Learning the Deep and the Shallow: Deep-Learning-Based Depth Phase Picking and Earthquake Depth Estimation

Abstract Automated teleseismic earthquake monitoring is an essential part of global seismicity analysis. Although constraining epicenters in an automated fashion is an established technique, constraining event depths is substantially more difficult. One solution to this challenge is teleseismic depth phases, but these can currently not be identified precisely by automatic detection methods. Here, we propose two deep-learning models, DepthPhaseTEAM and DepthPhaseNet, to detect and pick depth phases. For training the models, we create a dataset based on the ISC-EHB bulletin—a high-quality catalog with detailed phase annotations. We show how backprojecting the predicted phase arrival probability curves onto the depth axis yields accurate estimates of earthquake depth. Furthermore, we show how a multistation model, DepthPhaseTEAM, leads to better and more consistent predictions than the single-station model, DepthPhaseNet. To allow direct application of our models, we integrate them within the SeisBench library.