Rapid earthquake magnitude estimation combining a neural network and transfer learning in China: Application to the 2022 Lushan M6.1 earthquake

Abstract

Introduction: China is one of the most seismically active countries in the world. It is an important task for a Chinese earthquake early warning system to quickly obtain robust magnitude estimation. However, within the first few seconds after P-wave arrival, there is considerable scatter in magnitude estimation for traditional methods based on a single early warning parameter. Methods: To explore the feasibility of using a convolutional neural network for magnitude estimation in China, establish a magnitude estimation model suitable for China and provide more robust magnitude estimation based on strong-motion data from China, we propose a new approach combining a convolutional neural network and transfer learning (TL) to construct a magnitude estimation model (TLDCNN-M) in this study. Results and Discussion: Our results show that for the same test dataset, in terms of the mean absolute error and standard deviation of magnitude estimation errors, the TLDCNN-M model has better performance than traditional methods and convolutional neural network models without using TL. Meanwhile, we apply the method to the 2022 Lushan M6.1 earthquake occurred in Sichuan province, China. At 3 s after the earliest P phase, the magnitude estimation error is less than 0.5. With the increase in time after the earliest P phase, the magnitude estimation is close to the catalog magnitude; at 10 s after the earliest P phase, the magnitude estimation error is less than 0.2.