SES-05. Seismic Event Detection in Volcanic Regions using Machine Learning
Volcanic eruptions are usually accompanied by seismic events with a variety of temporal signatures: volcano-tectonic earthquakes, low-frequency earthquakes, tremors, and explosions. The detection of these events in seismograph data has been done manually for several decades now. This process produces accurate results but is time and labor-intensive. A variety of machine-learning methods have been developed over the past decade to automate elements of this task. Few of these methods incorporate expert knowledge about the process; rather, they train a model with a large labeled data set and then simply work with the raw seismograph data. Our goal is to build and train a model that combines supervised machine learning and expert knowledge to aid in this process. The main technical challenge here is feature engineering: identification of task-specific, salient, higher-level attributes that can be extracted from the raw data and used as inputs to the model. A well-chosen feature set can greatly improve both training and deployment. Through this research, we describe a neural network that identifies the arrival times of the P and S waves of small earthquakes preceding the 2012 New Zealand Mount Tongariro/Te Maari craters eruption. We utilize seismic waveform and analyst-reviewed phase arrival data, which are available through GeoNet New Zealand. The initial implementation involves training a neural network that contains convolutional and fully connected layers on the raw waveform data. The subsequent implementation involves training a similar neural network model, without the convolutional layers, using a set of engineered features. We compare the performance of both models to determine the efficiency and effectiveness of using a feature set as opposed to using raw waveform data.