Transformer Graph Convolutional Network for Relative Travel-Time Shift Prediction

Abstract

Abstract Predicting surface-wave travel-time shifts is valuable for analyzing potential effects caused by changes in medium properties, station clock errors, instrument response errors, and other factors. Many current neural networks used in seismology are single-station models trained using single-station (pair) data. However, most seismic methods require knowledge of the spatial positions between multiple stations. Multiple stations contain rich interrelationships and spatial information that cannot be exploited by single-station models. We proposed a multistation neural network structure Transformer Graph Convolutional Network (TGCN) that utilizes temporal attention and spatial attention to capture spatiotemporal information for predicting relative travel-time shifts. Before that, we introduced a method that treats station pairs as nodes and constructs a graph with multiple station pairs. We collected original ambient noise waveforms from 2017 to 2019 in the Alaska region and 2010 to 2014 in the southern California region to obtain relative travel-time shift sequences of station pairs for model training and testing. To showcase the improvement of spatial information to the model, we compared TGCN with two other baseline single-station models—temporal convolutional network and long short-term memory. Our proposed method predicted travel-time values more accurately than the two baseline models, and it also exhibited slower decay in performance when predicting over larger intervals. We also found that the number of station pairs has an impact on the model. When there are a sufficient number of station pairs, the model can effectively utilize the rich spatial information and achieve higher accuracy. Our approach, which incorporates spatiotemporal information, provides outputs that are more efficient and accurate compared with the traditional single-station (pair) method that only considers temporal information, suggesting that spatial information does enhance the performance of the model.