globe. The detection of low-magnitude seismic events is essential for microseismic monitoring of hydraulic fracturing, carbon capture and storage, and geothermal operations for hazard detection and mitigation. Moreover, the detection of micro-earthquakes is crucial to understanding the underlying mechanisms of larger earthquakes. Various algorithms, including deep learning methods, have been proposed over the years to detect such low-magnitude events. However, there is still a need for improving the robustness of these methods in discriminating between local sources of noise and weak seismic events. In this study, we propose a convolutional neural network (CNN) to detect seismic events from shallow borehole stations in Groningen, the Netherlands. We train a CNN model to detect low-magnitude earthquakes, harnessing the multi-level sensor configuration of the G-network in Groningen. Each G-network station consists of four geophones at depths of 50, 100, 150, and 200 m. Unlike prior deep learning approaches that use 3-component seismic records only at a single sensor level, we use records from the entire borehole as one training example. This allows us to train the CNN model using moveout patterns of the energy traveling across the borehole sensors to discriminate between events originating in the subsurface and local noise arriving from the surface. We compare the prediction accuracy of our trained CNN model to that of the STA/LTA and template matching algorithms on a two-month continuous record. We demonstrate that the CNN model shows significantly better performance than STA/LTA and template matching in detecting new events missing from the catalog and minimizing false detections. Moreover, we find that using the moveout feature allows us to effectively train our CNN model using only a fraction of the data that would be needed otherwise, saving plenty of manual labor in preparing training labels. The proposed approach can be easily applied to other microseismic monitoring networks with multi-level sensors.

Deep vs. Shallow Learning: A Benchmark Study in Low Magnitude Earthquake Detection

“…Using data from the Groningen Gas Field in the Netherlands, we build on the logistic regression model of [2] by the addition of four further features. We then evaluate the performance of the augmented model relative to the CNN of [3], pre-trained on the Groningen data, on progressively increasing noise-to-signal ratios. We discover that, for each ratio, our logistic regression model correctly detects every earthquake, while the deep model fails to detect nearly 20 % of seismic events.…”

“…Groningen Gas Field. Each G-network station has four geophones, between 50 m and 200 m, and we gather data from all four geophone levels as the comparison model of [3] requires this. We downloaded seismograms for event examples from web services hosted by the Royal Netherlands Meteorological Institute (KNMI), applying the same date and magnitude (≥ 0.2, resulting mean 1.05) selection criteria as in [2].…”

“…This resulted in 2300 seismograms recording 47 events. We were unable to obtain 4000 noise samples, as used in [2], from the 2017-18 period of the studies of [2] and [3] and so used the same time of year in 2020-21. Additional negative examples for the increasing noise-to-signal phase of this study were drawn from this same period, with 38100 new samples collected.…”

“…Additional negative examples for the increasing noise-to-signal phase of this study were drawn from this same period, with 38100 new samples collected. All waveform data were detrended, demeaned, and bandpass-filtered to frequencies of 5-25 Hz as in [3]. In addition, in order to be compatible with the CNN of [3], the resulting seismograms were then downsampled by a factor of two.…”

“…All waveform data were detrended, demeaned, and bandpass-filtered to frequencies of 5-25 Hz as in [3]. In addition, in order to be compatible with the CNN of [3], the resulting seismograms were then downsampled by a factor of two. We chose to segregate our train (60 %), validation (20 %), and test (20 %) datasets by event rather than seismogram (seismogram-based segregation having been used in the studies of [2] and [3]) in order to prevent data leakage from test to train datasets.…”