Classifying earthquake and explosion events using a dual-stage trained neural network

Misclassifying anthropogenic explosions as natural earthquakes contaminates seismic catalogs, distorts b -value estimates, and undermines foreshock–aftershock discrimination—compromising the reliability of seismic hazard forecasting. Conventional methods suffer from two core limitations: they either process each seismic trace independently and ignore multi-station physical correlations within a single event, or rely on heuristic aggregation rules such as majority voting that cannot adapt to epicentral-distance-dependent signal degradation and fail to correct erroneous single-trace predictions. This work proposes a dual-stage three-branch convolutional neural network for event-level binary discrimination between natural earthquakes and anthropogenic explosions, integrating raw waveforms, time–frequency spectrograms, and numerical physical features. The framework advances the field in three respects. Our framework couples trace-level feature extraction with event-level decision aggregation through a dual-stage end-to-end training paradigm. For multi-station fusion, we developed a self-supervised quality-aware attention mechanism that weights traces by their prediction confidence alone, removing the need for ground-truth quality labels. We further demonstrated robust out-of-distribution generalization across geologically independent datasets, from hard-rock underground quarry blasts to unconsolidated-sediment surface military explosions.The first stage learns intrinsic source-mechanism features via three parallel branches (1D-CNN, 2D-CNN, and a fully connected branch), while the second applies a quality-aware attention module to dynamically upweight reliable records and suppress ambiguous ones. This hierarchical design bridges single-station feature extraction to multi-station evidence aggregation, and end-to-end event-level training further allows encoders to capture physically meaningful, distance-varying seismic signatures, improving far-field robustness. Evaluated on a balanced ComCat test set (865 earthquakes, 865 explosions), the event-level model achieves 96.41% accuracy, 98.08% precision, 94.68% earthquake recall, and 98.15% explosion recall, with only 46 false negatives and 16 false positives. Cross-domain tests yield 84.83% accuracy on hard-rock quarry blasts and 84.72% accuracy on military surface explosions in unconsolidated sediments, verifying strong out-of-distribution generalization capability. Score-CAM visualizations confirm that the model focuses on P/S wave arrivals and the P–S energy gap, shifting systematically with epicentral distance in a manner consistent with seismological wave attenuation theory. These results demonstrate that deep learning can achieve both high classification accuracy and physical interpretability, providing a deployable system for operational seismic monitoring and nuclear test-ban verification.