Accurate classification of seismic events is essential to prevent confusion in catalogs, which can distort seismic statistics and hazard analyses. Monitoring small-scale explosions aids in safety supervision and the detection of illegal mining, while rapid event identification facilitates timely emergency responses. With the proliferation of seismic monitoring stations and increased human activities such as mining, tunnel construction, and island development, the number of non-natural seismic events recorded by seismic networks has surged. Therefore, developing automatic classification methods for seismic events is vital for enhancing safety and disaster prevention efforts. This study utilized 28,421 seismic events and 172,214 waveform data from the Fujian Seismic Network (2010-2022). Three neural network models, VGG, AlexNet, and ResNet, were used to develop a deep-learning model for seismic event classification. We analyzed factors influencing model accuracy, such as signal-to-noise ratio, epicentral distance, magnitude, and source depth. The Grad-CAM and Activation Maximization method was employed to identify the primary features learned by the models. All three networks achieved over 95% accuracy, with VGG reaching 97%. In classifying earthquake events in Fujian in 2023, the model achieved 94.9% accuracy by individual stations and 98.8% by network average. Natural earthquake recall was 94.1%, and explosion recall was 95%. Generalization tests showed 85.9% accuracy in Shandong and 87.9% in Utah. This study demonstrates deep convolutional neural networks’ high accuracy in classifying earthquake events from raw waveforms, with good generalization capabilities.