Siamese networks for triggered earthquakes detection

Knowledge of triggered seismic events helps foster an understanding of seismic hazard and seismic stability in general. We explore whether machine-learning, in particular deep learning, can facilitate and/or automate detection of dynamically triggered earthquakes. Triggered local earthquakes coincide in space and time with the passage of surface waves from strong, distant earthquakes and are triggered by the transient stresses transmitted by these waves. Currently, detecting triggered earthquakes is done by human experts. Researchers must comb through thousands of seismograms per earthquake, often adjusting the time window and frequency band multiple times per seismogram. The amount of available data from seismograms all over the world and its ongoing growth makes this manual processing a very daunting task. Automating this process, even partially, could save thousands of hours of expert time. For this study, human experts identified triggered earthquakes on a dataset of over two thousand seismograms from eight large earthquakes from the Pacific and Indian Ocean rims. We then trained a deep learning system to identify triggered earthquakes on this data. Since deep learning systems typically require many more than two thousand examples for training, we applied a Siamese training approach, where the network was trained to classify pairs of examples as "same" or "different". This let us generate n²/2 pairs from n examples, thus increasing the number of elements in our training data set. Once the network is trained, it can be used to reduce a seismogram to a point in a 10-dimensional space. The data set is then encoded in this space and a linear classifier is used to learn a decision boundary between triggered events and other events. The Siamese network performs well at detecting triggered earthquakes, however we are still exploring the robustness of the detections and the sensitivities of the network, as well as comparing the network's performance against that of efficient non-machine learning automated detection methods, such as using a signal to noise ratio. Detecting and understanding dynamically triggered earthquakes are likely helpful means to detecting and understanding all types of earthquakes. Our way of automating such detections might likewise be helpful to automating the detection of other types of seismic events.