Abstract Reliable discrimination between underground nuclear explosions (UNEs) and earthquakes (EQs) is paramount for nuclear nonproliferation. Recently, deep learning has demonstrated significant potential in seismic event classification due to its powerful feature extraction capabilities. However, the performance of these models is heavily dependent on the quality of the training datasets. Given the characteristics of existing UNE records, this study investigates the impact of training waveform duration and the geographical location of EQs on the performance of convolution neural network classifiers. Results demonstrate that although all models generalize well and consistently to unseen UNEs, their generalization performance in identifying EQs varies significantly depending on the training data. We find that models trained on EQs near nuclear test sites (NTSs) generalize better than those trained on unconstrained ones, and that longer training waveform durations yield superior performance. The geographical advantage stems from training on the spatial clustering of EQs and UNEs, which suppresses the learning of path-dependent features unrelated to the source type. The performance improvement from using longer waveforms may be attributed to the enriched information in the dataset, which enables the model to implicitly learn and correct for path effects. We recommend prioritizing EQs from near NTSs when training the UNEs classifier. However, the trade-off between optimal waveform duration and the resulting reduction of the training dataset remains an open question.
Zhang et al. (Wed,) studied this question.