Abstract
Abstract
Deep learning earthquake signal denoisers commonly operate directly on waveform data or on time‐frequency representations, yet the impact of this design choice has not been systematically assessed under controlled conditions. Here we design and benchmark four comparable deep learning architectures to assess the effect of input data representation: a time‐domain denoiser, a time‐frequency denoiser–the Earthquake Seismogram (EQS) Denoiser–and two hybrid models processing both domains in parallel or sequentially. We evaluate waveform quality as a direct measure of denoising performance and assess the impact on signal detection and phase picking tasks using standard phase pickers. We further examine how denoising enhances signal onsets using a dedicated first‐motion polarity model, benchmark the best‐performing model against state‐of‐the‐art denoisers, and evaluate its generalization to out‐of‐distribution data. Our results show that the time‐domain model produces the most accurate waveforms at high signal‐to‐noise ratios (SNRs), but fails to detect event signals at low SNRs. The time‐frequency denoiser captures event signals down to lower SNRs but can introduce artifacts such as smeared onsets due to the time‐frequency binning. Hybrid models mitigate shortcomings of single‐domain approaches. The hybrid sequential model , which extends the EQS architecture, demonstrates strong signal detection at low SNRs while preserving sharp onsets, improving phase picking and extending first‐motion polarity determination to lower SNRs compared to raw or bandpass‐filtered data. Overall, surpasses benchmark models across every waveform and downstream task metric for both in‐ and out‐of‐distribution data, demonstrating that hybrid domain processing is a powerful strategy for accurate seismic denoising.