Attention-based LSTM-FCN for earthquake detection and location

Abstract

Earthquake detection and localization are challenging since the seismic signal usually is noisy and the microearthquakes are hidden in the seismic noise. Traditional detection and localization methods often rely on manually picked phases or computationally intensive algorithms. Inspired by the successful application of a deep learning model, ConvNetQuake, in detecting and locating the seismic events, we train an attention-based long short-term memory fully convolutional network (LSTM-FCN) model to improve the detection and location accuracy on the same data set. We use a parallel structure of FCN and LSTM to extract different features separately and merge them as a vector for better classification. In particular, FCN is used to extract high-level features and, similarly, LSTM is employed to model the temporal dependences. Besides, an attention mechanism is added to the LSTM to select a significant input segment along with a squeeze-and-excitation block in FCN to enhance useful feature maps for classification. We show that the trained model has a classification accuracy of 89.1 per cent, which represents 14.5 per cent improvement compared to the ConvNetQuake model. Moreover, the ConvNetQuake model only considered classifying seismic events roughly into one of the six geographic regions. But our model can locate the seismic events with a higher resolution by classifying them into as a greater number of regions as to 15 while a relatively high accuracy is maintained. We also demonstrate that the incorporated attention mechanism can effectively improve the classification performance by automatically and selectively enhancing the significant feature maps and inputs.