BazNet: A Deep Neural Network for Confident Three-component Backazimuth Prediction

Abstract

As the Treaty Monitoring Community seeks to lower detection thresholds across its sparse sensor network, single-station location estimates and accurate backazimuth predictions become increasingly important. Accurate backazimuth predictions are traditionally limited to array stations, where beamforming provides high-confidence backazimuth prediction that can be reliably passed on to the associator. Three-component stations, on the other hand, rely on polarization analysis for backazimuth prediction, which suffers from both high error and low confidence. As such, very few three-component backazimuth predictions are passed on to the association algorithm. This study presents BazNet, a deep neural-network that takes in a three-component seismogram and produces both a backazimuth prediction and corresponding certainty measure. For existing stations with ample historical training data, the technique achieves an overall median absolute deviation of around $$14^{\circ }$$ , a modest improvement over the $$15^{\circ }$$ achieved by polarization. More importantly, each estimate is accompanied by a robust certainty measure, allowing the selection of high-confidence predictions to be passed on to the associator. Using the BazNet certainty measure, roughly 60% of all three-component predictions can be selected with a median absolute deviation of just $$6^{\circ }$$ , which is on par with the predictions from a full beamformed seismic array. This represents a sevenfold improvement over the 8% of signals similarly selectable via polarization analysis. BazNet performance is demonstrated against 10 years of waveform data from 561,154 cataloged arrivals across nine stations selected from the global IMS Network: STKA, CPUP, VNDA, LPAZ, AAK, BOSA, ULM, BATI, INK.