Improving template matching detections using a Convolutional Neural Network

The use of template matching to detect previously missed small earthquakes is widespread in seismology, due to its power in searching for similar signals. The newly detected earthquakes improve understanding of the geology, seismo-tectonics and seismogenesis of the area explored. The usefulness of template matching has sparked the development of many software tools (e.g. QuakeMatch; Toledo et al., 2024) that allow seismologists to easily apply them to their area of interest. 

Like every detection technique, the performance of template matching shows a tradeoff between sensitivity and false detection rate that is dependent on the choice of the detection threshold. The value of the correlation coefficient between two earthquake signals is highly dependent on several properties such as the distance of the earthquake from the station, the noise level at the station, the magnitude of the earthquake, the focal mechanism of the earthquake, etc, making the selection of a specific correlation coefficient threshold hard. To detect a larger number of earthquakes, researchers often use a lower correlation coefficient detection threshold and manually inspect the detected events to classify them as true events (Toledo et al., 2024). This is, however, a tedious task, especially when using a large number of templates. To reduce the human workload, which can be especially important during evolving earthquake sequences, we employ a Convolutional Neural Network (CNN) to discriminate between earthquakes and noise using the template matching detections as input. We use the data from several microearthquake natural and induced Swiss sequences (Simon et al., 2024, in prep.), to train and test the developed CNN model. Our CNN model uses single-station 3-component waveforms of any length and outputs an earthquake detection score. We demonstrate that the developed CNN can be used to significantly reduce the human workload with high accuracy, allowing the use of low correlation coefficient value thresholds for template matching detections. Furthermore, we show the implementation of the method inside the QuakeMatch software (Toledo et al., 2024).