On-site instrumental seismic intensity prediction for China via recurrent neural network and transfer learning

China is a seismically active country. Rapidly and accurately predicting instrumental seismic intensity at recording sites is important for China to mitigate earthquake disasters. According to the peak ground acceleration (PGA) and peak ground velocity (PGV) at recording stations, instrumental seismic intensity for China is measured. Here, for the robust and rapid on-site instrumental seismic intensity prediction, we propose a method combining recurrent neural network (RNN) and transfer learning to predict on-site PGA and PGV for China. For the same test dataset from China, our results indicate that at 3 s after P-wave arrival, the RNN models using transfer learning have better performance on PGA and PGV prediction than the baseline models, which include traditional methods based on the single parameter and RNN models without using transfer learning. Additionally, according to the predicted PGA and PGV of the RNN models using transfer learning, we statistically analyze the alarm performance based on the predicted on-site instrumental seismic intensity. Meanwhile, according to the proposed method in this paper, we test five destructive earthquake events (M ≥ 6.6) occurred in China. The results show that at 3 s after the P-wave arrival, the predicted instrumental seismic intensity is almost consistent with the observed instrumental seismic intensity, the predicted instrumental seismic intensity error is mainly within ± 1, and the mean absolute error is 0.78. Meanwhile, for the area near the epicenter, the percentage of successful alarms reaches 90%, and the percentage of false alarms is 0%.