Seismo-Acoustic Fusion With Multi-Head Cross-Attention for Discriminating Earthquakes and Explosions

We propose a seismo-acoustic fusion framework based on multi-head cross-attention neural networks. A dual-branch CNN–TCN encoder extracts event-level representations from seismic spectrograms and array-derived infrasound features. Seismic embeddings serve as queries and infrasound embeddings as keys and values, enabling conditional introduction of infrasound information. A learnable gating mechanism further regulates the contribution of cross-modal features to suppress unreliable infrasound observations.