Multimodal deep learning network for fast seismic discrimination and magnitude classification

Abstract

Quickly and accurately identifying seismic events from background noise and classifying earthquake magnitudes is extremely important for improving the performance of earthquake early warning (EEW) systems. Microtremors are weak nonearthquake-induced vibrations that may trigger EEW systems, leading to false alarms and causing unnecessary public concern. Moreover, quickly predicting whether an earthquake event is of low or high magnitude is important for EEW systems to determine the potential earthquake damage and the alert area. Here, we develop a multimodal deep learning network (MDLNet) that can identify seismic events while determining whether an earthquake is of low magnitude (M < 5.5) or high magnitude (M ≥ 5.5). MDLNet can handle multimodal data and uses time-domain and spectrum encoders to extract features. Then, the features extracted by these encoders are fused with ground-motion parameter data. We train MDLNet using multimodal data from seismic event signals and microtremor signals recorded by the Japanese Kyoshin Network. We demonstrate that using data from the 3-s period following the onset of P-waves, MDLNet can recognize 99.92% of microtremor signals, 96.65% of low-magnitude seismic signals and 90.68% of high-magnitude seismic signals, values higher than those for single-mode data. Multimodal deep learning techniques can provide new insight into seismology and EEW.