SEQNet: A Multiscale Feature Attention and Global Feature Aggregation Network for Seismic Waveform Denoising

ABSTRACT

Deep-learning denoising is increasingly used for earthquake monitoring, and time–frequency mask-reconstruction models such as DeepDenoiser have shown strong potential for seismic waveform recovery. However, many existing approaches still rely on U-Net-like backbones, leaving the contribution of denoising-specific architectural design insufficiently isolated. Here, we introduce SEQNet, an architecture-focused seismic waveform denoising model that integrates multiscale feature attention (MFA) and global feature aggregation (GFA) into a phase-sensitive time–frequency mask-reconstruction framework. MFA refines encoder features through multiscale context and attention, whereas GFA incorporates bottleneck-level global summaries before decoding. Using paired event–noise waveforms from INSTANCE and STEAD with physical-event-level separation, matched ablation experiments identify MFA as the most strongly supported architectural contribution, while the aggregate effect of GFA is smaller and metric-dependent. Compared with DeepDenoiser and representative conventional methods, SEQNet shows a more favorable reference-based waveform-fidelity profile, whereas DeepDenoiser often produces stronger event-to-pre-event contrast. Trace-level analysis further shows that contrast enhancement and waveform fidelity can disagree on the same records, and a four-day continuous Lushan aftershock test examines downstream phase-picking behavior in continuous data. These results show that targeted architectural modifications can improve waveform-oriented seismic denoising and that module attribution should distinguish waveform fidelity from contrast enhancement.