Bridging Supervised and Unsupervised Learning to Build Volcano Seismicity Classifiers at Kilauea Volcano, Hawaii

Abstract

Abstract Real-time classification of volcano seismicity could become a useful component in volcanic monitoring. Supervised learning provides a powerful means to achieve this but often requires a large amount of manually labeled data. Here, we build supervised learning models to discriminate volcano tectonic events (VTs), long-period events (LPs), and hybrid events in Kilauea by training with pseudolabels from unsupervised clustering. We test three different supervised models, and all of them achieve >93% accuracy. We apply the model ensemble to the six-day seismicity during the eruption in 2018 and show that they were mainly VTs (62%), in comparison with the dominance of LPs prior to the eruption (68%). The success of our method is aided by the accuracy of the majority of pseudolabels and the consistency of the three models’ performance. Using Shapley additive explanations, we show that the frequency contents at 1–4 Hz are the most important to differentiate volcano seismicity types. This work, together with our previous clustering analysis, provides an example of bridging unsupervised and supervised learning to construct potential real-time seismic classifiers from scratch.