Abstract: This paper proposes a new method for seismic event classification based on Gram's angle field and multiscale residual neural network using natural earthquakes, artificial blasts and collapse events collected and labeled by Jiangsu Seismic Network Center as experimental samples. In this paper, we first perform pre-processing operations such as filtering and normalization on the waveform data, and then realize a two-dimensional encoded image of the waveform data by Gram's angle field conversion, which is used as the input of the multiscale residual neural network and yields the classification results. The classification accuracy of 1078 natural seismic events, 981 blast events and 830 collapse events are 92.55% for single waveform and 96.36% for event-based classification. The experimental results show that the Gram-angle field-based and multiscale residual neural networks have good classification effects for seismic classification.