Feature-Based Magnitude Estimates for Small Earthquakes in the Yellowstone Region

ABSTRACT Accurately computing the magnitude of small earthquakes during periods of high seismicity rates can be a challenging problem. Here, we introduce a machine learning method which uses features derived from the waveforms of individual phase arrivals and event source parameters to predict the local magnitude (ML) of earthquakes recorded by the Yellowstone seismic network. Our approach works for events with small temporal separation, does not require three-component broadband stations, and produces magnitudes that are consistent with the existing Yellowstone earthquake catalog. We train one support vector machine (SVM) per station–phase pair, resulting in 34 models using P features and 18 models using S features. Producing a model for each station is a straightforward approach to ensure the SVMs learn individual station corrections. In addition, we can easily interrogate, update, and add individual station models going forward. For each station–phase pair, we introduce a recursive feature elimination (RFE) algorithm that simplifies and slightly improves the overall predictive performance of the models. For P and S features, our RFE algorithm reduces 45 potential features to a set of seven common features that work well for nearly all models. In addition, our RFE algorithm limits selection bias and can be used for different monitoring regions, regression algorithms, magnitude types, and features. Using a simple network average of the magnitude predictions of each station model, we achieve excellent ML prediction (R2∼0.95) for two distinct testing sets in the Yellowstone region. Overall, our approach produces accurate ML estimates in the magnitude range of ∼0–3.5 and increases the number of stations available to compute ML in the Yellowstone region from 14 to 34, lowering the variance of ML estimates. With our approach, we can efficiently compute ML for swarm events and significantly lower the ML magnitude of completeness in the region.