Multi-scale convolution networks for seismic event classification with windowed self-attention

The classification of seismic events is important for earthquake emergency warnings and earthquake catalog database establishment. In this paper, we developed a multiscale convolution and window self-attention network for seismic event classification by combining multiscale convolution with inductive bias capability and self-attention mechanism with long-range information capture capability. This paper employed a pre-processing strategy to acquire the complete seismic waveform and separate seismic data in each component. Additionally, a voting mechanism is proposed to integrate data from three components for classification, improving overall accuracy. The experimental results showed that the overall classification accuracy is 94.02% when considering seismic data from a single component only. However, after incorporating a voting mechanism, the classification accuracy increases to 97.56%, which outperforms other methods. The results demonstrated that the multi-scale convolutional and windowed self-attention networks can effectively and significantly improve the accuracy of seismic event classification, which get a good result in seismic event classification.